You are on page 1of 171

Understanding the determinants of Land-Use and Land-Cover

change as a basis for conservation management in the Warren


River Catchment

Achieve International Excellence

Thayse Nery de Figueiredo

M.Sc. Ecology

University of Basel, Switzerland

This thesis is presented for the degree of

Doctor of Philosophy

of

The University of Western Australia

Faculty of Science

School of Agriculture and Environment

2018
STATEMENT OF CANDIDATE CONTRIBUTION

This thesis contains papers prepared for publication, some of which have been co-

authored. The bibliographical details of the work and where it appears in the thesis are

outlined below.

Chapter 2: Nery, T., Sadler, R., Aulestia, M.S., White B., and Polyakov, M.

Discriminating native and plantation forests in a Landsat time-series for land use policy

design. Under review: International Journal of Remote Sensing 11-Aug-2017. Student

contribution: 70%.

Chapter 3: Nery, T., Sadler, R., White B., and Polyakov, M. Predicting future plantatio n

forest development in response to policy initiatives: A case study of the Warren River

Catchment in Western Australia. Under review: Journal of Environmental Science &

Policy 10-Apr-2018. Student contribution: 80%.

Chapter 4: Nery, T., Polyakov, M., Sadler, R., and White B. Policy influence on forest

plantation development in the Warren River Catchment of Western Australia. Under

review: Journal of Environmental Management 06-Apr-2018. Student contribution: 80%.

I have the permission of all co-authors to include this work in my thesis.

Student’s signature:

Coordinating supervisor’s signature:

iii
CERTIFICATION

I certify that this thesis has been substantially completed during the course of enrolme nt

in this degree at The University of Western Australia and has not previously been

submitted or accepted for a degree at this or any other institution. I certify that help s

received in preparing this thesis and all sources used have been acknowledged.

Thayse Nery de Figueiredo

Perth, August 2018

iv
ACKNOWLEDGEMENTS

I gratefully acknowledge the funding received towards my PhD from the Brazilia n

Federal Support and Evaluation of Graduate Education Agency (CAPES) through the

Ciência Sem Fronteiras Project (CsF). I feel very privileged to have had this financ ia l

support to undertake this study.

I would particularly like to thank my supervisors Associate Prof. Ben White and Dr.

Maksym Polyakov for their guidance, wisdom, time, and most importantly, their

encouragement throughout this journey. Dr. Morteza Chalak’s critical comments and

support also helped a great deal.

My special thanks also goes to Dr. Rohan Sadler who has been a tremendous mentor for

me, thank you for your meticulous support by providing constructive solution and

statistical analysis.

I express my sincere thanks to UWA for accepting me as a PhD candidate in the School

of Agriculture and Environment. I also take this opportunity to thank the entire academic

staff of the School for helping students through periodic postgraduate enhanceme nt

program and interaction opportunities. I feel proud to be part of this School.

I acknowledge my PhD fellows for their support and friendship along this challenging

journey. My special thanks to Maria Solis Aulestia for the pleasant time and valuable

contributions at various stages of this study. Deborah Swindells, Emma Smith, Theresa

Goh and Heather Gordon are acknowledged for administrative support. I extend my

v
thanks to Warren Catchment Council in Manjimup for providing information on river

action plans within the catchment.

I also, and wholeheartedly, acknowledge family support from my beloved mother, my

siblings, and my niece for their love, patience and unending encouragement during this

entire journey. Finally, I would like to thank my adorable daughter Amanda, and my

husband Andre Gomig, who has been an invaluable partner in my achievements. This

accomplishment would not have been possible without (all of) you. Thank you!

vi
DEDICATED TO

My mother, Claudia Batista Nery,


who has continuously inspired and motivated me.

My beloved daughter, Amanda Nery Gomig, who has made me stronger,

better and more fulfilled than I could have ever imagined.

&

My husband, Andre Gomig, for his endless love, sacrifices, support and

advice.

vii
ABSTRACT

Land use policies can have a profound effect on spatial and temporal patterns of Land

Use and Land Cover (LULC) change across a landscape. However, the contribution of

policies in driving landscape changes remains a largely hidden and untested element in

policy analysis. This is especially the case in the Warren River Catchment (WRC) in the

south-west of Western Australia which has undergone extensive LULC conversions in

response to marked changes in economic conditions, biophysical drivers, and land use

policy. The key issue for the WRC has been conversion of natural forested ecosystems to

agricultural land, which peaked at 30% of the total area in 1979. This forest clearance has

come at a high environmental cost, resulting in biodiversity loss, development of dryland

salinity, and consequent degradation of water quality.

In an effort to better understand the factors affecting LULC changes in the WRC and the

potential impact of different policies on land use decisions, this thesis expands upon

existing studies by: (1) providing the first historical analysis of LULC changes in the

catchment to include two spectrally similar but ecologically distinct forest classes, namely

plantation and native forest; (2) identifying the major changes in land use since 1979,

including a comprehensive set of transitions among six LULC classes (i.e., agriculture,

water, sand dunes, native forest, harvested native forest, plantation forest); (3) forecasting

the likely future area of plantation forest under the current policy scenario; and, (4)

determining the dominant driving forces of landscape change over the past 35 years.

By integrating remote sensing data, geographical information systems, and machine

learning algorithms, we reconstructed the trajectories of landscape changes in the

catchment from 1979 to 2014 using aerial photography and Landsat imagery. In

viii
classifying the imagery the Support Vector Machine (SVM) presented an improved fit

over two alternative algorithms: Random Forests (RF), and Classification and Regression

Trees (CART). The most successful approach for consistent historical processing has

involved the combination of SVM classifier, four bands held in common between the four

Landsat Sensors (MSS 1979; TM 1992; ETM+ 2003; OLI 2014), and principa l

components analysis (PCA) of texture features. Further, the sampling strategy used to

collect the training data may also explain the high classification accuracy obtained across

four Landsat sensors. Eucalypt plantation forests reaching full canopy cover presented

the highest rates of misclassification because they share spectral properties with the

Eucalypt dominant native forest. However, the methodology presented here provides a

consistent and accurate means for distinguishing plantation and native forest, particula r ly

where they share similar dominant genera.

Mapping and forecasting the expansion or contraction of plantations is critical if we are

to unpack unintended policy effects (e.g., rising salinity levels in water resources due to

harvesting of plantations). Thus, we also predicted the spatial distribution of LULC

changes for 2025 by integrating Markov transition probabilities with a multila yer

perceptron neural network, cellular automata and spatially explicit socio-economic and

biophysical data. Although a series of policy initiatives have provided subsidies through

tax breaks for plantation forest, this land use was predicted to decrease under the current

policy measures. Large areas of plantations are thus likely to be reconverted to agriculture

due to the harvesting of existing plantations and a lack of new plantings. This finding may

be attributable to the uncertainty surrounding the changes in the taxation laws governing

the Managed Investment Schemes (MIS), the lower profitability of the forestry MIS

compared to agriculture, and the onset of the Global Financial Crisis. These estimates

imply that the government reliance on plantation forest to increase the timber supply and

ix
help arrest the rise in salinity is no longer viable. Ideally, a catchment specific policy

should have been implemented to encourage the retention of plantation forest to manage

salinity.

In addition to the simulation of landscape changes, we estimate an econometric model of

major drivers of LULC changes in the WRC using a spatially explicit Multinomial Logit

Model. Understanding the land use responses to policy is important because many

policies influence land allocation. The key empirical finding was that investment in

forestry is vulnerable to policy changes. While the unprecedented expansion of plantatio n

forest in the late 1990s was spurred by financial incentives (e.g., Plantations for

Australian: The 2020 Vision, and The Managed Investments Act 1998), the considerably

changes in tax deductibility for MIS had reduced investor interest in forestry by the late

2000s. Despite a reduction of stream salinity, this potential co-benefit from plantatio n

seems not to be a driver of plantation expansion as it usually does not contribute to direct

cash-flow and thus, remains largely unvalued by the landholder. This shows that the co-

benefits from growing plantations (e.g., carbon sequestration and water) are not captured

by current policies.

The outcomes of this thesis provide a basis for demonstrating empirically the effects of

past policy initiatives. This evidence-based approach can therefore equip policy-makers

with precise and up-to-date information, and so support the design of new policy

instruments to improve the sustainable use of timber resources and the provision of non-

timber benefits such as salinity mitigation.

x
TABLE OF CONTENTS

Statement of Candidate Contribution ................................................................................iii


Certification...................................................................................................................... iv
Acknowledgements ........................................................................................................... v
Abstract ........................................................................................................................... viii
Table of Contents ............................................................................................................. xi
List of Tables...................................................................................................................xiv
List of Figures ................................................................................................................. xv
Abbreviations and Acronyms ..........................................................................................xvi

1. INTRODUCTION ...................................................................................................... 1
1.1. Background............................................................................................................ 1
1.2. Economic and policy drivers of land use change in Western Australia ................ 6
1.3. Study area ............................................................................................................ 15
1.4. Research objectives and questions ...................................................................... 18
1.5. Thesis Outline...................................................................................................... 19
1.6. References ........................................................................................................... 20

2. DISCRIMINATING NATIVE AND PLANTATION FORESTS IN A


LANDSAT TIME-SERIES FOR LAND USE POLICY DESIGN ........................... 27
Abstract ........................................................................................................................... 27
2.1. Introduction ......................................................................................................... 28
2.2. Study Area and Data Sources .............................................................................. 32
2.2.1. Study Area ..................................................................................................... 32
2.2.2. Data Sources.................................................................................................. 34
2.3. Methods ............................................................................................................... 35
2.3.1. Imagery Pre-Processing................................................................................. 35
2.3.2. Imagery Classification................................................................................... 38
2.3.2.1. Unsupervised Classification................................................................... 38
2.3.2.2. Texture Measurement ............................................................................. 39
2.3.2.3. Image Segmentation ............................................................................... 40
2.3.2.4. Machine Learning Algorithms ............................................................... 41
2.3.2.5. Accuracy Assessment.............................................................................. 44
2.4. Results and Discussion ........................................................................................ 45
2.4.1. Classification Performance ........................................................................... 45
2.4.2. LULC Changes in the WRC.......................................................................... 50
2.4.3. Implications for Land Use Policy.................................................................. 51
2.5. Conclusion ........................................................................................................... 53
2.6. Reference ............................................................................................................. 54
Appendix A2. Sampling Strategy: Two Strata Sampling. .............................................. 60

xi
Appendix B2. Overall classification performance for SVM, RF and CART classifie rs
per Landsat sensors and Models C1 to C4, including Kappa, F-score, Producer (PA)
and user (UA) accuracies. ............................................................................................... 63
Appendix C2. McNemar’s test results for comparison of the difference between
Model C1 and the other classifications schemes (i.e., Models C2, C3 and C4) using
RF, SVM, and CART classifiers..................................................................................... 65

3. PREDICTING FUTURE PLANTATION FOREST DEVELOPMENT IN


RESPONSE TO POLICY INITIATIVES: A CASE STUDY OF THE
WARREN RIVER CATCHMENT IN WESTERN AUSTRALIA .......................... 67
Abstract ........................................................................................................................... 67
3.1. Introduction .......................................................................................................... 68
3.2. Study area and policy context .............................................................................. 71
3.3. Methods................................................................................................................ 75
3.3.1. Data ............................................................................................................... 76
3.3.2. Transition probability matrix ........................................................................ 78
3.3.3. Transition potential modelling using a Multilayer Perceptron (MLP) ......... 79
3.3.4. Cellular Automata (CA) simulation .............................................................. 80
3.4. Results and Discussion ........................................................................................ 82
3.4.1. Analysis of LULC Markov transition probabilities ...................................... 82
3.4.2. Model Calibration and Validation ................................................................ 86
3.4.3. Future Prediction Using Multilayer Perceptron CA-Markov Model (MLP-
CA-Markov)............................................................................................................ 87
3.5 Conclusion ............................................................................................................ 95
3.6. References ............................................................................................................ 96

4. POLICY INFLUENCE ON FOREST PLANTATION DEVELOPMENT IN


THE WARREN RIVER CATCHMENT OF WESTERN AUSTRALIA ............. 104
Abstract ......................................................................................................................... 104
4.1. Introduction ........................................................................................................ 105
4.2. Background: Plantation forest development in Western Australia .................... 107
4.3. Study Area.......................................................................................................... 110
4.4. Materials and Methods....................................................................................... 112
4.4.1. Modelling approach: Multinomial Logit Model (MNL) ............................ 112
4.4.2. Data sources ................................................................................................ 115
4.4.2.1. Land‐Use and Land-Cover dataset ...................................................... 115
4.4.2.2. Other data used for modelling ............................................................. 116
4.4.3. Sample size ................................................................................................. 118
4.4.4. Selection of variables and multicollinearity analysis.................................. 119
4.5. Results ................................................................................................................ 121
4.5.1. LULC trajectories ....................................................................................... 121
4.5.2. Multinomial Logit Model (MNL) of land use change ................................ 122
4.6. Discussion .......................................................................................................... 126
4.7. Conclusion ......................................................................................................... 131
4.8. References .......................................................................................................... 132

xii
5. CONCLUSIONS ..................................................................................................... 138
5.1. Overview ........................................................................................................... 138
5.2. Research Findings ............................................................................................. 139
5.2.1. Research Question 1 .................................................................................... 139
5.2.2. Research Question 2 .................................................................................... 141
5.2.3. Research Question 3 .................................................................................... 143
5.2.4. Research Question 4 .................................................................................... 145
5.2.5. Research Question 5 .................................................................................... 147
5.3. Study limitations and future research directions ................................................ 150

xiii
LIST OF TABLES

Table 2.1. Satellite data for the study area. ..................................................................... 35

Table 2.2. Landsat spectral bands, principal components analysis (PCA) of texture
features based on grey-level co-occurrence matrices, and NDVI used in each model. .. 37

Table 2.3. Land use classification system....................................................................... 38


Table 2.4. Texture measurements extracted from the Landsat spectral bands. The
subscripts 𝒊 and 𝒋 give the horizontal and vertical coordinates of the co-occurrence
matrix respectively (i.e., a Moore neighbourhood), p is the cell value and µ is the
mean cell value over a local window (Haralick et al. 1973). .......................................... 39

Table 2.5. Final input parameters used to train the algorithms and the out-of-bag
(OOB) estimate of the prediction error for RFs. The CART complexity parameter
(cp) was given as cp = 0.008 across all Models.............................................................. 43
Table 2.6. SVM estimates of historical land use and land cover in the Warre n
Catchment. ...................................................................................................................... 51
Table 3.1. Transitional probability matrix derived from the LULC map for the periods
1979-1992, 1992-2003 and 2003-2014........................................................................... 82
Table 3.2. Comparison of classified and projected LULC classes in 2014 .................... 87

Table 3.3. Estimated area of each LULC class in the Warren River Catchment. ........... 88

Table 3.4. Change in areal proportion of each land use between periods. ..................... 88

Table 4.1. List of prospective explanatory variables for the multinomial logit model. 117
Table 4.2. Marginal Effects of explanatory variables on the transition probabilities of
LULC change. ............................................................................................................... 123

xiv
LIST OF FIGURES

Figure 1.1. Simplified historical overview of the major policies affecting land
management in south-west of Western Australia............................................................ 13
Figure 1.2. Location of the Warren River Catchment in south-east of Western
Australia. ......................................................................................................................... 16
Figure 2.1. Location of the study area within the coastal range of Western Australia. .. 33

Figure 2.2. Flow diagram of pre-processing, processing, classification and evaluatio n


of classifier performance. ................................................................................................ 36
Figure 2.3. Comparison of overall classification performance for CART, RF and
SVM algorithms for each Landsat sensor across Models C1 to C4. ............................... 45
Figure 2.4. Classified imagery for each Landsat sensor (SVM – Model C1). ................ 50

Figure 3.1. Location of the Warren River Catchment in Western Australia. .................. 72

Figure 3.2. Overview of the methodology applied in this study. .................................... 78


Figure 3.3. The most representative areas of changes between 1992 and 2014. ............ 84

Figure 3.4. LULC maps in different years: (a) 1979, (b) 1992, (c) 2002, (d) 2014 and
(e) 2025. .......................................................................................................................... 90
Figure 4.1. Historical land use policies and its effects on LULC changes in WRC. .... 110
Figure 4.2. Location of the study area........................................................................... 111

Figure 4.3. Trends in surface areas and percentage for the three major LULC classes
in the study area............................................................................................................. 122

Figure 4.4. Map of LULC showing the expansion of plantation forest from 1992 to
2014. .............................................................................................................................. 126

xv
ABBREVIATIONS AND ACRONYMS

AIC Akaike Information Criterion


ALUM Australian Land Use and Management Classification Maps
ATO Australian Taxation Office
CA Cellular Automata
CART Classification and Regression Trees
DEM Digital Elevation Model
DOS Dark-Object Subtraction
DT Decision trees
EROS Earth Resources Observation and Science
ETM+ Enhanced Thematic Mapper Plus
FSS First Sampling Stratum
GIS Geographic Information Systems
Gl Gigalitres
GLCM Grey-Level Co-occurrence Matrices
GRASS Geographic Resources Analysis Support System
GSS Group Settlement Scheme
IID Identically and Independently Distributed
ISODATA Iterative Self-Organizing Data Analysis Technique
(ISODATA)
ITAA Income Tax Assessment Act
LSMS Large-Scale Mean Shift Segmentation
LULC Land Use and Land Cover
MIA Managed Investments Act
MIS Managed Investment Schemes
MLA Machine Learning Algorithm
MLP Multilayer Perceptron
MNL Multinomial Logit Model
MOLUSCE Modules for Land Use Change Evaluation

xvi
MSS Multispectral Scanner
NDVI Normalized Difference Vegetation Index
OBIA Object-Based Image Analysis
OLI Operational Land Imager
OOB Out-Of-Bag
PA Producer Accuracy
PCA Principal Components Analysis
PRS Product Ruling System
RF Random Forest
RFA Regional Forest Agreement
RPART Recursive Partitioning and Regression Trees
SAP Salinity Action Plan
SLIP Shared Location Information Platform
SRTM Shuttle Radar Topography Mission
SSS State Salinity Strategy
SVM Support Vector Machine
TDS Total Dissolved Solids
TM Thematic Mapper
TR Taxation Ruling
UA User Accuracy
USGS United States Geological Survey
VIF Variance inflation factors
WA Western Australia
WRC Warren River Catchment
WRRC Water Resource Recovery Catchments
WRS World Reference System
WSLSS War Service Land Settlement Scheme

xvii
Chapter 1. Introduction

1. INTRODUCTION

1.1. B ACKGROUND

The rapid growth in the world’s population over the last century has resulted in an

increased exploitation of natural resources (Lambin, Geist and Lepers 2003; Meyer and

Turner 1992). Indeed, with natural ecosystems degrading at unprecedented rates, a clear

scientific consensus has emerged that Land Use and Land Cover1 (LULC) has reduced

the capacity of land to provide ecosystem services at a global scale (Lambin et al. 2001;

Lambin et al. 2003; Manandhar, Odeh and Pontius 2010). The majority of LULC change

is anthropogenically driven and includes deforestation through the expansion of

agricultural and urban lands (Barron et al. 2012; Etter et al. 2006). LULC change is thus

the key source of anthropogenic change on the environment (Polyakov, Majumdar and

Teeter 2008; Meyer and Turner 1992).

LULC has also transformed Australian ecosystems. Despite most of Australia’s land area

(~75%) being deserts or arid lands usually unsuitable for forest growth, a rapid decline in

forest cover has been observed in the most fertile regions close to the coastal periphery

(Bradshaw 2012; Lunt 1998). Those LULC changes have produced significant economic

effects such as the development of agricultural and forestry enterprises. However, there

are also adverse social and environmental impacts, including habitat fragmentation, loss

1Land Cover is defined as the biophysical state of the earth (e.g., water and native forest); whereas Land
Use is the underlying exploitation of land cover (e.g., forest used for timber production; Braimoh 2004).

1
Chapter 1. Introduction

of biodiversity, decline in soil fertility and water quality. These changes have several

policy implications, including mitigation of dryland salinity, protection of wildlife

habitat, and carbon sequestration. Policies related to agricultural production, native forest

management and plantation forestry have attracted significant attention in Australia. The

Plantations for Australian (The 2020 Vision; Commonwealth of Australia 2002)

exemplifies it well due to its role in promoting an increase in Australia’s plantatio n

resources. However, in addition to the environmental impacts that differ for each land

class conversion, the costs of achieving a given policy goal such as The 2020 Vision

might vary depending on whether policy instruments target land retention, conversion, or

a combination of both (Plantinga and Ahn 2002). Therefore, understanding the drivers of

human decisions to alter the landscape as well as responses to the policies intended to

change these decisions is important for the development of more targeted environme nta l

and land use policies (Hobbs et al. 2016).

The conversion of the south-western Australian landscape from deep-rooted native

vegetation to shallow-rooted crops and pastures has come at a high environmental cost

(George et al. 2012). The effects of regional-scale deforestation have led not only to

widespread degradation of water quality, but have also contributed to significa nt

biodiversity loss (Myers et al. 2000), the development of dryland salinity (Graham,

Pannell and White 2010; Pannell and Roberts 2010) and the subsequent loss of productive

farming lands (Ridley and Pannell 2005). This is especially the case in the south-west of

Western Australia which experienced its most rapid deforestation rates between the 1920s

and 1980s, prior to its designation as a Global Biodiversity Hotspot due to its high number

of endemic and threatened plant species (Myers et al. 2000; Bradshaw 2012). Despite the

region’s high biodiversity and economic value, the dynamics of its LULC is not well-

2
Chapter 1. Introduction

understood. Burvill (1979) provides a broad summary of land clearing in south-west of

Western Australia.

Understanding the causes and the effects of LULC changes on the environment depends

on the availability of accurate historical information (Lambin et al. 2001), including the

detection of the types of changes occurring, where and when they occur, and the rates at

which changes occur (Lambin 1997). Currently, remote sensing and geographic

information system (GIS) technologies have been widely used as a base tool to identify,

monitor and analyse the spatio-temporal dynamics of LULC conversions. Also, various

change detection techniques have been developed to identify the extent and location of

environmental changes, with the conventional method involving a cross-tabulation matrix

obtained from the comparison of two or more temporal LULC maps (Braimoh 2006).

A key component in the production of LULC maps for change analysis is a machine

learning algorithm (MLA), or classifier, to predict LULC for the whole region from a

subset of ground-truthed data. A number of recently developed MLAs provide accurate

predictions for LULC mapping. Examples include Support Vector Machines (SVM),

Random Forests (RF) and Classification and Regression Trees (CART) (Kotsiantis 2007;

Otukei and Blaschke 2010; Sesnie et al. 2010; Shao and Lunetta 2012; Zheng et al. 2015).

Importantly, the selection of a suitable classifier is essential to improve upon

classification results (Lu and Weng 2007; Otukei and Blaschke 2010).

However, in spite of the development of MLAs, the classification of spectrally similar

classes using moderate resolution, remote sensing imagery (e.g., Landsat and MODIS

imagery) remains a challenging task (Marcos-Martinez and Baerenklau 2015; Wardlow

and Egbert 2008; Melgani and Bruzzone 2004). For instance, distinguishing between

3
Chapter 1. Introduction

forest types in south-western Australia. Mapping and evaluation of spectrally similar but

ecologically distinct classes is crucial for many ecological and forestry applicatio ns

(Shang and Chisholm 2014), and therefore should be considered as an essential

component of land use and conservation monitoring and management.

As we seek to develop our further understanding of the land use changes we are

confronted with the question of what are the major drivers of LULC changes. The basic

premise is that the landowners allocate private land among alternative uses to maximise

returns, which are determined by biophysical, economic, and policy factors. The changes

in land allocation are driven by the interaction and combination of various actors (e.g.,

landholders, policy-makers) and factors (e.g., policies, climate) at different spatial and

temporal scales (Lambin et al. 2001; Veldkamp and Lambin 2001). In this regard,

understanding and predicting the dynamic processes and driving forces of LULC change

implies a multi-disciplinary approach that integrates the knowledge of biophysical, socio-

economic and ecological factors (Irwin and Geoghegan 2001; Wu et al. 2008; Agarwal et

al. 2002). Although the design of spatially explicit models of LULC change has recently

gained considerable attention in a number of fields (Paiboonvorachat and Oyana 2011),

less thought has been given to the development of these models to understand the

economic and political processes that underlay landscape changes (Irwin and Geoghegan

2001).

In this context, economic theory is often used to support the model development,

including the selection of explanatory variables. However, to date, there is no single

unifying theory of LULC change. The two principal economic theories to explain spatial

patterns of LULC changes have been the Ricardian and von Thünen theories. Both the

Ricardian and von Thünen models are based on land rent theories that state that any parcel

4
Chapter 1. Introduction

of land will be used in the way that earns the highest rent. The Ricardian theory states that

land use relates mainly to the physical and immutable qualities of land (e.g., soil fertility;

Ricardo 1817). Therefore, the input of labour per hectare is greatest on land parcels that

possess the highest intrinsic land quality. In contrast, the von Thünen theory extends the

Ricardian’s proposition by adding location-specific characteristics and transportatio n

costs to the model. In this model, rent (i.e., profit) is primarily a function of distance from

the center of economic activity (Thünen 1966), and assumes that the landscape is a

homogenous and “featureless plane”. Especially in cases of large public land, we can

assess the distance from public transportation infrastructure.

A further in-depth analysis would also consider the causal mechanisms behind land use

transitions as involving two forces. The first force is motivated by the desire to achieve

new economic opportunities through socio-economic changes, innovations or policy

incentives. The second force is driven by the need to respond to resource scarcities and

ecological constraints that were caused by past land use practices (Lambin and Meyfroidt

2010). Based on those theories, LULCs are therefore driven by landholders’ choices

among economic opportunities, political factors, specific characteristics of a landscape,

or a landholder’s attitude to risk and financial situation. Such considerations should not

be overlooked when a structural framework is designed to address the subtle political and

economic processes underpinning land conversions. Hence, it is important to develop

modelling techniques that are sufficiently general that they can be used to test alternative

paradigms.

The Multinomial Logit Model (MNL) allows flexibility in behavioural assumptions while

explaining complex spatial-temporal patterns in the distribution of a set of land use

alternatives (Bockstael 1996; Chomitz and Gray 1996; Lin et al. 2014; Polyakov and

5
Chapter 1. Introduction

Zhang 2008). In general, MNL allows researchers to focus on developing a model of the

individual landowner’s choice. As such, an MNL is appropriate for understanding the

driving forces of landscape changes in the south-west of Western Australia as it can be

structured to account for location in a landscape and the response of LULC change to

spatial variation in other variables of interest (Chomitz and Gray 1996; Irwin and

Geoghegan 2001). The MNL can, therefore, be used to build an understanding of human

decisions in response to a set of predictor variables (drivers). This step of model building

is critical to LULC responses to policies, not only to predict their potential impact on the

environment but also to avoid repeating past policy mistakes.

1.2. ECONOMIC AND POLICY DRIVERS OF LAND USE CHANGE IN WESTER N

AUSTRALIA

Regions of south-western Australia experienced extensive clearance stimulated initia lly

by the Group Settlement Scheme (GSS) in the 1920s, and later on by the War Service

Land Settlement Scheme (WSLSS) in the 1950s. The GSS was an assisted migratio n

scheme instigated after World War I, and was aimed at developing dairy farms within

Australia's South West, hence reducing the reliance of geographically isolated Western

Australians on imported dairy products (Brunger and Selwood 1997). Similarly, the

WSLSS was created to accommodate and provide employment to soldiers returning from

service in World War II. Originally, it was intended to develop farms on already cleared

land that was deemed to be suitable for agriculture (Sanders 2005). However, governme nt

investment through various taxation concessions, training, infrastructure development

6
Chapter 1. Introduction

and loan incentives provided economic drivers for further land clearing (Australia n

Greenhouse Office 2000).

As a result, deforestation and forest degradation issues have been the focus of a number

of policy initiatives in Western Australia, as exemplified by The State Salinity Strategy

(SSS) and The Country Areas Water Supply Act of 1947 – Part IIA “Control of

Catchment Areas” (1947 Act). The 1947 Act was amended in 1976 and 1978 by the

Western Australian State Government to control further clearing of native vegetation in

five potential water supply catchments in the south-west: the catchments of Wellingto n

Dam, Mundaring Weir, Denmark River, Kent River Water Reserve, and Warren River

Water Reserve (Smith et al. 2006; Government of Western Australia 2015). In 1990 the

Harris River Dam Catchment was also proclaimed under the 1947 Act (Government of

Western Australia 2015).

Initially referred to as Clearing Controlled Catchments, and in 1996 renamed as Water

Resource Recovery Catchments (WRRC), portions of those key catchments affected by

salinity were designated as a part of a state-level salinity strategy with a 30-year vision

termed the Salinity Action Plan (SAP; Government of Western Australia 1996).

However, through a process of review of the SAP and public consultation, the strategy

was redeveloped to become the State Salinity Strategy (also known as Western Australia

Salinity Strategy) in 2000 (State Salinity Council 2000).

As salinity is the greatest threat to sustainable land use and conservation of biodivers ity

in south-west Western Australia (Wallace 2001), the main goal of the State Salinity

Strategy (SSS) is to reduce the economic damage caused by salinity to agriculture,

buildings and infrastructure in rural areas, and the environment. The SSS is managed by

7
Chapter 1. Introduction

the Recovery Teams through a strategic partnership approach, led by the Department of

Water and State Salinity Council, providing key stakeholders with a role in salinity

management within their catchments (State Salinity Council 2000).

Research has shown that significant investment is needed to reduce, or stop, the impact

of salinity on water resources across south-west Western Australia, with 1.8 millio n

hectares already affected by salinity (Wallace 2001). The scale and type of interventio n

required to achieve a potable water standard of 500mg/L Total Dissolved Solids (TDS) is

unique for each catchment. The interventions usually address the following manage me nt

options: deep-rooted perennial shrubs and trees, including large-scale planting of

previously pastoral land with commercial or non-commercial shrub or tree plantatio ns;

phase farming (rotations which include a perennial phase); pumping groundwater; and,

divert saline water through shallow drains and other means (Government of Western

Australia 1996). Currently, the state government contributes about $40 million a year to

salinity management, and the Commonwealth Government has also contributed

significantly. However, the largest investment has been funded by private landholders

(State Salinity Council 2000).

Although LULC changes have been recognised as a key element that substantially affects

ecosystem functioning, accurate understanding of historical landscape changes is not

usually addressed when an agri-environmental policy is implemented. For instance, the

SSS for the Warren River Catchment was developed in 2000 to recommend mechanis ms

to achieve the stream salinity goal of 500mg/L TDS by 2030. However, according to

Smith et al. (2006) the Warren River will only achieve the stated goal if additiona l

interventions, such as upper catchment reforestation or engineering works, are applied.

More in-depth analysis revealed that this policy was implemented without knowing the

8
Chapter 1. Introduction

historical rate of land transitions, and without the accurate use of socio-economic

information.

One of the key management options proposed by the SSS was commercial plantatio n

forest. In the late 1990s commercial plantations spread across previously cleared lands in

south-west Western Australia. A number of contingent circumstances and policies

combined to drive this plantation establishment, as illustrated by the Income Tax

Assessment Act of 1997 (ITAA 1997), Plantations for Australian (The 2020 Vision), and

The Managed Investments Act 1998 (MIA 1998). The ITAA 1997 was released in 1997

to replace the Income Tax Assessment Act 1936, and is one of the main regulations under

which income tax is calculated in Australia. The Commonwealth Government recognised

in the ITAA 1997 that the main impediment to strengthening plantation forests within

Australia was that investment in plantations would not produce revenue for between 8

and 25 years. Consequently, the Government established immediate tax deductibility for

investments in plantation forest. Like forestry agribusiness, non-forestry investme nts

were also structured around a tax benefit received at the point of initial investment (ATO

2017a). However, the lead time for returns on non-forestry investments was much shorter

than for investments in forestry.

The 2020 Vision was subsequently launched in 1997 by the Ministerial Council on

Forestry, Fisheries and Aquaculture, and aimed to treble Australia’s plantation area by

2020 to a total of 30,000 km2 . The benefits of The 2020 Vision would be a reduction of

Australia’s reliance on native forest products sourced either domestically or overseas, an

enhanced capability to compete internationally for a share of the wood product market,

and a reduction of the trade deficit of around $2 billion in wood and paper products in

1995 (Commonwealth of Australia 1995). Additionally, by increasing plantation forest

9
Chapter 1. Introduction

environmental benefits such as salinity mitigation would be favoured as an indirect effect

of landscape-scale reforestation (Townsend et al. 2012; Department of Water 2012). To

achieve this target, diverse initiatives to stimulate plantation development were proposed,

including tax incentives (e.g., ITAA 1997; Commonwealth of Australia 2002) as a well-

known practice to encourage the supply of desirable ecosystem services from agro-

ecosystems through incentivising changes in land use (Bryan 2013; Plantinga and Wu

2003).

Similarly, the MIA 1998 responded to the timber supply shortage by creating a retail

investment structure whereby investors could achieve a personal tax deduction for the

expense of investing in agribusiness development activities through Managed Investme nt

Schemes (MIS). Agribusiness MIS at this time were assigned to one of two categories:

forestry or non-forestry2 . In general, an MIS arrangement is where investors are brought

together to pool their funds for the common interest of financing an agribusiness project.

The funds are then managed by a 'responsible entity' who operates the scheme

(Mackarness and Malcolm 2006). Currently, Australia’s total plantation area is about

20,125 km2 , and Western Australia contains the nation’s second largest area of plantatio ns

(4,029km2 ; Gavran 2014). Nonetheless, a secondary objectives of the tax scheme in

promoting plantations (e.g., Eucalyptus globulus - Tasmanian bluegum) was the

enhancement of rural economies through the provision of employment, and to potentially

meet social and environmental objectives, such as enhancing biodiversity through

2Non-forestry MIS activities are primarily focused on the horticulture industry (e.g., wine grapes, almonds
and olives), but also encompass other primary industries such as beef cattle, aquaculture and poultry.

10
Chapter 1. Introduction

reconnecting a fragmented landscape of remnant vegetation, and water quality through

forestry based mitigation of landscape-scale salinity.

Furthermore, another relevant factor to the afforestation investments was the introductio n

of the Product Ruling System (PRS) by the Australian Taxation Office (ATO) in 1998.

The PRS clarifies the taxpayer’s obligation, and specifies the tax deductibility status of

each afforestation project for investors. Before PRS was introduced, investors relied

heavily upon a seller’s description of the taxation benefits of a particular investment. If

the seller was wrong, taxation penalties would be applied to investors. Thus, the PRS has

become a crucial component in MIS by providing tax certainty to both investors and MIS

industry, given the scheme proceeds according to the plan submitted to the ATO (The

Treasury 2008). A further event with an effect on MIS investments was the release of Tax

Ruling TR 2000/8 in 2000, which focused almost exclusively on investment schemes

(ATO 2000). Hereupon, the ITAA 1997, PRS and TR 2000/8 provided certainty to

investors that their investment in MISs were an allowable tax deduction, in alignme nt

with The 2020 Vision.

As the agroforestry investments were deemed to be profitable under the new tax regime ,

several companies became involved in domestic agroforestry activities, both Australia n

and overseas owned3 . Most Eucalyptus plantations have been undertaken on previous

agricultural lands regardless of their actual potential for forest productivity, and therein

lies a problem. Significantly, MIS companies had to acquire land assets which led to

substantial bank debt as tax deductibility under the MIA was not applicable for buying

3The main companies involved are New Forests, Integrated Tree Cropping Ltd., Green Triangle Plantation
Forest Company of Australia Pty Ltd. (owned by a Japanese co nsortium), and Global Forest Partners.

11
Chapter 1. Introduction

land. What followed was a rapid expansion of forestry MIS, with demand from investors

exceeding planned expectations. In this case, commercial plantation forests was used as

a potential response strategy to reduce salinity and enhance water quality in the WRRC.

The forestry MISs were then sold as ethical investments, whereby desirable

environmental goals could be satisfied through forestry MIS over the long-term.

Between 1997 and 2007, the expense of establishing, managing, harvesting and

transporting the outputs of forestry and non-forestry investments were considered as a

non-capital business expenditure and were thus fully tax deductible in the year the

business incurred the expense (under Division 8 Section 1: General Deductions of the

ITAA 1997). However, over the past ten years, the taxation laws governing MIS

operations in Australia have been subjected to successive changes. These changes

generated significant uncertainty among MIS investors, with the eventual downfall of the

MIS industry attributable in part to this uncertainty. In the mid-2000s the Governme nt

questioned the effectiveness of the MIS, as the tax-driven strategies of the MIS activities

could lead to irrational investment and perverse outcomes (e.g., new investors were the

only source of revenue for previous investors until the trees matured). Thereafter, in 2006,

the MIS industry was dealt a severe blow when the ATO notified that any contributio ns

paid into non-forestry and forestry MIS instruments would no longer be tax deductible

from 1 July 2007 (this was later postponed to 1 July 2008).

After much lobbying by the MIS industry, Division 394 was inserted into the ITAA 1997

in 2007 to provide tax certainty for forestry MIS investors (ATO 2017b). This tax change

allowed immediate tax deductions for expenditure incurred in forestry MIS, provided the

schemes complied with certain conditions. From this date, any change in the general

taxation rules did not apply to forestry MISs as this industry is protected under Divis io n

12
Chapter 1. Introduction

394 of the ITAA 1997. The main reason for providing this concession was to continue

the expansion of Australia’s plantation forestry. Controversially, MIS for non-forestry

agribusiness was to be taxed under the general income taxation rule (The Treasury 2008).

Those changes led to the development of a new TR 2007/8, replacing the TR 2000/8, and

also the release of the Product Ruling PR 2007/71 in 2007.

However, following the outcome of a test case in 2009 regarding non-forestry

agribusiness the Government withdrew its decision, and TR 2007/8 was abolished.

Subsequently, PR 2007/71 Addendum A1 was provided to re-allow non-forestry

agribusiness MIS an up-front deduction for their investments, subject on the facts of each

situation, under Section 8-1 of the ITAA 1997 (ATO 2017a). In contrast, forestry MIS

continues to rely on Division 394 to the present day (ATO 2017b). The foregoing is

summarised in the timeline presented in Figure 1.1.

Figure 1.1. Simplified historical overview of the major policies affecting land
management in south-west of Western Australia.

13
Chapter 1. Introduction

Before those major tax revisions were introduced, MIS investments were at their peak,

with AUD 1.2 billion placed in MIS in 2006 (The Treasury 2008). However, MIS sales

had decreased drastically by 2010 to AUD 100 million (NewForests 2015), leading to the

near collapse of agribusiness MIS in Australia. The drivers of this collapse may be

attributable to the uncertainty surrounding the changes in MIS legislation, the lower than

expected profitability of the forestry MIS, poorer quality wood due to underinvestment in

good management practice, an underestimation of the component costs of plantatio n

forestry, and the onset of the Global Financial Crisis. The risk of investing in forestry

MIS had been highlighted since the late 2000s (ASIC 2012; Brown, Trusler and Davis

2010), with six high-profile agribusiness companies operating in Australia having failed. 4

This is particularly the case for Great Southern Group and Timbercorp Limited, the two

major agribusiness companies operating in the south-west of Western Australia. Both

were estimated as having around 40% of all MIS business in Australia (Brown et al.

2010), but they folded in 2009 with a multi- million dollar debt. These agribusiness

companies focused on negotiating the tax benefits rather than on profitable commodity

production, resulting in an unsustainable business model that was supported solely by

MIS concessions. Investments driven by the demand for tax minimisation, and not by

sound plantation management practices, are more closely associated with a higher risk of

collapse (Ajani 2008).

Now more than ever the question arises as to the future of Western Australia’s plantatio n

forestry following the downfall of MIS investments. The serious flaws intrinsic to the

4Environinvest Limited, FEA Plantations Limited, Great Southern Group Limited, Rewards Projects
Limited, Timbercorp Limited, and Willmott Forests Limited.

14
Chapter 1. Introduction

MIS instruments resulted in severe losses for investors, jeopardising not only the future

of MIS plantations in Western Australia, but also the efforts to decrease the water salinity

levels of the recovery catchments, especially the Warren River Catchment. The collapse

of the industry, and the lack of economic certainty given possible future changes of

forestry MIS tax laws by the government, is likely a deterrent to further investment in

forestry MIS, at least not at the scale observed prior to 2007. This has significa nt

consequence for policy outcomes under the Commonwealth Government’s 2020 Vision

and the state government’s SSS. These policy outcomes may largely be measured through

the LULC changes towards agroforestry with the rise and fall of forestry MIS

investments. Studying the LULC changes as a response to the policy will assist in

evaluating the role of government policy incentives on land manageme nt.

1.3. STUDY AREA

This thesis uses a case-study area in the south-west of Western Australia to examine the

implication of past and present LULC change patterns (Figure 1.2). The Warren River

Catchment (WRC) is located about 300 km south-east of Perth, and extends from near

Kojonup on the north-east of the catchment to the south coast of the state, encompassing

the small towns of Manjimup and Pemberton. The region experiences a warm temperate

Mediterranean climate, with long, hot, dry summers and cool, wet winters. The WRC is

one of the largest surface water resources in the south west of Western Australia with a

mean annual flow of 291 Gl (1990-2001; at the Barker Road Crossing gauging station;

Smith et al. 2006). The mean annual rainfall ranges from 550 mm/year in the north-east

of the catchment to over 1200 mm/year in the coastal south-west, with most rain falling

15
Chapter 1. Introduction

during winter and early spring (Townsend et al. 2012). Situated in a high rainfall zone,

the WRC is highly suitable for investments in MIS forestry, especially for the Tasmania n

bluegum species.

Figure 1.2. Location of the Warren River Catchment in south-east of Western Australia.

The WRC covers an area of 4,416 km2 which is geographically divided into upper and

lower catchments. The upper catchment comprises the Tone River, Perup River and

Wheatley Farm sub-catchments. Tone and Perup rivers rise in farmland in the south-west

of Kojonup, and both join to form the Warren River east of Manjimup. The lower

catchment resides below this confluence, and the river then flows through mainly native

forest before reaching the Southern Ocean south-west of Pemberton (Smith et al. 2006).

16
Chapter 1. Introduction

The soil types in the upper catchment are typically gravelly ironstone over a hard lateritic

duricrust. Around Pemberton, in the lower catchment, the soil type changes to heavy red,

or karri, loams, which significantly favours horticulture and supports tall karri forest

interspersed with jarrah and marri open forest (Munro 2006).

State forest reserves that are subject to commercial logging are the main LULC in the

lower catchment, alongside agriculture, horticulture and plantation forest. In contrast, the

upper catchment has largely been cleared primarily for farmland, including broad-acre

agriculture (annual crops and livestock). The development of the agricultural land use

began in the 1920s as part of the GSS (Munro 2006) and continued in the 1950s with the

WSLSS. Cleared areas for agricultural use jumped from about 20% in 1950 (Collins and

Barrett 1980) to 30% in 1979 (see 2.4.2. LULC Changes in the WRC). The pattern of

deforestation in the upper catchment has resulted in the fragmentation of the native forest

across the landscape.

The WRC provides a number of highly prized ecosystem services, such as its water

resources, rich biodiversity, and the surrounding forests that support both timber

production and a vibrant tourist industry. Therefore, since 1978, the WRC is recognised

as a clearing control catchment under the Country Areas Water Supply Act 1947 and is

referred to as a WRRC (Government of Western Australia 2015). Additionally, major

timber plantations began in the 1990s, spurred on by the changes to the tax legisla tio n

described above (e.g., ITAA 1997, PRS, and TR 2000/8) and the introduction of two

federal policy initiatives: The 2020 Vision and MIA 1998. Since then, there have been

high levels of conversion of agricultural land into commercial plantation forest.

As a result of this extensive clearing, salinity in the Warren River exceeded 500 mg/L

17
Chapter 1. Introduction

TDS in the 1960s (Munro 2006). Between 1993 and 2004, the average annual salinity in

the upper parts of the catchment was around 1015 mg/L, with most of the salt load, around

74%, coming from the Tone and Perub Rivers (Platt 2007). Despite its high salinity level,

the Warren catchment remains significant for the provision of both economic and

ecosystem services. Unfortunately, those essential services remain under threat from

dryland salinity and the consequent stream quality. Land-use policies and incentives

promoting the sustainable use of the catchment’s resources can potentially influe nce

future LULC change and the impact of that LULC change on the quality of those

resources.

1.4. RESEARCH OBJECTIVES AND QUESTIONS

Given the importance of the Warren River as the third largest river and potential water

supply for the south-west of Western Australia, there is an urgent need to identify the

processes driving the evolution of the current pattern of LULC. Therefore, the objective

of this dissertation is twofold. First, to develop methods to classify and simulate spatial-

temporal Land Use and Land Cover change patterns, targeting two spectrally similar but

ecologically distinct LULC classes namely plantation forest and native forest. The second

objective is to identify the dominant driving forces behind the historical evolution of the

LULC pattern, whilst focusing on the effects of government policies influencing land use

management. These objectives will be achieved through addressing the following five

research questions:

18
Chapter 1. Introduction

1. Which classifier exhibits better learning rates for discriminating spectrally

similar classes of a time-series of Landsat imagery? (That is, between plantatio n

forest and native forest).

2. What are the main changes in LULC that have occurred since 1979? (That is,

since the start of the Landsat record).

3. Given the current policy scenario, what is the likely future of plantation forest?

4. What were the main drivers of LULC change over the past 35 years (i.e., 1979-

2014)?

5. What is the role of government policy incentives in the establishment of

plantation forest?

This research will go beyond previous analyses of landscape changes conducted in the

WRC by linking the historical spatio-temporal LULC transformation with their main

drivers. In so doing, economic theories of agricultural land use allocation will have been

included in the model. In this context, this study will identify some of the main societal

responses in terms of LULC transitions in response to policy incentives, and supply policy

makers with a better understanding and framing of what has historically driven land use

change within the Warren River Catchment.

1.5. THESIS OUTLINE

This thesis consists of 5 chapters based on papers. This introductory chapter provides a

broad overview of the thesis, of the problem statement and of the research questions to

19
Chapter 1. Introduction

be addressed. Chapters 2 to 4 are based on scientific papers which have either been

published or have been submitted for publication in refereed journals. Each research

chapter contains an independent introduction, data description, methodology, results and

discussion, followed by a conclusion. In addition, each chapter is independently

referenced. As the papers are all closely connected by the overall research objective s,

there is some unavoidable repetition within thesis across the different chapters,

particularly with regard to descriptions of the data and the case study area. Finally, chapter

5 provides a summary of the main conclusions and the implications of the study for

agroforestry policy.

1.6. REFERENCES

Agarwal, C., G.M. Green, J.M. Grove, T.P. Evans, and C.M. Schweik. 2002. “A review
and assessment of land-use change models: Dynamics of space, time, and human
choice.” General Technical Report NE-297:61.
Ajani, J. 2008. “Australia’s Transition from Native Forests to Plantations: The
Implications for Woodchips, Pulpmills, Tax Breaks and Climate Change.” Agenda
15(3):21–38.
ASIC - Australian Securities and Investments Commission. 2012. “Agribusiness
managed investment schemes: Improving disclosure for retail investors.” :41.
Available at: http://download.asic.gov.au/media/1246962/ris- for-rg232.pd f
[Accessed January 18, 2016].
ATO - Australian Taxation Office. 2017a. “Income Tax Assessment Act 1997 No. 38.
Chapter 1 Introduction and core provisions. Division 8: Deductions. Vol 1.” :27–28.
Available at:
https://www.legislation.gov.au/Details/C2017C00155/Html/Volume_1 [Accessed
July 10, 2017].
ATO - Australian Taxation Office. 2017b. “Income Tax Assessment Act 1997 No. 38.
Chapter 3 Specialist liability rules: Division 394: Forestry managed investme nt
schemes. Vol 7.” :343–351. Available at:
https://www.legislation.gov.au/Details/C2017C00155/Html/Volume_7 [Accessed
July 10, 2017].

20
Chapter 1. Introduction

ATO - Australian Taxation Office. 2000. “Taxation Ruling TR 2000/8.” :79. Availab le
at:
http://law.ato.gov.au/atolaw/view.htm?locid=%27TXR/TR20008/NAT/ATO%27&
PiT=20010516000001 [Accessed May 20, 2016].
Australian Greenhouse Office. 2000. “Land Clearing: A Social History - Technica l
Report No. 4.”
Barron, O., R. Silberstein, R. Ali, R. Donohue, D.J. McFarlane, P. Davies, G. Hodgson,
N. Smart, and M. Donn. 2012. “Climate change effects on water-dependent
ecosystems in south-western Australia.” Journal of Hydrology 434–435:95–109.
Bockstael, N.E. 1996. “Modelling Economics and Ecology: The Importantance of a
Spatial Perspective.” American Journal of Agricultural Economics 78(5):1168–
1180.
Bradshaw, C.J.A. 2012. “Little left to lose: Deforestation and forest degradation in
Australia since European colonization.” Journal of Plant Ecology 5(1):109–120.
Braimoh, A.K. 2004. “Modeling land-use change in the Volta Basin of Ghana.” Ecology
and Development Series No. 14:175.
Braimoh, A.K. 2006. “Random and systematic land-cover transitions in northern Ghana.”
Agriculture, Ecosystems & Environment 113(1–4):254–263.
Brown, C., C. Trusler, and K. Davis. 2010. “Managed Investment Scheme regulatio n:
Lessons from the Great Southern failure.” Jassa - The Finsia Journal of Applied
Finance (2):23–28.
Brunger, A.G., and J. Selwood. 1997. “Settlement and land alienation in western
Australia: the shire of Denmark.” Journal of Historical Geography 23(4):478–495.
Bryan, B.A. 2013. “Incentives, land use, and ecosystem services: Synthesizing complex
linkages.” Environmental Science and Policy 27:124–134.
Burvill, G. H. 1979. “The development of light lands”. In Agriculture in Western
Australia, ed. G. H. Burvill. Nedlands: University of Western Australia Press.
Chomitz, K.M., and D.A. Gray. 1996. “Roads, land use, and deforestation: A spatial
model applied to Belize.” The World Bank Economic Review 10(3):487–512.
Collins, P.D.K., and D.F. Barrett. 1980. “Shannon, Warren and Donnelly River Basins -
water resources survey - Report No. WRB 6.”
Commonwealth of Australia. 2002. “Plantations for Australia : The 2020 Vision. ”
Canberra, Australia:24. Available at: http://www.agriculture.gov.au/S tyle
Library/Images/DAFF/__data/assets/pdffile/0009/2398185/plantations-australia-
2020-vision.pdf [Accessed April 10, 2015].
Commonwealth of Australia. 1995. Wood and paper industry strategy. Camberra.
Department of Water. 2012. “Warren–Donnelly surface water allocation plan methods
report.”

21
Chapter 1. Introduction

Etter, A., C. McAlpine, D. Pullar, and H. Possingham. 2006. “Modelling the conversion
of Colombian lowland ecosystems since 1940: drivers, patterns and rates.” Journal
of environmental management 79(1):74–87.
Gavran, M. 2014. “Australian plantation statistics 2014 update. Technical report 14.2.”
:14. Available at:
http://data.daff.gov.au/data/warehouse/aplnsd9ablf002/aplnsd9ablf201409/AustPla
ntationStats_2014_v.1.0.0.pdf [Accessed February 21, 2016].
George, S.J., R.J. Harper, R.J. Hobbs, and M. Tibbett. 2012. “Agriculture , Ecosystems
and Environment A sustainable agricultural landscape for Australia : A review of
interlacing carbon sequestration , biodiversity and salinity management in
agroforestry systems.” Agriculture, Ecosystems and Environment 163:28–36.
Government of Western Australia. 2015. “Country Areas Water Supply Act 1947.
Control of catchment areas Part IIA.” :19. Available at:
https://www.slp.wa.gov.au/pco/prod/filestore.nsf/FileURL/mrdoc_25672.pdf/$FIL
E/Country Areas Water Supply Act 1947 - %5B10-00-00%5D.pdf?OpenEle me nt
[Accessed December 19, 2015].
Government of Western Australia. 1996. Western Australian Salinity Action Plan. Perth,
Australia: Prepared by Agriculture Western Australia, Department of Conservation
and Land Management, Department of Environmental Protection and Water and
Rivers Commission.
Graham, T., D.J. Pannell, and B. White. 2010. “On-site and off-site economic benefits of
dryland salinity mitigation resulting from establishment of perennial vegetation on
farms: a breakeven analysis.” Australasian Journal of Environmental Management
17(2):112–124.
Hobbs, T.J., C.R. Neumann, W.S. Meyer, T. Moon, and B.A. Bryan. 2016. “Models of
reforestation productivity and carbon sequestration for land use and climate change
adaptation planning in South Australia.” Journal of Environmental Management
181:279–288.
Irwin, E.G., and J. Geoghegan. 2001. “Theory, data, methods: developing spatially
explicit economic models of land use change.” Agriculture, Ecosystems &
Environment 85:7–23.
Kotsiantis, S.B. 2007. “Supervised machine learning: A review of classifica tio n
techniques.” Informatica 31:249–268.
Lambin, E.F. 1997. “Modelling and monitoring land-cover change processes in tropical
regions.” Progress in Physical Geography 21(3):375–393.
Lambin, E.F., H.J. Geist, and E. Lepers. 2003. “Dynamics of land-use and land-cover
change in tropical regions.” Annual Review of Environment and Resources
28(1):205–241.
Lambin, E.F., and P. Meyfroidt. 2010. “Land use transitions: Socio-ecological feedback
versus socio-economic change.” Land Use Policy 27(2):108–118.

22
Chapter 1. Introduction

Lambin, E.F., B.L. Turner, H.J. Geist, S.B. Agbola, A. Angelsen, J.W. Bruce, O.T.
Coomes, R. Dirzo, G. Fischer, C. Folke, P.S. George, K. Homewood, J. Imbernon,
R. Leemans, X. Li, E.F. Moran, M. Mortimore, P.S. Ramakrishnan, J.F. Richards,
H. Skånes, W. Steffen, G.D. Stone, U. Svedin, T.A. Veldkamp, C. Vogel, and J. Xu.
2001. “The causes of land-use and land-cover change: Moving beyond the myths.”
Global Environmental Change 11(4):261–269.
Lin, Y., X. Deng, X. Li, and E. Ma. 2014. “Comparison of multinomial logistic regression
and logistic regression: which is more efficient in allocating land use?” Frontiers of
Earth Science 8(4):512–523.
Lu, D., and Q. Weng. 2007. “A survey of image classification methods and techniques
for improving classification performance.” International Journal of Remote Sensing
28(5):823–870.
Lunt, I.D. 1998. “Two hundred years of land use and vegetation change in a remnant
coastal woodland in wouthern Australia.” Australian Journal of Botany 46(6):629–
647.
Mackarness, P., and B. Malcolm. 2006. “Public policy and managed investment schemes
for hardwood plantations.” Extension Farming System Journal 2(1):105–116.
Manandhar, R., I.O.A. Odeh, and R.G. Pontius. 2010. “Analysis of twenty years of
categorical land transitions in the Lower Hunter of New South Wales, Australia. ”
Agriculture, Ecosystems & Environment 135(4):336–346.
Marcos-Martinez, R., and K.A. Baerenklau. 2015. “Controlling for misclassified land use
data: A post-classification latent multinomial logit approach.” Remote Sensing of
Environment 170:203–215.
Melgani, F., and L. Bruzzone. 2004. “Classification of hyperspectral remote sensing
images with support vector machines.” IEEE Transactions on Geoscience and
Remote Sensing 42(8):1778–1790.
Meyer, W.B., and B.L. Turner. 1992. “Human population growth and global land-use.”
Annual Review of Ecology and Systematics 23:39–61.
Munro, J. 2006. “Lower Warren River Action Plan.” Manjimup Land Conservation
District Committee.
Myers, N., R.A. Mittermeier, C.G. Mittermeier, G.A.B. da Fonseca, and J. Kent. 2000.
“Biodiversity hotspots for conservation priorities.” Nature 403(6772):853–858.
NewForests. 2015. “Rationalising Timberland Managed Investment Schemes: The
changing landscape of Australia’s forestry investment sector.” :1–4. Available at:
https://www.newforests.com.au/wp-content/uploads/2015/06/New-Forests-MIS-
Review.pdf [Accessed December 20, 2015].
Otukei, J.R., and T. Blaschke. 2010. “Land cover change assessment using decision trees,
support vector machines and maximum likelihood classification algorithms. ”
International Journal of Applied Earth Observation and Geoinformation 12(1):27–
31.

23
Chapter 1. Introduction

Paiboonvorachat, C., and T.J. Oyana. 2011. “Land-cover changes and potential impacts
on soil erosion in the Nan watershed, Thailand.” International Journal of Remote
Sensing 32(21):6587–6609.
Pannell, D.J., and A.M. Roberts. 2010. “Australia’s National Action Plan for salinity and
water quality: A retrospective assessment.” Australian Journal of Agricultural and
Resource Economics 54(4):437–456.
Plantinga, A.J., and S. Ahn. 2002. “Efficient policies for environmental protection: An
econometric analysis of incentives for land conversion and retention.” Journal of
Agricultural and Resource Economics 27(1):128–145.
Plantinga, A.J., and J. Wu. 2003. “Co-Benefits from carbon sequestration in forests:
evaluating reductions in agricultural externalities from an afforestation policy in
Wisconsin.” Land Economics 79(1):74–85.
Platt, J. 2007. “Warren River - Revised management options. Why update the Salinity?”
Polyakov, M., I. Majumdar, and L. Teeter. 2008. “Spatial and temporal analysis of the
anthropogenic effects on local diversity of forest trees.” Forest Ecology and
Management 255(5–6):1379–1387.
Polyakov, M., and D. Zhang. 2008. “Property Tax Policy and Land-Use Change.” Land
Economics 84(3):396–408.
Ricardo, D. 1817. The principles of political economy and taxation. London, United
Kingdom: John Murray.
Ridley, A.M., and D.J. Pannell. 2005. “The role of plants and plant-based research and
development in managing dryland salinity in Australia.” Australian Journal of
Experimental Agriculture 45(11):1341–1355.
Sanders, D. 2005. From Colonial Outpost to Popular Tourism Destination: an Historical
Geography of the Leeuwin-Naturaliste Region 1829-2005, Unpublished Doctoral
thesis. Murdoch University.
Sesnie, S.E., B. Finegan, P.E. Gessler, S. Thessler, Z.R. Bendana, and A.M.S. Smith.
2010. “The multispectral separability of Costa Rican rainforest types with support
vector machines and Random Forest decision trees.” International Journal of
Remote Sensing 31(11):2885–2909.
Shang, X., and L.A. Chisholm. 2014. “Classification of Australian native forest species
using hyperspectral remote sensing and machine- learning classification algorithms. ”
IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
7(6):2481–2489.
Shao, Y., and R.S. Lunetta. 2012. “Comparison of support vector machine, neural
network, and CART algorithms for the land-cover classification using limited
training data points.” ISPRS Journal of Photogrammetry and Remote Sensing 70:78–
87.
Smith, M.G., R.N.M. Dixon, L.H. Boniecka, M.L. Berti, T. Sparks, M.A. Bari, and J.
Platt. 2006. “Salinity Situation Statement: Warren River. Water Resource Technica l
24
Chapter 1. Introduction

Series No. WRT 32.”


State Salinity Council. 2000. “Natural Resource Management in Western Australia: The
Salinity Strategy.”
The Treasury. 2008. “Review of Non-Forestry Managed Investment Schemes.” Availab le
at: https://archive.treasury.gov.au/documents/1401/PDF/No n-
Forestry_Managed_Investment_Schemes_Issues_Paper.pdf [Accessed December 1,
2016].
Thünen, J.H. von. 1966. Von Thünen’s isolated State: an English edition of Der isoliere
Staat 1st ed. Oxford, United States: Pergamon Press.
Townsend, P. V., R.J. Harper, P.D. Brennan, C. Dean, S. Wu, K.R.. J. Smettem, and S.E.
Cook. 2012. “Multiple environmental services as an opportunity for watershed
restoration.” Forest Policy and Economics 17:45–58.
Veldkamp, A., and E.F. Lambin. 2001. “Predicting land-use change.” Agriculture,
Ecosystems & Environment 85(1–3):1–6.
Wallace, K.J. 2001. “State Salinity Action Plan 1996: Review of the Department of
Conservation and Land Management ’ s programs. January 1997 to June 2000.”
2001:179. Available at: https://www.dpaw.wa.gov.au/images/conservatio n-
management/salinity/salinity_report_june2001.pdf [Accessed May 22, 2015].
Wardlow, B.D., and S.L. Egbert. 2008. “Large-area crop mapping using time-series
MODIS 250m NDVI data: An assessment for the U.S. Central Great Plains.” Remote
Sensing of Environment 112(3):1096–1116.
Wu, X., Z. Shen, R. Liu, and X. Ding. 2008. “Land Use/Cover Dynamics in Response to
Changes in Environmental and Socio-Political Forces in the Upper Reaches of
Yangtze River, China.” Sensors 8(12):8104–8122.
Zheng, B., S.W. Myint, P.S. Thenkabail, and R.M. Aggarwal. 2015. “A support vector
machine to identify irrigated crop types using time-series Landsat NDVI data.”
International Journal of Applied Earth Observation and Geoinformation 34(1):103–
112.

25
Chapter 2. Discriminating native and plantation forests
Chapter 2. Discriminating native and plantation forests

2. DISCRIMINATING NATIVE AND PLANTATION FORESTS IN

A LANDSAT TIME-SERIES FOR LAND USE POLICY DESIGN

Abstract

The Warren River Catchment of south-western Australia is an area of high biodivers ity

threatened by the loss of native vegetation and dryland salinity. Over the last 20 years it

has been the target of a series of policies that encourage conversion of agricultural land

to plantation forest. Remote sensing has a key role in measuring trends in the area of

plantation forest observed across the landscape and hence the effectiveness of policy

initiatives. Despite its importance to land use policy, accurate data on historical land use

and land cover (LULC) dynamics of two spectrally similar but ecologically distinct forest

types such as native forest and plantation forest are not readily available for south-western

Australia, largely due to prohibitive data delivery costs. However, we argue that regular

low cost monitoring of long-term change in the spatial distribution of plantation forest

through remote sensing is a critical input into environmental policy for the catchment. To

this end, a 35 year time-series of Landsat imagery was acquired, and three differe nt

classifiers were tested (Support Vector Machines – SVM; Random Forests – RF; and

Classification and Regression Trees – CART) on spectral and textural indices applied to

four spectral bands. The six major LULC classes considered were agriculture, water,

native forest, sand dunes, plantation forest and harvested native forest. In classifying the

imagery the SVM and RF outperformed the CART across all classes. However, the SVM

classifier gave a slightly higher F-score for most individual classes than the RF. Eucalypt

dominated plantation forest reaching full canopy cover was subject to the highest rates of

27
Chapter 2. Discriminating native and plantation forests

misclassification inasmuch as they share spectral properties with the Eucalypt dominant

native forest. When applied to Landsat time-series imagery, SVM classifier combined

with four bands held in common between the four Landsat Sensors (MSS 1979; TM 1992;

ETM+ 2003; OLI 2014), and derived textures metrics are valuable in classifying

plantation and native forest, particularly where these share similar species. The

differences in prediction accuracy when including additional Landsat bands were not

statistically significant, as demonstrated by the McNemar test. Thus, we achieved a trade-

off in reducing processing time without significantly impacting on classification accuracy

(≥86%). The relatively high accuracy of the proposed method enables the effects of past

policy initiatives to be observed, and hence the efficient design of environmental and

conservation policy in future.

Keywords: Land Use Policy; Plantation Forest; Support Vector Machines; Landsat;

LULC Classification.

2.1. INTRODUCTION

Sustainable land management is critical to protect ecosystems and provide ecosystem

services (Lambin et al. 2001; Lambin, Geist and Lepers 2003; Manandhar, Odeh and

Pontius 2010). An essential element of an effective land management policy is Land Use

and Land Cover (LULC) monitoring. Of the tools available, there is a scientific consensus

that satellite data provide the most cost-effective and comprehensive means of observing

LULC change. There are of course many different types of satellite sensors, whose

capabilities to monitor the land surface have been increasing rapidly in recent years.

28
Chapter 2. Discriminating native and plantation forests

A common problem in LULC monitoring is the difficulty to distinguish between

spectrally similar classes (Marcos-Martinez and Baerenklau 2015; Chan and Paelinckx

2008). Very high-resolution (<5 m resolution) and/or hyperspectral imagery (>10 spectral

bands) can provide better discrimination than medium-resolution imagery (30m

resolution; Govender, Chetty, and Bulcock 2007; Melgani and Bruzzone 2004). However,

these technologies are relatively recent and do not provide the length of time-series or

global coverage required for a historical analysis of the slow rates of LULC change found

in agricultural regions such as the Warren River Catchment in south-western Australia.

Instead, the Landsat family of moderate resolution sensors (4-11 bands, 30m resolutio n)

provides the longest and most consistent archive of historical imagery, and so this archive

is often the only one available to measure long-term LULC changes (Wulder et al. 2016;

Sesnie et al. 2010).

The sensors impose constraints on how well LULC (e.g., commercial plantation forest

and native forest) with similar spectral signatures can be differentiated, as the sensors

map the full spectral signature into a finite number of bands, thereby limiting the level of

differentiation that can practically be achieved. Whether this ‘upper’ constraint on

achievable classification accuracy is realised depends much on the application of Machine

Learning Algorithms – MLA (or classifiers). In general, classifiers do not make

distributional assumptions about the data, and have been shown to significantly reduce

computational effort in developing class predictions when data measurement spaces are

large and complex (Foody 2003). See Kotsiantis (2007) for a general review of these

algorithms and Lu and Weng (2007) and Rogan et al. (2008) for a review of their

application to LULC assessments. Recent research has demonstrated that non-parametric

classifiers (e.g., Support Vector Machines – SVM; Random Forests – RF; and

29
Chapter 2. Discriminating native and plantation forests

Classification and Regression Trees – CART) compare favourably with more established

parametric classifiers such as maximum likelihood (Otukei and Blaschke 2010; Rogan et

al. 2008; Shao and Lunetta 2012; Szuster, Chen and Borger 2011). While SVM, RF and

CART have been applied in a wide range of classification problems (Adam et al. 2014;

Herold, Koeln and Cunnigham 2003; Lemon et al. 2003; Sharma, Ghosh and Joshi 2013;

Vayssières, Plant and Allen-Diaz 2000; Wardlow and Egbert 2008), it is only recently

that they have been trialled for discriminating spectrally similar land cover classes from

medium resolution data (Brandt et al. 2012; Rodriguez-Galiano et al. 2012).

Most tests of SVM, RF and CART to date have been applied to either only a single time

period (Li et al. 2014), or to analyse changes in cropping patterns (Devadas, Denham and

Pringle 2012; Löw et al. 2013; Wardlow and Egbert 2008; Zheng et al. 2015; Shao and

Lunetta 2012), and natural vegetation classes (Singh et al. 2014; Im, Beier and Li 2013).

Few previous published studies focusing on the classification of plantation forest has been

mostly constrained to a single set of imagery (Sesnie et al. 2010), to high resolutio n

imagery of limited spatial extent (Fagan et al. 2015; Rapinel et al. 2014; Adam et al.

2014), or not include the most recent Landsat OLI sensor (Brandt et al. 2012). As to our

knowledge, there have been no studies focused on the performance of these algorithms in

classifying Eucalyptus plantation with low inter-class separability from its surrounding

native forest, as captured by Landsat time-series imagery (i.e., MSS; TM; ETM+; OLI).

Consequently, this study sought to explore the complex process associated with mapping

a Eucalypt dominated plantation forest from adjacent Eucalypt native forest across the

landscape and through time.

LULC classes that are difficult to distinguish may be important to regional and local land

use policy. This is the case for the Mediterranean landscapes of south-western Australia,

30
Chapter 2. Discriminating native and plantation forests

a Global Biodiversity Hotspot that hosts a large number of endemic and threatened plant

species (Myers et al. 2000) and sustained large-scale destruction and fragmentation of

forest cover over the past 60 years (Bradshaw 2012). Regions within this Mediterranean

biome, such as the Warren River Catchment (WRC), were widely cleared for agriculture

in the 1950s (Smith et al. 2006). Consequently, these regions have been the focus of a

number of policy initiatives to primarily address the resulting biodiversity loss and land

degradation (e.g., Country Areas Water Supply Act 1947 in 1978 and Water Resource

Recovery Catchments in 1996). More recently, pulpwood plantations were established on

private land across the WRC following the introduction of government tax subsidies (e.g.,

Plantations for Australia – The 2020 Vision in 1997; and The Managed Investments Act

1998 – MIA 1998). A secondary objective of the tax scheme in promoting plantatio ns

(predominantly Eucalyptus globulus - Tasmanian bluegum) was to reduce stream salinity

and, thereby enhance the ecosystem health of neighbouring Eucalypt dominated forest

and aquatic ecosystem (Smith et al., 2006). Measuring the LULC change observed under

the policy initiatives may be used to inform future land use policy for the region. Since

the majority of plantations were established on private lands, they were not systematica lly

mapped by the government agencies. Most LULC change maps developed for Australia

at regional scales (ABARES 2016) do not provide consistent monitoring of a region

through time (Caccetta et al. 2012), and thus are not suited to studying the historica l

development of a plantation forest locally. For instance, the availability of detailed land

use maps in Australia was limited before the Australian Collaborative Land Use Mapping

Program was established in 2000.

Despite the limitations of remote sensing data (given spectral, spatial, and radiometr ic

resolutions), its use is growing in environmental policy analysis (e.g., Pandit, Polyakov,

31
Chapter 2. Discriminating native and plantation forests

and Sadler 2014; White and Sadler 2012; Martinez et al. 2017; Summers et al. 2015) and

there is a need to develop robust classifiers to make important distinctions between

spectrally similar but ecologically distinct LULCs. Hence, the main objective of this

paper is to classify 35 years of LULC change in the Warren River Catchment. In

undertaking the classification we will assess: (i) the ability of Landsat time-series imagery

to discriminate plantation from native forest; (ii) the consistency and accuracy of three

different classifiers (SVM, RF and CART) in addressing the low inter-class separability

between native and plantation forest; (iii) different combinations of spectral bands for

enhancing the classification accuracy; and, (iv) the possible implications of classifier

performance for land management policy. The open-source methods applied here are

sufficiently general to have potential application to similar scenarios elsewhere.

2.2. STUDY AREA AND DATA SOURCES

2.2.1. Study Area

The Warren River has a mean annual flow of 291 Gl (1990–2001; Smith et al. 2006;

Figure 2.1), and annual rainfall ranging from 550 mm/year in the north-east of the

catchment to over 1200 mm/year in the coastal south-west (Townsend et al. 2012). The

WRC covers 4,416 km2 , and includes the small towns of Manjimup and Pemberton. At

the 2011 census, these towns had a population in the urban centre of 4,164 and 777

respectively (ABS 2011).

32
Chapter 2. Discriminating native and plantation forests

Figure 2.1. Location of the study area within the coastal range of Western Australia.

Extensive LULC change intensified in the 1950s following the War Service Land

Settlement Scheme and was concentrated in the upper (north-east) Warren catchment.

This LULC change consisted primarily of the conversion of native woodlands, scrublands

and heath to rain-fed cereal production, followed by sheep and cattle pastures. The

primary land use in the lower (south-west) Warren catchment has, for over a hundred

years, been the selective logging and clear-felling of native forest, and significa nt

horticultural production around Manjimup and Pemberton. However, with increasing

environmental concerns over the loss of native vegetation and increasing dryland salinity

in the upper catchment, the State Government of Western Australia identified the Warren

River as a potential source of potable water. Hence the State Government enacted clearing

control legislation in 1978 (Country Areas Water Supply Act – Part IIA; Government of

33
Chapter 2. Discriminating native and plantation forests

Western Australia 2015) to prevent further loss of native vegetation, and thereby reduce

the threat of salinization of the river. In the 1990s, large-scale tree planting had

commenced on previously cleared lands in response to new tax incentives provided within

managed investment schemes. These tax incentives had largely been withdrawn by the

late 2000s. Logging of native forest is enacted through the Conservation and Land

Management Act (WA), and through the Western Australian Regional Forest Agreement

(RFA) since 1999.

2.2.2. Data Sources

An image series was constructed over the study area from four Landsat sensors: Landsat

3 Multispectral Scanner (MSS), Landsat 5 Thematic Mapper (TM), Landsat 7 Enhanced

Thematic Mapper (ETM+) and Landsat 8 Operational Land Imager (OLI; A in Figure

2.2). The Landsat imagery were obtained from the United States Geological Survey

(USGS) Centre for Earth Resources Observation and Science (EROS). Up to two scenes

were selected from the World Reference System (WRS2; Table 2.1). Cloud-free images

taken in the summer (December to February) were selected to maximize the potential

separation between native forest, commercial plantation forest and dryland agriculture.

Other reference data (A in Figure 2.2) includes: a digital elevation model (DEM) from

the Shuttle Radar Topography Mission (SRTM; 1 Arc-Second Global available by

USGS); the Australian Land Use and Management Classification Maps (ALUM;

1:1000,000 scale; ABARES 2016); plantation forest maps from the Department of Water

WA and Forest Products Commission WA from 1978 onwards (accuracy mostly of +/-

15 m from ortho-photo updates); and, aerial photographs taken in 2000 (10m resolutio n),

34
Chapter 2. Discriminating native and plantation forests

2004, 2013 and 2015 (5m resolution) from Landgate's Shared Location Informatio n

Platform (SLIP).

Table 2.1. Satellite data for the study area.

Acquisition data Sensor Path/Row Resolution


(meters)
10/01/1979 MSS 119/84 60
30/01/1992 TM 112/84 30
07/01/1992 111/84
05/02/2003 ETM+ 112/84 30
13/02/2003 111/84
27/02/2014 OLI 112/84 30
04/02/2014 111/84

2.3. METHODS

2.3.1. Imagery Pre-Processing

All Landsat images underwent atmospheric correction to convert digital number values

into surface reflectance using dark-object subtraction (DOS 4; Moran et al. 1992).

Topographic correction was through the Minnaert algorithm as applied to the SRTM

elevation data (Moran et al. 1992; B in Figure 2.2). Both correction methods were applied

using GRASS GIS (Geographic Resources Analysis Support System; GRASS 2015).

After these corrections, the MSS image was resampled from the original 60m to a 30m

spatial resolution. From a statistical and policy analysis perspective, it is preferable that

the data are as consistent across the different sensors as possible.

35
Chapter 2. Discriminating native and plantation forests

Figure 2.2. Flow diagram of pre-processing, processing, classification and evaluation of


classifier performance.

36
Chapter 2. Discriminating native and plantation forests

Hence, firstly, only the four bands from each sensor sharing similar spectral ranges to the

MSS sensor were considered (Model C1 and C4 in Figure 2.2). Then, Model C1 was

augmented by grey-level co-occurence matrix (GLCM) derived textural metrics applied

to those bands (Haralick, Shanmugam and Dinstein 1973; Hurni et al. 2013; Lu and Weng

2007), as described in section 2.3.2.2 below. The accuracy of the four-band classificatio ns

was then compared to classifications applied to imagery with up to 7 Landsat bands (i.e.,

bands 1 to 5 and 7 of the TM and ETM+ sensors, and bands 1 to 7 of the Landsat OLI;

Table 2.2), including derived textures metrics (Model C2 in Figure 2.2). To suppleme nt

these raw Landsat bands, the Normalized Difference Vegetation Index (NDVI) was

included to examine the ability of NDVI to differentiate two spectrally similar classes

(i.e., native and plantation forest; Model C3 – D in Figure 2.2), over and above the non-

linear transforms applied to the raw bands by the SVM and tree-based classifiers. The

steps in image pre-processing, sampling and classification are summarized in Figure 2.2.

Table 2.2. Landsat spectral bands, principal components analysis (PCA) of texture
features based on grey-level co-occurrence matrices, and NDVI used in each model.

Model Spectral Bands 5 PCAs NDVI Total of


Textures Bands used
MSS 4–7 Yes No 9
C1 TM / ETM+ 1–4 9
OLI 2–5 9
TM / ETM+ 1 – 5, 7 Yes No 11
C2 OLI 1–7 12
MSS 4–7 Yes Yes 10
C3 TM / ETM+ 1 – 5, 7 12
OLI 1–7 13
MSS 4–7 No No 4
C4 TM / ETM+ 1–4 4
OLI 2–5 4

37
Chapter 2. Discriminating native and plantation forests

2.3.2. Imagery Classification

2.3.2.1. Unsupervised Classification

Initially, an unsupervised classification based on the ISODATA algorithm (a Self-

Organizing Data Analysis Technique; Ball and Hall 1967) was performed with a 95%

convergence threshold, to identify potential target LULC classes (E in Figure 2.2). Our

target classes were then utilised in our stratified random sampling strategy for test and

training data to ensure an appropriate coverage over the range of possible spectral values.

Twenty unsupervised classes were allowed, with each class assigned to one of the six

targeted LULC classes through visual inspection of aerial photos and overlay of a coarse-

scale land use map (Table 2.3). The plantation forest class was not present in the MSS

1979 imagery, with the 1979 classification providing a base-line reference to the other

classifications containing plantation forestry. Urban areas were not included as a target

class as they represented a small proportion of the landscape (<0.4%). The sampling

strategy is discussed in detail in the Appendix A2.

Table 2.3. Land use classification system.

Class Description
Agriculture Cleared areas for pasture or cropping.
Water Water bodies including ephemeral water in lakes, rivers and dams.
Native Forest Indigenous plants (e.g., trees and shrubs).
Sand Dunes Ridge of sand especially on the coast.
Plantation Forest Commercial plantation trees consisted of closed monocultures
Harvested Native Forest Native forest recently cleared for timber extraction (clearcut logging) or to
improve forest health.

38
Chapter 2. Discriminating native and plantation forests

2.3.2.2. Texture Measurement

To increase the separability between LULC classes, texture metrics were calculated for

each model (i.e., Models C1 to C3) using grey-level co-occurrence matrices (GLCM;

Zvoleff 2015; G in Figure 2.2), computed using R (R Core Team 2016). The seven texture

metrics applied are described in Table 2.4, and in more detail by Haralick et al. (1973).

Table 2.4. Texture measurements extracted from the Landsat spectral bands. The
subscripts 𝒊 and 𝒋 give the horizontal and vertical coordinates of the co-occurrence matrix
respectively (i.e., a Moore neighbourhood), p is the cell value and µ is the mean cell value
over a local window (Haralick et al. 1973).

Texture measures Equation

Mean 𝑀𝐸 = ∑ 𝑖(𝑝𝑖,𝑗 ) (2.1)


𝑖,𝑗

Variance 𝑉𝐴 = ∑(𝑖 − 𝜇) 2 𝑝(𝑖, 𝑗) (2.2)


𝑖,𝑗

Homogeneity 1 (2.3)
𝐻𝑂 = ∑ 𝑝(𝑖, 𝑗)
1 − 𝑖 − 𝑗) 2
(
𝑖,𝑗

Contrast 𝐶𝑂 = ∑|𝑖 − 𝑗|2 𝑝(𝑖, 𝑗) (2.4)


𝑖,𝑗

Dissimilarity 𝐷𝐼 = ∑|𝑖 − 𝑗|𝑝(𝑖, 𝑗) (2.5)


𝑖,𝑗

Entropy 𝐸𝑁 = ∑ 𝑝 (𝑖, 𝑗) 𝑙𝑜𝑔(𝑝(𝑖, 𝑗)) (2.6)


𝑖,𝑗

Second Moment 𝑆𝑀 = ∑{𝑝(𝑖, 𝑗)}2 (2.7)


𝑖,𝑗

39
Chapter 2. Discriminating native and plantation forests

Each set of texture images was calculated for 3×3 (90 x 90m) and 7×7 (210 x 210m)

window sizes to potentially capture texture at two different level of details (~1 ha and ~4

ha, respectively). Textural metrics captured at different scales have been shown to

improve low inter-class separability and consequently, increase classification accuracy

(Hurni et al. 2013; Rodriguez-Galiano et al. 2012). Small windows sizes provide a more

accurate representation of the local variance, allowing us to capture plantations of

different spatial extent. The texture calculation generated 14 additional bands for each

spectral band per capture date. Since the dimensionality of the imagery was significa ntly

increased by applying textures, the texture metrics were reduced to the first five

components of a principal component analysis (PCA; G in Figure 2.2), representing more

than 98% of the total variation across the textural bands. These five components were

then stacked with the raw spectral bands (Table 2.2), on which the image segmentatio n

was conducted.

2.3.2.3. Image Segmentation

An object-based image analysis (OBIA) was undertaken to segment the landscape into

homogeneous objects that represent real-world landscape elements (H in Figure 2.2). The

segmentation was conducted through mean shift clustering, as applied in the Large-Scale

Mean Shift segmentation workflow (LSMS) provided by the Orfeo Toolbox (Michel,

Youssefi, and Grizonnet 2015; OTB 2016). The segmentation quality was evaluated by

visually comparing reference aerial photos with corresponding image segments. Then,

values for three key parameters were selected, namely the spectral range (colour), range

40
Chapter 2. Discriminating native and plantation forests

radius (pixel) and minimum region size. This single set of parameter values were applied

to the segmentation of the images at each capture date.

2.3.2.4. Machine Learning Algorithms

Model building, parameter tuning and accuracy assessments for each of the SVM, RF and

CART models were performed through the Caret package implemented in the R Software

environment (Kuhn 2016). Here, the methods 'rpart' (Recursive Partitioning and

Regression Trees) and 'rf' fit the classical Breiman’s CART and RF respectively, whereas

'svmRadial' implements a classic SVM model (K in Figure 2.2). Tuning parameters for

each algorithm were selected through a grid search seeking to minimise the prediction

error, as derived from 15-fold cross-validation (repeated 5 times with the folds selected

randomly).

SVMs generate a hyperplane that optimally splits different classes by a decision boundary

defined by a kernel function that is applied to the training data (Vapnik 2000). The

decision boundary is specified by the subset of the training sample that lies closest to the

decision boundary (i.e., the support vectors), yet which maximise s the distance between

the boundary and the support vectors in regularized space. For a Gaussian kernel two

pairs of tuning parameters, namely cost 𝐶 and the hyperparameter kernel specific functio n

Σ, need to be defined: C is a penalty parameter that controls the trade-off between training

error and model complexity; whereas Σ controls the shape of the kernel. Generally, these

parameters are found through a cross-validation procedure. Multi-class classifica tio n

problems are emulated by a voting rule, given the SVM is a binary classifier. Here, a one-

versus-one approach is adopted in which a SVM is developed for each pairwise

41
Chapter 2. Discriminating native and plantation forests

comparison between any two LULC classes; each SVM then assigns a datum to one of

two classes, termed a vote, and the class with the most votes across the different SVMs

assigns its class label to the datum (Hsu and Lin 2002). For the SVM algorithm, the final

range of tuning parameters were defined as Σ ∈ [1,4] and ∈ [1,35], with the Gaussian

radial basis function selected as the SVM kernel.

Decision trees (DT) are one of the most widely used methods for inductive inference

(Mitchell 1997). A DT is made of nodes and branches. A node is a point where a choice

must be made; whereas decision branches are generated by a node’s choice. Each branch

represents one set of possible preferences, with the set of alternatives mutually exclusive

and collectively exhaustive. There are many algorithms for constructing decision trees,

and, we apply Breiman’s Classification and Regression Trees (CART; Breiman et al.

1984). CART has been one of the most popular classifiers for MODIS and Landsat

imagery (Herold et al. 2003; Shao and Lunetta 2012). The advantage of this algorithm is

that significant predictor variables are automatically determined through entropy

reduction. To improve the CART model, the complexity parameter (cp) was tuned to

optimize the value by which splitting a node improved the relative error, with cp=0.008

selected.

A Random Forest (RF) is an ensemble of decision trees built from multiple bootstrapped

(i.e., random) samples of the training data. Bagging in RF automatically generates new

training sets by randomly resampling with replacement 63.2% of N, with N being the

number of samples in the original training set. The RF can be reduced to two tuning

parameters, namely the number of trees (ntree) and the number of predictor variables

considered for each split (mtry). Three different ntree values were tested for each Landsat

imagery (𝑛𝑡𝑟𝑒𝑒 = {100,500,1000}), along with the full set of mtry numbers (e.g.,

42
Chapter 2. Discriminating native and plantation forests

𝑚𝑡𝑟𝑦 = {1,2, . . . ,9}). Table 2.5 summarises the final input parameters selected to train

the algorithms to classify the six LULC classes.

Table 2.5. Final input parameters used to train the algorithms and the out-of-bag (OOB)
estimate of the prediction error for RFs. The CART complexity parameter (cp) was given
as cp = 0.008 across all Models.

SVM RF
Model 𝑪 mtry OOB %
MSS 24 2 7.94
C1 TM 35 4 10.46
ETM+ 34 3 17.86
OLI 34 6 19.08
TM 15 2 10.19
C2 ETM+ 19 1 19.27
OLI 32 4 17.84
MSS 30 2 8.76
C3 TM 20 1 11.05
ETM+ 35 5 18.94
OLI 32 2 15.78
MSS 35 2 12.16
C4 TM 34 2 14.44
ETM+ 35 4 25.90
OLI 34 2 20.53

An out-of-bag (OOB) estimate of the prediction error is then constructed over the

expected 36.8% of observations that will not be sampled with each resampling instance.

Through a majority vote the predictions of the individual random forest trees are

combined and a final class value is determined. The OOB estimate value was used to

select the ‘best’ model. Further, the prediction error of this ‘best’ model was

independently tested against the test set for the performance comparison of the differe nt

classifiers.

43
Chapter 2. Discriminating native and plantation forests

2.3.2.5. Accuracy Assessment

Evaluation criteria to compare the performance of the SVM, RF and CART classifiers are

required to be unbiased and independent of the method used to develop a particular model

(Kotsiantis 2007; Vayssières et al. 2000). Thus, performance of the classifiers were

evaluated on a test set that was independent of the training set on which those classifiers

were developed. The ground-truth data were randomly split into training and test sets,

comprising 80% and 20% of the available ground-truth data, respectively (J in Figure

2.2). The performance metrics applied included the Kappa coefficient, overall accuracy,

McNemar's test, F-score, the producer’s and the user’s accuracy (L in Figure 2.2; Foody

2002; Singh et al. 2014; Smits, Dellepiane, and Schowengerdt 1999). These metrics were

derived from the error (or confusion) matrices resulting from each of the LULC maps that

were produced: MSS 1979, TM 1992, ETM+ 2003, and OLI 2014. Performance metrics

take on the range of values [0, 1], where 1 indicates perfect classification performance for

all metrics apart from the p-value of the McNemar’s test.

We hypothesised that four bands held in common between the four Landsat Sensors (MSS

1979; TM 1992; ETM+ 2003; OLI 2014), and their derived textures (Model C1) are

capable of consistently classifying historical landscape transitions, particularly wher e

these share similar species. Thus, the alternative hypothesis implies that neither the

incorporation of more information (i.e., additional spectral bands or NDVI – Models C2

and C3) nor the classification of only four bands (Model C4) significantly increase

classification accuracy of a time series Landsat imagery. For this reason, we assessed the

superiority of the proposed null hypothesis against the alternative hypotheses through

McNemar’s test.

44
Chapter 2. Discriminating native and plantation forests

2.4. RESULTS AND DISCUSSION

2.4.1. Classification Performance

Landsat time-series imagery combined with RF and SVM produce satisfactory results for

monitoring historical landscape changes at regional scale (Figure 2.3; Appendix B2),

including in the case of low inter-class separability (i.e., native and plantation forest;

overall accuracy > 75%). Generally, SVM performed similarly across Models C1 to C3

(accuracy >84%), and performed slightly better than RF (accuracy >80%). CART

produced the lowest classification performance across all Models (> 69%).

Figure 2.3. Comparison of overall classification performance for CART, RF and SVM
algorithms for each Landsat sensor across Models C1 to C4.

45
Chapter 2. Discriminating native and plantation forests

At individual class levels, SVM, RF and CART varied in their ability to classify

commercial plantation forest from adjacent native forest. However, the SVM classifier

produced a slightly higher F-score for most individual classes (F-score ≥0.734 and ≥0.829

for plantation forest and native forest respectively). The F-score, and user’s and

producer’s accuracy measurements are related to commission and omission errors at the

individual class level (Appendix B2).

Both RF and SVM were able to successfully separate agriculture and sand dunes from the

other LULC types for all scenes (F-score ≥0.767 and ≥0.824 for agriculture and sand

dunes, respectively), except for Model C4. These results support previous studies that

have demonstrated the potential of SVMs to identify different crops (Bahari, Ahmad, and

Aboobaider 2014; Devadas, Denham, and Pringle 2012; Löw et al. 2013; Zheng et al.

2015).

Although the SVM classifier achieved a higher performance for most individual classes,

the overall accuracy statistics did not clearly reveal major map differences of one thematic

map over another (i.e., Models C1 to C3). Thus, the McNemar’s test was applied to

evaluate the statistical significance of the difference in accuracy performance (Appendix

C2; Foody 2004). The McNemar’s test suggests that the SVM classifications derived

from Model C1 generated significantly (p < 0.05) better performance outcomes than the

CART algorithm across all models and sensors.

For the sensors MSS 1979 and TM 1992, slightly higher overall accuracies or kappa

values were obtained with Model C1 classified with SVM. However, the McNemar’s test

did not indicate significantly different performances at the 5% significance level between

the different models, for both the RF and SVM classifiers. This suggests that includ ing

46
Chapter 2. Discriminating native and plantation forests

more spectral bands from the latest Landsat sensors, or their derivatives such as texture

or NDVI, do not increase dramatically the classification accuracy of those sensors when

discriminating between the LULC classes considered here.

The LULC maps from the ETM+ 2003 and OLI 2014 sensors classified with SVM and

Model C1 were found to have a significantly higher accuracy (McNemar’s test p < 0.05)

in comparison to their corresponding RF classifications. This confirms the consistency of

the SVM classifier in providing superior classification accuracy for spectrally similar

vegetation covers (i.e., native and plantation forest). Appropriate selection of machine

learning algorithm, rather than of more spectral bands, has been found here to be more

important in the successful classification of historical land use and land cover.

As demonstrated by the results, we achieved a useful trade-off by reducing the number of

spectral bands, and hence processing time, without significantly impacting on

classification accuracy. Sometimes a slightly improvement in performance outcome can

be obtained by using more Landsat bands (i.e., maximum of 1% enhancement of overall

accuracy and 2% kappa value for ETM+ 2003 using SVM – Model C3 instead of Model

C1). However, this marginal improvement is not statistically significant. Additionally, we

minimise the bias in any following statistical modelling that would be engendered by the

different classification accuracies arising from applying of different band combinatio ns

across multiples classifications. Most Landsat time-series studies to date have utilised the

maximal number of spectral bands available to increase classification accuracy (Ramita

Manandhar, Odeh, and Ancev 2009; Zheng et al. 2015; Muller and Zeller 2002; Zhang,

Zhengjun, and Xiaoxia 2009). However, in the historical analysis of landscape change

the consistency of the underlying data through time may be argued to be just as important

as the need for greater spectral resolution (i.e., more spectral bands) to improve

47
Chapter 2. Discriminating native and plantation forests

classification accuracy. In the simplest case, the error term in a statistical model can

account for much of the error arising from misclassification of the LULC classes, but only

if an assumption of homogeneity (i.e., consistency) of error terms through time can be

applied.

Following the widespread establishment of plantation forests during the late 1990s it is

apparent that overall classification accuracy declines, and there is an increased differe nce

in accuracy between the classifiers (Figure 2.3). For instance, user's accuracy decreased

by 0.049 for native forest from 1979 to 2014 for the SVM classifier ( Model C1), whereas

user’s accuracy for plantation forest decreased from 1.000 in 1992 to 0.797 in 2014

(Appendix B2) as the plantations from the 1990s matured (with similar trends in

classification success over time for CART and RF). The decline in performance may be

explained by the absence of the plantation forest class from the base-line MSS 1979

imagery, resulting in higher separability between the target LULC classes for the base-

line classification. Evidently, the spectral variation found within a given land use class is

a function of its intrinsic physical and spectral attributes, and of the environme nta l

context. For example, the spectral response of plantation forest in an image may vary as

a function of time since last fire, density of cover, seasonal drought stress, disease events,

topographic position and sensor view angle. In contrast, commercial plantation forest is

usually comprised of tree monocultures that are uniform across space; however, when

plantations are at maturity they can be dominated by dense-canopied trees, showing only

minor spectral differences from mature native forest, particularly when dominant species

are of the same genus (i.e., Eucalyptus). Further, E. globulus has quite different juvenile

and adult foliage, which difficult classification. Little harvesting of the plantation forest

had occurred before 2003, given commercial tree plantations were widely established in

48
Chapter 2. Discriminating native and plantation forests

the late 1990s, and the key plantation species (Tasmanian bluegums) tends to be harvested

for pulpwood on a 10–12 year rotation (Smith et al. 2006). Once plantations were

introduced into the landscape the spectral similarity of plantation and native forest has

led to greater overall misclassification.

This difficulty in discriminating between the plantation forest and native forest classes

also explains the differential performance of the machine learners, with the benefit of

deploying SVMs over RFs increasing as the separation problem becomes more diffic ult.

The RFs are constructed to minimize the overall error rate, therefore, it tends to focus

more on the prediction accuracy of the majority classes such as agriculture and native

forest (which together account for more than 90% of the landscape; Figure 2.4),

potentially masking the poor discrimination of the minority LULC classes, such as

plantation forest. We believe that SVM’s superior predictive ability was due to its

capability in selecting a more efficient weighting of the training samples in the

classification. SVMs assume that training samples lying on the class boundaries are more

spectrally similar than those lying further away, therefore those border training samples

are the most useful for SVM classification (Foody and Mathur 2004). Hence, the high

weighting attributable to the data points comprising each support vector, relative to the

RF, generates a hyperplane that more efficiently separates the plantation forest land use

from the native forest. This may explain in part the low relevance of NDVI to successful

classification in general, as at least part of the non-linearity captured by a vegetation index

may be supplanted by the non-linearity of the support vectors themselves. Machine

learners may thus be largely agnostic to the theoretical importance of vegetation indices

in informing vegetation state, if the raw bands underpinning those indices are made

49
Chapter 2. Discriminating native and plantation forests

available to the learner. Figure 2.4 shows the LULC classification maps resulted from the

best performing classifier (SVM – Model C1).

Figure 2.4. Classified imagery for each Landsat sensor (SVM – Model C1).

2.4.2. LULC Changes in the WRC

The overall pattern of the LULC change observed between 1992 and 2003 is consistent

with the introduction of the 2020 Vision and MIA 1998. Both policy initiatives were

supported by a favourable federal government tax treatment associated with forestry

Management Investment Schemes (MIS; Warman and Nelson 2016). Plantation forest

areas subsequently increased from 0.44% in 1992 to 4.36% in 2003 and 2014 due to

extensive plantations in the upper Warren. Table 2.6 presents the estimated areas of

LULC classes in WRC, using Model C1 classified with SVMs at four points in time.
50
Chapter 2. Discriminating native and plantation forests

Table 2.6. SVM estimates of historical land use and land cover in the Warren Catchment.
1979 1992 2003 2014
LULC class Area Area Area Area
km2 % km2 % km2 % km2 %
Agriculture 1326.52 30.04 1234.72 27.96 1176.53 26.64 1112.56 25.19
Water 24.52 0.56 18.15 0.41 22.86 0.52 21.54 0.49
Native Forest 2996.79 67.85 3059.00 69.26 2911.84 65.93 3023.08 68.45
Sand Dunes 38.82 0.88 30.30 0.69 35.27 0.80 25.40 0.58
Harvested Native Forest 29.86 0.68 54.98 1.24 77.44 1.75 41.41 0.94
Plantation Forest 0.00 0.00 19.36 0.44 192.58 4.36 192.53 4.36

Total area 4416.52 4416.52 4416.52 4416.52

The proportion of cleared areas for agriculture in the catchment has declined from 30%

in 1979 to 25% in 2014 (Table 2.6). The direction of change is consistent with that

reported by other authors (Collins and Barrett 1980; Smith et al. 2006), while the

magnitude of the estimates differs. Specifically, Smith et al. (2006) reported the decline

of the proportion of cleared lands from 36% in 1978 to 24% in 2000. However, their

estimates for different years come from different sources and may be inconsistent, while

our estimates are obtained by applying the same method to each snapshot in time. In

addition to an improved consistency of classification across time, the benefits of applying

machine learners within a largely automated workflow enables rapid updating of LULC

maps, at relatively low cost.

2.4.3. Implications for Land Use Policy

The historical land use and land cover dynamics are not well reported at local level,

despite their recognized importance for land use policy. This lack of information is not

51
Chapter 2. Discriminating native and plantation forests

only related to LULC changes, as Harper, Sochacki, and McGrath (2017) have described

the limitations of inventory data around carbon forestry programs in Australia.

Although considerable effort has been made to reduce stream salinity in the WRC to

potable standards (500mg/L) by 2030 (Smith et al. 2006), the State Salinity Strategy

(Government of Western Australia 1996; State Salinity Council 2000) does not seek to

refer to updated and more accurate LULC maps when designing management actions. It

is possible that local governments (e.g., Shire of Manjimup) have more detailed

information, however it is not publicly available. Plantation extent may be considered a

key performance indicator on which evidence-based policy can be formed, given the area

of plantations are a key surrogate variable for improvement in water potability (more

forest areas increase groundwater draw down, leading to a reduction in dryland salinity).

Classifying the land use class associated with the recent tax subsidy scheme is of great

significance for the design of future land use policy within this region. For instance, if the

forest cover classification is oversimplified into one broad forest class (and not separate

native and plantation forest classes) then key ecological differences that inform effective

forest and biodiversity management in the region may well be ignored. The approach

presented in this paper is suitable not only for identification of new forest plantation areas

through ongoing monitoring, but also for calculating their life span. Harvested forest areas

are readily detected within a couple of years of harvesting, even if harvested areas are

located within an agricultural land use. Estimating the rate of plantation harvest is just as

important as estimating the rate of plantation establishment for policy directed towards

timber supply and/or delivering environmental co-benefits. Hence, remote monitoring of

the spatial and temporal changes in commercial plantation forest should be considered a

crucial component of forest management at the regional scale. Only then may we

52
Chapter 2. Discriminating native and plantation forests

understand the role of plantation forest development in combatting a variety of large-scale

environmental issues, and how best to design environmental policy and tax incentives to

achieve desirable land use objectives.

2.5. CONCLUSION

Although the classification of plantation forest by itself is of great significance for the

design and application of regional and local land use policies, much of the LULC research

to date has not addressed the spectral separability of plantation forest from their

neighbouring environment. We have therefore presented methods to discriminate

plantation forest from native forest. These methods address the increasing importance of

developing robust and cost-effective methods to monitor the historical development of

commercial timber plantations. Successful discrimination required a minimal informa tio n

set when SVMs and RFs were applied, as the McNemar tests demonstrated that accessing

further spectral bands from the Landsat OLI offered little improvement in classifica tio n

success. Importantly, the SVM classifier outperformed the RF machine learner when

considering the low separability of the plantation and native forest classes. Future

improvements in classification accuracy above our baseline will likely come from higher

spatial and spectral resolution imagery. Accurate classification of LULC classes from

snapshots enables the monitoring of LULC change. Thus we consider accurate LULC

classification a necessary first step in designing efficient policy targeting both timber

supply and environmental co-benefits. Subsequent remote sensed monitoring of policy

outcomes will then embed future policy development within a strict evidence-based

framework. In Western Australia, information regarding the historical plantation forest

53
Chapter 2. Discriminating native and plantation forests

development is publicly lacking. In evaluating future plantation policy there is an urgent

need to better coordinate the existing datasets and analysis across different agencies (i.e.,

local, State and Federal governments) to improve the delivery of information on policy

outcomes.

2.6. REFERENCE

ABARES - Australian Bureau of Agricultural and Resource Economics and Sciences.


2016. “Land use data download.” Canberra ACT. Available at:
http://www.agriculture.gov.au/abares/aclump/pages/land- use/data-download.aspx
[Accessed May 20, 2016].
ABS - Australian Bureau of Statistics. 2011. “2001.0 - Census of population and housing:
Basic community profile.” Canberra, Australia. Available at:
http://www.abs.gov.au/websitedbs/censushome.nsf/home/quickstats [Accessed
September 20, 2016].
Adam, E., O. Mutanga, J. Odindi, and E.M. Abdel-Rahman. 2014. “Land-use/cover
classification in a heterogeneous coastal landscape using RapidEye imager y:
evaluating the performance of random forest and support vector machines
classifiers.” International Journal of Remote Sensing 35(10):3440–3458.
Ball, G.H., and D.J. Hall. 1967. “A clustering technique for summarizing multivar iate
data.” Behavioral science 12(2):153–155.
Bradshaw, C.J.A. 2012. “Little left to lose: Deforestation and forest degradation in
Australia since European colonization.” Journal of Plant Ecology 5(1):109–120.
Brandt, J.S., T. Kuemmerle, H. Li, G. Ren, J. Zhu, and V.C. Radeloff. 2012. “Using
Landsat imagery to map forest change in southwest China in response to the national
logging ban and ecotourism development.” Remote Sensing of Environment
121:358–369.
Breiman, L., J. Friedman, R.A. Olshen, and C.J. Stone. 1984. Classification and
regression trees. Belmont, California USA: Wadsworth Publishing Company.
Caccetta, P., C. Furby, J. Wallace, G. Richards, and R. Waterworth. 2012. “Long- term
monitoring of Australian land cover change using Landsat data: Developme nt,
implementation, and operation.” In Global Forest Monitoring from Earth
Observation. Florida USA: CRC Press, pp. 243–258.
Chan, J.C.W., and D. Paelinckx. 2008. “Evaluation of Random Forest and Adaboost tree-
based ensemble classification and spectral band selection for ecotope mapping using
airborne hyperspectral imagery.” Remote Sensing of Environment 112(6):2999–

54
Chapter 2. Discriminating native and plantation forests

3011.
Collins, P.D.K., and D.F. Barrett. 1980. “Shannon, Warren and Donnelly River Basins -
water resources survey - Report No. WRB 6.”
Devadas, R., R.J. Denham, and M. Pringle. 2012. “Support Vector Machine classifica tio n
of object-based data for crop mapping, using multi-temporal Landsat imagery.” In
ISPRS - International Archives of the Photogrammetry, Remote Sensing and Spatial
Information Sciences. Melbourne, Australia: XXII ISPRS Congress, pp. 185–190.
Fagan, M.E., R.S. DeFries, E. Sesnie, S, J.P. Arroyo-Mora, C. Soto, A. Singh, P.A.
Townsend, and R.L. Chazdon. 2015. “Mapping species compositio n of forests and
tree plantations in northeastern Costa Rica with an integration of hyperspectral and
multitemporal landsat imagery.” Remote Sensing 7(5):5660–5696.
Foody, G.M. 2002. “Status of land cover classification accuracy assessment.” Remote
Sensing of Environment 80(1):185–201.
Foody, G.M. 2003. “Uncertainty, knowledge discovery and data mining in GIS.”
Progress in Physical Geography 27(1):113–121.
Foody, G.M., and A. Mathur. 2004. “Toward intelligent training of supervised image
classifications: Directing training data acquisition for SVM classification.” Remote
Sensing of Environment 93(1–2):107–117.
Govender, M., K. Chetty, and H. Bulcock. 2007. “A review of hyperspectral remote
sensing and its application in vegetation and water resource studies.” Water SA
33(2):145–151.
Government of Western Australia. 2015. “Country Areas Water Supply Act 1947.
Control of catchment areas Part IIA.” :19. Available at:
https://www.slp.wa.gov.au/pco/prod/filestore.nsf/FileURL/mrdoc_25672.pdf/$FIL
E/Country Areas Water Supply Act 1947 - %5B10-00-00%5D.pdf?OpenEle me nt
[Accessed December 19, 2015].
GRASS. 2015. “Geographic Resources Analysis Support System (GRASS) Software.
Version 7.0.” Open Source Geospatial Foundation. Available at:
http://grass.osgeo.org [Accessed March 17, 2015].
Haralick, R.M., K. Shanmugam, and I. Dinstein. 1973. “Textural Features for Image
Classification.” IEEE Transactions on Systems, Man, and Cybernetics SMC-
3(6):610–621.
Harper, R. J., S.J. Sochacki, and J.F. McGrath. 2017. “The Development of Reforestatio n
Options for Dryland Farmland in South-Western Australia: A Review.” Southern
Forests 79(3): 185–196.
Herold, N.D., G. Koeln, and D. Cunnigham. 2003. “Mapping impervious surfaces and
forest canopy using classification and regression tree (CART) analysis.” In the
American Society for Photogrammetry and Remote Sensing (ASPRS) 2003 -
Conference Proceedings . Anchorage, Alaska. Available at:
http://nemo.uconn.edu/tools/impervious_surfaces/pdfs/Herold_etal_2003.pdf%5Cn
http://web.uconn.edu/nemo/tools/impervious_surfaces/pdfs/Herold_etal_2003.pdf

55
Chapter 2. Discriminating native and plantation forests

[Accessed December 21, 2014].


Hsu, C., and C. Lin. 2002. “A comparison of methods for multiclass support vector
machines.” Neural Networks, IEEE Transactions on 13(2):415–425.
Hurni, K., C. Hett, M. Epprecht, P. Messerli, and A. Heinimann. 2013. “A texture-based
land cover classification for the delineation of a shifting cultivation landscape in the
lao PDR using landscape metrics.” Remote Sensing 5(7):3377–3396.
Im, J., C. Beier, and M. Li. 2013. “Machine learning approaches for forest classifica tio n
and change analysis using multi-temporal Landsat TM images over Huntingto n
Wildlife Forest.” Giscience & Remote Sensing 50(4):361–384.
Kotsiantis, S.B. 2007. “Supervised machine learning: A review of classifica tio n
techniques.” Informatica 31:249–268.
Kuhn, M. 2016. “Caret: Classification and Regression Training. R package version 6.0-
70.” Available at: https://cran.r-project.org/package=caret [Accessed January 22,
2016].
Lambin, E.F., H.J. Geist, and E. Lepers. 2003. “Dynamics of land-use and land-cover
change in tropical regions.” Annual Review of Environment and Resources
28(1):205–241.
Lambin, E.F., B.L. Turner, H.J. Geist, S.B. Agbola, A. Angelsen, J.W. Bruce, O.T.
Coomes, R. Dirzo, G. Fischer, C. Folke, P.S. George, K. Homewood, J. Imbernon,
R. Leemans, X. Li, E.F. Moran, M. Mortimore, P.S. Ramakrishnan, J.F. Richards,
H. Skånes, W. Steffen, G.D. Stone, U. Svedin, T.A. Veldkamp, C. Vogel, and J. Xu.
2001. “The causes of land-use and land-cover change: Moving beyond the myths.”
Global Environmental Change 11(4):261–269.
Lemon, S.C., J. Roy, M.A. Clark, P.D. Friedmann, and W. Rakowski. 2003.
“Classification and regression tree analysis in public health: methodological review
and comparison with logistic regression.” Annals of behavioral medicine : a
publication of the Society of Behavioral Medicine 26(3):172–181.
Li, C., J. Wang, L. Wang, L. Hu, and P. Gong. 2014. “Comparison of Classifica tio n
Algorithms and Training Sample Sizes in Urban Land Classification with Landsat
Thematic Mapper Imagery.” Remote Sensing 6(2):964–983.
Löw, F., U. Michel, S. Dech, and C. Conrad. 2013. “Impact of feature selection on the
accuracy and spatial uncertainty of per-field crop classification using Support Vector
Machines.” ISPRS Journal of Photogrammetry and Remote Sensing 85:102–119.
Lu, D., and Q. Weng. 2007. “A survey of image classification methods and techniques
for improving classification performance.” International Journal of Remote Sensing
28(5):823–870.
Manandhar, R., I.O.A. Odeh, and R.G. Pontius. 2010. “Analysis of twenty years of
categorical land transitions in the Lower Hunter of New South Wales, Australia. ”
Agriculture, Ecosystems & Environment 135(4):336–346.
Marcos-Martinez, R., and K.A. Baerenklau. 2015. “Controlling for misclassified land use
data: A post-classification latent multinomial logit approach.” Remote Sensing of
56
Chapter 2. Discriminating native and plantation forests

Environment 170:203–215.
Marcos-Martinez, R., B.A. Bryan, J.D. Connor, and D. King. 2017. “Land Use Policy
Agricultural land-use dynamics : Assessing the relative importance of
socioeconomic and biophysical drivers for more targeted policy.” Land Use Policy
63:53–66.
Melgani, F., and L. Bruzzone. 2004. “Classification of hyperspectral remote sensing
images with support vector machines.” IEEE Transactions on Geoscience and
Remote Sensing 42(8):1778–1790.
Michel, J., D. Youssefi, and M. Grizonnet. 2015. “Stable mean-shift algorithm and its
application to the segmentation of arbitrarily large remote sensing images.” IEEE
Transactions on Geoscience and Remote Sensing 53(2):952–964.
Mitchell, T.M. 1997. “Decision Tree Learning.” In Machine Learning. New York, New
York, USA: McGraw-Hill Education, pp. 52–80.
Moran, M.S., R.D. Jackson, P.N. Slater, and P.M. Teillet. 1992. “Evaluation of simplif ied
procedures for retrieval of land surface reflectance factors from satellite sensor
output.” Remote Sensing of Environment 41(2–3):169–184.
Myers, N., R.A. Mittermeier, C.G. Mittermeier, G.A.B. da Fonseca, and J. Kent. 2000.
“Biodiversity hotspots for conservation priorities.” Nature 403(6772):853–858.
OTB - Orfeo ToolBox. 2016. “The Orfeo ToolBox Cookbook.” Available at:
https://www.orfeo-toolbox.org/CookBook/CookBook.html [Accessed June 1,
2016].
Otukei, J.R., and T. Blaschke. 2010. “Land cover change assessment using decision trees,
support vector machines and maximum likelihood classification algorithms. ”
International Journal of Applied Earth Observation and Geoinformation 12(1):27–
31.
Pandit, R., M. Polyakov, and R. Sadler. 2014. “Valuing public and private urban tree
canopy cover.” Australian Journal of Agricultural and Resource Economics
58(3):453–470.
R Core Team. 2016. “R: A language and environment for statistical computing. R
Foundation for Statistical Computing.” Available at: https://www.r-project.o r g/
[Accessed February 13, 2016].
Rapinel, S., B. Clément, S. Magnanon, V. Sellin, and L. Hubert-Moy. 2014.
“Identification and mapping of natural vegetation on a coastal site using a
Worldview-2 satellite image.” Journal of Environmental Management 144:236–
246.
Rodriguez-Galiano, V.F., M. Chica-Olmo, F. Abarca-Hernandez, P.M. Atkinson, and C.
Jeganathan. 2012. “Random Forest classification of Mediterranean land cover using
multi-seasonal imagery and multi-seasonal texture.” Remote Sensing of Environment
121:93–107.
Rogan, J., J. Franklin, D. Stow, J. Miller, C. Woodcock, and D. Roberts. 2008. “Mapping
land-cover modifications over large areas: A comparison of machine learning
57
Chapter 2. Discriminating native and plantation forests

algorithms.” Remote Sensing of Environment 112(5):2272–2283.


Sesnie, S.E., B. Finegan, P.E. Gessler, S. Thessler, Z.R. Bendana, and A.M.S. Smith.
2010. “The multispectral separability of Costa Rican rainforest types with support
vector machines and Random Forest decision trees.” International Journal of
Remote Sensing 31(11):2885–2909.
Shao, Y., and R.S. Lunetta. 2012. “Comparison of support vector machine, neural
network, and CART algorithms for the land-cover classification using limited
training data points.” ISPRS Journal of Photogrammetry and Remote Sensing 70:78–
87.
Sharma, R., A. Ghosh, and P.K. Joshi. 2013. “Decision tree approach for classification of
remotely sensed satellite data using open source support.” Journal of Earth System
Science 122(5):1237–1247.
Singh, S.K., P.K. Srivastava, M. Gupta, J.K. Thakur, and S. Mukherjee. 2014. “Appraisal
of land use/land cover of mangrove forest ecosystem using support vector machine. ”
Environmental Earth Sciences 71(5):2245–2255.
Smith, M.G., R.N.M. Dixon, L.H. Boniecka, M.L. Berti, T. Sparks, M.A. Bari, and J.
Platt. 2006. “Salinity Situation Statement: Warren River. Water Resource Technica l
Series No. WRT 32.”
Smits, P.C., S.G. Dellepiane, and R.A. Schowengerdt. 1999. “Quality assessment of
image classification algorithms for land-cover mapping: A review and a proposal for
a cost-based approach.” International Journal of Remote Sensing 20(8):1461–1486.
Summers, D.M., B.A. Bryan, M. Nolan, and T.J. Hobbs. 2015. “The costs of
reforestation: A spatial model of the costs of establishing environmental and carbon
plantings.” Land Use Policy 44:110–121.
Szuster, B.W., Q. Chen, and M. Borger. 2011. “A comparison of classification techniques
to support land cover and land use analysis in tropical coastal zones.” Applied
Geography 31(2):525–532.
Townsend, P. V., R.J. Harper, P.D. Brennan, C. Dean, S. Wu, K.R.. J. Smettem, and S.E.
Cook. 2012. “Multiple environmental services as an opportunity for watershed
restoration.” Forest Policy and Economics 17:45–58.
Van Niel, T.G., T.R. McVicar, and B. Datt. 2005. “On the relationship between training
sample size and data dimensionality: Monte Carlo analysis of broadband multi-
temporal classification.” Remote Sensing of Environment 98(4):468–480.
Vapnik, V.N. 2000. The nature of statistical learning theory 2nd ed. New York, New
York, USA: Springer.
Vayssières, M.P., R.E. Plant, and B.H. Allen-Diaz. 2000. “Classification trees: An
alternative non-parametric approach for predicting species distributio ns.” Journal of
Vegetation Science 11(5):679–694.
Wardlow, B.D., and S.L. Egbert. 2008. “Large-area crop mapping using time-series
MODIS 250m NDVI data: An assessment for the U.S. Central Great Plains.” Remote
Sensing of Environment 112(3):1096–1116.
58
Chapter 2. Discriminating native and plantation forests

Warman, R.D., and R.A. Nelson. 2016. “Forest conservation, wood production
intensification and leakage: An Australian case.” Land Use Policy 52:353–362.
White, B., and R. Sadler. 2012. “Optimal conservation investment for a biodiversity- r ic h
agricultural landscape.” Australian Journal of Agricultural and Resource Economics
56(1):1–21.
Wulder, M.A., J.C. White, T.R. Loveland, C.E. Woodcock, A.S. Belward, W.B. Cohen,
E.A. Fosnight, J. Shaw, J.G. Masek, and D.P. Roy. 2016. “The global Landsat
archive: Status, consolidation, and direction.” Remote Sensing of Environment
185:271–283.
Zheng, B., S.W. Myint, P.S. Thenkabail, and R.M. Aggarwal. 2015. “A support vector
machine to identify irrigated crop types using time-series Landsat NDVI data.”
International Journal of Applied Earth Observation and Geoinformation 34(1):103–
112.
Zvoleff, A. 2015. “Calculate Textures from Grey-Level Co-Occurrence Matrices
(GLCMs) in R. R package version 1.2.” Available at: http://cran.r-
project.org/package=glcm [Accessed April 13, 2016].

59
Chapter 2. Discriminating Native and Plantation Forests

Appendix A2. Sampling Strategy: Two Strata Sampling.

An augmented two strata sampling strategy was adopted to form the sample of ground -

truthed data, with the segments derived from the segmentation procedure applied as the

sampling frame (termed here as 𝑆; F in Figure 2.2). The first sampling stratum (FSS)

divides the study area into six large grid cells where 𝐺𝑘 is the 𝑘 𝑡ℎgrid cell, 𝑘𝜖{1,2, … , 𝐾}

and 𝐾 = 6. The sample size for each grid cell was calculated as proportional to the

intersection between these grid cells and the study area:

𝐴𝐺 𝑘∩𝑆
𝑁𝐺𝑘 = 𝑀𝑅𝑚𝑖𝑛 (A1)
𝐴𝑆

where 𝐴 is the area (km2 ), 𝑁 is the sample size, and 𝐺𝑘 ∩ 𝑆 is the grid cell constrained to

the study region 𝑆. To simplify the notation 𝐺𝑘 ∩ 𝑆 is give as 𝐺𝑘. The total number of

segments contained within the study region 𝑆 is given by 𝑀 = 256,704. 𝑅𝑚𝑖𝑛 is a

minimum sampling rate, and set here as 𝑅𝑚𝑖𝑛 = 0.5% of all segments falling within each

target class. The FSS ensures that all land uses will be represented as evenly as possible

across the landscape thus ensuring spatial evenness, given that land uses tend to cluster

spatially while image quality may vary spatially. This reduces the risk that a land use will

be sampled predominantly from a small number of locations.

The second stratum sampling was designed to ensure spectral evenness. The 20

unsupervised classes were initially assigned to one of the six target land use classes by

visual assessment for sampling purposes. Segments from these 20 unsupervised classes

were then pooled into these preliminary target classes to select proportionately a simple

random sample from each preliminary target class within each grid cell, based on the total

area of the segments falling within each class:

60
Chapter 2. Discriminating Native and Plantation Forests

𝐴𝐶𝑆 ∩𝐺 𝑘∨𝐶𝑆 =𝑐
𝑁𝐺𝑐𝑘 = 𝑚 𝑚
𝑁𝐺𝑘 (A2)
𝐴𝐺 𝑘

where 𝐶𝑆𝑚 is the land use class of a set of segments 𝑆𝑚 indexed by 𝑚 ∈ {1,2, … , 𝑀}, with

the realised class 𝑐 ∈{Agriculture, Water, Native Forest, Sand Dunes, Plantation Forest,

Harvested Native Forest}. Hence 𝐴 𝐶𝑆𝑚∩𝐺𝑘∨𝐶𝑆𝑚=𝑐 states the total area of all segments

belonging to class 𝑐 and contained within grid cell 𝐺𝑘.

This sample was then augmented in two ways. The first augmentation was to satisfy the

recommendation of Van Niel et al. (2005) to provide a training sample of size no less

than 10-30 times the number of bands for each target class. Hence, we define a minimum

training sample size of 𝑁𝑚𝑖𝑛 = 100 and augment 𝑁𝐺𝑐𝑘 with:

𝑐
𝑁𝐺 𝑁𝑚𝑖𝑛
⌈ 𝑘
⌉ , 𝑤ℎ𝑒𝑟𝑒 ∑𝐾 𝑐
𝑘=1 𝑁𝐺𝑘 < 𝑁𝑚𝑖𝑛
𝑁𝐺𝑐∗𝑘 = { ∑𝐾 𝑐
𝑘=1 𝑁𝐺 𝑘 (A3)
𝑁𝐺𝑐𝑘 , 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

where ⌈∙⌉ denotes the ceiling operator. The second augmentation addressed the fact that

native forest and plantation forest were difficult to distinguish spectrally. The visual

quality of the land use map, produced by the supervised classification described below,

was examined in an iterative manner with respect to these two land use classes by

overlaying higher resolution aerial imagery and LULC maps. The sample size for each of

these two land use classes was increased until the resulting land use map was deemed

suitable (i.e., without an excess of 'salt-and-pepper' noise). The total sample sizes for the

training and test sets are reported in Table A1.

Ground-truthing was then conducted, through interpretation of aerial imagery, to assign

a target land use to each segment comprising the training set (I in Figure 2.2). Of the

61
Chapter 2. Discriminating Native and Plantation Forests

training set, 20% was randomly set-aside as the test set to be subjected to the out-of-box

assessment of classification accuracy.

Table A1. Total sample sizes for the training and test sets.
Training set Test set
LULC class 1979 1992 2003 2014 1979 1992 2003 2014
Agriculture 222 232 285 353 66 52 71 95
Water 126 150 110 123 33 33 28 21
Native Forest 336 388 516 528 88 98 149 146
Sand Dunes 157 128 123 130 33 32 24 30
Harvested Native Forest 129 126 126 112 23 31 22 29
Plantation Forest 0 123 245 258 0 41 58 55
Total 970 1147 1405 1504 243 287 352 376

62
Chapter 2. Discriminating Native and Plantation Forests

Appendix B2. Overall classification performance for SVM, RF and CART classifiers

per Landsat sensors and Models C1 to C4, including Kappa, F-score, Producer (PA)

and user (UA) accuracies.

C1 C3 C4
F- F- F-
LULC classes
PA UA score PA UA score PA UA score
Agriculture 0.867 0.897 0.881 0.854 0.932 0.891 0.868 0.825 0.846
Water 0.933 1.000 0.966 0.846 0.917 0.880 0.789 0.938 0.857
Native Forest 0.952 0.878 0.913 0.924 0.871 0.897 0.914 0.892 0.902
SVM

Sand Dunes 1.000 0.946 0.972 1.000 0.931 0.964 0.958 1.000 0.979
Harvested Native Forest 0.829 0.967 0.892 0.826 0.826 0.826 0.862 0.862 0.862
Overall accuracy 92% 89% 89%
Kappa 89% 86% 85%
Agriculture 0.900 0.844 0.871 0.854 0.953 0.901 0.895 0.773 0.829
Water 0.933 0.966 0.949 0.846 0.957 0.898 0.789 0.938 0.857
MSS 1979

Native Forest 0.904 0.872 0.888 0.939 0.861 0.899 0.938 0.916 0.927
RF

Sand Dunes 0.971 0.971 0.971 1.000 0.964 0.982 0.833 1.000 0.909
Harvested Native Forest 0.800 0.966 0.875 0.826 0.792 0.809 0.828 0.857 0.842
Overall accuracy 90% 90% 88%
Kappa 87% 87% 84%
Agriculture 0.900 0.806 0.850 0.729 0.921 0.814 0.895 0.739 0.810
Water 0.833 1.000 0.909 0.885 0.920 0.902 0.579 1.000 0.733
Native Forest 0.916 0.864 0.889 0.939 0.827 0.879 0.901 0.869 0.885
CART

Sand Dunes 0.943 0.917 0.930 0.889 0.889 0.889 0.792 1.000 0.884
Harvested Native Forest 0.714 0.926 0.806 0.783 0.720 0.750 0.828 0.774 0.800
Overall accuracy 88% 85% 84%
Kappa 84% 81% 78%

C1 C2 C3 C4
F- F- F- F-
LULC classes
PA UA score PA UA score PA UA score PA UA score
Agriculture 1.000 0.938 0.968 0.915 0.977 0.945 0.956 0.956 0.956 1.000 0.891 0.942
Water 1.000 1.000 1.000 1.000 1.000 1.000 1.000 0.963 0.981 1.000 1.000 1.000
Native Forest 0.882 0.909 0.896 0.971 0.835 0.898 0.985 0.807 0.887 0.941 0.800 0.865
SVM

Sand Dunes 1.000 0.963 0.981 1.000 0.966 0.982 1.000 1.000 1.000 1.000 1.000 1.000
Harvested Native Forest 0.733 0.710 0.721 0.640 0.842 0.727 0.600 1.000 0.750 0.519 0.824 0.636
Plantation Forest 0.900 1.000 0.947 0.852 0.958 0.902 0.778 1.000 0.875 0.654 0.944 0.773
Overall accuracy 92% 91% 91% 88%
Kappa 90% 89% 89% 85%
Agriculture 0.983 0.952 0.967 0.979 0.902 0.939 0.978 0.936 0.957 0.959 0.887 0.922
Water 0.974 0.974 0.974 1.000 0.955 0.977 1.000 1.000 1.000 0.944 1.000 0.971
Native Forest 0.892 0.910 0.901 0.897 0.847 0.871 0.985 0.788 0.876 0.912 0.838 0.873
TM 1992

Sand Dunes 1.000 1.000 1.000 0.964 0.964 0.964 0.964 1.000 0.982 1.000 1.000 1.000
RF

Harvested Native Forest 0.700 0.656 0.677 0.560 0.667 0.609 0.480 0.923 0.632 0.593 0.762 0.667
Plantation Forest 0.933 1.000 0.966 0.815 1.000 0.898 0.778 1.000 0.875 0.731 0.826 0.776
Overall accuracy 92% 88% 90% 88%
Kappa 89% 85% 87% 84%
Agriculture 0.917 0.917 0.917 0.851 0.851 0.851 0.889 0.889 0.889 1.000 0.803 0.891
Water 0.949 0.974 0.961 0.857 1.000 0.923 0.846 0.917 0.880 0.889 1.000 0.941
Native Forest 0.853 0.870 0.861 0.868 0.843 0.855 0.868 0.766 0.814 0.912 0.747 0.821
CART

Sand Dunes 0.962 0.926 0.943 0.964 1.000 0.982 0.964 1.000 0.982 0.964 0.964 0.964
Harvested Native Forest 0.567 0.630 0.596 0.720 0.545 0.621 0.560 0.636 0.596 0.222 0.545 0.316
Plantation Forest 0.900 0.771 0.831 0.667 0.857 0.750 0.778 0.875 0.824 0.538 0.824 0.651
Overall accuracy 86% 83% 84% 81%
Kappa 83% 79% 79% 75%

63
Chapter 2. Discriminating Native and Plantation Forests

C1 C2 C3 C4
F- F- F- F-
LULC classes
PA UA score PA UA score PA UA score PA UA score
Agriculture 0.897 0.854 0.875 0.907 0.875 0.891 0.940 0.825 0.879 0.922 0.839 0.879
Water 0.875 0.913 0.894 0.556 0.833 0.667 0.611 0.917 0.733 0.692 0.900 0.783
Native Forest 0.829 0.877 0.852 0.880 0.786 0.831 0.902 0.856 0.878 0.909 0.769 0.833
SVM

Sand Dunes 0.920 0.920 0.920 0.952 1.000 0.976 0.950 0.950 0.950 0.955 0.913 0.933
Harvested Native Forest 0.818 0.900 0.857 0.727 0.762 0.744 0.870 0.833 0.851 0.778 1.000 0.875
Plantation Forest 0.842 0.750 0.793 0.778 0.875 0.824 0.783 0.923 0.847 0.622 0.875 0.727
Overall accuracy 86% 84% 87% 84%
Kappa 81% 79% 83% 79%
Agriculture 0.782 0.753 0.767 0.833 0.900 0.865 0.880 0.830 0.854 0.745 0.809 0.776
Water 0.875 0.875 0.875 0.556 1.000 0.714 0.556 0.909 0.690 0.538 0.700 0.609
ETM+ 2003

Native Forest 0.849 0.832 0.841 0.891 0.719 0.796 0.891 0.759 0.820 0.870 0.698 0.775
Sand Dunes 0.880 0.880 0.880 0.905 0.950 0.927 0.950 0.950 0.950 0.864 0.864 0.864
RF

Harvested Native Forest 0.818 0.857 0.837 0.727 0.889 0.800 0.913 0.875 0.894 0.889 0.941 0.914
Plantation Forest 0.632 0.692 0.661 0.689 0.775 0.729 0.630 0.879 0.734 0.489 0.647 0.557
Overall accuracy 80% 81% 82% 75%
Kappa 73% 74% 77% 67%
Agriculture 0.808 0.649 0.720 0.722 0.780 0.750 0.900 0.726 0.804 0.784 0.727 0.755
Water 0.792 0.826 0.809 0.556 1.000 0.714 0.556 0.833 0.667 0.385 0.714 0.500
Native Forest 0.760 0.782 0.771 0.837 0.636 0.723 0.793 0.702 0.745 0.870 0.620 0.724
CART

Sand Dunes 0.880 0.880 0.880 0.905 0.905 0.905 0.850 0.944 0.895 0.727 0.800 0.762
Harvested Native Forest 0.818 0.900 0.857 0.500 0.688 0.579 0.783 0.720 0.750 0.500 1.000 0.667
Plantation Forest 0.404 0.511 0.451 0.467 0.618 0.532 0.478 0.786 0.595 0.422 0.704 0.528
Overall accuracy 73% 70% 74% 69%
Kappa 63% 60% 66% 58%
Agriculture 0.902 0.943 0.922 0.837 0.854 0.845 0.804 0.804 0.804 0.872 0.891 0.882
Water 0.963 0.867 0.912 0.824 0.933 0.875 0.714 0.833 0.769 0.667 1.000 0.800
Native Forest 0.885 0.829 0.856 0.888 0.888 0.888 0.924 0.867 0.895 0.870 0.792 0.829
SVM

Sand Dunes 1.000 0.871 0.931 0.889 0.842 0.865 1.000 0.818 0.900 0.895 0.850 0.872
Harvested Native Forest 0.867 0.929 0.897 0.810 0.944 0.872 0.792 0.950 0.864 0.526 1.000 0.690
Plantation Forest 0.681 0.797 0.734 0.927 0.826 0.874 0.756 0.838 0.795 0.842 0.762 0.800
Overall accuracy 86% 87% 85% 83%
Kappa 82% 83% 80% 77%
Agriculture 0.837 0.917 0.875 0.796 0.929 0.857 0.870 0.870 0.870 0.851 0.870 0.860
Water 0.926 1.000 0.962 0.882 1.000 0.938 0.643 1.000 0.783 0.667 1.000 0.800
Native Forest 0.908 0.753 0.824 0.955 0.833 0.890 0.913 0.792 0.848 0.859 0.738 0.794
OLI 2014

Sand Dunes 0.889 0.800 0.842 0.778 0.875 0.824 0.944 0.850 0.895 0.842 0.842 0.842
RF

Harvested Native Forest 0.733 0.710 0.721 0.857 0.857 0.857 0.708 0.850 0.773 0.526 0.714 0.606
Plantation Forest 0.580 0.833 0.684 0.829 0.872 0.850 0.683 0.824 0.747 0.605 0.697 0.648
Overall accuracy 82% 87% 83% 78%
Kappa 76% 83% 77% 69%
Agriculture 0.837 0.906 0.870 0.796 0.796 0.796 0.696 0.800 0.744 0.894 0.808 0.848
Water 0.852 1.000 0.920 0.824 1.000 0.903 0.714 1.000 0.833 0.667 1.000 0.800
Native Forest 0.924 0.661 0.771 0.854 0.710 0.776 0.837 0.726 0.778 0.826 0.710 0.764
CART

Sand Dunes 0.852 0.793 0.821 0.667 0.750 0.706 0.889 0.640 0.744 0.684 0.929 0.788
Harvested Native Forest 0.533 0.667 0.593 0.571 0.706 0.632 0.500 0.667 0.571 0.316 0.500 0.387
Plantation Forest 0.348 0.750 0.475 0.561 0.719 0.630 0.610 0.694 0.649 0.526 0.588 0.556
Overall accuracy 76% 75% 73% 73%
Kappa 67% 66% 64% 62%

64
Chapter 2. Discriminating Native and Plantation Forests

Appendix C2. McNemar’s test results for comparison of the difference between

Model C1 and the other classifications schemes (i.e., Models C2, C3 and C4) using

RF, SVM, and CART classifiers.

MS S 1979 TM 1992 ETM+ 2003 OLI 2014


C1 C1 p-value C1 C1 p-value C1 C1 p-value C1 C1 p-value
SVM RF 0.3428 SVM RF 1.0000 SVM RF 0.0101 SVM RF 0.0141
SVM CART 0.0442 SVM CART 0.0080 SVM CART 0.0000 SVM CART 0.0000
C1 C3 C1 C2 C1 C2 C1 C2
SVM SVM 0.1003 SVM SVM 0.7277 SVM SVM 1.0000 SVM SVM 1.0000
SVM RF 0.1374 SVM RF 0.2002 SVM RF 0.4764 SVM RF 1.0000
SVM CART 0.0058 SVM CART 0.0072 SVM CART 0.0006 SVM CART 0.0014
C1 C4 C1 C3 C1 C3 C1 C3
SVM SVM 0.1508 SVM SVM 0.7277 SVM SVM 0.3960 SVM SVM 0.6025
SVM RF 0.1116 SVM RF 0.4047 SVM RF 0.8097 SVM RF 0.2530
SVM CART 0.0072 SVM CART 0.0072 SVM CART 0.0145 SVM CART 0.0004
C1 C4 C1 C4 C1 C4
SVM SVM 0.1547 SVM SVM 1.0000 SVM SVM 0.2530
SVM RF 0.1093 SVM RF 0.0279 SVM RF 0.0121
SVM CART 0.0008 SVM CART 0.0003 SVM CART 0.0002

65
Chapter 3. Predicting future plantation forest in response to policy
Chapter 3. Predicting future plantation forest in response to policy

3. PREDICTING FUTURE PLANTATION FOREST

DEVELOPMENT IN RESPONSE TO POLICY INITIATIVES: A

CASE STUDY OF THE WARREN RIVER CATCHMENT IN

WESTERN AUSTRALIA

ABSTRACT

The Warren River Catchment in the south-west of Western Australia exemplifies how

inconsistent forestry investment policies can have adverse environmental effects

especially related to dryland salinity. This region experienced a rapid expansion of

plantation forest in the late 1990s mainly due to tax incentives for forestry Managed

Investments Schemes. Thus, understanding how the current policy scenario can impact

the future of plantation forest is necessary to adequately inform future policy makers. The

objective of this study is to address this gap in knowledge by applying a spatial modelling

procedure that integrates Markov transition probabilities, Multilayer Perceptron neural

network and Cellular Automata to provide an accurate forecast of LULC change. In the

first stage, geospatial analysis determines the spatial drivers of LULC conversions.

Second, Markov transition probability matrices are estimated, and the Multila yer

Perceptron was trained to determine a model for every transition based on spatial drivers.

Finally, a Cellular Automata model was applied to forecast spatially explicit changes in

LULC to 2025 under the current policy regime. The predictive power of the model was

validated with a Kappa coefficient of 72% and a Kappa location of 74%. The simula tio n

forecasts an increase in the agricultural areas by 2025 compared to 2014; whereas

67
Chapter 3. Predicting future plantation forest in response to policy

harvested native forest areas were predicted to decrease, contributing to a slight increase

of the native forest areas. Despite government efforts to increase the areas of plantatio ns,

the model predicts a decrease in this land use. These results will assist decision- makers

in improving policy, by working through the long term implications of policy incentives

for forestry in terms of broader landscape objectives related to salinity and conservation.

Keywords: Land Use Policy; Plantation Forest; Multilayer Perceptron; Cellular

Automata; Western Australia.

3.1. INTRODUCTION

Land use and land cover (LULC) change is one of the most important threats to the natural

environment globally. LULC is driven by market forces and policy (Lambin et al. 2001).

The Warren River Catchment in the south-west of Western Australia has undergone rapid

plantation forest expansion since the introduction of the Federal Government’s policy

Plantations for Australia (The 2020 Vision; Commonwealth of Australia 2002) in 1997

and The Managed Investments Act 1998 (MIA 1998).

A key difficulty in measuring the effect of policy on LULC change is the need to

differentiate between policy responses and LULC changes resulting from other economic,

social and environmental drivers. Limited monitoring of both the change process and the

drivers can hamper the understanding of policy impacts. This key process becomes

increasingly difficult with low levels of spatial resolution of the available information and

increasing spatial extent of the study area (Veldkamp and Lambin 2001). Spatially

explicit LULC models are powerful tools for policy development and analysis by

68
Chapter 3. Predicting future plantation forest in response to policy

providing a solution to this attribution problem by quantifying policy responses, and by

predicting future LULC changes (Veldkamp and Lambin 2001; Li and Yeh 2002;

Corbelle-Rico et al. 2015).

A number of modelling approaches have been developed to analyse, understand, and

predict LULC changes. These models divide into three categories: machine learning,

statistical, and spatially disaggregated economic models (Otukei and Blaschke 2010;

Polyakov and Zhang 2008; Chomitz and Gray 1996; Ellis et al. 2010; Resler et al. 2014;

Lewis and Plantinga 2007); agent based and Cellular Automata (CA) models (Arsanjani,

Kainz and Mousivand 2011); and integrated models (e.g., Conversion of Land Use and

its Effect, and Land Change Modeler) (Han, Yang and Song 2015; Megahed et al. 2015;

Veldkamp and Fresco 1996). However, many CA models may be considered to be

Markov chain models, with a probabilistic rule set defining class transitions in a local

neighbourhood. A review of the most commonly used approaches can be found in Aburas

et al. (2016), Agarwal et al. (2002), Baker (1989), Irwin and Geoghegan (2001), and Sohl

and Claggett (2013).

Markov chain models, characterised by class-dependent transition matrices, have been

widely used to determine the probabilities of LULC change between two time periods.

These models assume that the future class of a given LULC type is entirely determined

by its immediately preceding class (i.e., the process is memoryless) (Baker 1989; Bell

and Hinojosa 1977; Murayama 2012). However, Markov chains in their simplest forms

are not spatially explicit and thus cannot provide a spatial distribution of future LULC

changes (Halmy et al. 2015; Li and Yeh 2002). In contrast, CA are spatially explicit, with

future classes of a landscape element dependent only on the previous class of the elements

contained within its local neighbourhood. Markov chains can be integrated into CAs (i.e.,

69
Chapter 3. Predicting future plantation forest in response to policy

termed stochastic CAs, or CA-Markov) (Al-sharif and Pradhan 2014; Halmy et al. 2015)

to include a stochastic component into the CA models (Myint and Wang 2006).

CA-Markov models incorporating remote sensed LULC data have been used to forecast

urban land use changes in Portugal (Araya and Cabral 2010), Japan (Guan et al. 2011)

and Australia (Deilami and Kamruzzaman 2017). Kamusoko et al. (2009) used a similar

approach when forecasting future LULC changes in rural areas of Zimbabwe in southern

Africa. However, a crucial issue remains in the definition of the CA-Markov transitio n

rules (Park et al. 2011; Li and Yeh 2001). Namely that when multiple LULC classes are

present, the transition rules of conventional CA models become substantially more

complex: the model includes a much larger set of uncertain parameters defining the

transition rules, and requires additional spatial variables to be included in the model (Li

and Yeh 2002).

To address this curse-of-dimensionality problem, machine learning methods such as

Multilayer Perceptron (MLP) neural networks may be combined with CA-Markov to

identify the most likely transition rules between LULC classes. The advantages are two-

fold: firstly, neural networks such as the MLP can effectively capture non-linear, complex

features in modelling processes to increase prediction accuracy (Li and Yeh 2002;

Pijanowski et al. 2002; Cooner, Shao and Campbell 2016); and, secondly, the relative

importance of various biophysical, social and economic factors in LULC change can be

quantified through spatially explicit transition potential maps, which are the derivation of

weights representing the relative importance of each factor in relationship to a given

LULC change (Hosseinali and Alesheikh 2008). The optimal weighting of these transitio n

potential maps, through the MLP’s backpropagation algorithm, is applied to determine

the overall probability of transition between any two LULC classes within the CA-

70
Chapter 3. Predicting future plantation forest in response to policy

Markov model, thus leading to spatially explicit LULC predictions (Baysal 2013; Mishra

and Rai 2016; Ahmed and Ahmed 2012; Wang and Li 2011; Park et al. 2011; Li and Yeh

2002).

In this research, the Warren River Catchment in the south-west of Western Australia was

chosen as a case study due to the unprecedented and rapid expansion of plantation forest

since the late 1990s. The objectives of this study are therefore: i) to understand the pattern

of LULC change in the catchment from 1979 to 2014; and, ii) forecast future LULC

change for 2025 through the CA-Markov-MLP modelling, assuming only that the land

use policies in operation in 2014 continue over the subsequent 11 years. Predicting land

use changes in environmentally relevant areas is essential because there are some land

policies that influence those decisions over an extended time period (Bockstael 1996;

Bustamante et al. 2014).

3.2. STUDY AREA AND POLICY CONTEXT

The Warren River Catchment (4,416 km2 ) is the third largest river in the south-west of

Western Australia with a mean annual streamflow of 291 Gl (1990-2001; at the Barker

Road Crossing gauging station; Smith et al. 2006), and annual rainfall ranging from 550

mm/year in the north-east of the catchment to over 1200 mm/year in the coastal south-

west (Townsend et al. 2012). The catchment encompasses the small towns of Manjimup

and Pemberton (Figure 3.1). From the 2011 census, these towns had a population of 4,164

and 777 (ABS 2011). The study area is in a Biodiversity Hotspot which hosts a large

number of endemic and threatened plant species (Myers et al. 2000). The economy of the

Warren largely depends upon tourism, agriculture, and plantation forest, with the latter

71
Chapter 3. Predicting future plantation forest in response to policy

two largely supported by Managed Investment Schemes (MIS).

Figure 3.1. Location of the Warren River Catchment in Western Australia.

The catchment is composed of two distinct areas: nearby the coast (south-west) is

dominated by native forests (i.e., jarrah and marri forest) alongside agriculture,

horticulture and plantation forest, while the north-east has been subject to large-scale

clearing and fragmentation of forest cover over the past century as the catchment

transitioned from wilderness to a mixed agricultural and forestry landscape. The

expansion of extensive agriculture in the north-east was accelerated in the 1920s with the

Group Settlement Scheme and later in the 1950s with the War Service Land Settlement

Scheme. These policy driven LULC changes consisted primarily of the conversion of

72
Chapter 3. Predicting future plantation forest in response to policy

native woodlands, scrublands and heath to rain-fed cereal production and grassland. As a

result, cleared areas for agricultural use increased from about 20% in 1950 (Collins and

Barrett 1980) to 30% in 1979 (see 2.4.2. LULC Changes in the WRC).

Since the late 1970s this region has been the focus of policy initiatives to primarily

address the resulting biodiversity loss, land degradation and water salinity issues (e.g.,

Country Areas Water Supply Act 1947, known as 1947 Act, amended in 1978, and Water

Resource Recovery Catchments in 1996). For instance, the 1947 Act was created to

control further clearing of native vegetation in five potential water supply catchments in

the south-west, which includes the Warren (Smith et al. 2006). By the early 1980s it had

become apparent that clearing controls alone were not capable of maintaining potable

water supplies in most catchments, and thus, a series of reforestation programs were

implemented within the Water Resource Recovery Catchments. In the late 1990s,

commercial tree plantations were widely established on previously cleared agricultura l

lands mainly boosted by two federal policies directed at increasing the area of plantatio n

forest: The 2020 Vision and MIA 1998.

The 2020 Vision aims to treble Australia’s plantation area to 30,000 km2 by 2020

(Commonwealth of Australia 2002). Western Australia’s forest industry supplies over

20% of the nation's timber, and includes the largest area of hardwood plantations in

Australia with 2,989 km2 (31%; Gavran 2014). The most widely planted hardwood

species is the Tasmanian bluegum (Eucalyptus globulus) a fast growing species suitable

for high-rainfall zones (> 600 mm per year) such as the Warren. It is managed for

woodchips on a 10-12 year rotation, while other hardwood species (Eucalyptus

cladocalyx, Eucalyptus saligna or Corymbia maculata) are managed for sawlogs on a 20-

25 year rotation.

73
Chapter 3. Predicting future plantation forest in response to policy

The MIA 1998 incentivises commercial tree plantations by creating a retail investme nt

structure that entitles investors to claim an up-front tax deduction for the expense of

investing in timber plantations and agribusiness development activities through MIS. As

a structure, MIS allows a large number of investors to pool funds, or invest in a common

enterprise that are administered by the Managed Investment Act (Parliament of Australia

2017). Under this scheme, the MIS manager usually takes little risk whereas the investor

is exposed to different types of uncertainty (e.g., market and production risks) (Lacey,

Watson and Crase 2006). The Australian Taxation Office considered that investors in

most MIS are carrying on a business and thus able to claim a full tax deduction for

expenditure, provided the Scheme’s Responsible Authority conforms to the Scheme’s

Product Disclosure Statement (Mackarness and Malcolm 2006). Both The 2020 Vision

and the MIA 1998 were designed to address a national shortage in native timber supply,

while delivering indirectly positive environmental externalities such as river salinity

mitigation.

Indeed, the tax advantages assured by the Taxation Ruling 2000/8 was crucial for the

attractiveness to investors of all MIS schemes in the 2000s (ATO 2000). However, in the

mid-2000s the government debated the appropriateness of the tax treatment of MIS, and

decided to curtail this sector. In 2006 the Australian Taxation Office revised its view of

the MIS’s deductibility for forestry and non-forestry, and notified that from 1 July 2008

contributions paid by MIS investors would no longer be tax deductible. However, after

much lobbying by the forestry industry, in 2007 the government withdrew its decision

and Division 394 was inserted into the Income Tax Assessment Act 1997 to support

forestry MIS (ATO 2017b; The Treasury 2008). In 2007 under the revised policy scenario

a statutory deduction for expenditure incurred in forestry MIS could be claimed, while

74
Chapter 3. Predicting future plantation forest in response to policy

non-forestry MIS investors were deprived of it as they were considered passive investors

in a trust (The Treasury 2008).

Currently, an initial participant in both MIS investments can claim tax deductions in the

year of payment provided the schemes complies with certain conditions (ATO 2017a;

ATO 2017b; Thompson 2010). While investors in forestry MIS enjoys a statutory

deduction under Division 394 to the present, investors in non-forestry rely on the Section

8-1 under the Income Tax Assessment Act 1997 since 2009. Based on this current policy

scenario, we simulated the effects of MIS tax incentives on land use changes in terms of

quantity of future change and spatial distribution.

3.3. METHODS

This study integrates Markov transition probabilities, Multilayer Perceptron (MLP)

neural network, Cellular Automata (CA) and spatially referenced socio-economic and

biophysical data to quantify historical LULC change and predict the spatial distributio n

of future LULC changes. The methodological steps are illustrated in Figure 3.2. Firstly,

Markov transition matrices were calculated from the LULC maps derived from remotely

sensed imagery covering the period from 1979 to 2014. Secondly, random training

samples from multiple classes of LULC and factor variables were drawn. Third, using the

extracted variables as input, the network model was trained to determine the optimal

transition potential maps. Finally, future LULC classes for 2014 and 2025 were projected

using the CA approach. The simulation of future LULC was made available in Quantum

GIS using the plugin MOLUSCE (Modules for Land Use Change Evaluation; NextGIS

2016; QGIS Development Team 2016).

75
Chapter 3. Predicting future plantation forest in response to policy

3.3.1. Data

A multi-temporal remote sensing dataset was constructed of imagery captured from the

Multispectral Scanner, Thematic Mapper, Enhanced Thematic Mapper, and Operational

Land Imager. Imagery (30m resolution) were acquired across the summer (December to

February), to increase the separability between agricultural and forestry land uses, and

over four years (1979, 1992, 2003, 2014). The imagery from 1979 is the base-line map to

assess potential forest clearance after the amendment of the Act 1947; whereas we

selected imagery from 1992, 2003 and 2014 as they are evenly 11 years apart and thus

enabled land change patterns to be monitored before and after the introduction of the key

policies: The 2020 Vision in 1997, the Managed Investments Act in 1998, and the revised

Taxation Ruling 2007/8 for MIS in 2007. Limited resources were available for image

processing in this study, and hence only four images were captured. If the imagery had

been acquired more frequently we would have observed more closely the rate of LULC

change following the MIS. However, this finer rate of change would have been diffic ult

to reliably measure given the relatively high error rates in the LULC classification itself.

A supervised LULC object-based classification approach was conducted using a Support

Vector Machine (SVM) classifier (A in Figure 3.2). The overall accuracies for the

classifications were ≥ 86% (Kappa ≥ 81%). Details about the classification are given in

the section 2.3 Methods. The LULC maps discriminate the following classes: agriculture

(cleared areas for pasture or cropping), water bodies (lake, river and dams), native forest

(indigenous plants such as trees and shrubs), sand dunes (ridge of sand especially on the

coast), harvested native forest (clearcut logging that is not yet regenerated), and plantatio n

forest (commercial plantation trees).

76
Chapter 3. Predicting future plantation forest in response to policy

In this study, six biophysical and socio-economic factors were identified as potential

drivers of LULC transitions in the study area. Respective spatial datasets were derived

from the publicly available vector and raster data: (1) the distance from each map location

to the nearest road (including railways, highways and country roads; Government of

Western Australia 2016), representing the potential accessibility of a location; (2)

distance from each map location to the nearest water body (BOM 2014) which may

indicate the potential water availability for a new land use; (3) organic carbon (CSIRO

2016) which is an indicator of soil health, representing the potential establishment of

agricultural land use; (4) annual rainfall (long term average from 1961 to 1990; BOM

2016) which shows the potential for conversion to land uses suitable for rainfall zones

above 600 mm (e.g., Tasmanian bluegum plantation); (5) slope which may constrain some

land use practices as its gradient can impact drainage, crop productive and soil proprieties;

and finally (6) water flow direction, representing the variation of water direction across

land use types. Both of these latter variables were calculated from a digital elevation

model (DEM; 1 Arc-Second Global) with the R statistical software (R Core Team 2016)

to represent the geomorphology of the study area (B in Figure 3.2.).

77
Chapter 3. Predicting future plantation forest in response to policy

Figure 3.2. Overview of the methodology applied in this study.

3.3.2. Transition probability matrix

The first modelling step was to develop a transition probability matrix for each of the

three intervals between image captures: 1979-1992, 1992-2003, 2003-2014 (E in Figure

3.2). Markov transition matrices have been widely used to quantify LULC changes

between discrete time periods, and to describe the transfer rate among different land use

classes (Wu et al. 2006; Lambin 1997; Mirkatouli, Hosseini and Neshat 2015; Sang et al.

2011). The Markov transition matrices were used in the validation (1992 to 2003), and in

78
Chapter 3. Predicting future plantation forest in response to policy

the prediction (2003 to 2014) of the CA model to constrain the maximum number of

transitions (36) between any two states.

3.3.3. Transition potential modelling using a Multilayer Perceptron (MLP)

There are a number of artificial neural networks types, of which Multilayer Perceptron

is one of the most commonly used to forecast LULC changes (Mas et al. 2014; Megahed

et al. 2015; Bernetti and Marinelli 2010; Pijanowski et al. 2002; Khoi and Murayama

2012). A neural network consists of interconnected processing elements (neurons),

arranged in layers (Hassoun 1995; Lipták 1999). An MLP is a feed-forward (i.e., directed)

neural network which may include more than one hidden layers between the input data

and the machine learning output (i.e., prediction of LULC class) (C in Figure 3.2;

Kotsiantis 2007). In this application the MLP calculates a set of ‘transition potential

maps’ (TPM; D in Figure 3.2), which is a weighted score of how likely a map location

transitions to one of six different states from the current state. This probability is based

on both the normalised values of the driving factors in the local size 2 neighbour hood

(i.e., 5x5 pixels), and the current LULC states observed in that neighbourhood, thereby

forming the input layer of the MLP (C in Figure 3.2). Based on the first law of geography

pixels nearby have a greater influence on a land cover class than pixels at a greater

distance (i.e., land use/neighbourhood interactions), thus we applied a neighbourhood size

of 5x5 pixels to provide a more accurate representation of the local variance, allowing us

to capture land uses at small scales (~2 ha).

The MLP calibration begins with assigning values to the parameters (i.e., momentum,

learning rate, neighbourhood and weights that connect the different layers). Then a

79
Chapter 3. Predicting future plantation forest in response to policy

sample of 25,000 randomly selected pixels was divided into a test subset (20%) and a

training subset (80%). The MLP model operates in loop for calibrations, and an evaluatio n

is performed over the independent test subset. The initial weights are updated with each

iteration of the back propagation algorithm, either until the error of the model achieve a

certain threshold or until it reaches the maximum number of iterations.

A hidden layer consisting of 6 neurons was chosen, for parsimony of computation and

with minimal loss in prediction accuracy. The number of neurons in the output layer

equate to the six possible LULC states. The logistic function formed the activatio n

function of the MLP. Within the MLP network, the neurons of each layer have a weighted

connection to neurons in the immediately previous and/or following layers. The

weightings are multiplied by the input values relating to a neuron’s incoming connections,

with the product summed over those connections. The activation function then transforms

this summed product to a value between 0 and 1, and for the output layer this defines the

TPM for each LULC state.

3.3.4. Cellular Automata (CA) simulation

The CA model for a given time period consists of four elements: 𝐶𝐴 ~(𝐶, 𝑛, 𝐾, 𝑅) ; where

cells 𝐶 are represented by pixels, with each pixel having a 5x5 Moore neighbourhood of

size 𝐾 = 2; a finite number of possible LULC classes 𝑛 = 6; and 𝑅 are the TPMs as an

ad hoc measure of the likelihood of state change from 𝑛𝑡 to 𝑛𝑡+1 .

The CA model identifies for each pixel the two states 𝑖 and 𝑗 with the highest TPM values

(G in Figure 3.2). A ‘certainty’ map that is the difference in TPM values for those two

80
Chapter 3. Predicting future plantation forest in response to policy

states are then calculated. The pixels are ordered in terms of their certainty, and a number

of pixels are assigned a state change 𝑖−> 𝑗, equal to the number implied by the

corresponding transition probability given by the Markov transition matrix associated

with that period of transition. Hence, we do not use explicit user-defined transition rules

for the CA simulation (Li and Yeh 2001), but instead applied the MLP to develop a ranked

scoring by which LULC transitions were assigned. This ranked scoring depends on the

non-linear interaction between driving factors and the array of LULC states over a local

neighbourhood. This model of transitions may be repeatedly applied to generate a set of

LULC predictions.

To validate the CA predictions, the total number of conversions for each transition 𝑖−> 𝑗

was given by the Markov transition matrix for the 1992-2003 period (F in Figure 3.2).

LULC predictions were then constructed for 2014, and the Kappa statistic for the

predicted map compared to that of the classified LULC map for 2014 (H and I in Figure

3.2). Once the model validation process was assured, LULC map was predicted for 2025

using the same set of parameter values and including the transition probabilities from

2003 and 2014 as initial and final grid input respectively. A prediction for 2025 is then

performed similarly, but without the validation, by applying the MLP-CA trained on the

2003-2014 data, and with the 2003-2014 Markov transition matrix and 2014 LULC map

as inputs (J and K in Figure 3.2). Note that the period of 1979-1992 was not included in

the MLP-CA-Markov model because the plantation forest class was not classified for the

1979 imagery as there were no visible instances of this class at that time.

81
Chapter 3. Predicting future plantation forest in response to policy

3.4. RESULTS AND DISCUSSION

3.4.1. Analysis of LULC Markov transition probabilities

Most areas under native forest cover remained stable from 1979 to 1992 and from 2003

to 2014 (94%), with only a slight decrease of 4% observed between 1992 and 2003 (Table

3.1). This decrease in native forest cover was reflected in the probability of this class

being converted to either agriculture (6%) or harvested native forest (2%). The most likely

transition among all three periods is from the class harvested native forest to native forest

(above 85%), due to natural regrowth following timber harvesting.

Table 3.1. Transitional probability matrix derived from the LULC map for the periods
1979-1992, 1992-2003 and 2003-2014.

LULC 1992
Harvested
Native Sand Native Plantation
LULC 1979 Agriculture Water Forest Dunes Forest Forest
Agriculture 0.8435 0.0019 0.1462 0.0000 0.0012 0.0071
water 0.0469 0.4180 0.5052 0.0237 0.0009 0.0052
Native Forest 0.0380 0.0017 0.9399 0.0001 0.0170 0.0033
Sand Dunes 0.0095 0.0063 0.2272 0.7556 0.0013 0.0000
Harvested Native Forest 0.0147 0.0000 0.9078 0.0000 0.0775 0.0000
Plantation Forest 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000

LULC 2003
Harvested
Native Sand Native Plantation
LULC 1992 Agriculture Water Forest Dunes Forest Forest
Agriculture 0.7948 0.0018 0.0799 0.0000 0.0017 0.1217
Water 0.1409 0.5311 0.3179 0.0075 0.0001 0.0024
Native Forest 0.0610 0.0033 0.9007 0.0022 0.0222 0.0105
Sand Dunes 0.0032 0.0191 0.0471 0.9306 0.0000 0.0000
Harvested Native Forest 0.0171 0.0010 0.8456 0.0008 0.1350 0.0004
Plantation Forest 0.2577 0.0069 0.2103 0.0000 0.0019 0.5232

82
Chapter 3. Predicting future plantation forest in response to policy

LULC 2014
Harvested
Native Sand Native Plantation
LULC 2003 Agriculture Water Forest Dunes Forest Forest
Agriculture 0.8027 0.0047 0.1275 0.0004 0.0061 0.0586
Water 0.0919 0.5525 0.3363 0.0006 0.0037 0.0149
Native Forest 0.0380 0.0011 0.9402 0.0003 0.0087 0.0117
Sand Dunes 0.0001 0.0021 0.3184 0.6786 0.0007 0.0000
Harvested Native Forest 0.0357 0.0003 0.8580 0.0000 0.1058 0.0002
Plantation Forest 0.2725 0.0012 0.2603 0.0000 0.0028 0.4632

Despite Lothian and Conacher (2005) affirmation that the clearing controls embedded by

the 1947 Act have not been effectively enforced, and so far, only one landholder has been

fined for illegal clearing, our classification did not identify any major clearing of native

forest in the catchment since the introduction of the amendment to the 1947 Act in 1978.

This is consistent with Smith et al. (2006) who reported a decrease in cleared areas of

around 36% in 1980 to 24% in 2000. For monitoring the dynamics of LULC changes at

smaller spatial scales, satellite or aerial imagery with higher spatial and spectral resolutio n

would better identify where illegal clearing may have taken place.

Prior to the introduction of policy incentives for plantation forest there were very few

plantations in the catchment at the time the 1979 base-line image was captured. From

1992 onwards, we observed a substantial increase in the conversion of agriculture to

plantation forest. As can be seen in Figure 3.3, the establishment of forest plantations was

centred on the north-east of the catchment where most of the land clearing for agriculture

had occurred since the 1950s. This area of plantation forest development is associated

with higher salinity concentrations than elsewhere because of the widespread clearing for

agriculture (Munro 2006; Platt 2007).

83
Chapter 3. Predicting future plantation forest in response to policy

Figure 3.3. The most representative areas of changes between 1992 and 2014.

Salinity in the Warren River Catchment exceeded 500 mg/L Total Dissolved Solids

(TDS) in the 1960s (Munro 2006), and achieved its highest average annual salinity of

1015 mg/L between 1993 and 2004. Most of the salt load, around 74%, comes from the

Tone and Perup Rivers located in the upper catchment (Platt 2007). This spatial pattern

of plantation forest development is compatible with the State Salinity Strategy which

suggests commercial plantation forest in the already cleared areas as a key manage me nt

option in reducing salinity within the catchment.

The important transitions in the 1992-2003 period from a policy perspective were from

agriculture to plantation forest (12%), and from plantation forest to agriculture (26%).

Plantation forests increase significantly in extent during this period, with the majority of

the plantation forests established only in the 1990s (Smith et al. 2006) following the

84
Chapter 3. Predicting future plantation forest in response to policy

introduction of the tax incentives. Thus, the 11 year period between image captures is

sufficient for a significant proportion of the 1992 plantations to be both harvested and re-

planted by 2003.

For the period 2003 to 2014 the rate of transition from agriculture to plantation forest

slowed to 6%, corresponding to the changes to the tax inducements associated with the

MIS schemes between 2006 and 2009. Further evidence of this is the strong decline in

MIS subscription from its peak in 2006 (Brown, Trusler and Davis 2010). The rate of

harvesting of commercial plantation forest and reconversion to agriculture mainta ined

itself at 27%. The net change of conversion of water bodies to other LULC classes

represented only 0.11% between 1992 and 2003, and -0.03% between 2003 and 2014

(Table 3.4). This result is likely a reflection of the total water exposure at time of acquiring

the Landsat imageries. For instance, variation in rainfall and drought are natural events

which can lead to a decrease in water availability, and consequently, more exposure of

low lying riparian vegetation. Moreover, the continued decrease in water body extents

may well be a reflection of long-term drying associated with global warming as well as

with decrease in rainfall and water supply in other areas of Australia’s south-west (Gove

et al. 2013).

A possible confounding feature of the analysis is the misclassification error resulting from

the derivation of LULC maps from remotely sensed imagery (Marcos-Martinez and

Baerenklau 2015). It is uncertain whether transitions from sand dunes to native forest

(32% transition rate between 2003 and 2014) are in part an artefact of misclassificatio n,

despite dune systems potentially being highly mobile over time relative to other

ecological systems. However, the net change of conversion revealed that less than 0.25%

of this land use actually transitioned to other LULC classes. More significantly, the

85
Chapter 3. Predicting future plantation forest in response to policy

classification accuracy rate for plantation forest is low for the 2003 and 2014 imagery

(~60%; see 2.4. Results and Discussions), with mature plantations likely confounded with

native vegetation given both are dominated by eucalypt species.

3.4.2. Model Calibration and Validation

Tuning parameters for the MLP–CA–Markov model, namely the learning parameter (set

to 0.01) and the maximum number of iterations (500), were chosen by maximising the

overall classification accuracy. The model was validated by cross comparing of actual

(classified) and simulated LULC maps for 2014. The simulated map for the year 2014

was evaluated based on the Kappa statistics to check the validity in terms of quantity and

location (Pontius 2000). The MLP-CA-Markov model performed well with a Kappa

standard coefficient of 72%, and a Kappa location coefficient of 74%.

Table 3.2 compares the area of each LULC for both the simulated and classified maps for

2014. In general, we consider the simulated map provides an acceptable forecast (less

than 5% difference between maps on average) of the classified map. However, it is

important to note there is some discrepancy between the classified and simulated LULC,

most significantly for the dune systems. Apart from misclassification errors between

remotely sensed scenes, the source of the discrepancies may include the available

selection of factors for the transition potential maps or the size of the neighbour hood

windows, which are both crucial for success in forecasting LULC classes. Nevertheless,

we conclude the MLP-CA-Markov model produced satisfactory results that permit LULC

changes to be projected for 2025.

86
Chapter 3. Predicting future plantation forest in response to policy

Table 3.2. Comparison of classified and projected LULC classes in 2014

2014 Classified 2014 Simulated Absolute Error


LULC class Area Area Area
km2 % km2 % km2 %
Agriculture 1112.56 25.19 1170.74 26.51 58.18 5.23
Water 21.54 0.49 20.72 0.47 -0.81 -3.78
Native Forest 3023.08 68.45 2956.20 66.94 -66.88 -2.21
Sand Dunes 25.40 0.58 35.25 0.80 9.84 38.75
Harvested Native Forest 41.41 0.94 41.73 0.94 0.33 0.79
Plantation Forest 192.53 4.36 191.88 4.34 -0.65 -0.34

3.4.3. Future Prediction Using Multilayer Perceptron CA-Markov Model (MLP-CA-

Markov)

The MLP-CA-Markov model was used to predict the six different LULC classes for 2025

in response to six potential driving factors, based on LULC changes from 2003 to 2014.

All the projections from the model are only one of multiple possible scenarios under the

current policy regime. Table 3.3 reports the classified area of each LULC class, and

includes the prediction results for 2025. The 2025 estimate for native forest was 69.17%,

revealing a 1.31% increase compared to the 1979 base-line. This increase may be

explained in part by the decrease in harvested native forest areas which were 0.94% in

2014 and are expected to decline to 0.14% in 2025. Table 3.4 presents the percentage of

change between each period for ease of comparison.

87
Chapter 3. Predicting future plantation forest in response to policy

Table 3.3. Estimated area of each LULC class in the Warren River Catchment.

1979 1992 2003 2014 2025


LULC class Area Area Area Area Area
km2 % km2 % km2 % km2 % km2 %
Agriculture 1326.52 30.04 1234.72 27.96 1176.53 26.64 1112.56 25.19 1183.42 26.80
Water 24.52 0.56 18.15 0.41 22.86 0.52 21.54 0.49 17.94 0.41
Native Forest 2996.79 67.85 3059.00 69.26 2911.84 65.93 3023.08 68.45 3054.84 69.17
Sand Dunes 38.82 0.88 30.30 0.69 35.27 0.80 25.40 0.58 25.36 0.57
Harvested Native Forest 29.86 0.68 54.98 1.24 77.44 1.75 41.41 0.94 6.01 0.14
Plantation Forest 0.00 0.00 19.36 0.44 192.58 4.36 192.53 4.36 128.95 2.92

Table 3.4. Change in areal proportion of each land use between periods.

Change in Change in Change in Change in


LULC class % % % %
1979-1992 1992-2003 2003-2014 2014-2025
Agriculture -2.08 -1.32 -1.45 1.60
Water -0.14 0.11 -0.03 -0.08
Native Forest 1.41 -3.33 2.52 0.72
Sand Dunes -0.19 0.11 -0.22 0.00
Harvested Native Forest 0.57 0.51 -0.82 -0.80
Plantation Forest 0.44 3.92 0.00 -1.44

Historically, the occurrence of clearcut logging was concentrated in the south-west of the

catchment, where the main land cover was state forests, and relatively small extents of

agriculture and plantation forest. In the catchment, timber harvesting operations are

carried out within state forests in line with the State Government’s Forest Management

Plan 2014-2023 (Conservation Commission of Western Australia 2013), in consultatio n

with the Forest Products Commission – the statutory authority tasked with delivering a

sustainable timber supply. The potential decrease in harvesting of native forest likely

reflects the increasing conflict between timber production and environmental values such

as biodiversity, clean air and water. For example, the Conservation Commission of

Western Australia (2013) emphasizes the need to address the potential adverse effects of

88
Chapter 3. Predicting future plantation forest in response to policy

the harvesting of native forest on stream salinity, by recommending that harvesting

activities be located in parts of the landscape that do not risk an increase in groundwater

salinity level.

The model also revealed a slight decrease in water area (0.41%) in 2025. However,

significant inter-year variation in yearly and summer rainfall has led to the area of

standing water in natural wetlands and artificial dams varying widely among the four

different dates of image capture. There are 921 farm dams documented for irrigation and

stock use in the catchment (Bari, Silberstein and Aryal 2012). The results also

demonstrated that sand dunes will occupy 0.57% of the Warren landscape in 2025. High

Pleistocence era sand dunes extend for approximately 8km inland from the coast, with

the majority of the sand dunes located within the D'Entrecasteaux National Park near the

mouth of the Warren river (Munro 2006).

The key change predicted for 2025 across the catchment is a substantial reduction of

plantation forest area (Figure 3.4). The simulation predicts that this land use will decrease

from 192.53 km2 in 2014 to 128.95 km2 in 2025. This decline is reflected in an increase

of the area of agricultural land use, which will reach 1183.42 km2 . As a result, the area of

forest plantations is predicted to be the smallest since the introduction of the governme nt

tax incentives for forestry MIS in the late 1990s.

89
Chapter 3. Predicting future plantation forest in response to policy

Figure 3.4. LULC maps in different years: (a) 1979, (b) 1992, (c) 2002, (d) 2014 and (e)
2025.

In Western Australia between 2012 and 2013 a decline in plantation areas (0.55%) was

already reported by Gavran (2014), unless action is taken this trend is forecast to continue

90
Chapter 3. Predicting future plantation forest in response to policy

(Forest Industries Federation 2015). This concern extends to the whole of Australia where

existing area of plantations is under threat with a contraction in the area of short rotation,

and no increase in long rotations (FWPA 2012). Although studies have shown that MIS

companies were financially unsustainable in the long-run (Brown, Trusler and Davis

2010; Mackarness and Malcolm 2006), the successive reformulation of MIS legisla tio n

might make the scheme vulnerable to policy changes, leading to a weakness in investor

interest in forestry MIS. Thus, given the tax-driven nature of investment in MIS, and the

lack of government action on further enhancing investment security in forestry, modelling

the spatial distribution of different land uses by classes utilizing historical land use data

is necessary to forecast future land use changes in the catchment.

The MIS sector initially grew quickly in Australia as it offered full tax deductions up-

front for investment in timber plantation and agribusiness development activities (The

Treasury 2008). The tax treatment was especially attractive for high income earners.

However, the alteration of the MIS legislation combined with a stock market contraction

contributed to a near collapse of the MIS industry in Australia (NewForests 2015;

Thompson 2010), and lead to uncertainty about forestry MIS. This was the case of the

multi-billion dollar failure in 2009 of two high-profile agribusiness companies operating

in the south-west of Western Australia: the Great Southern Group and Timbercorp

Limited. These two companies were estimated as having captured 40% in total of all MIS

business in Australia (Brown, Trusler and Davis 2010).

With the predicted decline of plantation forest areas a key environmental concern for

clean water may regain prominence through increased dryland salinity – large areas of

plantation forest can lower the water table at the landscape scale, thereby keeping the

highly saline ground-water from mixing with fresh water river flows. Importantly, Cleary

91
Chapter 3. Predicting future plantation forest in response to policy

et al. (2010) have modelled the impact of plantation forest on stream salinity reductions.

Their results suggest that strategically tree planting in the upper Warren Catchment would

be effective in reducing the groundwater levels and thus lower stream salinity in the study

area. Townsend et al. (2012) extended this analysis by valuing the restored water quality

from increased plantation activity. Currently, the catchment is managed under State

Salinity Strategy to achieve potable water levels of 500 mg/L TDS by 2030 (State Salinity

Council 2000). However, additional management strategies are required for the

successful achievement of this potable water standard, with the State Salinity Strategy

having identified increased commercial plantation forest as a potential aid in reducing

stream salinity. The State Salinity Strategy forecast of a continuous increase in plantatio n

forest within the catchment (State Salinity Council 2000) now looks to be overly

optimistic, given the strategy forecast predates the collapse of investments in plantatio n

forest following the failure of the MIS.

A decrease in plantation forest areas will not only impact Western Australia’s pulpwood

production, but may also jeopardize efforts to reduce annual average water salinity levels

over the long-term. Significantly, the State Salinity Strategy has not been updated to

consider the failure of MIS-driven plantation forest, with the State Salinity Strategy

assumption of relying on commercial plantation forest to deliver success in reducing

water salinity levels in the catchment now out of date. Land use policy in Australia has

been criticised previously for the lack of reliable information (Hobbs et al. 2016), and for

not considering detailed subject knowledge when a policy is designed (Pannell and

Roberts 2009). In this instance, current land use policy may be failing due to policy not

being adapted in a timely fashion to changing patterns of land use and policy incentives.

Moreover, the fact that MIS-driven investments for forestry did not persist for much more

92
Chapter 3. Predicting future plantation forest in response to policy

than a decade suggests that more rigour is required in the scenario testing of key

assumptions underpinning policies impacting on land use change, and on their possible

follow-on effects, such as the reduction in water salinity envisaged under the State

Salinity Strategy.

The Australian government has typically used incentives such as personal tax deduction

and low interest loans to accelerate plantation and agriculture development, thus MIS

structures have been the biggest source of investment in Australian agriculture and

forestry in the past decade (Commonwealth of Australia 2002; The Treasury 2008;

Mackarness and Malcolm 2006). Although forestry MIS arrangements were designed to

provide an ongoing mechanism for attracting investors, there is speculation that at least

30% of plantations may not be re-planted after harvesting (FWPA 2011). This might be

the case of the Warren River Catchment which has recently not established any significa nt

area of new plantations, and the simulation revealed a potential increase in agricultura l

areas following plantation harvesting. Plantation MIS is quite different to some non-

forestry MIS investments, with large up-front expenses at the time of establishment and

lengthy periods between planting and harvesting. By contrast, some agricultura l

commodities have annual production cycles which favours cash-flow pattern and

adoption of non-forestry MIS. Thus, the net present value of forestry investment without

the tax break is less than the present value of returns from agriculture.

In the next few years, it is unlikely that the government will take action toward improving

the actual MIS taxation to avoid the decline in establishment of new plantations (MIS

related) and retention of existing ones due to relatively profitability based on yield and

production price. Thus, our model is based on the current policy scenario where investors

in forestry MIS relies on a statutory deduction under Division 394 (ATO 2017b).

93
Chapter 3. Predicting future plantation forest in response to policy

Nevertheless, in the past few decades Australia has seen a range of public policies

emerging that target environmental benefits such as climate change mitigation (e.g.,

Carbon Farming Initiative, and Direct Action Plan), and salinity abatement (e.g., State

Salinity Strategy). Those policies might re-incentivise plantation forestry. However,

external benefits from growing plantations are not captured by current policies

(Thompson 2010). Western Australia attracted around $64m for reforestation project, but

this was not focused on the WRRC. An obvious policy would be to compensate

landowners for not clear-felling plantation forestry especially where hydrologica l

modelling indicates that it would have an adverse effect on salinity. Future policies will

need to address the economic aspects of plantations, while also targeting multiple benefits

(e.g., carbon sequestration, and water quality), and community goals (Barlow and Cocklin

2003; Williams 2014). Townsend et al. (2012) suggested charging those who caused

salinity in the catchment while using the proceeds to pay for local reforestation.

By taking a strategic partnership between the forest industry and government, plantatio n

forests could be seen as a valued land use that could provide both commercial and

environmental benefits, the tax subsidy may be viewed as a contribution to the external

benefits. These results highlight the need for policy makers to be cautious when relying

on temporary incentives to achieve long term aims, such as salinity abatement. Pannell

and Roberts (2010) emphasized that where conservation practices are sufficie ntly

attractive to landholders, government does not need to employ incentive-based

mechanisms to encourage adoption. Specifically, past investment in plantation forest in

the catchment appears to have been driven more by a demand for tax minimization than

by market returns on investment.

94
Chapter 3. Predicting future plantation forest in response to policy

3.5 CONCLUSION

In this paper we assessed and modelled the future trend of LULC change in the Warren

River Catchment using an integrated approach combining Markov transitio n

probabilities, Multilayer Perceptron and Cellular Automata. The simulation technique

employed produced satisfactory results allowing us to project LULC changes for 2025.

Results predict that agriculture areas will increase slightly in the future, while plantatio n

forest areas will decrease. Significantly, large areas of plantation forest are likely to be

reconverted to agriculture as a legacy of failed forestry MIS investments driven mainly

by taxation benefits and optimistic claims about profitability.

There is an urgent need for balancing the expectations of temporary tax incentives with

long-term conservation strategies in the design of effective land use policies. MIS is

undoubtedly an important vehicle for the establishment of new plantations, however

periodic changes in policy might create serious problems for the forestry sector, while

also impacting management strategies for preventing and managing dryland salinity.

Future conservation planning will need to be more sensitive to case-specific

circumstances when targeting environmental goals, as is the case for landscape-scale

plantation development via a tax incentive, to ensure effective and persistent results with

low conservation impacts. More specifically, understanding how the potential future

changes in the landscape might influence river and forest health can likely assist future

conservation planning for the third largest river in the south-west of Western Australia,

the Warren. Further detailed studies are needed to document the effect on water and

dryland salinity of the widespread reclamation of agricultural lands from diminis hing

plantation forests if plantation forest is a linchpin for reducing that salinity.

95
Chapter 3. Predicting future plantation forest in response to policy

3.6. REFERENCES

ABS - Australian Bureau of Statistics. 2011. “2001.0 - Census of population and housing:
Basic community profile.” Canberra, Australia. Available at:
http://www.abs.gov.au/websitedbs/censushome.nsf/home/quickstats [Accessed
September 20, 2016].
Aburas, M.M., Y.M. Ho, M.F. Ramli, and Z.H. Ash’aari. 2016. “The simulation and
prediction of spatio-temporal urban growth trends using cellular automata models:
A review.” International Journal of Applied Earth Observation and Geoinformation
52:380–389.
Agarwal, C., G.M. Green, J.M. Grove, T.P. Evans, and C.M. Schweik. 2002. “A review
and assessment of land-use change models: Dynamics of space, time, and human
choice.” General Technical Report NE-297:61.
Ahmed, B., and R. Ahmed. 2012. “Modeling urban land cover growth dynamics using
multi‑temporal satellite images: A case study of Dhaka, Bangladesh.” ISPRS
International Journal of Geo-Information 1(1):3–31.
Al-sharif, A.A.A., and B. Pradhan. 2014. “Monitoring and predicting land use change in
Tripoli metropolitan city using an integrated Markov chain and cellular automata
models in GIS.” Arabian Journal of Geosciences 7(10):4291–4301.
Araya, Y.H., and P. Cabral. 2010. “Analysis and Modeling of Urban Land Cover Change
in Setúbal and Sesimbra, Portugal.” Remote Sensing 2(6):1549–1563.
Arsanjani, J.J., W. Kainz, and A.J. Mousivand. 2011. “Tracking dynamic land-use change
using spatially explicit Markov Chain based on cellular automata: the case of
Tehran.” International Journal of Image and Data Fusion 2(4):329–345.
ATO - Australian Taxation Office. 2017a. “Income Tax Assessment Act 1997 No. 38.
Chapter 1 Introduction and core provisions. Division 8: Deductions. Vol 1.” :27–28.
Available at:
https://www.legislation.gov.au/Details/C2017C00155/Html/Volume_1 [Accessed
July 10, 2017].
ATO - Australian Taxation Office. 2017b. “Income Tax Assessment Act 1997 No. 38.
Chapter 3 Specialist liability rules: Division 394: Forestry managed investme nt
schemes. Vol 7.” :343–351. Available at:
https://www.legislation.gov.au/Details/C2017C00155/Html/Volume_7 [Accessed
July 10, 2017].
ATO - Australian Taxation Office. 2000. “Taxation Ruling TR 2000/8.” :79. Availab le
at:
http://law.ato.gov.au/atolaw/view.htm?locid=%27TXR/TR20008/NAT/ATO%27&
PiT=20010516000001 [Accessed May 20, 2016].
Baker, W.L. 1989. “A review of models of landscape change.” Landscape Ecology
2(2):111–133.

96
Chapter 3. Predicting future plantation forest in response to policy

Bari, M.A., R.P. Silberstein, and S.K. Aryal. 2012. “Strategic tree planting and climate
change - Simulating future water availability at the Warren River catchment,
Western Australia.” Engineers Australia:167–174.
Barlow, K., and C. Cocklin. 2003. “Reconstructing rurality and community: Plantatio n
forestry in Victoria, Australia.” Journal of Rural Studies 19(4):503–519.
Baysal, G. 2013. Urban land use and land cover change analysis and modeling. A case
study area Malatya, Turkey. Master Thesis. Westfälische Wilhelms- Universität.
Münster, Germany: Westfälische Wilhelms- Universität.
Bell, E.J., and R.C. Hinojosa. 1977. “Markov analysis of land use change: continuo us
time and stationary processes.” Socio-Economic Planning Sciences 11(1):13–17.
Bernetti, I., and N. Marinelli. 2010. “Evaluation of landscape impacts and land use
change: a Tuscan case study for CAP reform scenarios.” Aestimum - Firenze
University Press (56):1–29.
Bockstael, N.E. 1996. “Modelling Economics and Ecology: The Importantance of a
Spatial Perspective.” American Journal of Agricultural Economics 78(5):1168–
1180.
BOM - Bureau of Meteorology. 2014. “Australian Hydrological Geospatial Fabric
(Geofabric). Geofabric Surface Catchments V2.1.1.” Commonwealth of Australia.
Available at: http://www.bom.gov.au/water/geofabric/download.shtml [Accessed
June 1, 2015].
BOM - Bureau of Meteorology. 2016. “Maps of average conditions.” Available at:
http://www.bom.gov.au/climate/averages/maps.shtml [Accessed March 10, 2016].
Brown, C., C. Trusler, and K. Davis. 2010. “Managed Investment Scheme regulatio n:
Lessons from the Great Southern failure.” Jassa - The Finsia Journal of Applied
Finance (2):23–28.
Bustamante, M., C. Robledo‐Abad, R. Harper, C. Mbow, N.H. Ravindranat, F. Sperling,
H. Haberl, A. Siqueira Pinto, and P. Smith. 2014. “Co-benefits , trade-offs , barriers
and policies for greenhouse gas mitigation in the agriculture , forestry and other land
use (AFOLU) sector.” Global Change Biology 20(10):3270–3290.
Chomitz, K.M., and D.A. Gray. 1996. “Roads, land use, and deforestation: A spatial
model applied to Belize.” The World Bank Economic Review 10(3):487–512.
Cleary, S., M. Bari, and K. Smettem. 2010. “Targeting strategic tree and perennial pasture
plantings to reduce stream salinity in the Warren River Recovery Catchment.” In
Ozwater. Brisbane, Australia, pp. 1–8.
Collins, P.D.K., and D.F. Barrett. 1980. “Shannon, Warren and Donnelly River Basins -
water resources survey - Report No. WRB 6.”
Commonwealth of Australia. 2002. “Plantations for Australia : The 2020 Vision. ”
Canberra, Australia:24. Available at: http://www.agriculture.gov.au/S tyle
Library/Images/DAFF/__data/assets/pdffile/0009/2398185/plantations-australia-
2020-vision.pdf [Accessed April 10, 2015].
97
Chapter 3. Predicting future plantation forest in response to policy

Conservation Commission of Western Australia. 2013. “Forest management plan 2014-


2023.”
Cooner, A., Y. Shao, and J. Campbell. 2016. “Detection of urban damage using remote
sensing and Machine Learning Algorithms: Revisiting the 2010 Haiti earthquake.”
Remote Sensing 8(10):868.
Corbelle-Rico, E., V. Butsic, M.J. Enríquez-García, and V.C. Radeloff. 2015.
“Technology or policy? Drivers of land cover change in northwestern Spain before
and after the accession to European Economic Community.” Land Use Policy
45:18–25.
CSIRO - Commonwealth Scientific and Industrial Research Organisation. 2016. “Soil
and Landscape Grid National Soil Attribute Maps.” The CSIRO Data Access Portal.
Available at: https://data.csiro.au/dap/search?q=TERN+Soil [Accessed December
10, 2016].
Deilami, K., and M. Kamruzzaman. 2017. “Modelling the urban heat island effect of
smart growth policy scenarios in Brisbane.” Land Use Policy 64:38–55.
Ellis, E.A., K.A. Baerenklau, R.M. Martinez, and E. Chávez. 2010. “Land use/land cover
change dynamics and drivers in a low-grade marginal coffee growing region of
Veracruz, Mexico.” Agroforestry Systems 80(1):61–84.
Forest Industries Federation. 2015. Western Australia Plantations: The Missing Piece of
the Puzzle? Bentley WA.
FWPA - Forest and Wood Products Australia. 2011. Review of policies and investment
models to support continued plantation investment in Australia. Project No:
PRA189-1011. R. Fegely, M. Stephens, and A. Hansard, eds. Melbourne, Australia.
FWPA - Forest and Wood Products Australia. 2012. “The case for renewed development
in plantations Identifying forest values and the constraints to attainment – Stage
one.” :93.
Gavran, M. 2014. “Australian plantation statistics 2014 update. Technical report 14.2.”
:14. Available at:
http://data.daff.gov.au/data/warehouse/aplnsd9ablf002/aplnsd9ablf201409/AustPla
ntationStats_2014_v.1.0.0.pdf [Accessed February 21, 2016].
Gove, A.D., R. Sadler, M. Matsuki, R. Archibald, S. Pearse, and M. Garkaklis. 2013.
“Control charts for improved decisions in environmental management: A case study
of catchment water supply in south-west Western Australia.” Ecological
Management and Restoration 14(2):127–134.
Government of Western Australia. 2017. “Road Network.” Available at:
https://catalogue.data.wa.gov.au/dataset/ntwk-iris-road-
network/resource/7fa83c82-e1d5-3deb-9b2a-4f95e1ccca7a [Accessed June 1,
2017].
Guan, D., H. Li, T. Inohae, W. Su, T. Nagaie, and K. Hokao. 2011. “Modeling urban land
use change by the integration of cellular automaton and Markov model.” Ecological

98
Chapter 3. Predicting future plantation forest in response to policy

Modelling 222(20–22):3761–3772.
Halmy, M.W.A., P.E. Gessler, J.A. Hicke, and B.B. Salem. 2015. “Land use/land cover
change detection and prediction in the north-western coastal desert of Egypt using
Markov-CA.” Applied Geography 63:101–112.
Han, H., C. Yang, and J. Song. 2015. “Scenario simulation and the prediction of land use
and land cover change in Beijing, China.” Sustainability 7(4):4260–4279.
Hassoun, M.H. 1995. Fundamentals of artificial neural networks. Cambridge, United
States: MIT Press.
Hobbs, T.J., C.R. Neumann, W.S. Meyer, T. Moon, and B.A. Bryan. 2016. “Models of
reforestation productivity and carbon sequestration for land use and climate change
adaptation planning in South Australia.” Journal of Environmental Management
181:279–288.
Hosseinali, F., and A.A. Alesheikh. 2008. “Weighting spatial information in GIS for
copper mining exploration.” American Journal of Applied Sciences 5(9):1187–1198.
Irwin, E.G., and J. Geoghegan. 2001. “Theory, data, methods: developing spatially
explicit economic models of land use change.” Agriculture, Ecosystems &
Environment 85:7–23.
Kamusoko, C., M. Aniya, B. Adi, and M. Manjoro. 2009. “Rural sustainability under
threat in Zimbabwe - Simulation of future land use/cover changes in the Bindura
district based on the Markov-cellular automata model.” Applied Geography
29(3):435–447.
Khoi, D.D., and Y. Murayama. 2012. “Multi- layer Perceptron Neural Networks in
Geospatial Analysis.” In Y. Murayama, ed. Progress in Geospatial Analysis. Tokyo:
Springer Japan, pp. 125–141.
Kotsiantis, S.B. 2007. “Supervised machine learning: A review of classifica tio n
techniques.” Informatica 31:249–268.
Lacey, R., A. Watson, and J. Crase. 2006. “Economic effects of income-tax law on
investments in australian agriculture: With particular regard to new and emerging
industries.”
Lambin, E.F. 1997. “Modelling and monitoring land-cover change processes in tropical
regions.” Progress in Physical Geography 21(3):375–393.
Lambin, E.F., B.L. Turner, H.J. Geist, S.B. Agbola, A. Angelsen, J.W. Bruce, O.T.
Coomes, R. Dirzo, G. Fischer, C. Folke, P.S. George, K. Homewood, J. Imbernon,
R. Leemans, X. Li, E.F. Moran, M. Mortimore, P.S. Ramakrishnan, J.F. Richards,
H. Skånes, W. Steffen, G.D. Stone, U. Svedin, T.A. Veldkamp, C. Vogel, and J. Xu.
2001. “The causes of land-use and land-cover change: Moving beyond the myths.”
Global Environmental Change 11(4):261–269.
Lewis, D.J., and A.J. Plantinga. 2007. “Policies for habitat fragmentation: Combini ng
econometrics with GIS-based landscape simulations.” Land Economics 83(2):109–
127.
99
Chapter 3. Predicting future plantation forest in response to policy

Li, X., and A.G. Yeh. 2001. “Calibration of cellular automata by using neural networks
for the simulation of complex urban systems.” Environment and Planning A
33(8):1445–1462.
Li, X., and A.G. Yeh. 2002. “Neural-network-based cellular automata for simula ting
multiple land use changes using GIS.” International Journal of Geographical
Information Science 16(4):323–343.
Lipták, B.G. 1999. “Control Theory.” In Instrument engineers’ handbook: process
control. Boca Raton, United States: CRC Press, pp. 1–159.
Lothian, J., and A. Conacher. 2005. “Managing secondary salinity of rivers in South-
Western Australia: An evaluation of the water resource recovery catchment
approach.” Land Degradation & Development 16(2):189–200.
Mackarness, P., and B. Malcolm. 2006. “Public policy and managed investment schemes
for hardwood plantations.” Extension Farming System Journal 2(1):105–116.
Marcos-Martinez, R., and K.A. Baerenklau. 2015. “Controlling for misclassified land use
data: A post-classification latent multinomial logit approach.” Remote Sensing of
Environment 170:203–215.
Mas, J.F., M. Kolb, M. Paegelow, M.T. Camacho Olmedo, and T. Houet. 2014.
“Inductive pattern-based land use/cover change models: A comparison of four
software packages.” Environmental Modelling and Software 51(December):94–111.
Megahed, Y., P. Cabral, J. Silva, and M. Caetano. 2015. “Land cover mapping analysis
and urban growth modelling using remote sensing techniques in Greater Cairo
region—Egypt.” ISPRS International Journal of Geo-Information 4(3):1750–1769.
Mirkatouli, J., A. Hosseini, and A. Neshat. 2015. “Analysis of land use and land cover
spatial pattern based on Markov chains modelling.” City, Territory and Architecture
2(1):4.
Mishra, V.N., and P.K. Rai. 2016. “A remote sensing aided multi- layer perceptron-
Markov chain analysis for land use and land cover change prediction in Patna district
(Bihar), India.” Arabian Journal of Geosciences 9(4).
Munro, J. 2006. “Lower Warren River Action Plan.” Manjimup Land Conservation
District Committee.
Murayama, Y. 2012. Progress in Geospatial Analysis Y. Murayama, ed. Tokyo: Springer
Japan.
Myers, N., R.A. Mittermeier, C.G. Mittermeier, G.A.B. da Fonseca, and J. Kent. 2000.
“Biodiversity hotspots for conservation priorities.” Nature 403(6772):853–858.
Myint, S.W., and L. Wang. 2006. “Multicriteria decision approach for land use land cover
change using Markov chain analysis and a cellular automata approach.” Canadian
Journal of Remote Sensing 32(6):390–404.
NewForests. 2015. “Rationalising Timberland Managed Investment Schemes: The
changing landscape of Australia’s forestry investment sector.” :1–4. Available at:

100
Chapter 3. Predicting future plantation forest in response to policy

https://www.newforests.com.au/wp-content/uploads/2015/06/New-Forests-MIS-
Review.pdf [Accessed December 20, 2015].
NextGIS. 2016. “Landscape change analysis with MOLUSCE - methods and algorithms. ”
Available at: http://wiki.gis-
lab.info/w/Landscape_change_analysis_with_MOLUSCE_-
_methods_and_algorithms [Accessed July 1, 2016].
Otukei, J.R., and T. Blaschke. 2010. “Land cover change assessment using decision trees,
support vector machines and maximum likelihood classification algorithms. ”
International Journal of Applied Earth Observation and Geoinformation 12(1):27–
31.
Pannell, D.J., and A.M. Roberts. 2010. “Australia’s National Action Plan for salinity and
water quality: A retrospective assessment.” Australian Journal of Agricultural and
Resource Economics 54(4):437–456.
Pannell, D.J., and A.M. Roberts. 2009. “Conducting and delivering integrated research to
influence land-use policy: salinity policy in Australia.” Environmental Science and
Policy 12(8):1088–1098.
Park, S., S. Jeon, S. Kim, and C. Choi. 2011. “Prediction and comparison of urban growth
by land suitability index mapping using GIS and RS in South Korea.” Landscape
and Urban Planning 99(2):104–114.
Parliament of Australia. 2017. “Managed investment schemes.” Available at:
http://www.aph.gov.au/Parliamentary_Business/Committees/Senate/Economics/MI
S/Report/c02 [Accessed January 10, 2017].
Pijanowski, B., D.G. Brown, B.A. Shellito, and G.A. Manik. 2002. “Using Neural
Networks and GIS to forecast land use changes: A land transformation model.”
Computers, Environment and Urban Systems 26:553–575.
Platt, J. 2007. “Warren River - Revised management options. Why update the Salinity?”
Polyakov, M., and D. Zhang. 2008. “Population growth and land use dynamics along
urban-rural gradient.” Journal of Agricultural and Applied Economics 40(2):649–
666.
Pontius, R.G. 2000. “Quantification error versus location error in comparison of
categorical maps.” Photogrammetric Engineering and Remote Sensing 66(8):1011–
1016.
QGIS Development Team. 2016. “QGIS Geographic Information System. Open Source
Geospatial Foundation Project.” Available at: http://www.qgis.org/en/s ite/
[Accessed November 4, 2016].
R Core Team. 2016. “R: A language and environment for statistical computing. R
Foundation for Statistical Computing.” Available at: https://www.r-project.o r g/
[Accessed February 13, 2016].
Resler, L.M., Y. Shao, D.F. Tomback, and G.P. Malanson. 2014. “Predicting functio na l
role and occurrence of whitebark pine (Pinus albicaulis) at alpine Treelines: Model
101
Chapter 3. Predicting future plantation forest in response to policy

accuracy and variable importance.” Annals of the Association of American


Geographers 104(4):703–722.
Sang, L., C. Zhang, J. Yang, D. Zhu, and W. Yun. 2011. “Simulation of land use spatial
pattern of towns and villages based on CA-Markov model.” Mathematical and
Computer Modelling 54(3–4):938–943.
Smith, M.G., R.N.M. Dixon, L.H. Boniecka, M.L. Berti, T. Sparks, M.A. Bari, and J.
Platt. 2006. “Salinity Situation Statement: Warren River. Water Resource Technica l
Series No. WRT 32.”
Sohl, T.L., and P.R. Claggett. 2013. “Clarity versus complexity: Land-use modeling as a
practical tool fordecision-makers.” Journal of Environmental Management
129:235–243.
State Salinity Council. 2000. “Natural Resource Management in Western Australia: The
Salinity Strategy.”
The Treasury. 2008. “Review of Non-Forestry Managed Investment Schemes.” Availab le
at: https://archive.treasury.gov.au/documents/1401/PDF/No n-
Forestry_Managed_Investment_Schemes_Issues_Paper.pdf [Accessed December 1,
2016].
Thompson, D. 2010. “Management Plantation investment models and forestry policy.”
Australian Forest Growers Conference:21–36.
Townsend, P. V., R.J. Harper, P.D. Brennan, C. Dean, S. Wu, K.R.. J. Smettem, and S.E.
Cook. 2012. “Multiple environmental services as an opportunity for watershed
restoration.” Forest Policy and Economics 17:45–58.
Veldkamp, A., and L.O. Fresco. 1996. “CLUE: A conceptual model to study the
Conversion of Land Use and its Effects.” Ecological Modelling 85(2–3):253–270.
Veldkamp, A., and E.F. Lambin. 2001. “Predicting land-use change.” Agriculture,
Ecosystems & Environment 85(1–3):1–6.
Wang, Y., and S. Li. 2011. “Simulating multiple class urban land-use/cover changes by
RBFN-based CA model.” Computers and Geosciences 37(2):111–121.
Williams, K.J.H. 2014. “Public acceptance of plantation forestry: Implications for policy
and practice in Australian rural landscape.” Land Use Policy 38:346–354.
Wu, Q., H. Li, R. Wang, J. Paulussen, Y. He, M. Wang, B. Wang, and Z. Wang. 2006.
“Monitoring and predicting land use change in Beijing using remote sensing and
GIS.” Landscape and Urban Planning 78(4):322–333.

102
Chapter 4. Policy influence on forest plantation development
Chapter 4. Policy influence on forest plantation development

4. POLICY INFLUENCE ON FOREST PLANTATION

DEVELOPMENT IN THE WARREN RIVER CATCHMENT OF

WESTERN AUSTRALIA

ABSTRACT

The Warren River Catchment is a region within the south-west Australia Biodivers ity

Hotspot that has undergone a profound landscape transformation since the 1920s. In the

last two decades, plantation forest has emerged as a major land use within the catchment

in response to a mix of land use policies and tax inducements, with the established broad-

acre agriculture and pastoral land use shifting towards predominantly Eucalyptus

plantations. However, the spatial effects of government policy incentives on land

management is not well understood, and our intention here was to elucidate the linkage

between policy and land use change. This paper presents an empirical analysis of major

drivers of Land Use and Land Cover (LULC) changes in the Warren River Catchment

from 1979 to 2014, using a spatially explicit Multinomial Logit Model. The results

indicate that land use policies associated with tax benefits, phosphorus and soil organic

carbon in the topsoil layer, distance to nearest town, past land uses, spatial lags and private

land tenure were strong influences in the policy-driven transition towards plantatio n

forest. Regardless, the restructuring of tax deductions from the Management Investme nt

Schemes investments in the late 2000s slowed the conversion of agriculture to plantatio n

forest and reduced the retention of existing plantation forest areas. This illustrates the

sensitivity and vulnerability of investments in plantation forest to policy, with respect to

104
Chapter 4. Policy influence on forest plantation development

changes in the tax incentives. Understanding the extent of policy-driven LULC change in

the catchment will be crucial for the effective design of future public policies.

Keywords: Multinomial Logit Model; LULC; Land Use Policy; Plantation Forest;

Western Australia.

4.1. INTRODUCTION

Broad-scale changes in Land Use and Land Cover (LULC) are usually the result of

complex, multi-dimensional interactions that produce important effects with implicatio ns

for climate change, soil degradation, biodiversity loss, watershed hydrology, and human

welfare (Lambin et al. 2001; Lambin, Geist and Lepers 2003; Lubowski, Plantinga and

Stavins 2008). An understanding of such interactions can be described in terms of drivers

(e.g., climate change, soil degradation, and socioeconomic) of LULC changes, and is

necessary for policymakers concerned with a variety of land use management, and

environmental planning policies (Bell and Irwin 2002; Marcos-Martinez et al. 2017).

Since the begin of nineteenth-century LULC changes occurring in the south-west of

Western Australia were primarily stimulated by the introduction of several land use and

land management policies, as exemplified by the Group Settlement Scheme in the 1920s,

which was largely focused on agricultural expansion to reduce dependence on food

imports from the interstate. While the outcomes of land clearing lead to economic

development (e.g., farm production, rural towns, and improved livelihoods), it resulted in

extensive biodiversity loss (Myers et al. 2000), land degradation, dryland salinity

(Graham, Pannell and White 2010; Pannell and Roberts 2010), and decreases in water

105
Chapter 4. Policy influence on forest plantation development

quality. Consequently, since 1970s this region has been the focus of a number of

management strategies and policies to reduce the magnitude of actual and threatened

impacts of salinity on water and soil quality (e.g., Country Areas Water Supply of 1947

– 1947 Act amended in the 1970s, and Water Resource Recovery Catchments – WRRC

of 1996 ).

The last two decades has seen a rapid expansion of commercial tree plantations in

Western Australia, especially of the Tasmanian bluegum (Eucalyptus globulus). As a

result, Western Australia now hosts the largest area of hardwood plantations in Australia,

containing 31% (2,989 km2 ) of the nation's hardwood plantations in 2013 (Gavran 2014).

This expansion has been particularly pronounced within the Warren River Catchment

(WRC) that feeds the third largest river in the south-west of Western Australia. From

2002 to 2014, the area of forest plantations increased from 19 km2 to 193 km2 in the WRC

(see 2.4.2. LULC Changes in the WRC ). This unprecedented expansion has followed the

introduction in the late 1990s of the Federal Government’s policy Plantations for

Australia (The 2020 Vision; Commonwealth of Australia, 2002), and extensive tax

incentives offered by The Managed Investments Act 1998 (MIA 1998) to investors in

Management Investment Schemes (MIS) (Cummine 2008; Ajani 2008; FWPA 2011).

However, while some studies have examined LULC transitions in the catchment (Smith

et al. 2006), and the impact of reforestation on water resources (Bari, Silberstein and

Aryal 2012; Townsend et al. 2012), none have provided robust information on the role

of those policies in driving the historical LULC change.

To address these issues, this study is partially framed within the classic nineteenth-centur y

land use theory that attributes changes in land use to the difference in relative rents of

alternative uses. These rents vary according to intrinsic land characteristics (e.g., soil

106
Chapter 4. Policy influence on forest plantation development

fertility) and location (e.g., accessibility to the market) as first emphasized by David

Ricardo (1817) and Johann von Thünen (1966), respectively. Other factors (e.g., socio-

economic factors, and policy variables) could also influence land use decision for a given

land parcel (Chakir and Parent 2009), and so flexibility in behavioural assumptions for

LULC modelling is required. This flexibility may be achieved through the use of

Multinomial Logit (MNL) models.

This paper provides novel insights regarding the determinants of LULC change in the

WRC by identifying the major driving forces of plantation land use over the last 35 years.

We hypothesise that the conversion of agricultural land to plantation forest was spurred

by government policies and incentives associated with the introduction of the MIS.

Subsequently, a decline in plantation forest may be related with the restructuring of these

policies incentives in the late 2000s. Therefore, the specific aims of this paper are, firstly,

to develop a MNL model to identify the spatial determinants of LULC changes between

1979 and 2014, and secondly to examine the role of government policy incentives in land

management. A good understanding of how government policies influence land use

patterns over time would help policy-makers and other stakeholders to better design new

public policies.

4.2. B ACKGROUND: P LANTATION FOREST DEVELOPMENT IN WESTER N

AUSTRALIA

Following the Group Settlement Scheme (GSS) in the 1920s and later on the War Service

Land Settlement Scheme (WSLSS) in the 1950s, regions within south-west of Western

Australia (e.g., Wheatbelt and Warren) experienced extensive LULC change with the

107
Chapter 4. Policy influence on forest plantation development

clearing of native forest for the establishment of agriculture, and the accelerating

expansion of urban areas (Australian Greenhouse Office 2000). As a result, the Warren

River Catchment was declared a clearing control catchment under the 1947 Act in 1978

to prevent the rise in salinity (Government of Western Australia 2015). In 1996 this key

catchment was nominated as WRRC, and later in 2000 were prioritised for action as a

part of the State Salinity Strategy (SSS; or Western Australia Salinity Strategy), owing to

continuing salinity and resultant water quality concerns (State Salinity Council 2000).

By the late 1990s plantation forest in the WRC, which was one of the management options

proposed by SSS, had begun to spread rapidly across previously cleared agricultura l

lands. This LULC conversion was a result of substantial plantation investment in the

region from international groups (i.e, Oji International), and also due to an agribusiness

scheme market proposed by the Government, with the establishment of The 2020 Vision

in 1997 and MIA 1998, and their associated income tax subsidies introduced by the

Income Tax Assessment Act of 1997 (ITAA 1997) and Product Ruling System (PRS) of

1998. In this sense, The 2020 Vision and MIA 1998 were directed to increase a domestic

plantation forest industry, enhance rural economies and potentially assist in salinity

amelioration, carbon sequestration and other environmental issues. However, despite the

potential of reforestation to provide a set of environmental services (George et al. 2012),

these possible co-benefits generated by tree plantations were not clearly emphasized in

those policies.

For instance, The 2020 Vision aims to treble the area of Australia’s plantation forest

between 1997 and 2020 to a total of 30,000 km2 . To achieve this target, diverse initiatives

to stimulate plantation development have been proposed, including tax inducements (e.g.

Taxation Ruling TR 2000/8; Commonwealth of Australia 2002). On the other hand, the

108
Chapter 4. Policy influence on forest plantation development

MIA 1998 created a retail investment structure that allowed investors to take an up-front

tax deduction for the expense of investing in timber plantations and agribusiness

development activities through Managed Investment Schemes (MIS). Hereupon, the

ITAA 1997, PRS and TR 2000/8 lent support to The 2020 vision by providing certainty

to investors that investments in MIS were an allowable deduction, whilst simultaneo us ly

ensuring the continued expansion of agribusiness and plantation forest.

Consequently, the MIS industry was seen as a direct product of tax legislation with

demand from investors above government expectations. Therefore, the Federal

Government of Australia attempted to limit the growth of MIS schemes by revising the

deductibility of MIS investor contributions through the Australian Taxation Office (ATO)

in 2006. According to this revision, from 1 July 2008 contributions paid by non-forestry

and forestry MIS investors would no longer be tax deductible. However, in 2007, the

Government revised its decision about forestry MIS, and Division 394 was inserted into

the ITAA 1997 to provide tax certainty for forestry MIS arrangements (ATO 2017b; The

Treasury 2008). In the same year, the TR 2007/8 was released replacing the TR 2000/8.

Whereas investors in forestry MIS could still claim an immediate tax deduction for

expenditure incurred in the scheme under Division 394, non-forestry MIS relied on the

same tax legislation as other businesses.

In 2009, once again, the Government withdrew its decision, and since then investors in

non-forestry MIS can continue to claim deductions in accordance with the relevant terms

under section 8-1 of the ITAA 1997, depending on the facts of each situation (ATO

2017a). Although forestry MIS is protected under the Division 394 of ITAA 1997 since

2007, past adjustments to the taxation status of these MIS have now generated significa nt

uncertainty among investors. The subsequent lack of growth in MIS may be attributab le,

109
Chapter 4. Policy influence on forest plantation development

at least in part, to reduced investor confidence resulting from the continued restructur ing

of the MIS. To summarise the foregoing, a timeline of selected land use policies, taxation

rules and historical aspects of LULC changes in the WRC is presented in Figure 4.1.

Figure 4.1. Historical land use policies and its effects on LULC changes in WRC.

4.3. STUDY AREA

The Warren River Catchment is about 300 km south-east of Perth in Western Australia,

with the Tone and Perup Rivers as headwaters (Figure 4.2). It is located within the

Southwest Australia Ecoregion, representing one of the 34 Global Biodiversity Hotspot

(Myers et al. 2000), and covers an area of 4,416 km2 which extends from near Kojonup

on the north-east of the catchment to the south-coast of the state encompassing the small

towns of Manjimup and Pemberton. The region experiences a warm temperate

Mediterranean climate, with long, hot, dry summers and cool, wet winters. It has a mean

annual flow of 291 Gl (1990-2001; at the Barker Road Crossing gauging station; Smith

et al. 2006) and the mean annual rainfall ranges from 550 mm/year to over 1200 mm/year,

110
Chapter 4. Policy influence on forest plantation development

with most rain falling during winter and early spring (Townsend et al. 2012). Situated in

a high rainfall zone, the WRC is considered highly suitable for a predominate ly

Tasmanian Bluegum (Eucalyptus globulus) hardwood forestry.

Figure 4.2. Location of the study area.

Native forest of marri and wandoo dominated the upper catchment until the 1920s (Beard

1981). However, following the GSS, and later on with the WSLSS, extensive land

clearing occurred in this region, resulting in a significant increase in stream salinity. Prior

to clearing, salinity levels varied between 120 to 350 mg/L TDS (Total Dissolved Solids;

Smith et al. 2006). However, the salinity level exceeded 500 mg/L TDS by the late 1960s

(Munro 2006), with the current average annual salinity of around 1000 mg/L at Barker’s

111
Chapter 4. Policy influence on forest plantation development

Road Crossing gauging station (Townsend et al. 2012). Most of this salt load, around

74%, is from the Tone and Perup rivers located in the heavily cleared upper catchment

(Platt 2007).

In the late 1990s, an extensive and growing Eucalyptus plantation industry emerged on

already cleared lands in WRC, following the introduction of tax concessions and

incentives by the Federal Government. Since then, the proportion of cleared areas for

agriculture dropped from a total of 30% in 1979 to 25% in 2014 (see 2.4.2. LULC

Changes in the WRC). These changes, especially the conversion of agricultural and

pastoral areas to plantation forest, have the potential to significantly reduce the water

salinity in the catchment over the long term (Cleary et al. 2010; Smith et al. 2006), to

support the SSS. As demonstrated, the WRC offers a unique context in which to

disentangle the major drivers of landscape changes due to its complex mosaic of protected

natural forests and agriculture, very high biodiversity values, and shifting land use and

land management policies over time.

4.4. MATERIALS AND METHODS

4.4.1. Modelling approach: Multinomial Logit Model (MNL)

Our modelling approach assumes that LULC spatial patterns and their changes result from

landowners’ choices, in which they select an optimal LULC to maximise his or her utility.

We assume that utility 𝑊𝑛𝑡𝑗 , derived from the parcel of land 𝑛 in land use 𝑗 at time 𝑡 is

equal to the present value of future streams of return to this parcel of land. This return

depends on four groups of factors: biophysical land features (e.g., phosphorus), locations

112
Chapter 4. Policy influence on forest plantation development

(e.g., distance to the nearest town), socio-economics (e.g., land use policies), and spatial-

temporal controls (e.g., previous LULC). If the landowner converts a parcel of land from

use 𝑖 to any alternative use, they will incur conversion cost 𝐶 𝑛𝑡𝑗|𝑖 , which depends on

biophysical, locational and socioeconomic factors, and is specific to land uses 𝑖 and 𝑗.

The landowner will chose to convert the parcel of land currently in land use 𝑖 to a new

land use 𝑗 if 𝑊𝑛𝑡𝑗 − 𝐶𝑛𝑡𝑗|𝑖 ≥ 𝑊𝑛𝑡𝑘 − 𝐶𝑛𝑡𝑘|𝑖 for all possible uses 𝑘, including the current

𝑖 (Bockstael 1996).

Here we define utility of converting a parcel 𝑛 from land use 𝑖 to land use 𝑗 as 𝑈𝑛𝑡+1𝑗|𝑖 =

𝑊𝑛𝑡𝑗 − 𝑊𝑛𝑡𝑖 − 𝐶𝑛𝑡𝑗|𝑖 . The expected utility of land use conversion is not observable

directly, but can be expressed as a representative utility 𝑉𝑛𝑡+1𝑗 |𝑖 that captures the factors

affecting utility, and random part 𝜖𝑛𝑡𝑗 . Assuming these errors have an IID (identically and

independently distributed) Type 1 Gumbel distribution, the probability can be specified

as a multinomial logit model (MNL; McFadden 1973). The observed response 𝑦𝑛 is a

vector of probabilities (of length 𝐽 = 3), where the landowner has an option to allocate a

given parcel of land n to one of 𝐽 possible LULC types, 𝑗 ∈ {agriculture, native forest,

plantation forest}. The probability that a parcel 𝑛 that is currently in land use i will be

allocated to land use j is:

exp( 𝛼𝑖𝑗 + β′𝑗x𝑛 )


𝑝(𝑦𝑛 = 𝑗|𝑖) = ∑𝐽−1 exp(𝛼𝑖𝑘+ β′𝑘x𝑛)
𝑖 = 0, … 𝐽 − 1, 𝑗 = 0, … 𝐽 − 1, β𝑟 = 0, 𝛼𝑖𝑟 = 0 (4.1)
𝑘=0

where 𝛼𝑖𝑗 is the transition-specific constant for each combination of initial and final

LULC (Polyakov and Zhang 2008), β𝑗 is a vector of coefficients of the biophysical land

features, locations, socio-economics, and spatial-temporal factors specific to LULC

alternative 𝑗, and x𝑛 is the vector of explanatory variables for each grid cell 𝑛. In order

113
Chapter 4. Policy influence on forest plantation development

for coefficients to be estimated in MNL, they need to be normalised by setting β𝑟 = 0 and

𝛼𝑖𝑟 = 0, where r is a reference LULC category. Therefore, in the case with 𝐽 possible

outcomes, 𝐽 − 1 sets of coefficients 𝛽 and 𝐽 × (𝐽 − 1) of coefficients 𝛼 need to be

estimated.

Parameter estimates from the MNL model are not readily interpretable. Therefore

corresponding marginal effects were computed. We evaluated the marginal effects for

each initial land use as the partial derivatives of the predicted probabilities with respect

to the vector of predictors:

𝜕𝑝( 𝑦=𝑗|𝑖) 𝐽−1


= 𝑝 (𝑦 = 𝑗|𝑖)(β𝑗 − ∑𝑘=0 𝑝(𝑦 = 𝑗|𝑖)β𝑘 ) 𝑖 = 0, … 𝐽; 𝑗 = 0, … 𝐽 (4.2)
𝜕𝐱

The standard error of marginal effects were estimated using the Delta Method (Greene

2003).

There are econometric issues caused by the spatial nature of the data. One issue arises

from spatially correlated unobservable variables. If they are spatially correlated with the

dependent variable, the estimates of the coefficients can be biased. A second common

challenge facing MNL is spatial-autocorrelation, which may emerge when the LULC of

nearby grid cells directly affects the LULC at each location (Irwin and Bockstael 1998).

For example, conversions to plantation forest may be more likely near existing forest

plantations, due to the adoption of similar land use practices by neighbouring landowners

with similar land characteristics. The random sampling strategy for acquiring data should,

in principle, downplay the effect of spatial autocorrelation (see section 4.4.3. Sample

Size). However, to account for this bias, we include variables representing spatially-

temporally LULC type in previous period. This allows us to capture the potential impact

of the proximity of neighbouring land use change (Plant 2012; Irwin and Bockstael 1998;

114
Chapter 4. Policy influence on forest plantation development

Polyakov and Zhang 2008). The equation to calculate spatiotemporal lag 𝑙𝑛𝑡𝑗 of land use

j for parcel n at time t is given by:

∑𝑚(𝑦𝑚,𝑡−1=𝑗)/𝑑𝑛,𝑚
𝑙 𝑛𝑡𝑗 = ∑𝑚 1/𝑑𝑛,𝑚
∀ 𝑚: 𝑑𝑛,𝑚 ≤ 3 𝑘𝑚 (4.3)

where 𝑦𝑚,𝑡−1 is land use of parcel 𝑚 in period 𝑡 − 1 ; 𝑑𝑛,𝑚 is distance between parcels

𝑛 and 𝑚; 𝑚: 𝑑𝑛,𝑚 ≤ 3 𝑘𝑚 indicates that we use all parcels 𝑚 within distance 3 km of

parcel n. In this study, we use the ‘vglm’ function in the VGAM package to estimate

MNLs in the R software (Yee 2010; R Core Team 2016).

4.4.2. Data sources

4.4.2.1. Land‐Use and Land-Cover dataset

The LULC database was developed in Chapter 2 using multi-temporal remote sensing

imagery from Landsat Multispectral Scanner (MSS), Thematic Mapper (TM), Enhanced

Thematic Mapper (ETM+), and Operational Land Imager (OLI), for four years (1979,

1992, 2003, and 2014) with a spatial resolution of 30 meters. In addition to these four

existing LULC maps, this study included a LULC classification based on Landsat ETM+

acquired on 01 February 2010 (Path 111/Row 84) and on 24 February 2010 (Path

112/Row 84). This 2010 classification enabled observation of LULC changes

immediately following the amendment in 2007 of tax arrangements for forestry MIS. A

supervised object-based classification approach was conducted using a Support Vector

Machine (SVM) classifier, and was implemented using the same methodology given in

Chapter 2. The LULC maps discriminate between the following classes: agriculture,

water, native forest, sand dunes, harvested native forest and plantation forest.

115
Chapter 4. Policy influence on forest plantation development

4.4.2.2. Other data used for modelling

Table 4.1 shows a list of potential explanatory variables for the MNL. The variables that

might drive LULC change trajectories in the WRC were chosen in part based on the land-

rent theories postulated by Ricardo (1817) and von Thünen (1966). For instance, we

hypothesised that biophysical measures of land productivity (i.e., Ricardo) and location

(i.e., von Thünen) affect returns to land uses and, therefore, decisions of LULC allocatio n.

Moreover, we believe that policies play an important role in shaping farmers’ land use

choices; therefore, two key changes in tax incentives applied to commercial plantatio n

forest were included in the model as explanatory variables.

In this study, the explanatory variables are classified into four groups, namely:

biophysical land features, location, socio-economic and spatial-temporal controls

variables. A digital elevation model (DEM) from the Shuttle Radar Topography Mission

(SRTM 2017); 1 Arc-Second Global available by United States Geological Survey

USGS) was acquired and, six terrain characteristics (e.g. slope, aspect) were calculated

using the ‘Raster’ package (Hijmans 2016). Distances from each pixel to all major towns

were calculated. Furthermore, fourteen soil attributes from the Soil and Landscape Grid

National Soil Attribute Maps (3" resolution; CSIRO 2016), and climatic variables

available from BOM (2016) were included as potential predictors of LULC changes. Each

soil attribute was estimated at each of six depth intervals (0-5cm, 5-15cm, 15-30cm, 30-

60cm, 60-100cm and 100-200cm), as specified in the GlobalSoilMap specificatio ns

(Science Committee 2015).

116
Chapter 4. Policy influence on forest plantation development

Table 4.1. List of prospective explanatory variables for the multinomial logit model.
Explanatory Variables Units Attribute Description
Biophysical land Terrain model (30m) Terrain Ruggedness Index (TRI) m TRI represents the amount of elevation difference between adjacent cells of a DEM.
attributes Extracted from SRTM Topographic Position Index (TPI) m TPI is the difference between the value of a cell and the mean value of its 8 neighbouring pixels.
(2017) Roughness m Roughness is the difference between the maximum and the minimum value of a cell and a specified surrounding cell.
Slope Degree Slope is the degree of incline of a surface.
Aspect Degree Aspect is the orientation of slope, measured clockwise in degrees from 0 to 360.
Flow direction of Water The direction of the greatest drop in elevation. They are encoded as powers of 2 (0 to 7).
Soil attribute Total Phosphorus (PTO) % Mass fraction of total phosphorus in the soil by weight.
(CSIRO 2016) Total Nitrogen (NTO) % Mass fraction of total nitrogen in the soil by weight.
Available Water Capacity (AWC) % AWC is computed for each of the specified depth increments.
Organic Carbon (SOC) % Mass fraction of carbon by weight in the < 2 mm soil material as determined by dry combustion at 900° C.
Bulk Density - Whole Earth (BDw) g/cm3 BDw is calculated for the whole soil (fine and coarse texture fractions) in mass per unit volume.
Bulk Density - Fine Earth (BDf) g/cm3 BDf is calculated for the soil fine earth fraction (< 2mm).
Clay (CLY) % CLY is calculated as % of the <2mm fraction with a grain size of less than 2 μm.
Electrical Conductivity (ECD) dS/m ECD samples are analysed in a 1:5 saturated paste.
pH Water (pHw) NONE pHw is calculated of a 1:5 soil water solution.
Plant Exploitable Depth (DPE) cm Reflects the common depth of agricultural crops in Western Australia.
Depth to Rock (DPR) cm The estimated depth to hard rock down to 200cm
Sand (SND) % SND is based on the <2mm fraction with a grain size of 20 μm to 2mm.
Silt (SLT) % SLT is based on the <2mm fraction with a grain size between 2 - 20 μm.
Coarse Fragments (CFG) % CFG is based on the <2mm fraction with a grain size of less than 2 μm.
Climate conditions Rainfall wet season mm The average seasonal rainfall over the period 1961 to 1990 - Wet season between November and March.
(BOM 2016) Rainfall dry season mm The average seasonal rainfall over the period 1961 to 1990 - Dry season between April and October.
Temperature minimum °C The average annual distributions of temperature over the period 1961 to 1990.
Temperature maximum °C The average annual distributions of temperature over the period 1961 to 1990.
Temperature average °C The average annual distributions of temperature over the period 1961 to 1990.
Average annual sunshine hours Annual sunshine duration over the period 1900 to 2003.
Humidity annual average % The 9am and 3pm average annual relative humidity over the period 1976 to 2005.
Average annual evaporation mm Average amount of water which evaporates from an open pan annually over the period 1975 to 2005.
Evapotranspiration annual average mm Based on the transfer and balance of energy in the environment over the period 1961 to 1990.
Location Accessibility Distance to Manjimup m Log of the distance from each individual LULC pixel to the nearest town.
Distance to Pemberton m
Distance to Kojonup m
Socio-economic Policy Policies 1998 ITAA (1997), PRS (1998), the 2020 Vision (1997) and MIA (1998) responded to the timber supply shortage.
Policies 2007 The revised deductibility of MIS investments between 2006 and 2007 leading to the development of the TR2007.
Tenure (Conservation State Forest Forest administered or protected by Western Australian government.
Commission of Western Conservation Areas Western Australian protected areas.
Australia 2013). Private Land Property owned by non-governmental legal entities.
Spatial-temporal Spatial lag Previous LULC %% LULC Proportion of each of LULC in previous period within 3km weighted by inverse distance.
controls Previous LULC Previous LULC Dummy Type of preceding LULC.
Time Time since last classification’ Years The total years between classifications.

117
Chapter 4. Policy influence on forest plantation development

The first policy factor included was based on the introduction of PRS, ITAA 1997, The

2020 Vision and MIA 1998. Since they were all introduced in the late 1990s and were

covered by the same Taxation Rules (TR2000/8) from 2000 to 2007, we decided to

combine them into one explanatory variable called ‘Policies 1998’. The second policy

factor was the major review of MIS investments between 2006 and 2007 (Policies 2007),

where the TR 2000/8 was replaced by the TR 2007/8. Additionally, to test the impact of

policies on private land and state forests we included three tenure types, termed state

forest, conservation area, and private land (Conservation Commission of Western

Australia 2013). We also added a time variable namely ‘time since last classification’ to

account for different re-measurement periods (Polyakov, Wear and Huggett 2010).

Hence, the prospective explanatory variables included: (i) biophysical land features

divided into 84 potential soil variables (14 soil attributes x 6 depth intervals), 9 climate

conditions, and 6 terrain characteristics; (ii) 3 location attributes (in terms of distance to

town); (iii) 5 socio-economic factors, including the 2 policy sets targeting plantatio n

forest; and (iv) 3 spatial temporal controls (Table 4.1). The most appropriate combinatio n

of explanatory variables was selected based on a preliminary data analysis (e.g. selection

of variables and multicollinearity analysis, see Section 4.4.4.).

4.4.3. Sample size

Five LULC maps (1979, 1992, 2003, 2010, and 2014) allow LULC transitions to be

identified for four periods (1979-1992, 1992-2003, 2003-2010, and 2010-2014) for each

pixel within study area. We then randomly drew a sample of 54,000 (30x30 m) pixels

118
Chapter 4. Policy influence on forest plantation development

representing 0.5% of total pixels of the study area. The selected pixels comprise 13,500

observations for each transition period with only one observation per location.

Further, we excluded observations that had LULC classes “water”, “sand dunes” and

“harvested native forest” (clearcut) from our samples as they are not directly impacted by

the policies targeting plantation forest, and occupy only a very small portion of the

landscape (< 3%). The remaining observations had one of three LULC classes

(agriculture, native forest, plantation forest) at the beginning and end of the transitio n

period. Further attributes were assigned to each observation from the explanatory

variables presented in Table 4.1.

To select an appropriate sample size for which there is minimal improvement in

prediction accuracy when sample size is increased, we randomly divided our samples into

six training subsets of sizes {1000, 5000, 10000, 20000, 30000, 40000}. Test subsets were

drawn independently of the training set, numbering 30% of the size of the training subset

(i.e., {300, 1500, 3000, 6000, 9000, 12000}, respectively). The ideal sample size for the

MNL model was calculated with a margin of error no more than 5%, and was achieved

with 30,000 data observations. Once the ideal sample size value was selected, it was used

in the variables selection for the MNL model.

4.4.4. Selection of variables and multicollinearity analysis

LULC choices can be influenced by numerous factors, of which many may be highly

correlated. Multicollinearity in this instance would result in biased coefficients estimates

and unreliable and unstable model predictions (Alin 2010). To overcome

multicollinearity, we applied both all-subset regression and multicollinearity analysis.

119
Chapter 4. Policy influence on forest plantation development

Firstly, all-subset regression was applied to the groups of soils parameters to generate 14

MNLs – one for each pairwise combination (i.e., six depth intervals for each soil attribute)

of soil parameters. Of these the best three models were selected post-hoc by minimis ing

the Akaike's Information Criterion (AIC) scores (Goodenough, Hart and Stafford 2012;

Munroe and Muller 2007).

Variance inflation factors (VIF) were then computed independently for each subset of

variables (e.g., the three selected soil parameters, climate conditions, and accessibility –

Table 4.1) (Zuur et al. 2009), and the backward elimination process was applied to remove

the explanatory variables that were causing collinearity in each subset. Next, the VIF

values were calculated again for the reduced set of variables until each subset of variables

have all explanatory variable with the 𝑉𝐼𝐹𝑖 ≤ 2. The VIF quantifies how much of the

variance (the square of the standard error) associated with a given coefficient is ‘infla ted’

because of a linear dependence on the other predictors in the model. VIF values of ≥5

(moderate multicollinearity) or ≥10 (high multicollinearity) implies that the associated

regression coefficients are poorly estimated because of multicollinearity (Montgomer y,

Peck and Vining 2001). Thus, in this paper, it was desirable to state a cut-off value of

𝑉𝐼𝐹𝑖 ≤ 2, suggesting that no more than 50% of the estimated variance of a parameter

coefficient may be attributable to the remaining explanatory variables.

Finally, all predictors retained within each of the seven regression models, following

removal of multicollinear variables, were included in one single MNL model. Hence, only

16 variables remained in the pool of potential explanatory variables: Phosphorus and

organic carbon at 0-5cm and 100-200cm each, nitrogen (5-15cm and 100-200cm); slope;

distance to Manjimup and Kojonup; past state; spatial lag; time since last classificatio n;

private land; and the two set of policy variables (Policies 1998 and Policies 2007).

120
Chapter 4. Policy influence on forest plantation development

4.5. RESULTS

4.5.1. LULC trajectories

To provide context to our multinomial logit model results, figure 4.3 illustrate the

observed changes in the extent of the three LULC classes over the five data captures:

1979, 1992, 2003, 2010 and 2014. The area of native forest amounted to approximate ly

68% (3,023 km2 ) of the study area in 2014, and has remained relatively stable since 1979.

Hence, the control on clearing embedded by the 1947 Act appears to have been effective

in the WRC in minimising the clearing of native vegetation.

Broad-acre agriculture is the second largest land use in the catchment, occupying

approximately 25% of the landscape. However, this extent has decreased by 5% since

1979. Conversely, the plantation forest industry was almost non-existent in 1992, growing

to about 192 km2 in 2003, and reaching its largest area in 2010 (238 km2 ). From late 1990s

onwards, Eucalyptus plantations have been mainly established on the north-east of the

catchment and formerly occupied primarily by broad-acre cropping and pastures.

However, a slight decrease in plantation forest was observed in 2014 compared to 2010.

This is potentially a consequence of investors’ concerns regarding the proposed changes

of tax treatment for MIS investments in 2007.

121
Chapter 4. Policy influence on forest plantation development

Figure 4.3. Trends in surface areas and percentage for the three major LULC classes in
the study area.

4.5.2. Multinomial Logit Model (MNL) of land use change

The MNL model of LULC change explained 62.2% of the observed variation in the data

(McFadden’s pseudo-R2 ; Domencich and McFadden 1975). This model may be explained

in terms of its marginal effects (Table 4.2). For continuously-valued independent

variables the marginal effect measures the instantaneous rate of change in the probability

of moving towards a new land use; whereas discrete change was measured through

indicator variables.

122
Chapter 4. Policy influence on forest plantation development

Table 4.2. Marginal Effects of explanatory variables on the transition probabilities of


LULC change.

Final LULC type (Alternative J)

Explanatory variables Marginal Effect S td. Marginal Effect S td. Marginal Effect S td.
Agriculture Err. Native Forest Err. Plantation Forest Err.
Initial Land Use Agriculture
Private land 0.2128 *** 0.0271 -0.3058 *** 0.0132 0.0930 *** 0.0232
Policies 1998*Private land -0.1332 *** 0.0164 0.0071 0.0124 0.1261 *** 0.0108
Policies 2007*Private land 0.1417 *** 0.0344 -0.0161 0.0206 -0.1256 *** 0.0261
log(distance from M anjimup) 0.0357 *** 0.0090 -0.0332 *** 0.0068 -0.0025 0.0058
log(distance from Kojonup) -0.0473 * 0.0199 -0.0349 * 0.0153 0.0822 *** 0.0129
Phosphorus 0-5 cm -6.9018 *** 1.4223 4.3034 *** 1.0803 2.5984 ** 0.8274
Organic carbon 0-5 cm 0.0036 0.0054 -0.0136 ** 0.0043 0.0100 *** 0.0030
Nitrogen 5-15cm 2.9188 *** 0.2207 -1.3715 *** 0.1673 -1.5473 *** 0.1335
Phosphorus 100-200cm 24.1830 *** 3.5321 -16.9471 *** 2.6655 -7.2360 *** 2.0763
Organic carbon 100-200cm -1.1413 *** 0.1558 1.1888 *** 0.1145 -0.0475 0.0939
Nitrogen 100-200cm -1.3384 2.1212 0.1493 1.5718 1.1890 1.2939
Slope 0.0030 ** 0.0010 -0.0025 ** 0.0008 -0.0006 0.0006
log(time last classification) 0.1058 *** 0.0300 0.0085 0.0170 -0.1143 *** 0.0237
Spatial lag Agriculture 0.1759 *** 0.0265 -0.1597 *** 0.0200 -0.0162 0.0159
Spatial lag Plantation -0.1920 *** 0.0529 -0.0535 0.0414 0.2455 *** 0.0232
Initial Land Use Native Forest
Private land 0.1119 *** 0.0052 -0.1518 *** 0.0080 0.0399 *** 0.0044
Policies 1998*Private land -0.0101 * 0.0048 -0.0120 0.0075 0.0221 *** 0.0021
Policies 2007*Private land 0.0135 0.0082 0.0079 0.0092 -0.0214 *** 0.0048
log(distance from M anjimup) 0.0129 *** 0.0026 -0.0149 *** 0.0032 0.0021 0.0011
log(distance from Kojonup) 0.0086 0.0059 -0.0260 *** 0.0079 0.0174 *** 0.0026
Phosphorus 0-5 cm -1.8025 *** 0.4215 1.6617 *** 0.5044 0.1408 0.1630
Organic carbon 0-5 cm 0.0046 ** 0.0017 -0.0075 *** 0.0021 0.0028 *** 0.0006
Nitrogen 5-15cm 0.6163 *** 0.0649 -0.4423 *** 0.0897 -0.1741 *** 0.0264
Phosphorus 100-200cm 6.9239 *** 1.0404 -6.9078 *** 1.2636 -0.0160 0.4072
Organic carbon 100-200cm -0.4534 *** 0.0448 0.5520 *** 0.0592 -0.0986 *** 0.0182
Nitrogen 100-200cm -0.1265 0.6142 -0.0758 0.7266 0.2023 0.2518
Slope 0.0010 ** 0.0003 -0.0011 ** 0.0004 0.0001 0.0001
log(time last classification) 0.0034 0.0068 0.0178 * 0.0075 -0.0212 *** 0.0043
Spatial lag Agriculture 0.0622 *** 0.0078 -0.0714 *** 0.0098 0.0092 ** 0.0031
Spatial lag Plantation 0.0062 0.0162 -0.0544 ** 0.0209 0.0482 *** 0.0046
Initial Land Use Plantation Forest
Private land -0.0186 0.0467 -0.7388 *** 0.0660 0.7574 *** 0.1136
Policies 1998*Private land -0.2516 *** 0.0235 -0.3204 *** 0.0354 0.5720 *** 0.0526
Policies 2007*Private land 0.2556 *** 0.0535 0.3049 *** 0.0742 -0.5605 *** 0.1215
log(distance from M anjimup) 0.0227 0.0120 -0.0471 * 0.0184 0.0244 0.0273
log(distance from Kojonup) -0.1428 *** 0.0272 -0.2728 *** 0.0421 0.4155 *** 0.0622
Phosphorus 0-5 cm -7.4190 *** 1.7761 0.1066 2.6728 7.3124 3.8862
Organic carbon 0-5 cm -0.0124 0.0066 -0.0484 *** 0.0102 0.0608 *** 0.0143
Nitrogen 5-15cm 3.7780 *** 0.2875 1.8599 *** 0.4464 -5.6379 *** 0.6495
Phosphorus 100-200cm 23.3265 *** 4.4507 -8.3111 6.6693 -15.0154 9.7546
Organic carbon 100-200cm -0.5453 ** 0.2029 2.0448 *** 0.2945 -1.4994 *** 0.4433
Nitrogen 100-200cm -2.4172 2.7497 -2.8899 4.0704 5.3071 6.0431
Slope 0.0024 0.0013 -0.0025 0.0020 0.0001 0.0029
log(time last classification) 0.2201 *** 0.0484 0.3147 *** 0.0664 -0.5348 *** 0.1105
Spatial lag Agriculture 0.1176 *** 0.0339 -0.2151 *** 0.0512 0.0974 0.0754
Spatial lag Plantation -0.4537 *** 0.0571 -0.7328 *** 0.0814 1.1866 *** 0.1078
* Significant at 5%, ** Significant at 1%, and *** Significant at 0.1%

123
Chapter 4. Policy influence on forest plantation development

The estimated marginal effects revealed that the probability of conversion from

agriculture (P < 0.001) or native forest (P < 0.001) to plantation forest, or the retention in

plantation forest (P < 0.001), is greatest for those parcels located near Kojonup. These

probabilities decrease with proximity to Manjimup. Conversely, the probability of

transition from native forest to agriculture (P < 0.001), or retention in agriculture (P <

0.001) increases with proximity to Manjimup. Any conversion to native forest is

negatively related to the proximity to urban centres. These statistically significant effects

support the von Thünen model of rural development concentrated around urban centres.

Ricardian land rent theory is also supported by the data, with soil characteristics found to

significantly influence the probability of transitioning from one land use to another.

Parcels with high levels of nitrogen in the topsoil layer (5-15cm) were correlated with a

high probability of remaining in or transitioning to agriculture (P < 0.001); and negative ly

correlated with transitions to plantation forest (P < 0.001). By contrast, phosphorus (P <

0.01) and organic carbon (P < 0.001) in the topsoil (0-5cm) was positively and

significantly correlated with the transition from agriculture to plantation forest, and

negatively correlated to transitions from plantation forest to agriculture (Phosphorus P <

0.001). More specifically topsoils with high organic carbon (0-5cm) were more likely to

be managed for agriculture instead of native forest. Organic carbon at greater depth (100-

200cm) is associated with the presence of native forest (P < 0.001) due to deeper root

profile. Phosphorus at greater depth is positively related only to the conversion from any

LULC to agriculture (P <0.001). Slope is another parcel attribute with an influence on

LULC transitions; transitions to agriculture occur more frequently on higher sloped land

than compared to native forest (P <0.01).

124
Chapter 4. Policy influence on forest plantation development

Spatial dependence is also captured in the model, with the probability of transitio ning

from agriculture to plantation forest (P <0.001) greater when more Eucalyptus plantatio ns

are present within a 3 km radius. Similarly, a return to agriculture from plantation forest

(P <0.001) is more likely when agriculture predominates in the local landscape.

Land tenure has significant implications for LULC change in the WRC, with transition to

agriculture (P <0.001) or to plantation forest (P <0.001) more likely on private land, while

transition to native forest or staying in native forest is more likely on public land. As MIS

policies only apply to private land, the interaction between policy and land tenure was

explicitly modelled. Policies 1998 and private land was correlated with an increased

probability of transition to plantation forest from agriculture (P <0.001). This result

indicates that federal initiatives including the associated changes in tax rules (PRS and

TR 2000/8) has encouraged the expansion of plantation forest on private land within

WRC (Figure 4.4). In contrast, the transition rate from agriculture to plantation forest (P

<0.001) slowed following the major review of the taxation benefits associated with the

MIS (represented by the Policies 2007 variable). Moreover, the return of plantation forest

to broad-acre agriculture has also increased (P <0.001).

The time since last classification date was included to control for the unevenly spaced

time periods between different image captures. This variable was negatively correlated

with transitions towards plantation forest (P <0.001), accounting in part for the low

transition rate of the first period between image captures (13 years, 1979-1992). Despite

that, some land uses such as native forest remained stable over time; while others such as

plantation forest presented a higher rate of transition in a specific timeframe (i.e., 1992-

2003), with increased plantations depend somewhat on the previously cleared agricultura l

areas.

125
Chapter 4. Policy influence on forest plantation development

Figure 4.4. Map of LULC showing the expansion of plantation forest from 1992 to 2014.

4.6. DISCUSSION

The development of agricultural land use began around 1925 as part of postwar

rehabilitation GSS with the clearing of the Wilgarup, Dombakup and Lefroy sub-

catchments (Munro 2006). Since then, dairy farms, fruit and vegetable horticulture was

widespread developed near Manjimup, thus explaining the positive relationship between

126
Chapter 4. Policy influence on forest plantation development

this town and agriculture. In the 1950s following WSLSS a further increase in conversion

to agriculture was observed in the catchment on already cleared areas that were

considered to be suitable for those activities. Although the WSLSS aimed to settle new

farms on existing cleared land (Field 1963), an extensive removal of native vegetation

occurred in the upper catchment at that time, resulting in an increasing in river salinity

(Munro 2006).

Therefore, the areas in the north-east of the catchment, near Kojonup, represent a portion

of the landscape which has experienced the most landscape change since 1979,

specifically the conversion from agriculture to plantation forest. However, timber is

Manjimup’s major industry, as the rainfall near Manjimup better supports karri and jarrah

forests. The determinants of conversion model such as distance to the next town reinforce

evidence of von Thunen’s model, showing especially that new development such as

agriculture or plantation forest is more likely to locate close to urban places rather than

far away.

Plantation forest development was largely driven by government policies in the

catchment, with a rapid expansion following the MIS introduction in the late 1990s, and

subsequent stagnation with the policy changes in the late 2000s. The intentions of the tax

incentives have not achieved its purpose over the long term. Investors have been avoiding

investments in MIS since the late 2000s, in response to the considerable changes to the

structuring of the tax incentives over time (Brown, Trusler and Davis 2010; Summers et

al. 2015). Following the legislative changes of 2007 investments in MIS industry

decreased from approximately AU $1.2 billion in 2006 (The Treasury 2008) to AU $300

million in 2009, and in 2010 fell further to below AU $100 million (NewForests 2015).

Significantly, the uncertainty about the economic viability of tree plantations in Australia

127
Chapter 4. Policy influence on forest plantation development

contributed in part to the financial collapse in the last decade of at least five high-profile

agribusiness companies operating in Australia: Environinvest Ltd., FEA Plantations Ltd.,

Great Southern Group Ltd., Rewards Projects Ltd., and Timbercorp Ltd. (Thompson

2010). Since then, there has been only a small number of forestry MIS offered to

investors.

Consequently, the area of plantation forest in the catchment decreased in 2014 for the first

time since the introduction of the tax inducements for MIS investments, leading to an

increase in broad-acre agricultural areas formerly under plantation forest. This result is

consistent with the simulation of future LULC changes in the WRC proposed earlier (see

3.4.3. Future Prediction Using Multilayer Perceptron CA-Markov Model), with the

findings highlighting that Eucalyptus plantation areas will tend to decrease by 2025; while

agricultural areas may increase slightly with the reconversion of plantation forests to

agriculture. This concern extends to the whole Western Australia as regrettably some

plantations were not ideally located which lead to lower financial returns, and the likely

conversion of these unviable plantations to other land uses (Forest Industries Federation

2015).

This potential decrease in plantation area may impact not only timber production, but may

also threats the long-term efforts to decrease river salinity by 2030 to the State Salinity

Strategy’s desired potable standard (i.e., 500mg/L TDS). As already emphasised by

Bryan & Crossman (2013), failure to understand the interactions between tax incentives

and their impact across ecosystem services can result in inefficient policy outcomes, such

as are presented in this study. Our results highlighted the power of financial incentive

interventions for changing the landscape and confirmed our hypothesis that in the absence

128
Chapter 4. Policy influence on forest plantation development

or reduction of government inducements, forestry MIS would no longer be a profitable

‘business’ for investors, resulting in a decreasing of plantation areas.

Plantation forest as a mitigation strategy, when spatially targeted, can have a positive

impact on water resources (Bustamante et al. 2014). This co-benefit was also identified

by Townsend et al. (2012) who showed that reforestation could result in the restoration

of water quality in the WRC. Despite the economic benefits, there are at least three

justifications for incorporating plantations in the Australian landscape (climate change

mitigation, salinity abatement, and biodiversity conservation), however, currently these

issues are all considered separately (George et al. 2012). In the Warren River Catchment

the co-benefits generated by plantation investment incentives was neglected, and not

explicitly defined in the legislation. According to Plantinga & Wu (2003) multi-tar get

policies are more efficient than ones that target single benefits.

The consequences of basing policy on temporary incentives can clearly be seen in the

short-term increase in plantation forest within the WRC, followed by the collapse of two

of the largest MIS companies operating in Western Australia (Timbercorp Limited and

Great Southern Group Limited) when those tax incentives were partly removed through

an unanticipated change in the tax legislation. Notwithstanding, Lacey et al., (2006)

reported that most MIS sector continues to perform poorly with respect to realistic rates

of returns versus projected rates. This situation raises a crucial question why policies to

increase sustainable timber production and environmental co-benefits remain largely

unfulfilled despite large-scale government intervention. In fact, the financial collapse of

those MIS companies reflects the well-known issues of using temporary tax benefits to

fund an industry (NewForests 2015; Ajani 2008).

129
Chapter 4. Policy influence on forest plantation development

This study demonstrates how policy incentives impacting on LULC change can be

assessed with freely available remotely sensed data (e.g., United States Geological Survey

– USGS), and a range of open-source software (e.g., Quantum GIS). Moreover,

throughout the implementation of the MIS tax incentives, limited periodic assessment of

policy outcomes (e.g., timber production or co-benefits at catchment level) was

undertaken to assess program success, or whether tax changes were appropriate in

achieving long-term benefits.

One of the policy requirement underlined in our findings is that the outcomes of a policy

need to be assessed so that the policy or strategy measures can be adjusted to achieve the

policy objectives (i.e., sustainable timber production). Appropriate knowledge of the

historical drivers of LULC changes is essential because there are public policies that do

influence LULC decisions. A direction for future research is to evaluate policies that

target multiple benefits such as carbon sequestration and water quality. The governme nt

should take action towards more generous taxation breaks to encourage a multi-tar get

strategy and so avoid the decrease of plantation forest areas. To date, those environme nta l

co-benefits generated by plantations are not valued by most landowners as they do not

contribute to their revenue. Despite that, investors in plantation take the risk of investing

in a product that generates profits at mid to long term after the initial investment. In this

sense, agriculture can generally be seen as a more profitable farming system, with

potential annual production cycles and cash-flow.

However, agriculture is the main cause of the environmental degradation and saliniza tio n

of water resources. Therefore, the costs related to environmental degradation should be

borne by the agricultural sector, rather than transferred on to others including taxpayers.

In order to effectively design reliable land use policy that embraces more environmenta lly

130
Chapter 4. Policy influence on forest plantation development

friendly practices (i.e., multi-purpose plantation programs), causes of past land use

decision in response to policy initiative needs to be understood.

4.7. CONCLUSION

The introduction of The Managed Investments Act 1998 and The 2020 Vision, combined

with the associated benefits in taxation rules, has encouraged the expansion of Eucalyptus

plantations in private properties between the late 1990s and mid-2000s. Consequently,

those concessions promoted the expansion of Western Australia’s plantation forest and

reduced reliance on both native forests and imported timber. However, the change in tax

deduction caused uncertainty over the future tax treatment for MIS, discouraging the

investments in forestry schemes by the late 2000s.

The collapse of the two largest operators of agribusiness schemes in Western Australia

(Timbercorp Limited and Great Southern Limited) suggested that the MIS were never

entirely safe, long-term investments. In hindsight, these MIS were doomed to failure as

the investors were not driven by returns, but by up-front tax deductions in year one of the

investment. Profitable harvesting of the plantations a few years later would, in essence,

simply provide a windfall to the investor. Our results show the extent to which those

policies and associated tax inducements can impact LULC patterns, and also suggest that

tax-effective schemes should not merely be framed to attract people into an industry, but

be grounded in sound market principles to avoid failure. Pannell (2008) recommended

that incentives will be more effective if, once adopted, the practice will be able to mainta in

itself, otherwise it will inevitably be disadopted. Further, land use policies driven by

mitigation concerns should target multiple benefits (e.g. environmental, commercial and

131
Chapter 4. Policy influence on forest plantation development

community values), to support efficient use of land resources. A thorough understanding

of how different factors influence LULC patterns across space and time should be

considered as a crucial component in the assessment of successful policy outcomes for

both environmental and land use policy design.

4.8. REFERENCES

Ajani, J. 2008. “Australia’s Transition from Native Forests to Plantations: The


Implications for Woodchips, Pulpmills, Tax Breaks and Climate Change.” Agenda
15(3):21–38.
Alin, A. 2010. “Multicollinearity.” Wiley Interdisciplinary Reviews: Computational
Statistics 2(3):370–374.
ATO - Australian Taxation Office. 2017a. “Income Tax Assessment Act 1997 No. 38.
Chapter 1 Introduction and core provisions. Division 8: Deductions. Vol 1.” :27–28.
Available at:
https://www.legislation.gov.au/Details/C2017C00155/Html/Volume_1 [Accessed
July 10, 2017].
ATO - Australian Taxation Office. 2017b. “Income Tax Assessment Act 1997 No. 38.
Chapter 3 Specialist liability rules: Division 394: Forestry managed investme nt
schemes. Vol 7.” :343–351. Available at:
https://www.legislation.gov.au/Details/C2017C00155/Html/Volume_7 [Accessed
July 10, 2017].
Australian Greenhouse Office. 2000. “Land Clearing: A Social History - Technica l
Report No. 4.”
Bari, M.A., R.P. Silberstein, and S.K. Aryal. 2012. “Strategic tree planting and climate
change - Simulating future water availability at the Warren River catchment,
Western Australia.” Engineers Australia:167–174.
Beard, J.S. 1981. Swan — The vegetation of the Swan area, Vegetation survey of Western
Australia, 1:1 000 000 Vegetation Series, Explanatory notes to Sheet 7. Perth,
Australia: University of Western Australia Press.
Bell, K.P., and E.G. Irwin. 2002. “Spatially explicit micro-level modelling of land use
change at the rural-urban interface.” Agricultural Economics 27:217–232.
Bockstael, N.E. 1996. “Modelling Economics and Ecology: The Importantance of a
Spatial Perspective.” American Journal of Agricultural Economics 78(5):1168–
1180.

132
Chapter 4. Policy influence on forest plantation development

BOM - Bureau of Meteorology. 2016. “Maps of average conditions.” Available at:


http://www.bom.gov.au/climate/averages/maps.shtml [Accessed March 10, 2016].
Brown, C., C. Trusler, and K. Davis. 2010. “Managed Investment Scheme regulatio n:
Lessons from the Great Southern failure.” Jassa - The Finsia Journal of Applied
Finance (2):23–28.
Bryan, B.A., and N.D. Crossman. 2013. “Impact of multiple interacting financ ia l
incentives on land use change and the supply of ecosystem services.” Ecosystem
Services 4:60–72.
Bustamante, M., C. Robledo‐Abad, R. Harper, C. Mbow, N.H. Ravindranat, F. Sperling,
H. Haberl, A. Siqueira Pinto, and P. Smith. 2014. “Co-benefits , trade-offs , barriers
and policies for greenhouse gas mitigation in the agriculture , forestry and other land
use (AFOLU) sector.” Global Change Biology 20(10):3270–3290.
Chakir, R., and O. Parent. 2009. “Determinants of land use changes: A spatial
multinomial probit approach.” Papers in Regional Science 88(2):327–344.
Cleary, S., M. Bari, and K. Smettem. 2010. “Targeting strategic tree and perennial pasture
plantings to reduce stream salinity in the Warren River Recovery Catchment.” In
Ozwater. Brisbane, Australia, pp. 1–8.
Commonwealth of Australia. 2002. “Plantations for Australia : The 2020 Vision. ”
Canberra, Australia:24. Available at: http://www.agriculture.gov.au/S tyle
Library/Images/DAFF/__data/assets/pdffile/0009/2398185/plantations-australia-
2020-vision.pdf [Accessed April 10, 2015].
Conservation Commission of Western Australia. 2013. “Forest management plan 2014-
2023.”
CSIRO - Commonwealth Scientific and Industrial Research Organisation. 2016. “Soil
and Landscape Grid National Soil Attribute Maps.” The CSIRO Data Access Portal.
Available at: https://data.csiro.au/dap/search?q=TERN+Soil [Accessed December
10, 2016].
Cummine, A. 2008. “New tax arrangements for forestry Managed Investment Schemes -
why and how , and what they mean.” In Australian Forest Growers Biennial
Conference. Canberra, Australia: Australian Forest Growers (AFG), pp. 58–64.
Domencich, T., and D.L. McFadden. 1975. “Statistical Estimation of Choice Probability
Functions.” In Urban Travel Demand: A Behavioral Analysis. New York: North-
Holland Pub. Co., pp. 101–125.
Field, T.P. 1963. Postwar land settlement in Western Australia. Lexington: Univers ity
Press of Kentucky.
Forest Industries Federation. 2015. Western Australia Plantations: The Missing Piece of
the Puzzle? Bentley WA.
FWPA - Forest and Wood Products Australia. 2011. Review of policies and investment
models to support continued plantation investment in Australia. Project No:
PRA189-1011. R. Fegely, M. Stephens, and A. Hansard, eds. Melbourne, Australia.

133
Chapter 4. Policy influence on forest plantation development

Gavran, M. 2014. “Australian plantation statistics 2014 update. Technical report 14.2.”
:14. Available at:
http://data.daff.gov.au/data/warehouse/aplnsd9ablf002/aplnsd9ablf201409/AustPla
ntationStats_2014_v.1.0.0.pdf [Accessed February 21, 2016].
George, S.J., R.J. Harper, R.J. Hobbs, and M. Tibbett. 2012. “Agriculture , Ecosystems
and Environment A sustainable agricultural landscape for Australia : A review of
interlacing carbon sequestration , biodiversity and salinity management in
agroforestry systems.” Agriculture, Ecosystems and Environment 163:28–36.
Goodenough, A.E., A.G. Hart, and R. Stafford. 2012. “Regression with empirical variable
selection: Description of a new method and application to ecological datasets.” PLoS
ONE 7(3):1–10.
Government of Western Australia. 2015. “Country Areas Water Supply Act 1947.
Control of catchment areas Part IIA.” :19. Available at:
https://www.slp.wa.gov.au/pco/prod/filestore.nsf/FileURL/mrdoc_25672.pdf/$FIL
E/Country Areas Water Supply Act 1947 - %5B10-00-00%5D.pdf?OpenEle me nt
[Accessed December 19, 2015].
Graham, T., D.J. Pannell, and B. White. 2010. “On-site and off-site economic benefits of
dryland salinity mitigation resulting from establishment of perennial vegetation on
farms: a breakeven analysis.” Australasian Journal of Environmental Management
17(2):112–124.
Greene, W.H. 2003. “Models for Discrete Choice.” In Econometric analysis. Upper
Saddle River, N.J.: Prentice Hall, pp. 663–755.
Hijmans, R.J. 2016. “Raster: Geographic Data Analysis and Modeling.” Available at:
https://cran.r-project.org/package=raster [Accessed September 8, 2016].
Irwin, E.G., and N.E. Bockstael. 1998. “Interacting agents, spatial externalities and the
evolution of residential land use patterns.”
Lacey, R., A. Watson, and J. Crase. 2006. “Economic effects of income-tax law on
investments in australian agriculture: With particular regard to new and emerging
industries.”
Lambin, E.F., H.J. Geist, and E. Lepers. 2003. “Dynamics of land-use and land-cover
change in tropical regions.” Annual Review of Environment and Resources
28(1):205–241.
Lambin, E.F., B.L. Turner, H.J. Geist, S.B. Agbola, A. Angelsen, J.W. Bruce, O.T.
Coomes, R. Dirzo, G. Fischer, C. Folke, P.S. George, K. Homewood, J. Imbernon,
R. Leemans, X. Li, E.F. Moran, M. Mortimore, P.S. Ramakrishnan, J.F. Richards,
H. Skånes, W. Steffen, G.D. Stone, U. Svedin, T.A. Veldkamp, C. Vogel, and J. Xu.
2001. “The causes of land-use and land-cover change: Moving beyond the myths.”
Global Environmental Change 11(4):261–269.
Lubowski, R.N., A.J. Plantinga, and R.N. Stavins. 2008. “What drives land-use change
in the United States? A national analysis of landowner decisions.” Land Economics
84(4):529–550.

134
Chapter 4. Policy influence on forest plantation development

Marcos-Martinez, R., B.A. Bryan, J.D. Connor, and D. King. 2017. “Land Use Policy
Agricultural land-use dynamics : Assessing the relative importance of
socioeconomic and biophysical drivers for more targeted policy.” Land Use Policy
63:53–66.
McFadden, D. 1973. “Conditional logit analysis of quantitative choice models.” In Paul
Zarembka, ed. Frontiers in Econometrics. New York: Academic Press, pp. 105–142.
Montgomery, D.C., E.A. Peck, and G.G. Vining. 2001. Introduction to Linear Regression
Analysis 3rd ed. New York, USA: John Wiley & Sons.
Munro, J. 2006. “Lower Warren River Action Plan.” Manjimup Land Conservation
District Committee.
Munroe, D.K., and D. Muller. 2007. “Issues in spatially explicit statistical land-use/cover
change (LUCC) models: Examples from western Honduras and the Central
Highlands of Vietnam.” Land Use Policy 24(3):521–530.
Myers, N., R.A. Mittermeier, C.G. Mittermeier, G.A.B. da Fonseca, and J. Kent. 2000.
“Biodiversity hotspots for conservation priorities.” Nature 403(6772):853–858.
NewForests. 2015. “Rationalising Timberland Managed Investment Schemes: The
changing landscape of Australia’s forestry investment sector.” :1–4. Available at:
https://www.newforests.com.au/wp-content/uploads/2015/06/New-Forests-MIS-
Review.pdf [Accessed December 20, 2015].
Pannell, D.J. 2008. “Public Benefits, Private Benefits, and Policy Mechanism Choice for
Land-Use Change for Environmental Benefits.” Land Economics 84(2):225–240.
Pannell, D.J., and A.M. Roberts. 2010. “Australia’s National Action Plan for salinity and
water quality: A retrospective assessment.” Australian Journal of Agricultural and
Resource Economics 54(4):437–456.
Plant, R.E. 2012. “Measures of Spatial Autocorrelation.” In Spatial Data Analysis in
Ecology and Agriculture Using R. Boca Raton, Florida, United States: CRC Press,
pp. 95–124.
Plantinga, A.J., and J. Wu. 2003. “Co-Benefits from carbon sequestration in forests:
evaluating reductions in agricultural externalities from an afforestation policy in
Wisconsin.” Land Economics 79(1):74–85.
Platt, J. 2007. “Warren River - Revised management options. Why update the Salinity?”
Polyakov, M., D.N. Wear, and R.N. Huggett. 2010. “Harvest choice and timber supply
models for forest forecasting.” Forest Science 56(4):344–355.
Polyakov, M., and D. Zhang. 2008. “Population growth and land use dynamics along
urban-rural gradient.” Journal of Agricultural and Applied Economics 40(2):649–
666.
R Core Team. 2016. “R: A language and environment for statistical computing. R
Foundation for Statistical Computing.” Available at: https://www.r-project.o r g/
[Accessed February 13, 2016].

135
Chapter 4. Policy influence on forest plantation development

Ricardo, D. 1817. The principles of political economy and taxation. London, United
Kingdom: John Murray.
Science Committee. 2015. “Specifications Tiered GlobalSoilMap products, Release 2.4.”
GlobalSoilMap project:52. Available at: http://www.globalsoilmap.net/ [Accessed
March 14, 2016].
Smith, M.G., R.N.M. Dixon, L.H. Boniecka, M.L. Berti, T. Sparks, M.A. Bari, and J.
Platt. 2006. “Salinity Situation Statement: Warren River. Water Resource Technica l
Series No. WRT 32.”
SRTM - Shuttle Radar Topography Mission. 2017. “Digital Elevation - 1 Arc-Second
Global.” Available at: https://earthexplorer.usgs.gov/ [Accessed January 17, 2017].
State Salinity Council. 2000. “Natural Resource Management in Western Australia: The
Salinity Strategy.”
Summers, D.M., B.A. Bryan, M. Nolan, and T.J. Hobbs. 2015. “The costs of
reforestation: A spatial model of the costs of establishing environmental and carbon
plantings.” Land Use Policy 44:110–121.
The Treasury. 2008. “Review of Non-Forestry Managed Investment Schemes.” Availab le
at: https://archive.treasury.gov.au/documents/1401/PDF/No n-
Forestry_Managed_Investment_Schemes_Issues_Paper.pdf [Accessed December 1,
2016].
Thompson, D. 2010. “Management Plantation investment models and forestry policy.”
Australian Forest Growers Conference:21–36.
Thünen, J.H. von. 1966. Von Thünen’s isolated State: an English edition of Der isoliere
Staat 1st ed. Oxford, United States: Pergamon Press.
Townsend, P. V., R.J. Harper, P.D. Brennan, C. Dean, S. Wu, K.R.. J. Smettem, and S.E.
Cook. 2012. “Multiple environmental services as an opportunity for watershed
restoration.” Forest Policy and Economics 17:45–58.
Yee, T.W. 2010. “The VGAM Package for Categorical Data Analysis.” J. Statist. Soft.
32(10):1–34.
Zuur, A.F., E.N. Ieno, N. Walker, A.A. Saveliev, and G.M. Smith. 2009. Mixed Effects
Models and Extensions in Ecology with R. Dordrecht: Springer.

136
Chapter 5. Conclusions
Chapter 5. Conclusions

5. CONCLUSIONS

5.1. OVERVIEW

Land Use and Land Cover changes are influenced by a variety of natural and

anthropogenic processes and can generate significant economic, social and environme nta l

impacts that may become the foci of a wide range of policy issues. There has been much

debate in Australia recently over the merits of tax incentive policies to support the

plantation forest industry. While this land use has been recognised as an important

component of Australia’s economy by supplying forest products, as well as by potentially

helping to solve critical issues in natural resource management for example dryland

salinity, there has been some controversy surrounding the efficiency of the taxation

system for forestry Managed Investment Schemes (MIS).

Within an evidence-based framework to land-use policy development the design of

efficient management policies requires knowledge of the drivers of land use changes. In

turn, the impact of policy reforms on land use change can only be unpacked if the

consequent land use change is monitored. Such up-to-date and accurate information about

the expansion of plantation forest in the south-west of Western Australia is limited,

despite a strong emphasis on conservation and sustainable land use planning in the region.

The analysis of remote sensing imagery presents a viable approach to developing a better

understanding of the history and dynamics of a changing landscape in response to

policies. Developing methods to analyse these data and extract relevant information is,

138
Chapter 5. Conclusions

therefore, a fundamental step in supporting policy-makers and researchers to design more

targeted land use policies.

From this perspective, the overall objective of this thesis was to increase our

understanding of the historical landscape changes in the Warren River Catchment (WRC)

in Western Australia, and so identify the physical and political forces that drive those

changes. This thesis addressed questions about the nature of these land use changes by

integrating remote sensing, geographical information system, machine learning

algorithms and econometrics. As such, the key research findings and policy implicatio ns

specific to each research questions are presented concisely below, followed by an

overview of the research contribution and overall significance of the thesis. We close this

dissertation with an evaluation of the limitations of our study and suggestions for future

research.

5.2. RESEARCH F INDINGS

5.2.1. Research Question 1

Which classifier exhibits better learning rates for discriminating spectrally similar

classes of a time-series of Landsat imagery? (That is, between plantation forest and native

forest).

Key Research Findings

This first research question is addressed in Chapter 2. Six LULC classes were classified :

agriculture, water, native forest, sand dunes, plantation forest, and harvested native forest.

139
Chapter 5. Conclusions

Distinct forest types including plantations and native forest are often classified as one

broad class, leading to an underestimate of commercial plantation trees. The comparison

among three MLAs namely, SVM, RF and CART, suggested that SVMs were more able

to distinguish between spectrally similar but ecologically distinct classes, such as native

forest and plantation forest, when using Landsat time-series imagery.

An important innovation in this study was the development of a consistent and accurate

method for distinguishing plantation forest from the native forest in Western Australia.

The combination of SVM classifier, four bands held in common between the four Landsat

Sensors (MSS 1979; TM 1992; ETM+ 2003; OLI 2014), and PCA of texture features

allowed a consistent procedure to be applied in the historical reconstruction of the extent

of plantations and native forest. Further, the method applied here is sufficiently general

to have potential application to similar LULC classification scenarios elsewhere. This

approach is suitable not only for identification of new forest plantation areas, but also for

calculating their life span through ongoing monitoring. Besides that, previous estimates

of landscape change in the catchment came from different sources and were usually

applied to a single moment in time. These estimates might, therefore, be inconsistent as

they were derived from different classification schemes and image processing methods

through time. The results presented here consistently show that regardless of the spectral

similarity between both Eucalypt forests (i.e., plantation and native), they can be

successfully classified using SVMs, and should therefore not be oversimplified into one

broad forest class during classification.

140
Chapter 5. Conclusions

Policy Implications

The major policy benefit of this study is the development of an acceptably robust method

to separate Eucalypt dominated plantation forests from adjacent Eucalypt native forest

across the landscape and through time. This allows the distribution of plantation forests

throughout the landscape to be tracked through time in response to policy innovations, as

a key component in delivering evidence-based policy concerning sustainable timber and

conservation management. For example, correctly classyfing plantation forest as one of

the major land uses associated with the tax subsidy schemes in Western Australia (e.g.,

tThe 2020 Vision) provides an understanding of the role of plantation forest development

in combatting a variety of environmental issues such as salinity mitigation, soil erosion,

and carbon sequestration. Both plantation and native forest types deliver distinct

economic and ecological benefits, and thus should be considered separately during

classification for the effective design of more targeted policy.

5.2.2. Research Question 2

What are the main changes in LULC that have occurred since 1979? (That is, since the

start of the Landsat record).

Key Research Findings

Chapter 3 describes the trajectory of landscape changes in the catchment from 1979 to

2014. Empirical evidence shows that no major clearing of native forest was identified in

the WRC since 1979. By contrast, an in-depth analysis of the transition matrix revealed

141
Chapter 5. Conclusions

that the most dominant LULC change was the conversion of agriculture to plantatio n

forest, especially after the late 1990s. Plantation forest areas had grown from 0.44% in

1992 to 4.36% in 2014. These plantation forests were mostly centred on the upper north-

east of the catchment associated with greater rates of dryland salinity.

Broad-acre agriculture had consequently declined from 30% in 1979 to 25% in 2014. The

extent of harvested native forest has been maintained, increasing marginally from 0.7%

in 1979 to 0.9% in 2014. So far as we are aware, this analysis provides the most up-to-

date and detailed historical information of LULC change in the Warren River Catchment

Policy Implications

The key result is that the total area of cleared land for agriculture has decreased over time.

The Country Areas Water Supply Act 1947 has been in place in the WRC since 1978, and

so far has been effective in preventing further wide-spread clearing of native vegetation.

Improving our knowledge of this Act is of interest to researchers and policy-makers

concerned with the effectiveness of public policies, including any unanticipated indirect

benefits.

One of the main constraints faced by policy-makers is a lack of reliable information on

plantation forest development. The historical LULC changes estimated in this thesis

sought to fill this information gap. The historical reconstruction of LULC change can be

referenced when designing new policy instruments. Further, the spatial pattern of

plantation development have largely been consistent with the State Salinity Strategy,

which recommends commercial plantation forest in already cleared areas as a key

management option in reducing the salinity level of the region’s water resources. Our

142
Chapter 5. Conclusions

analysis of landscape change in the WRC is a step towards addressing a place-based

insight into economic and environmental changes.

5.2.3. Research Question 3

Given the current policy scenario, what is the likely future of plantation forest?

Key Research Findings

One of the main findings outlined in the Chapter 3 is the substantial reduction of

plantation forest areas forecasted for 2025 if the current trend of land use change is

maintained. Despite all government efforts to increase the areas of Eucalyptus

plantations, the model predicted growth of agricultural areas due to the ongoing

harvesting rates of existing plantation forest and a reduction in establishment rate of new

plantation forest. By 2025, plantation forest is predicted to have its smallest extent since

the introduction of government tax incentives for forestry in the 1990s. This reduction is

most likely related to the reformulations of the tax incentives supporting MIS in the late

2000s. The method presented here to predict future land use change trends shows that an

integrated approach combining Markov transition probabilities, Multilayer Perceptron

and Cellular Automata produced satisfactory results, based on Kappa statistics (Kappa

coefficient of 72% and Kappa location of 74%) for predicting LULC changes in the

catchment.

143
Chapter 5. Conclusions

Policy Implications

The findings of this chapter contribute to policy in two ways. Firstly, the estimates of

plantation forest areas allow researchers and policy-makers to better anticipate the likely

landscape change given the current policy scenario. In particular, this analysis provides

the first estimate of the future extent of plantation forest within the catchment. The

simulated results reveal that there will be a decline of plantations which will likely affect

Western Australia’s timber production. This decline will in turn threaten efforts to reduce

annual average water salinity over the long-term. So far, the government forecast of an

ongoing increase in plantation forest looks to be overly optimistic, and out of date.

Secondly, the lack of reliable information on tax incentives and their associated risks is

likely to constrain investments in MIS forestry within the catchment. Current forestry

MIS policy may be failing due to a focus on an allowance for tax minimisation, rather

than on realistic or actual rates of return. Regrettably, some plantations were poorly sited

in terms of climate and soil and/or poorly maintained, leading to an overestimate of the

rate of return. Further, high upfront costs combined with high opportunity costs due to

long time to reach investment maturity following harvest makes investment in long

rotation plantations economically challenging for many investors. Thus, our findings also

imply that forestry MIS will likely be doomed to fail if changes to the current policy

scenario do not take place. For instance, to prevent large areas of Eucalyptus plantatio ns

being reconverted to agriculture, the government could apply credits to the carbon

captured by these plantation forests. There is thus an urgent need to create economic

incentives to secure the benefits of plantations. Furthermore, the results highlight the

necessity of governmental and non-governmental organisation to be highly cautious when

144
Chapter 5. Conclusions

relying on tax inducements to design sustainable management strategies, if in part due to

the ease with which those inducements can be withdrawn from the market.

5.2.4. Research Question 4

What were the main drivers of LULC change over the past 35 years (i.e., 1979-2014)?

Key Research Findings

The MNL model presented in Chapter 4 reveals the main spatial determinants of LULC

changes in the catchment with the five major findings being: (1) biophysical land

attributes such as phosphorus and soil organic carbon in the topsoil layer (0-5cm)

influenced the transition to plantation forest, thus supporting the Ricardo land rent theory;

(2) the probability of conversion to agriculture or plantation forest is strongest for those

parcels located close to urban places, thereby reinforcing evidence for von Thunen’s

model; (3) a parcel of private land is more likely to be used for either plantatio n forest or

agriculture rather than for native forest; (4) the interaction between private land and the

introduction of the tax benefits for forestry MIS contributed significantly to the

development of Eucalyptus plantations in the catchment; (5) however, the revision of tax

deductibility for MIS investments was negatively associated to the conversion to and

retention of plantation forest areas.

145
Chapter 5. Conclusions

Policy Implications

The MNL model suggests that the rapid expansion, and subsequent stagnation of

plantation forest development was largely driven by land use policies in the catchment.

This finding underlines the important consequence of basing policy on temporary

incentives, namely when the rapid increase in plantation forest reversed to a decline

following the reduction in the associated tax incentive through sudden governme nt

reforms. This implication stems from: the results of the MNL model; the historica l

analysis of LULC changes which show that plantations are currently at a standstill in the

WRC; the predicted decline in plantation forest with modelling of the current rates of

LULC transition out to 2025; and, with no establishment of significant area of new

plantations having been documented since 2010.

Significantly, it does not appear that without tax incentives the current set of policies that

are in place will adequately meet policy targets for water quality, timber production and

indeed carbon sequestration. Presently, land use policies directed at plantation forest

development rely on the commercial benefit of plantations to sustain investment, and do

not have a provision for plantation owners to capture environmental and social co-benefits

to sustain this land use.

There is then a need to consider new strategies, such as focusing on potential co-benefits

associated with plantations (e.g., and mitigation of global climate change, dryland salinity

reduction, and biodiversity conservation), to attract investors into multi-purpose

plantation programs. These new strategies should strongly consider taxation reform,

given the rapid increase in plantation extent as a response to the tax incentive schemes of

the late 1990s.

146
Chapter 5. Conclusions

Industry also needs to be involved to ensure the design of a consistent policy, that reduces

the risk of unintended societal and environmental impacts, while supporting the continued

investment in long rotation plantations, given forestry MISs have been a contentious area

of investment, In all, the results presented in this chapter may assist policy makers in

developing policies that are more effectively targeted at sustaining development of long-

term, while embracing more environmentally friendly land use practices.

5.2.5. Research Question 5

What is the role of government policy incentives in the establishment of plantation forest?

Key Research Findings

A sequence of events led to the expansion of eucalyptus plantations in the catchment.

There is no doubt that plantations underwent a boom because of an agribusiness scheme

market (Managed Investment Schemes - MIS), which turned plantations into tax-effective

investments (following the findings in Chapters 3 and 4). By the late 1990s the spread of

plantation forest, especially of Tasmanian Bluegums, across previously cleared

agricultural lands was spurred by the MIA 1998 and The 2020 Vision. Furthermore, the

income tax subsidies afforded by the Income Tax Assessment Act of 1997 and Product

Ruling System of 1998 also supported the MIS sector by being aligned with the intent of

The 2020 Vision and MIA 1998. Overall, the tax incentives resulted in the rapid

establishment of plantation forests across a remarkably wide region of Australia.

147
Chapter 5. Conclusions

However, over the past decade, the taxation laws governing MIS operations have been

subjected to successive changes by the government. These changes have affected the tax

deductibility of plantation establishment costs (e.g., Taxation Ruling TR2007). Thus, a

notable decrease of investments in forestry was observed due to an increase of uncertainty

among MIS investors. This reduced level of investment constitutes a significant risk to

the fulfilment of The 2020 Vision, and for the future domestic supply of high value, long

rotation timber products. It is clear from the evidence presented that many of the current

plantations have, or will, revert to broad-acre agriculture. Overall, it seems that the rate

of development of forest MIS would have been considerably slower without the

substantial role of the Australian government in promoting and incentivising the growth

of the plantation industry.

Policy Implications

From an efficiency point of view, sustainable land use policies at the catchment and local

levels should involve a systematic analysis of investment options to encourage the

adoption and persistence of preferred land use. Indeed, the Warren River Catchment

exemplifies how unrealistic national policies and legislation, such as The 2020 Vision

and associated tax changes, have been in achieving the proposed local land-use and water

quality objectives. For instance, the State Salinity Strategy for the Warren relies heavily

on commercial plantation forest to reduce water salinity. However, plantations have

generated little worthwhile improvement in salinity mitigation within the catchment and

will have little lasting benefit insofar as the intention of the tax incentives to promote

plantations will not be achieved over the long term.

148
Chapter 5. Conclusions

If the government is intent on avoiding the decline in the area of long rotation plantatio n

then it needs to take action towards improving the MIS taxation arrangement. Moreover,

land-use policies should be revised urgently for the desired external benefits from

growing plantations to be realised. To date, the full range of co-benefits generated by

reforestation (e.g., carbon sequestration) is not explicitly defined in the relevant polic ies

(i.e., The 2020 Vision, and MIA 1998). It is then no surprise that these co-benefits remain

largely unvalued by the landholder as the co-benefits usually does not contribute to direct

cash-flow. For instance, landowners should be compensated for not clear-felling

plantation forestry especially where hydrological modelling indicates a potential adverse

effect on water salinity. The Commonwealth’s approach to plantation and water has

discussed multiple benefits of plantation. However, the National Water Initiative for

carbon plantation cut programs that intend to use reforestation for catchment water quality

improvement because of the negative impact in eastern Australia.

The potential effects in the landscape of land use policy innovations should always be

assessed as part of an evidence-based framework. This approach enables policy or

strategy measures to be adjusted to achieve a given policy goal (e.g., mitigation of dryland

salinity), by observing the landscape scale response to the policy measure as it evolves

over time. Furthermore, the forestry MIS needs to have a sound reputation as a feasible

investment to incentivise plantation adoption and persistence. A range of factors

contribute to the high risk associated with investing in MIS supported forestry when

compared to other non-forestry agribusiness. These factors include: the long wait before

returns are realised following plantation establishment; high upfront establishment costs;

a relatively low return compared to other industries; the volatility of commodity markets;

and, uncertainty about tax deductibility. While it is critical to address the economics of

149
Chapter 5. Conclusions

plantation forestry in future land use policies, so that plantation investments become more

viable, government institutions need to also target the multiple co-benefits that are

potentially available (e.g., water quality), and to align that policy with the aspirations of

the local community within which those future investments are to take place.

5.3. STUDY LIMITATIONS AND FUTURE RESEARCH DIRECTIONS

The methodological limitations of this study may feasibly be improved upon in future.

Firstly, the classification accuracy of the remote sensing component of this study may

readily be improved by sourcing high-resolution and/or hyperspectral multi-tempo ra l

images. Such imagery are critical to providing better discrimination of the spectrally

similar classes namely plantation forest and native forest. These data are not

retrospectively available for the study period starting in 1979, and so we applied the

Landsat family of moderate resolution sensors as the most consistent archive of historica l

imagery available. Despite this limitation in image resolution, we were able to generate

satisfactory results from the Landsat data, as described in Chapter 2. However, the future

monitoring of land-use change in response to policy innovations will benefit from the

application of the higher spatial and spectral resolution of the remotely sensed data now

available. This high-resolution imagery will also help in targeting policy to the local

context considering high discrimination will likely result. For example, different species

of forest trees within the native forest estate can potentially be identified, and related to a

set of ecological functions such as estimating the amount of carbon sequestered in the

catchment based on different native forest types.

150
Chapter 5. Conclusions

Precise and up-to-date information concerning the forest estate is really the only gateway

to improving our understanding of the causes underpinning the historical response of the

forest estate to economic, climatic and policy factors. Equipping government and non-

government organisations with this understanding enables a clearer identification of

potential land-use and ecological problems, and of the solutions required to address those

problems. The insight from this study identifies few limitations, and suggests a number

of future research priorities to further improve the effectiveness of land use policies.

In Chapter 3 we simulated the likely future extent of plantations under the current policy

settings. Further, in Chapter 4 we identified the main drivers of landscape change. More

may be done to support this evidence chain, such as the inclusion of a survey to explore

the specific socio-economic causes driving the land use change. These survey results may

be integrated into predicting the likely outcomes of alternative policy scenarios and

enable a better picture to be built of local community aspirations. The key reason for not

developing a survey instrument was the budgetary constraint of this study. This limita tio n

was partially addressed by utilising Government provided data, such as policies and land

tenure. Importantly, how the development of large-scale plantation forest impacts rural

communities should be addressed in future research. Such research would assist in

preserving pre-existing social, environmental and economic values by which rural

communities have historically operated within the catchment.

The third limitation of this research lies with the choice of policies that may be driving

plantation development, and the lack of a comprehensive policy brief. Over the past 20

years the forestry sector has been subjected to various government interventions, as

summarised in the Chapter 4. We concluded that the two main policies driving plantatio n

industry were the tax-effective Management Investment Schemes in support of The 2020

151
Chapter 5. Conclusions

Vision (Plantations for Australia). It would have been more useful if we could have

developed a more detailed historical analysis of policy decisions, and so simulate several

land use patterns according to different policies to reveal the implication of possible

policies. We acknowledge that a well-designed historical policy brief would contribute to

improving policy and decision-making process, while also distil lessons learned from this

research. Future research should therefore address this issue to ensure that the past

mistakes are taken into account and used in the development of a new policy.

One further research priority would be to develop a hedonic pricing model to explain

farmland characteristics and landholders’ preferences in the adoption and relinquishme nt

of plantations. The inclusion of historical land prices would help explain the preferences

of landholders among competing land uses, by detailing the commercial and nonmarket

benefits of plantation forestry and other land uses that influence the local community.

Understanding the dynamics of land prices in response to land use decisions, and vice

versa the land use decisions made in response to land prices would further socio-economic

questions, such as whether small and medium businesses family landholders are being

replaced by big companies and corporations, to be addressed. Again, this approach to

understanding historical prices as a driver of land use change due to the cost of the land

price data over and above the study’s budgetary constraint.

Finally, we felt that the policy challenges surrounding plantation forestry are related to

more deeply rooted issues than the profitability of sustaining commercial timber resource.

The expectations that have been placed on plantation forestry have been significant, and

include contributions to biodiversity conservation, carbon emissions and climate

warming, water quality and local community aspirations. Designing better land use

152
Chapter 5. Conclusions

policies will require further research to fill this gap by exploring the interactions that exist

between those issues.

153

You might also like