ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018) 22–37

Contents lists available at ScienceDirect

ISPRS Journal of Photogrammetry and Remote Sensing

journal homepage: www.elsevier.com/locate/isprsjprs

The impact of dataset selection on land degradation assessment T

a,b,⁎ a,b c
Arden L. Burrell , Jason P. Evans , Yi Liu
a
Climate Change Research Centre, School of Biological, Earth and Environmental Science, University of New South Wales, Australia
b
ARC Centre of Excellence for Climate Extremes, Australia
c
School of Geographic Sciences, Nanjing University of Information Science & Technology, Jiangsu 210044, China

A R T I C LE I N FO A B S T R A C T

Keywords: Accurate quantification of land degradation is a global need, particularly in the world’s dryland areas. However,
RESTREND there is a well-documented lack of field data and long-term observational studies for most of these regions.
BFAST Remotely sensed data offers the only long-term vegetation record that can be used for land degradation as-
Dryland degradation sessment at a national, continental or global scale. Both the rainfall and vegetation datasets used for land de-
NDVI
gradation assessment contain errors and uncertainties, but little work has been done to understand how this may
Trend analysis
AVHRR
impact results. This study uses the recently developed Time Series Segmented RESidual TREND
GIMMS (TSS-RESTREND) method applied to six rainfall and two vegetation datasets to assess the impact of dataset
TSS-RESTREND selection on the estimates of dryland degradation over Australia. Large differences in the data and methods used
to produce the precipitation datasets did not significantly impact results with the estimate of average change
varying by < 4% and a single dataset being sufficient to capture the direction of change in > 95% of regions. On
the other hand, the vegetation dataset selection had a much greater impact. Calibration errors in the Global
Inventory Monitoring and Modeling System Version 3 NDVI (GIMMSv3.0g) dataset caused significant errors in
the trends over some of Australia’s dryland regions. Though identified over Australia, the problematic calibra-
tion in the GIMMSv3.0g dataset may have effected dryland NDVI values globally. These errors have been ad-
dressed in the updated GIMMSv3.1g which is strongly recommended for use in future studies. Our analysis
suggests that using an ensemble composed of multiple runs performed using different datasets allows for the
identification of errors that cannot be detected using only a single run or with the data quality flags of the input
datasets. A multi-run ensemble made using different input datasets provides more comprehensive quantification
of uncertainty and errors in space and time.

1. Introduction
Drylands cover about 41% of the land surface and are characterised
by low annual precipitation (Ruppert et al., 2015) and large interannual
climate variability (Broich et al., 2014; Khishigbayar et al., 2015) which
result in significant natural variation in vegetation productivity (Broich
et al., 2014). Dryland degradation poses a serious threat to interna-
tional food security and has been identified by the United Nations (UN)
as an issue of global concern (MEA, 2005). At the 12th Conference of
the Parties to the UN Convention to Combat Desertification, nations
were called upon to implement plans to reach land degradation neu-
trality which they defined as “a state whereby the amount and quality of
land resources necessary to support ecosystem functions and services and
enhance food security remain stable or increase within specified temporal
and spatial scales and ecosystems” (Orr et al., 2017).
Despite the importance of dryland degradation monitoring (IPCC,
2017; MEA, 2005), current estimates of the scale of the problem have
been described as “highly unreliable and spurious” because they de-
pend heavily on small scale field studies as well as low spatial and
temporal resolution expert opinions (Higginbottom and Symeonakis,
2014). Over Australia, which is well studied by global standards, spatial
data products provide information about a range of environmental
parameters including land cover (Lymburner et al., 2011), vegetation
climate zones (BoM, 2012), land use (Australian Bureau of Agricultural
and Resource Economics and Sciences, 2015) and soil acidification
(State of the Environment 2011 Committee, 2011) (for a full list see
Lawley et al. (2016)). None of these products are suitable for change
detection or time series analysis, as most are a snapshot of a single point
in time or are produced using inconsistent methods (Caccetta et al.,
2012) from sparse and infrequent field data (Ludwig et al., 2007). This
lack of a long term monitoring program and the problems with existing
programs for land degradation detection has been extensively

⁎
Corresponding author at: Climate Change Research Centre, Level 4, Mathews Building, The University of New South Wales, Sydney NSW 2052, Australia.
E-mail address: arden.burrell@unsw.edu.au (A.L. Burrell).

https://doi.org/10.1016/j.isprsjprs.2018.08.017
Received 9 March 2018; Received in revised form 21 August 2018; Accepted 23 August 2018
0924-2716/ © 2018 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V. All rights reserved.
A.L. Burrell et al.
documented (Day et al., 2007; Eyre et al., 2011; Fisher and Kutt, 2006;
Healy et al., 2016; Ludwig et al., 2004; McAlpine et al., 2014; Pickup,
1998). Remotely sensed datasets offer the only viable way to assess
dryland degradation at national, continental or global spatial scales,
and over multi-decadal time periods.
There are two types of methods used to assess ecosystem changes in
dryland regions; those that analyse changes in the seasonal phenology
of vegetation, and those that look for changes in the relationship be-
tween vegetation and climate variables such as precipitation (Burrell
et al., 2017; Higginbottom and Symeonakis, 2014). The most widely
used method to assess changes in vegetation phenology with respect to
time is the Breaks For Additive Seasonal and Trend (BFAST) (Verbesselt
et al., 2010a, 2010b). BFAST decomposes the vegetation phenology
signal present in remotely sensed Vegetation Index (VI) time series data
into its seasonal, trend, and remainder components. This allows the
method to detect abrupt changes in both the trend and seasonal com-
ponents of the vegetation phenology (Kuenzer et al., 2015). In ecosys-
tems with low interannual climatic variability, the vegetation phe-
nology is relatively stable. This means that breakpoints detected in the
phenology cycle using BFAST can be attributed to ecosystem dis-
turbances (Hutchinson et al., 2015; Verbesselt et al., 2010a, 2010b). In
regions with high interannual climate variability, drought and flood
years can cause significant natural changes in the phenological cycle.
This makes the separation of natural variability from environmental
change problematic (Burrell et al., 2017; Fensholt et al., 2015; Kuenzer
et al., 2015; Watts and Laffan, 2014).
The second approach to dryland degradation analysis is to look at
changes in the relationship between climate variables and a VI such as
the Normalized Difference Vegetation Index (NDVI), which is a measure
of vegetation greenness and a proxy for ecosystem productivity. The
most widely used method based on this relationship is the Residual
Trend (RESTREND) method proposed by Evans and Geerken (2004).
RESTREND works by first performing an Ordinary Least Squares (OLS)
linear regression between annual peak NDVI and the relevant climate
variables (Reeves et al., 2015; Wang et al., 2012a; Wessels et al., 2007).
RESTREND is able to estimate the change in ecosystem productivity
that is not caused by interannual climatic variability (Evans and
Geerken, 2004; Wessels et al., 2007, 2012). Which climate variables are
used in RESTREND analysis varies regionally. In Australia and Africa,
where water is the primary limiting factor (Broich et al., 2014; Guan
et al., 2014; Zhang et al., 2018), a vegetation-precipitation relationship
(VPR) is calculated (Burrell et al., 2017; Evans and Geerken, 2004;
Wessels et al., 2012). In cold drylands like Mongolia, temperature also
plays a major role and, as such, is included as an additional climate
variable (Keenan and Riley, 2018; Liu et al., 2013). A trend analysis is
then performed on the VPR residuals, with a negative trend indicating
land degradation (Andela et al., 2013; Li et al., 2012). An example of a
RESTREND analysis is included in supplementary material.
Two of the three key assumption of the RESTREND method are that
there is a statistically significant VPR, and that this VPR remains
comparable for the entirety of the time series (Burrell et al., 2017;
Wessels et al., 2012). This means that in regions where a degradational
process is introduced or removed, leading to rapid ecosystem change,
RESTREND can fail to detect the change (Wessels et al., 2012). The
third assumption is that any trend in the residuals remains monotonic
for the entire time series. In a previous study of dryland degradation
over Australia, Burrell et al. (2017) found that at least one of the key
assumptions of the RESTREND method was violated in approximately
15% of pixels. In areas with documented examples of rapid ecosystem
change, RESTREND analysis alone was unable to capture the extent of
the changes. Similar problems with RESTREND were found in a study
over Kyrgyzstan (Eddy et al., 2017).
To address these problems, Burrell et al. (2017) proposed the Time
Series Segmented Residual Trends (TSS-RESTREND) method.
TSS-RESTREND is an extended version of RESTREND incorporating a
modified version of the BFAST method to look for rapid ecosystem
23
ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018) 22–37
changes that violate the key assumptions that underpin a standard
RESTREND (Burrell et al., 2017; Wessels et al., 2012). If a significant
breakpoint is found in either the VPR or the VPR-Residuals, TSS-RES-
TREND uses multivariate regression with an additional variable to ac-
count for the breakpoint. When applied to Australia, TSS-RESTREND,
was able to improve the detection of degraded areas compared to
RESTREND alone as well as to accurately detect both the timing and the
direction of change in two regions with known histories of degradation
(Burrell et al., 2017).
One of the main advantages of automated analysis of remotely
sensed data using methods like TSS-RESTREND, RESTREND or BFAST,
is that the analysis is easy to replicate. This is particularly true for
methods like TSS-RESTREND and BFAST, where the methods are
available as R packages that can be freely downloaded and im-
plemented (http://cran.rproject.org/package=bfast and (https://cran.
r-project.org/package=TSS.RESTREND). This facilitates the direct
comparison of different studies over different timescales and regions
which allows for the discussion of global dryland trends (Higginbottom
and Symeonakis, 2014). Inherent in the intercomparison of studies is
the assumption that any changes in vegetation and precipitation are
larger than the noise present in the datasets used to capture them
(Verbesselt et al., 2010b). That is, the signal to noise ratio is high en-
ough that it is possible to separate real variations in the vegetation from
those caused by the systematic errors and uncertainties in the datasets
(Scheftic et al., 2014; Verbesselt et al., 2010b).
A study by Ibrahim et al. (2015) in the Sub-Saharan region of West
Africa over the time period 1982–2012, found that using datasets with
different signal-to-noise ratios can have a substantial impact on the
results. The results of a RESTREND analysis between GIMMSv3.0g and
CRU3 were compared with one performed between GIMMSv3.0g and
the Soil Moisture Index produced by the Climate Prediction Centre. It
was found that the estimated vegetation trends differed across the en-
tire study area with RESTREND applied to a soil moisture dataset de-
tecting land degradation with greater consistency than the rainfall/
NDVI RESTREND. In addition, two other studies have examined the
impact of vegetation datasets on RESTREND analysis by comparing
NDVI to Enhanced Vegetation Index (EVI) (John et al., 2015) and the
passive microwave based Vegetation Optical Depth (VOD) (Andela
et al., 2013) over a common time period. All three studies found that
dataset selection significantly impacted results, but because the datasets
tested were not measuring the same things, it is impossible to know
how much of the difference is caused by errors in the datasets.
Climate controlled dryland vegetation trend analysis depends on
two datasets: a vegetation dataset (usually NDVI) and a climate dataset
(Wessels et al., 2007). The only globally consistent NDVI datasets with
more than 20 years temporal coverage are derived from NOAA Ad-
vanced Very High Resolution Radiometer (AVHRR) data (Yengoh et al.,
2015). Two versions of the AVHRR derived Global Inventory for Map-
ping and Modelling Studies (GIMMS) dataset (Pinzon and Tucker,
2014) are used in this study (GIMMSv3.0g and GIMMSv3.1g). GIMMS
NDVI is derived from multiple sensors and there are documented in-
consistencies across sensor transitions (see Section 2.2 for details).
Climate variability in dryland regions is often linked to multidecadal
climate processes like the El Niño–Southern Oscillation (ENSO) (Broich
et al., 2014) which operate on multiannual through multidecadal cy-
cles. This gives the GIMMS datasets a large advantage over shorter
temporal and higher spatial resolution products like those derived from
MODIS (Moderate Resolution Imaging Spectroradiometer). This is one
of the main reasons it remains in widespread use today in both vege-
tation assessment (Fensholt et al., 2009; Fensholt and Proud, 2012), as
well as in the testing and validation of land surface climate models
(Anav et al., 2013; Zhu et al., 2016).
Unlike vegetation datasets, there exists a range of global and na-
tional precipitation datasets that have been created using different data.
These datasets differ in spatial and temporal resolutions, data sources,
spatial coverage, temporal latency and design objective (Beck et al.,
A.L. Burrell et al.
2017b). Six of these datasets cover Australia for the GIMMSv3.0g era
(1982–2013) and have a spatial resolution of 0.5° or less. The limita-
tions of individual datasets as well as the discrepancies between dif-
ferent datasets have been extensively documented (Beck et al., 2017b;
Maggioni et al., 2016; Wu et al., 2018a). Despite the availability of
multiple rainfall datasets, no study has attempted to quantify how da-
taset selection impacts land degradation assessment.
This study has two aims. The first is to assess the impact of input
rainfall and vegetation dataset choice on the estimation of dryland
degradation. The second is to assess whether using an ensemble of runs
made using multiple equivalent datasets over a common time period
improves the robustness of the estimate of Australia’s dryland de-
gradation by quantifying the errors introduced by the input datasets.
2. Study area and data
2.1. Study area
The study area covers Australia, an area of approximately
7,692,000 km2 of which about 80% can be characterised as drylands
(Broich et al., 2014; Burrell et al., 2017). The climatic zones in Australia
range from equatorial and tropical in the north, through sub-tropical,
semiarid and arid in the centre, to temperate in the south (BoM, 2012)
as shown in Fig. 1.
2.2. Vegetation datasets
2.2.1. GIMMSv3.0g NDVI
Comparisons between different AVHRR derived NDVI datasets and
shorter temporal but high spatial resolution records like MODIS
Fig. 1. Map of the climate vegetation zones of Australia based on a modified Köppen climate classification (BoM, 2012). Australia’s drylands cover all of the
subtropical, grassland and desert regions as well as the non-coastal temperate and tropical areas.
24
ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018) 22–37
(Moderate Resolution Imaging Spectroradiometer), SPOT (Satellite
Pour l'Observation de la Terre) VEGETATION and Landsat have shown
that the datasets produced by the Global Inventory for Mapping and
Modelling Studies (GIMMS) group are consistently the most accurate
(Fensholt et al., 2012) particularly at inter-sensor transitions (Scheftic
et al., 2014). This is why the GIMMS datasets have been the mostly
widely used dataset for dryland degradation assessment (Fensholt et al.,
2009; Fensholt and Proud, 2012; Marshall et al., 2016; Zhu et al.,
2016).
In 2014, Global Inventory for Mapping and Modelling Studies
published the third version of their AVHRR derived NDVI dataset, re-
ferred to in this paper as GIMMSv3.0g. GIMMSv3.0g is a global non-
stationary NDVI dataset which has bi-monthly values starting in July
1981 and running through to December 2013 at a 1/12° (∼8 km at the
equator) spatial resolution. Details on the pre-processing, creation, ca-
libration and correction methods used to generate the GIMMSv3.0g
dataset are provided in Pinzon and Tucker (2014).
Despite being the most reliable AVHRR derived NDVI dataset,
GIMMSv3.0g has some documented problems that can affect values in
some dryland regions. These problems arise as a result of factors like
atmospheric contamination, orbital drift (Tian et al., 2015) as well as
geometric, spectral, and radiometric calibration errors (Huang et al.,
2013). Especially problematic for trend and breakpoint analysis are the
errors introduced by the transitions between the different National
Oceanographic and Atmospheric Administration (NOAA) satellites
which carry the AVHRR sensors. Whilst these problems are apparent at
every sensor change (Burrell et al., 2017; Tian et al., 2015), three
transition periods are particularly problematic (Burrell et al., 2017): (i)
the shifts in September 1994 and January 1995 after the unexpected
failure of NOAA-11; (ii) the transition from the AVHRR-2 sensor to the
A.L. Burrell et al. ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018) 22–37

AVHRR-3 sensor November 2000 (Fan and Liu, 2016; Tian et al., 2015);

Harris et al. (2014), Harris (2017),

Matsuura and Willmott (2007) and

Raupach et al. (2012, 2009) and
and (iii) the change in January 2009 which came within 4 years of the

Willmott and Matsuura (2001)

Data References & key papers

Schneider et al. (2015, 2016,

end of the time series and also corresponded to an El Niño in 2009 and

Funk et al. (2014, 2015)

La Niña in 2010 which impacted calibration (Tian et al., 2015). These

and New et al. (2000)

problems do not impact all ecosystem types and regions equally being

Beck et al. (2017a)

Jones et al. (2009)
particularly problematic for arid areas in the southern hemisphere
(Fensholt and Proud, 2012).

2014)
2.2.2. GIMMSv3.1g NDVI
In 2016 the GIMMS group released an updated version of their
GIMMSv3.0g NDVI dataset which they refer to as GIMMS NDVI 3g.v1

∼Annual (Harris, 2017)

∼3–4 years (Schneider
and is referred to in this paper as GIMMSv3.1g. GIMMSv3.0g has a
known error in the cross calibration with SeaWIFs data that tends to

Update Schedule
exaggerate variability in bare-grass areas (Pinzon, 2017, pers.comm.).

et al., 2016)
As well as extending the temporal record, GIMMSv3.1g also addresses

Frequently
Infrequent
Monthly

Monthly
this issue. In this study, the twice monthly GIMM3g (1982–2013) and
GIMMSv3.1g (1982–2015) datasets were aggregated into monthly data
using the mean of the values. Pixels with less than 85% valid data
points (‘Quality flag 1–2’ for GIMMSv3.0g and ‘Quality flag 0’ for

Station data gridded using a spherical version of Shepard’s algorithm

Weighted averaging of P anomalies produced using two interpolated

GIMMSv3.1g) were excluded (Detsch et al., 2016).

Rainfall estimates created from CCD-based satellite data based on

Station data gridded using a modified SPHEREMAP interpolation

gauge datasets, three satellite remote sensing datasets and two

Station data gridded using Angular Distance Weighting (ADW)
2.2.3. MODIS NDVI
To aid the interpretation of the differences between GIMMSv3.0g
and GIMMSv3.1g, the MODIS-Terra derived MOD13Q1 NDVI product
(Didan, 2015) was used for the period 2000–2013. Data over Australia

Station data gridded using the WaterDyn model

regression models blended with station data.

is available from http://www.auscover.org.au/purl/lpdaac-mosaic-
mod13q1-v5 (Paget and King, 2008). The data was aggregated from

atmospheric model reanalysis datasets.

its 16-day temporal resolution to a monthly time series using the mean
of the values and then interpolated from its native 250 m resolution to
the 1/12° grid of the GIMMSv3.0g and GIMMSv3.1g datasets using the
Gridding data and Method

First-Order Conservative Remapping (Jones, 1999) in CDO (CDO,

2018).

2.3. Precipitation datasets

Studies that have compared different precipitation datasets have

found that the accuracy of gridded precipitation datasets is strongly
affected by station density (Contractor et al., 2015) as well as factors
like elevation, the season and the climate region (Zhu et al., 2015). The
Spatial resolution and

spatial resolution of analysis can also impact results with resolutions of

0.05° quasi-Global

1° and finer resulting in larger interproduct spread (Herold et al., 2017).

0.05° Australia

0.25° Global
(50°S–50°N)

Determining which gridded dataset most accurately captures actual

0.5° Global

0.5° Global

0.5° Global
coverage

precipitation is difficult because large areas of the world including parts

of Africa, South America and Asia, have limited availability of high-
quality observational data (Yin et al., 2015). Rainfall in dryland regions
is often dominated by a small number of extreme events and it is large
1979–2015; 3-hourly
1900–2017; Monthly

1901–2013; Monthly

1901–2015; Monthly

1901–2014; Monthly

1981–2017; Monthly
Temporal range and

rainfall events where datasets have been shown to differ most sig-
nificantly (Contractor et al., 2015; King et al., 2013; Sylla et al., 2013).
See Beck et al. (2017b) for further examples.
frequency

Six precipitation datasets (5 global, 1 national) were selected for

intercomparison with the VI datasets (See Table 1 for details). Those
datasets with sub-monthly temporal frequency were aggregated to
University of East Anglia Climate Research Unit

monthly by taking the sum of the values on a per pixel basis. All da-
Multi-Source Weighted-Ensemble Precipitation
monthly precipitation Full data reanalysis

Precipitation with Stations v2.0 (CHIRPS)

Australian Water Availability Project (AWAP)

tasets where then regridded using a First-order Conservative Remap-

University of Delaware Precipitation V4.01
Overview of the precipitation datasets.

ping from their native grids to the 1/12° grid of the GIMMSv3.0g and
Global Precipitation Climatology Centre

GIMMSv3.1g datasets.
The Climate Hazards group Infrared

Because of the maximum accumulation period (12 months) and

maximum offset period used (3 months) in this study some pixels use
precipitation data from as early as October 1980. This is a problem for
TS v. 4.00 (CRU4)

the CHIRPS dataset which starts in January 1981. For the < 1.25% of
pixels that needed 1980s data to calculate the accumulated rainfall, it
V7 (GPCC)

was found that the rainfall in those regions had a near 1:1 correlation
(MSWEP)

with AWAP data and that the October-November-December values

(UDel)
Dataset

from 1980 would only have a small impact on the total accumulated
Table 1

rainfall for the first year (< 10% versus just using 1981 months). For
this reason, 3 months of AWAP data were appended to the start of the

25
A.L. Burrell et al.
CHIRPS dataset. CHIRPS and MSWEP are collectively referred to as
“merged” datasets because they combine remotely sensed and in-situ
based data (Contractor et al., 2015) compared to AWAP, GPCC, CRU
and UDel which only use station observations.
3. Methods
3.1. TSS-RESTREND
The TSS-RESTREND method used in this study is divided into two
parts. The first uses the complete time series of the VI to develop a list of
potential breakpoints on a per pixel basis. These breakpoints are loca-
tions where ecosystem degradation or restoration may have violated
one of the key assumptions of standard RESTREND analysis. The second
part tests if these breakpoints, which are statistically significant in the
complete time series, have a significant impact on the annual maximum
growing season NDVI (NDVImax) time series. It then uses that in-
formation to calculate the Total Change (TC) in ecosystem productivity.
Total Change (TC) is the change in expected VI between the start
and end of the time series given mean accumulated rainfall. TC is cal-
culated by finding the change in the VPR residuals between the be-
ginning and end of the time series with adjustment of any structural
changes in the VPR. It allows for the comparison of the changes in
pixels with breakpoints to those without. An overview of the individual
steps of the TSS-RESTREND method is included in supplementary ma-
terial section 1. For the full details of the method as well as assessments
of the method’s accuracy see Burrell et al. (2017).
Both the GIMMSv3.0g and GIMMSv3.1g datasets used in this ana-
lysis are composed of data from multiple sensors which has been shown
to cause breakpoint detection in BFAST analysis (Burrell et al., 2017; de
Jong et al., 2011; Tian et al., 2015). The total complete time series VPR-
breakpoints for the GIMMS NDVI datasets are shown in Fig. 2. The
sensor transitions from NOAA-11 to NOAA-9 then to NOAA-14 (Sep
1994–Jan 1995) and the change from NOAA-17 to NOAA-18 (Jan 2009)
coincided with significant spikes in the total number of breakpoints in
the GIMMS derived datasets. To account for these sensor transitions,
breakpoints detected in the two months either side of a sensor transition
were excluded from the analysis.
3.2. TSS-RESTREND runs
The TSS-RESTREND method uses one VI and one precipitation da-
taset. In this paper, a total of 20 runs were performed over Australia
(see Table 2). These runs can be broken up into three groups: dataset
intercomparison runs, current trend assessment runs and trend attri-
bution runs.
3.2.1. Dataset intercomparison runs
The dataset intercomparison runs (run no. 1–12 in Table 2) cover
the period 1982–2013 and are between the six precipitation datasets
(AWAP, GPCC, CRU4, Udel, CHIRPS, MSWEP) and the two NDVI da-
tasets (GIMMSv3.0g & GIMMSv3.1g). These GIMMSv3.0g (1–6) and
GIMMSv3.1g (7–12) ensembles will be referred to as G3.0 and G3.1
respectively to simplify discussion. To compare the impact of pre-
cipitation, the Total Change (TC) values of the individual runs were
used to calculate ensemble mean (TCmean) and range (TCrange). In ad-
dition, the number of significant runs (TCp < 0.05) and whether these
significant runs had the same direction (TCinv) were calculated on a per
pixel basis.
In order to assess the performance of the different precipitation
datasets the number of contrary results they produced was calculated. A
contrary result is defined here as any finding of a run, which includes
the presence or absence of a breakpoint (BP outlier), or the direction
(TC inversion outlier) and significance of an ecosystem change (TC
outlier), that contradicts the findings of every other run in the same
ensemble. In addition the correlation coefficients were calculated on
26
ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018) 22–37
the Total Change (TC) values produced using AWAP and the other
datasets within the same ensembles as AWAP is expected to produce the
most accurate results.
To compare the different NDVI datasets, the differences between the
total change of the matching G3.0 & 3.1 runs are investigated. The
intra-run mean and range was calculated on these difference values. To
aid the interpretation of the differences in spatio-temporal trends be-
tween G3.0 and G3.1 ensembles, the root-mean-square error (RMSE)
was calculated between the GIMMS datasets and the MODIS derived
NDVI product for the period 2000–2013.
To be able to draw meaningful conclusion from TSS-RESTREND
studies, three criteria need to be met:
1. The direction of trend should not vary significantly because of da-
taset choice. This is the most import criteria;
2. The range of total change (TCrange) values for a given pixel should
not vary considerably both in absolute terms and as a portion of the
mean of total change (TCmean);
3. The timing of breakpoints should be the same between runs;
3.2.2. Current trend assessment runs
The current trends in vegetation were then assessed with the full
temporal range of GIMMSv3.1g (1982–2015) (run no. 13–16 in
Table 2). It was calculated using the four precipitation datasets that
extend to the end of 2015 (AWAP, CRU4, CHIRPS, MWESP). As with the
G3.0 and G3.1 ensembles, the total change was calculated taking the
mean of the four 2015 runs. For comparisons between the 2015 en-
semble (G2015) and G3.1, only the AWAP, CRU4, CHIRPS and MWESP
runs were used (runs 7, 9, 11 & 12 in Table 2).
4. Results
4.1. The impact of precipitation datasets
The per pixel mean TC for both the G3.0 and G3.1 ensembles show a
general positive change over much of eastern Australia (G3.0:46.75%
increasing, 5.68% decreasing, 47.58% no change; and G3.1:59.54%
increasing, 5.28% decreasing, 35.18% no change). The ensemble means
for G3.0 and G3.1 are shown in Fig. 3a and b respectively. Inter-en-
semble range shows excellent correspondence with 86.01% of G3.0
pixel and 92.20% of G3.1 pixels having an NDVI total change
range < 0.05. When expressed as a percentage of the per pixel mean of
total change (TCmean), 62.94% of G3.0 pixels and 61.92% of G3.1 pixels
had a TCrange that was < 50% of the per pixel TCmean. As with range and
mean, both G3.0 and G3.1 show good inter-ensemble TC direction
consistency with 61.61% of G3.0 pixels and 66.22% of G3.1 pixels had
complete agreement in both direction and significance (MD (Matching
Directionality)). Less than 3.81% of G3.0 and 4.45% of G3.1 pixels had
significant runs with opposite directionality (DI - Direction Inversion)
and the spatial patterns of direction agreement are shown in Fig. 3c and
d. In places with contradictory results, 80% of G3.0 and 85% of G3.1
pixels had a max (TCmax) and min (TCmin) total change < ± 0.05
NDVI. This indicates that these areas may be regions with noisy input
data and small changes in vegetation.
Comparing the impact of the individual precipitation datasets, the
average difference between the TCmean estimates for the entire study
area was 3.76% for the G3.0 and 3.44% for the G3.1 ensembles. TC
values produced using AWAP were compared to the other runs in the
same ensemble using the Pearson’s Correlation Coefficients. GPCC had
the strongest correlation in both the G3.0 and G3.1 ensembles (0.888,
0.893) followed by Udel (0.884, 0.885), MSWEP (0.876, 0.880),
CHRIPS (0.850, 0.859) and CRU4 (0.824, 0.836). Breakpoints show less
agreement than TC. 67.2% of G3.0 pixels and 60.7% of G3.1 pixels have
complete run agreement about the presence or absence of a breakpoint.
For pixels where three or more runs that found significant breakpoints
in the NDVImax time series, 61.89% of G3.0 pixels and 58.96% of G3.1
A.L. Burrell et al. ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018) 22–37

Fig. 2. Count of the total number of breakpoints detected in each month of the complete time series VPR-Residuals for (a) GIMMSv3.0g, (b) GIMMSv3.1g
(1982–2013) and (c) GIMMSv3.1g (1982–2015). From left to right the blue bands indicate the change of AVHRR sensors to NOAA-11, NOAA-16 and NOAA-17. Red
bands indicate particularly problematic time periods. The left red bad is the sensor shift from NOAA-11 to older NOAA-9 in September 1994 followed by the jump to
NOAA-14 in January 1995. The right red band is the change from NOAA-17 to NOAA-18. (For interpretation of the references to colour in this figure legend, the
reader is referred to the web version of this article.)

27
A.L. Burrell et al. ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018) 22–37

Table 2
Summary of TSS-RESTREND runs performed.
Ensemble Vi dataset Run number and rainfall dataset Temporal range

Dataset intercomparison runs

G3.0 GIMMSv3.0g 1. AWAP 2. GPCC 3. CRU4 4. Udel 5. CHIRPS 6. MSWEP 1982–2013
G3.1 GiMMSv3.1g 7. AWAP 8. GPCC 9. CRU4 10. Udel 11. CHIRPS 12. MSWEP 1982–2013

Current trend assessment runs

G2015 GiMMSv3.1g 13. AWAP 14. CRU4 15. CHRIPS 16. MSWEP 1982–2015

pixels had all their breakpoints within a 3-year window. There was little
to no spatial correlation between areas with TCinv or large TCrange and
areas that had large differences on breakpoint timings and disagree-
ment between runs as to the presence or absence of breakpoint in both
the G3.0 and G3.1 ensembles. The full correlation table, as well as maps
of TCrange, breakpoint range, breakpoint numbers and breakpoint
clustering are included as supplementary material.
The number of times the different precipitation datasets resulted in
contradictory results is shown in Fig. 4. CRU4 had the highest number
of TC and BP outliers, but produced less of the problematic TCinv out-
liers than AWAP and CHRIPS in both the G3.0 and G3.1 ensembles.
4.2. The impact of the NDVI dataset
Land degradation, represented by a decrease TCmean, was found in
5.678% of pixels of the G3.0 ensemble. 47.578% of pixels had no
change and 46.744% were increasing. In the G3.1 ensemble 5.281% of
pixels decreased, 35.181% had no change and 59.538% increased. The
difference in TCmean between the G3.0 and G3.1 ensemble is shown in
Fig. 5a. There is good overall agreement between the G3.0 and G3.1
ensemble means with 64.363% of pixels having the same TCmean di-
rection and significance and only 2.930% having direction inversion.
The spatial distribution of TC direction agreement is show in Fig. 5b.
Comparing the intra-ensemble variance to the inter-ensemble variance,
70.215% of pixels had an ensemble TCrange that is smaller than the
difference between the TCmean of the two ensembles. Interestingly, for
pixels with a large ensemble TCrange, there is very little overlap between
ensembles. At a threshold of TCrange > 0.05, there are 11,487 pixels
but only 11.25% of those pixels have large range in both the G3.0 and
G3.1.

Fig. 3. (a–b) The mean of the total change of the significant runs of the (a) G3.0 and (b) G3.1 ensembles. Pixels with less than two significant runs (TCp < 0.05) are
masked in white. (c-d) The Inter-run directionality for (c) G3.0 and (d) G3.1. Red indicates that runs produced significant contradictory results (DI), light grey
indicates that some runs found a significant result and others found no significant change with no direction inversion (NDI), blue indicates pixels where all runs agree
about the direction and significance (MD). The dark grey are pixels with less than 85% valid data points and are masked in all figures. (For interpretation of the
references to colour in this figure legend, the reader is referred to the web version of this article.)

28
A.L. Burrell et al.
Fig. 4. The number of contradictory results produced by the different precipitation datasets in the G3.0 and G3.1 ensembles. A TC outlier (light and dark red) is a
value that disagrees in direction or significance with every other run. A TC Inversion (light and dark green) outlier is a type of TC outlier where the trend is opposite
to the one found in every other run. A BP outlier (light and dark blue) is an outlier in the detection or absence of a breakpoint in a given pixel. (For interpretation of
the references to colour in this figure legend, the reader is referred to the web version of this article.)
4.3. Extending the record through 2015
Overall there are similar trends observed in the G2015 (runs 13–16
in Table 2) and matching runs in the G3.1 ensemble (runs 7, 9, 11 & 12
in Table 2) with 57.496% of pixels in the G2015 ensemble showing a
positive TCmean, 37.643% stable and 4.861% decreasing compared to
53.976% increasing, 42.22% no change and 3.80% decreasing in match
G3.1 runs. The TCmean for the G2015 ensemble is shown in Fig. 6a and
the difference between the annual average Total change (TCmean/year)
of the G2015 and matched G3.1 ensembles is shown in Fig. 6b.
The runs in the G2015 ensemble had an average of 35,727 pixels
with significant breakpoints in the NDVImax timeseries compared to
28,544 for the matched G3.1 ensemble. The number of average
breakpoints per year for the matched runs in the G3.0, G3.1 and G2015
ensembles is shown in Fig. 7. For the period where all three ensembles
are able to detect breakpoints (1987–2008), there is little difference
between the number of breakpoints detected by G2015 and G3.1, ex-
cept in 2005 and 2007 where G2015 detected significantly fewer
breakpoints than G3.1. Almost all of the additional breakpoints in the
G2015 ensemble come from 2009 and 2010. While new breaks are
present over all of Australia, the largest cluster is over Central
Queensland. This same region had a significant number of breakpoints
change from 2005–2007 in G3.1 to 2009–2010 in G2015.
5. Discussion
Every run in this study found positive vegetation changes over more
than 55% of Australia. This greening trend has been observed in pre-
vious studies that have looked at Australia (Andela et al., 2013; Burrell
et al., 2017; Chen et al., 2014a; Gao et al., 2016; Zhang et al., 2017), as
well as other arid and semi-arid regions of the world including the Sahel
bioregion in northern Africa (Boschetti et al., 2013; Brandt et al., 2017;
Dardel et al., 2014; Fensholt et al., 2013, 2009), Asia (Chen et al.,
2014a, 2014b; Gao et al., 2016; He et al., 2014; John et al., 2015;
Keenan and Riley, 2018; Liu et al., 2013; Qiu et al., 2017; Wang et al.
(2012b)), North America (Chen et al., 2014a; Tang et al., 2015), and
South America (Baeza and Paruelo, 2018; Gao et al., 2016). This global
greening trend has been observed in multiple remotely sensed
29
ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018) 22–37
vegetation datasets (Kaptué et al., 2015; Wang et al., 2018; Zhang et al.,
2017; Zhu et al., 2016) as well as by field observations (Tian et al.,
2016). While the overall trends are consistent, there is considerable
variability between different studies. In order to draw meaningful
conclusions from comparisons of studies that have used different pre-
cipitation and VI datasets, three criteria need to be met. First, they
should not produce contradictory results for the same locations (very
little TCinv); second, the TCrange values for a given pixel should be small
both in absolute terms and as a portion of the mean of total change
(TCmean); and third, the timing of breakpoints should be consistent
between runs.
5.1. Comparing the rainfall dataset ensembles
In land degradation detection simulation analysis, it has been found
that detecting small changes is problematic because of the noise in the
climate and vegetation datasets (Wessels et al., 2012). For vegetation
trend analysis studies to be used to inform land management policy and
to be meaningfully compared to other studies, any changes in vegeta-
tion and precipitation must be larger than the noise inherent to the
datasets used to capture them (Scheftic et al., 2014; Verbesselt et al.,
2010b). For both the G3.0 and G3.1 run ensembles, there was good
overall agreement, with < 4.5% of pixels having runs that produced
contradictory results. Of these pixels with direction inversion (TCinv),
the TC was generally very small with ∼60% of both G3.0 and G3.1
having maximum absolute total change (|TCmax|) in the bottom quar-
tile.
As with trend directionality, the range of the TC values within the
ensembles also showed good correspondence, with 86% of G3.0 and
92% of G3.1 pixels with a TCrange < 0.05. However, there are clusters
in both the G3.0 and G3.1 ensembles that have much higher TCrange
values. It was hypothesised that if the primary cause of areas of large
variation in the TC values was large variations in the precipitation
datasets alone, then there would be significant overlap between areas of
high total change range in the G3.0 and G3.1 ensembles. It was ex-
pected that such areas would correlate to areas of low rainfall station
density where gridded precipitation datasets are known to differ the
most (Contractor et al., 2015; Herold et al., 2017; King et al., 2013; Zhu
A.L. Burrell et al. ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018) 22–37

Fig. 5. Comparing the G3.0 and G3.1 ensembles. (a) The mean of the difference between matched runs. The red box over central Australia highlights an area of large
disagreement and is discussed in the text below. Regions with no difference are masked in white. (b) The agreement in the TCmean Inter-ensemble directionality. Red
indicates that ensemble means had significant contradictory results (DI), light grey indicates that one ensemble found a significant result and other found no
significant change (no direction inversion (NDI)), blue indicates pixels where both ensembles agree about the direction and significance (MD). The black box over
central Queensland highlights an area of large disagreement and is discussed in the text below. The dark grey are pixels with less than 85% valid data points and are
masked in all figures. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

30
A.L. Burrell et al. ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018) 22–37

Fig. 6. (a) Mean of the total change (TCmean) of the G2015 ensemble. Regions with no significant change are are masked in white. (b) The difference between the
G2015 TCmean per year and the matched G3.1 TCmean per year. Regions with absolute rate difference < 0.00025 are masked in white. The dark grey are pixels with
less than 85% valid data points and are masked in all figures.

et al., 2015). Less than 5% of pixels with a large ensemble range pixels with an |TCrange/TCmean| > 0.5 present in both ensembles. This
(TCrange > 0.1) are present in both ensembles. This low spatial corre- unexpected result indicates that the variations in the precipitation da-
spondence is present even after controlling for the differences in the tasets alone cannot account for the large TCrange values, the cause of
mean of the total change on a per pixel basis, with < 19.75% of the which is discussed further in section 5.5.

31
A.L. Burrell et al.
Fig. 7. Breakpoints per year per run in the annual max NDVI time series of the G3.0, G3.1 and G2015 ensembles.
Breakpoint timing showed greater variability than TCrange or TCinv
with complete ensemble agreement on the presence or absence of a
breakpoint in 67.2% of G3.0 pixels and 60.7% of G3.1. This indicates
that the TSS-RESTREND methodology is not fully able to resolve the
documented issues of breakpoint detection in dryland regions (Burrell
et al., 2017; Fensholt et al., 2015). The agreement about breakpoints
improved considerably if at least one run found a breakpoint in the
VPR. This, combined with the fact that there is no spatial correlation
between areas with TCinv or large TCrange and areas that have low en-
semble agreement on breakpoints, indicates that large changes in eco-
systems are being captured accurately and that the disagreement stems
from smaller changes in the VPR residuals. The use of ensembles pro-
vides a way to find those breakpoints that correspond to the largest
ecological changes.
5.2. Comparing the performance of the rainfall datasets
While there was generally good ensemble agreement, the perfor-
mance of individual precipitation datasets varied. In both the G3.0 and
G3.1 ensembles, GPCC had the lowest number of outlier results and the
highest correlation with AWAP (r = 0.888), followed closely by Udel
(r = 0.884) and MSWEP (r = 0.876). The performance of GPCC is un-
surprising as the station data used in its creation is highly curated with
strict quality control standards (Funk et al., 2015). The largest dis-
crepancy between two runs was between AWAP and CRU4 in the G3.1
ensemble where the average TCmean varied by 11%. This is smaller than
might be expected given the known biases in the datasets (Beck et al.,
2017b)
Despite performing the best in both metrics, the usability of GPCC in
future studies is somewhat limited because it currently only extends to
2013 and is generally only updated every 4–6 years. This means that it
cannot be used to look at the complete GIMMSv3.1g or MODIS NDVI
record. This is a problem with all the purely station based gridded
datasets which generally have multiyear latencies to allow for increased
station reporting and curation (Beck et al., 2017b; Funk et al., 2015).
5.3. Comparing the NDVI ensembles
As with the intra-ensemble comparisons, there is generally good
32
ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018) 22–37
agreement between the G3.0 and G3.1 ensembles, with only 2.93% of
pixels experiencing direction inversion. The mean per pixel TCmean
difference between the two ensembles was only 0.00015 units NDVI.
There are notable exceptions in Central Australia, where G3.0 has large
positive change and G3.1 found no significant change, as well as over
much of central Queensland, where the G3.0 had lower TC values.
These areas also correspond to regions where there was the greatest
disagreement between the presence or absence of breakpoints between
the G3.0 and G3.1 ensemble.
BFAST, which is used to look for breakpoints in the TSS-RESTREND
method, uses a detection window (Verbesselt et al., 2010b) and as such
it cannot detect a breakpoint near the end of the timeseries (Verbesselt
et al., 2012). It was observed in this study (Fig. 2a and b) and by Burrell
et al. (2017) that TSS-RESTREND applied to the period 1982–2013
resulted in a large spike in breakpoints at the end of the time series.
Burrell et al. (2017) attributed this to the AVHRR sensor transition near
the end of the time series. It was hypothesised that extending the
temporal record would preserve the observed trends and reduce the
number of breakpoints. Extending the temporal record through to 2015
increased the total number of pixels with significant breakpoints
(Fig. 2a–c) which indicates that the spike in breakpoints is not the result
of the sensor transition alone. The cause of this, as well as the merits of
G3.0 vs G3.1 are discussed in Sections 5.4 and 5.5.
5.4. Problematic regions for land degradation assessment
Most of Australia meets the three criteria for the valid inter-
comparison of different studies outlined above. As mentioned in section
5.1 and 5.3, there are clusters of pixels that significantly violate these
assumptions. The two most notable examples are in Central Australia
(Region 1 Fig. 1) and the semiarid and subtropical shrub and grasslands
of central Queensland (Region 2 Fig. 1).
5.4.1. Region 1 – Central Australia
The region with the largest difference in inter-ensemble TCmean is a
cluster of pixels located in central Australia (Fig. 5a) with most pixels
having TCmean difference > 0.3 units NDVI and some as high as 0.5.
This area of Australia also had the highest TCrange in the G3.0 ensemble,
with many pixels having a TCrange > 0.3 units NDVI. An example NDVI
A.L. Burrell et al.
Fig. 8. The NDVI time series of a characteristic pixel from Region 1 comparing GIMMSv3.0g and GIMMSv3.1g. MODIS NDVI is included for comparison.
time series from a pixel in this region is shown in Fig. 8.
It is clear from Fig. 8 that there are differences between the
GIMMS3.0 and GIMMSv3.1g for the entire time series with GIMMS3.0
having exaggerated variability well above GIMMSv3.1g and MODIS.
This variability is probably caused by the errors is SEAWIFS calibration
that were addressed in version 3.1 (Pinzon, 2017, pers.comm.). Though
this problem is present throughout the time series it is particularly
pronounced after 2009. The 2009–2013 period is an interesting time
period for the GIMMS dataset (Poulter et al., 2014; Tian et al., 2015),
Australian vegetation in general and this region in particular.
In January 2009, there was a sensor transition NOAA-17 to NOAA-
18 (Tian et al., 2015) which occurred during one of the driest years of
Australia’s Millennium Drought ” (van Dijk et al., 2013). Then, starting
in November 2009, a La Niña event resulted in record breaking rainfall
which persisted though to 2012 causing widespread flooding which, in
turn, caused a massive vegetation green up (Leigh et al., 2015; Poulter
et al., 2014; Renzullo et al., 2011; van Dijk et al., 2013). In Region 1
there were 14 flood events between 1950 and 2016 of which five oc-
curred between February 2009 and March 2012 (BoM (2016)).
In region 1, these flooding events and the ecological response as-
sociated with it appears to have compounded the problems caused by
the 2009 sensor transition and the SEAWFIS calibration error to pro-
duce an extended period of extremely elevated NDVI in the
Fig. 9. RMSE between MODIS NDVI and (a) GIMMSv3.1g and (b) GIMMSv3.0g for the period 2000–2013.
33
ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018) 22–37
GIMMSv3.0g record. This error is also reflected in the RMSE between
GIMMSv3.0g and MODIS (Fig. 9b) with many values > 0.25 compared
to a mean pixel RMSE of 0.083. GIMMSv3.1g has a much lower RMSE
over this region (∼0.03). This strongly suggests that the high NDVI
values in the GIMMSv3.0g are the result of the SEAWFIS calibration
error. The presence of this error in GIMMSv3.0g at the end of the time
series explains the large TCrange in the G3.0 ensemble. It should be
noted that time series analysis that uses linear regression is sensitive to
outliers at the end of the time series (Wessels et al., 2012).
While the large variance in the TCrange in the G3.0 ensemble in
Region 1 is an example of the impact of the calibration error in the
NDVI datasets, it also points to a potential limitation of the
TSS-RESTREND method. The large NDVI change in the GIMMSv3.0g
record at the end of the time series caused the G3.0 ensemble members
TC results to diverge considerably. If this was a real environmental
change and not a data artefact, TSS-RESTREND results would not be
able to reliably capture the change. The change in the GIMMSv3.0g
time series is the equivalent of a tropical rainforest appearing in the
middle of a desert over the space of a single month. The authors know
of no processes that would cause variation on this scale at this rate in
dryland ecosystems, so this limitation may not be a problem in de-
tecting real change.
A.L. Burrell et al.
Fig. 10. The NDVI time series of a characteristic pixel from Region 2 comparing GIMMSv3.0g and GIMMSv3.1g. MODIS NDVI is included for comparison.
5.4.2. Region 2 – Queensland
The second region with large differences between the TCmean of the
G3.0 and G3.1 ensembles is in central Queensland. This also corre-
sponds to a cluster of pixels with direction inversion which is shown in
Fig. 5b. In this region, the G3.0 ensemble found large areas of negative
or no change while G3.1 found small but positive changes (Fig. 3a and b
respectively). Despite scale differences that make direct comparison
difficult, there is good correspondence between the areas with the
highest vegetation clearing rates (1988–2014) in central Queensland
(Held et al., 2015; Queensland Department of Science, Information
Technology and Innovation, 2015), a strong observed degradation
signal in the passive microwave based Vegetation Optical Depth (VOD)
(Andela et al., 2013) and the region of negative trend detected by other
studies conducted using GIMMSv3.0g or the original GIMMS datasets
(Andela et al., 2013; Bai et al., 2008a, 2008b; Burrell et al., 2017). This
makes the weak positive trend of the G3.1 ensemble surprising. An
example NDVI time series from this region is shown in Fig. 10.
Assuming that the reported rates of land clearing and associated
degradation are accurate (Queensland Department of Science,
Information Technology and Innovation, 2015), there are two possible
reasons that G3.1 did not detect it. The first is that the new SEAWIFS
correction in 3.1 has removed too much of the variability in the NDVI
signal. There is some evidence of this in both Figs. 8 and 10 with
GIMMS31 consistently below MODIS values in high growth periods.
However, it is the low NDVI months where the largest differences be-
tween GIMMS and MODIS are apparent and at these times GIMMSv3.1g
is a better match. In addition, the RMSE for GIMMSv3.1g is lower than
the GIMMSv3.0g for almost every pixel (Fig. 9). These factors indicate
that GIMMSv3.1g is the more accurate dataset overall. As such, the
more likely explanation for the greening is that the widespread flooding
that caused problems in Region 1 also caused rapid positive ecosystem
changes in region 2 that masked out the long-term negative trend
caused by land clearing (Andela et al., 2013; Queensland Department of
Science, Information Technology and Innovation, 2015). The spike in
breakpoints between 2010 and 2012 and the re-emergence of a nega-
tive trend over parts of Region 2 with active land clearing in the G2015
ensemble (Queensland Department of Science, Information Technology
and Innovation, 2015) supports this explanation.
5.5. Dataset selection and the advantages of using ensembles
A limitation of many studies that look for land degradation using
trend analysis is that there is a large gap between the end of the analysis
34
ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018) 22–37
and the publication of the study. For most studies its about 4–6 years
(Andela et al., 2013; Bai et al., 2008a; Burrell et al., 2017; Fensholt
et al., 2012; Li et al., 2012; Tong et al., 2017b; Wessels et al., 2007)
some having almost a decade lag (Huang and Kong, 2016;
Keersmaecker et al., 2017; Rishmawi and Prince, 2016). Part of the
reason for this is that they are dependent on the slow release of input
datasets. Datasets like MSWEP and CHIRPS use a limited number of
stations that report near real time data, combined with information
from satellites that have continuous global coverage (Yanto and
Rajagopalan, 2017) which means they can be updated much more
frequently (Funk et al., 2015). The results of this study indicate that, for
the purposes of vegetation trend analysis, these hybrid precipitation
datasets provide equivalent quality without the large temporal lag. This
finding is consistent with a recent precipitation data quality assessment
that looked at 22 gridded precipitation products and found MSWEP and
CHIRPS to be among the most accurate datasets (Beck et al., 2017b).
Unlike precipitation datasets, the NDVI dataset has a more sig-
nificant impact on results. This study provides clear evidence that that
the SEAWIFS calibration error on GIMMSv3.0g causes significant issues
in some of Australia’s drylands. The authors of this paper know of no
study that has performed a per pixel comparison of the two versions of
the GIMMS dataset which means the extent of the issue is unknown.
However, there is reason to think it has had a global impact. In the
supplementary material of their recent paper, Pan et al. (2018) included
a timeseries of the global annual mean NDVI for both GIMMSv3.0g and
GIMMSv3.1g. Both the annual mean global NDVI values and the rate of
global greening varied considerably between the two datasets. This may
be a problem because most of the recently published studies continue to
use GIMMSv3.0g, despite the availability of GIMMSv3.1g (Davis et al.,
2017; Feng et al., 2017; He et al., 2017; Liu et al., 2017; Pan et al.,
2018; Tong et al., 2017a; Wu et al., 2018b).
While this study is the first to apply an ensemble based approach to
dryland remote sensing degradation detection, ensembles are ex-
tensively used in other scientific fields like climate science (Haughton
et al., 2015), meteorology (Hagedorn et al., 2005) and hydrology
(Krueger et al., 2010) to improve accuracy and quantify errors. The
problematic regions identified above look normal in any individual run
and can only be observed by looking at the spread in the ensembles.
This, along with the growing number of regularly updated datasets like
MSWEP, CHIRPS and the newly released TerraClimate (Abatzoglou
et al., 2018) indicates future studies that want to quantify land de-
gradation should use an ensemble approach. The biggest limitation to
this approach is that it significantly increases the computational time
A.L. Burrell et al.
and cost though, given the documented importance of accurate land
degradation assessments (Bai et al., 2008a; Dudley and Alexander,
2017; Scholes et al., 2018), the expenditure seems warranted.
6. Conclusions
This paper examines the impact of dataset selection on land de-
gradation detection using the TSS-RESTREND method. All datasets
found a strong greening trend with NDVI increases in about 60% of
areas and decreases in about 5% of areas. While there are some dif-
ferences between the TSS-RESTREND runs performed using different
precipitation datasets, the differences were small enough that a single
run would be sufficient to capture the direction of change in almost all
regions. This, along with the low rate of direction inversions, supports
the validity of comparing existing studies that have used different
precipitation datasets and increases the confidence in future studies.
The fact that MSWEP and CHRIPS, which are low temporal lag hybrid
datasets, performed as well as the purely station-based datasets with
more stringent quality control is of particular note. Past land de-
gradation studies often have multiyear delays, in part because of the
infrequent update schedules of global dataset. Use of MSWEP and
CHIRPS opens the door for accurate near real time degradation detec-
tion. The availability and accuracy of these new precipitation products
also allows for an ensemble based global estimate of land degradation.
There is more variance in the NDVI datasets however. The extreme
precipitation/flood events in the last few years of the record, combined
with the SEAWIFS calibration error in GIMMSv3.0g, resulted in sig-
nificant inaccuracies in land degradation assessment for some dryland
areas. By using a multi-dataset ensemble, this study was able to identify
problematic regions and eliminate spurious trends caused by errors and
uncertainties within individual datasets. These problematic regions are
difficult to detect automatically using only a single run. Future studies
should adopt an ensemble approach to quantify errors and produce
robust estimates of dryland degradation. In addition, it is strongly re-
commended that all future studies in dryland regions use GIMMSv3.1g
rather than GIMMSv3.0g because of the extreme outliers that result
from the calibration errors. While this study has only looked at
Australia this problem may have effected dryland NDVI values globally
which in turn may have impacted many of the currently published es-
timates of dryland degradation.
7. Conflicts of interest
The authors declare no conflicts of interest.
Acknowledgements
This study was supported by the Australian Research Council (ARC)
Centre of Excellence for Climate Extremes (CE170100023). Arden
Burrell is supported by an Australian Postgraduate Award through the
University of New South Wales and a top-up scholarship from the
UNSW Climate Change Research Centre.
Appendix A. Supplementary material
Supplementary data associated with this article can be found, in the
online version, at https://doi.org/10.1016/j.isprsjprs.2018.08.017.
References
Abatzoglou, J.T., Dobrowski, S.Z., Parks, S.A., Hegewisch, K.C., 2018. TerraClimate, a
high-resolution global dataset of monthly climate and climatic water balance from
1958–2015. Sci. Data 5, 170191. https://doi.org/10.1038/sdata.2017.191.
Anav, A., Friedlingstein, P., Kidston, M., Bopp, L., Ciais, P., Cox, P., Jones, C., Jung, M.,
Myneni, R., Zhu, Z., 2013. Evaluating the land and ocean components of the global
carbon cycle in the CMIP5 earth system models. J. Clim. 26, 6801–6843. https://doi.
org/10.1175/JCLI-D-12-00417.1.
35
ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018) 22–37
Andela, N., Liu, Y.Y., van Dijk, A.I.J.M., de Jeu, R.A.M., McVicar, T.R., 2013. Global
changes in dryland vegetation dynamics (1988–2008) assessed by satellite remote
sensing: combining a new passive microwave vegetation density record with re-
flective greenness data. Biogeosci. Discuss 10, 8749–8797. https://doi.org/10.5194/
bgd-10-8749-2013.
Australian Bureau of Agricultural and Resource Economics and Sciences, 2015.
Catchment Scale Land Use of Australia – Updare March 2015.
Baeza, SS., Paruelo, J.M., 2018. Spatial and temporal variation of human appropriation of
net primary production in the Rio de la Plata grasslands. ISPRS J. Photogramm.
Remote Sens. https://doi.org/10.1016/j.isprsjprs.2018.07.014.
Bai, Z.G., Dent, D.L., Olsson, L., Schaepman, M.E., 2008a. Global assessment of land
degradation and improvement. 1. Identification by remote sensing. Wagening. Int.
Soil Ref. Inf. Cent. ISRIC.
Bai, Z.G., Dent, D.L., Olsson, L., Schaepman, M.E., 2008b. Proxy global assessment of land
degradation. Soil Use Manag. 24, 223–234. https://doi.org/10.1111/j.1475-2743.
2008.00169.x.
Beck, H.E., van Dijk, A.I.J.M., Levizzani, V., Schellekens, J., Miralles, D.G., Martens, B., de
Roo, A., 2017a. MSWEP: 3-hourly 0.25° global gridded precipitation (1979–2015) by
merging gauge, satellite, and reanalysis data. Hydrol. Earth Syst. Sci. 21, 589–615.
https://doi.org/10.5194/hess-21-589-2017.
Beck, H.E., Vergopolan, N., Pan, M., Levizzani, V., van Dijk, A.I.J.M., Weedon, G.P.,
Brocca, L., Pappenberger, F., Huffman, G.J., Wood, E.F., 2017b. Global-scale eva-
luation of 22 precipitation datasets using gauge observations and hydrological
modeling. Hydrol. Earth Syst. Sci. 21, 6201–6217. https://doi.org/10.5194/hess-21-
6201-2017.
BoM, 2016. Diamantina River Flood Warning System [WWW Document]. URL: http://
www.bom.gov.au/qld/flood/brochures/diamantina/diamantina.shtml#
PreviousFlooding (accessed 7.9.17).
BoM, 2012. Australian Climate Averages – Climate classifications [WWW Document].
Clim. Classif. Maps. http://www.bom.gov.au/jsp/ncc/climate_averages/climate-
classifications/index.jsp?maptype=kpngrp#info (accessed 8.3.16).
Boschetti, M., Nutini, F., Brivio, P.A., Bartholomé, E., Stroppiana, D., Hoscilo, A., 2013.
Identification of environmental anomaly hot spots in West Africa from time series of
NDVI and rainfall. ISPRS J. Photogramm. Remote Sens. 78, 26–40. https://doi.org/
10.1016/j.isprsjprs.2013.01.003.
Brandt, M., Tappan, G., Diouf, A.A., Beye, G., Mbow, C., Fensholt, R., 2017. Woody ve-
getation die off and regeneration in response to rainfall variability in the West African
Sahel. Remote Sens. 9, 39. https://doi.org/10.3390/rs9010039.
Broich, M., Huete, A., Tulbure, M.G., Ma, X., Xin, Q., Paget, M., Restrepo-Coupe, N.,
Davies, K., Devadas, R., Held, A., 2014. Land surface phenological response to dec-
adal climate variability across Australia using satellite remote sensing.
Biogeosciences 11, 5181–5198. https://doi.org/10.5194/bg-11-5181-2014.
Burrell, A.L., Evans, J.P., Liu, Y., 2017. Detecting dryland degradation using Time Series
Segmentation and Residual Trend analysis (TSS-RESTREND). Remote Sens. Environ.
197, 43–57. https://doi.org/10.1016/j.rse.2017.05.018.
Caccetta, P., Furby, S., Wallace, J., Wu, X., Richards, G., Waterworth, R., 2012. Long-term
monitoring of australian land cover change using landsat data. In: Global Forest
Monitoring from Earth Observation, Earth Observation of Global Changes CRC Press,
pp. 243–258. https://doi.org/10.1201/b13040-14.
CDO, 2018. Climate Data Operators.
Chen, B., Xu, G., Coops, N.C., Ciais, P., Innes, J.L., Wang, G., Myneni, R.B., Wang, T.,
Krzyzanowski, J., Li, Q., Cao, L., Liu, Y., 2014a. Changes in vegetation photosynthetic
activity trends across the Asia-Pacific region over the last three decades. Remote
Sens. Environ. 144, 28–41. https://doi.org/10.1016/j.rse.2013.12.018.
Chen, B., Zhang, X., Tao, J., Wu, J., Wang, J., Shi, P., Zhang, Y., Yu, C., 2014b. The impact
of climate change and anthropogenic activities on alpine grassland over the Qinghai-
Tibet Plateau. Agric. For. Meteorol. 189–190, 11–18. https://doi.org/10.1016/j.
agrformet.2014.01.002.
Contractor, S., Alexander, L.V., Donat, M.G., Herold, N., 2015. How well do gridded
datasets of observed daily precipitation compare over Australia? Adv. Meteorol.
2015, e325718. https://doi.org/10.1155/2015/325718.
Dardel, C., Kergoat, L., Hiernaux, P., Grippa, M., Mougin, E., Ciais, P., Nguyen, C.-C.,
2014. Rain-use-efficiency: what it tells us about the conflicting Sahel greening and
Sahelian paradox. Remote Sens. 6, 3446–3474. https://doi.org/10.3390/rs6043446.
Davis, C.L., Hoffman, M.T., Roberts, W., 2017. Long-term trends in vegetation phenology
and productivity over Namaqualand using the GIMMS AVHRR NDVI3g data from
1982 to 2011. South Afr. J. Bot. 111, 76–85. https://doi.org/10.1016/j.sajb.2017.03.
007.
Day, L., Knowledge, C., Springs, A., 2007. Status of Biodiversity Monitoring in
Rangelands. Australian Collaborative Rangelands Information System.
de Jong, R., Verbesselt, J., Schaepman, M.E., De Bruin, S., 2011. Detection of breakpoints
in global NDVI time series. In: Presented at the 34th International Symposium on
Remote Sensing of Environment (ISRSE), pp. 10–15.
Detsch, F., Otte, I., Appelhans, T., Hemp, A., Nauss, T., 2016. Seasonal and long-term
vegetation dynamics from 1-km GIMMS-based NDVI time series at Mt. Kilimanjaro,
Tanzania. Remote Sens. Environ. 178, 70–83. https://doi.org/10.1016/j.rse.2016.03.
007.
Didan, K., 2015. MOD13Q1 MODIS/Terra Vegetation Indices 16-Day L3 Global 250 m SIN
Grid V006. https://doi.org/10.5067/MODIS/MOD13Q1.006.
Dudley, N., Alexander, S., 2017. Global Land Outlook, first ed. UNCCD, Bonn, Germany.
Eddy, I.M.S., Gergel, S.E., Coops, N.C., Henebry, G.M., Levine, J., Zerriffi, H., Shibkov, E.,
2017. Integrating remote sensing and local ecological knowledge to monitor range-
land dynamics. Ecol. Indic. 82, 106–116. https://doi.org/10.1016/j.ecolind.2017.06.
033.
Evans, J., Geerken, R., 2004. Discrimination between climate and human-induced dry-
land degradation. J. Arid Environ. 57, 535–554. https://doi.org/10.1016/S0140-
1963(03)00121-6.
Eyre, T.J., Fisher, A., Hunt, L.P., Kutt, A.S., 2011. Measure it to better manage it: a
biodiversity monitoring framework for the Australian rangelands. Rangel. J. 33,
239–253.
A.L. Burrell et al.
Fan, X., Liu, Y., 2016. A global study of NDVI difference among moderate-resolution
satellite sensors. ISPRS J. Photogramm. Remote Sens. 121, 177–191. https://doi.org/
10.1016/j.isprsjprs.2016.09.008.
Feng, H., Zou, B., Luo, J., 2017. Coverage-dependent amplifiers of vegetation change on
global water cycle dynamics. J. Hydrol. 550, 220–229. https://doi.org/10.1016/j.
jhydrol.2017.04.056.
Fensholt, R., Horion, S., Tagesson, T., Ehammer, A., Grogan, K., Tian, F., Huber, S.,
Verbesselt, J., Prince, S.D., Tucker, C.J., Rasmussen, K., 2015. Assessing drivers of
vegetation changes in drylands from time series of earth observation data. In:
Kuenzer, C., Dech, S., Wagner, W. (Eds.), Remote Sensing Time Series, Remote
Sensing and Digital Image Processing. Springer International Publishing, pp.
183–202. https://doi.org/10.1007/978-3-319-15967-6_9.
Fensholt, R., Langanke, T., Rasmussen, K., Reenberg, A., Prince, S.D., Tucker, C., Scholes,
R.J., Le, Q.B., Bondeau, A., Eastman, R., Epstein, H., Gaughan, A.E., Hellden, U.,
Mbow, C., Olsson, L., Paruelo, J., Schweitzer, C., Seaquist, J., Wessels, K., 2012.
Greenness in semi-arid areas across the globe 1981–2007—an Earth Observing
Satellite based analysis of trends and drivers. Remote Sens. Environ. 121, 144–158.
https://doi.org/10.1016/j.rse.2012.01.017.
Fensholt, R., Proud, S.R., 2012. Evaluation of Earth Observation based global long term
vegetation trends—Comparing GIMMS and MODIS global NDVI time series. Remote
Sens. Environ. 119, 131–147. https://doi.org/10.1016/j.rse.2011.12.015.
Fensholt, R., Rasmussen, K., Kaspersen, P., Huber, S., Horion, S., Swinnen, E., 2013.
Assessing land degradation/recovery in the African Sahel from long-term earth ob-
servation based primary productivity and precipitation relationships. Remote Sens. 5,
664–686. https://doi.org/10.3390/rs5020664.
Fensholt, R., Rasmussen, K., Nielsen, T.T., Mbow, C., 2009. Evaluation of earth ob-
servation based long term vegetation trends—Intercomparing NDVI time series trend
analysis consistency of Sahel from AVHRR GIMMS, Terra MODIS and SPOT VGT data.
Remote Sens. Environ. 113, 1886–1898. https://doi.org/10.1016/j.rse.2009.04.004.
Fisher, A., Kutt, A., 2006. Biodiversity and land condition in tropical savanna rangelands:
summary report. Tropical Savannas CRC, Land & Water Australia, Darwin.
Funk, C., Peterson, P., Landsfeld, M., Pedreros, D., Verdin, J., Shukla, S., Husak, G.,
Rowland, J., Harrison, L., Hoell, A., Michaelsen, J., 2015. The climate hazards in-
frared precipitation with stations—a new environmental record for monitoring ex-
tremes. Sci. Data 2, 150066. https://doi.org/10.1038/sdata.2015.66.
Funk, C.C., Peterson, P.J., Landsfeld, M.F., Pedreros, D.H., Verdin, J.P., Rowland, J.D.,
Romero, B.E., Husak, G.J., Michaelsen, J.C., Verdin, A.P., 2014. A quasi-global pre-
cipitation time series for drought monitoring (Report No. 832), Data Series. Reston,
VA. https://doi.org/10.3133/ds832.
Gao, Q., Schwartz, M.W., Zhu, W., Wan, Y., Qin, X., Ma, X., Liu, S., Williamson, M.A.,
Peters, C.B., Li, Y., 2016. Changes in global grassland productivity during 1982 to
2011 attributable to climatic factors. Remote Sens. 8, 384. https://doi.org/10.3390/
rs8050384.
Guan, K., Medvigy, E.F.WoodD., Kimball, J., Pan, M., Caylor, K.K., Sheffield, J., Xu, X.,
Jones, M.O., 2014. Terrestrial hydrological controls on land surface phenology of
African savannas and woodlands. 2013JG002572. J. Geophys. Res. Biogeosci. 119.
https://doi.org/10.1002/2013JG002572.
Hagedorn, R., Doblas-Reyes, F.J., Palmer, T.N., 2005. The rationale behind the success of
multi-model ensembles in seasonal forecasting—I. Basic concept. Tellus Dyn.
Meteorol. Oceanogr. 57, 219–233. https://doi.org/10.3402/tellusa.v57i3.14657.
Harris, I., 2017. CRU TS v4.00 [WWW Document]. RELEASE NOTES CRU TS V400 31
January 2017. URL: https://crudata.uea.ac.uk/cru/data/hrg/cru_ts_4.00/Release_
Notes_CRU_TS4.00.txt (accessed 6.2.17).
Harris, I., Jones, P.D., Osborn, T.J., Lister, D.H., 2014. Updated high-resolution grids of
monthly climatic observations – the CRU TS3.10 Dataset. Int. J. Climatol. 34,
623–642. https://doi.org/10.1002/joc.3711.
Haughton, N., Abramowitz, G., Pitman, A., Phipps, S.J., 2015. Weighting climate model
ensembles for mean and variance estimates. Clim. Dyn. 45, 3169–3181. https://doi.
org/10.1007/s00382-015-2531-3.
He, C., Tian, J., Gao, B., Zhao, Y., 2014. Differentiating climate- and human-induced
drivers of grassland degradation in the Liao River Basin. China. Environ. Monit.
Assess. 187, 1–14. https://doi.org/10.1007/s10661-014-4199-2.
He, Y., Lee, E., Warner, T.A., 2017. A time series of annual land use and land cover maps
of China from 1982 to 2013 generated using AVHRR GIMMS NDVI3g data. Remote
Sens. Environ. 199, 201–217. https://doi.org/10.1016/j.rse.2017.07.010.
Healy, M.-A., Forrest, K., Bastin, G., 2016. The role of a knowledge broker in improving
knowledge and understanding of climate change in the Australian rangelands.
Rangel. J. 37, 541–554. https://doi.org/10.1071/RJ15060.
Held, A., Phinn, S., Soto-Berelov, M., Jones, S., 2015. AusCover good practice guidelines:
a technical handbook supporting calibration and validation activities of remotely
sensed data products (Technical Report).
Herold, N., Behrangi, A., Alexander, L.V., 2017. Large uncertainties in observed daily
precipitation extremes over land. 2016JD025842. J. Geophys. Res. Atmosp. 122.
https://doi.org/10.1002/2016JD025842.
Higginbottom, T.P., Symeonakis, E., 2014. Assessing land degradation and desertification
using vegetation index data: current frameworks and future directions. Remote Sens.
6, 9552–9575. https://doi.org/10.3390/rs6109552.
Huang, S., Kong, J., 2016. Assessing land degradation dynamics and distinguishing
human-induced changes from climate factors in the three-north shelter forest region
of China. ISPRS Int. J. Geo-Inf. 5, 158. https://doi.org/10.3390/ijgi5090158.
Huang, W., Huang, J., Wang, X., Wang, F., Shi, J., 2013. Comparability of red/near-
infrared reflectance and NDVI based on the spectral response function between
MODIS and 30 other satellite sensors using rice canopy spectra. Sensors 13,
16023–16050. https://doi.org/10.3390/s131216023.
Hutchinson, J.M.S., Jacquin, A., Hutchinson, S.L., Verbesselt, J., 2015. Monitoring ve-
getation change and dynamics on U.S. Army training lands using satellite image time
series analysis. J. Environ. Manage. 150, 355–366. https://doi.org/10.1016/j.
jenvman.2014.08.002.
Ibrahim, Y.Z., Balzter, H., Kaduk, J., Tucker, C.J., 2015. Land degradation assessment
using residual trend analysis of GIMMS NDVI3g, soil moisture and rainfall in Sub-
36
ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018) 22–37
Saharan West Africa from 1982 to 2012. Remote Sens. 7, 5471–5494. https://doi.
org/10.3390/rs70505471.
IPCC, 2017. Decision and Outline of the Special Report on climate change, desertification,
land degradation, sustainable land management, food security, and greenhouse gas
fluxes in terrestrial ecosystems (SIXTH ASSESSMENT REPORT (AR6) PRODUCTS).
Intergovernmental Panel on Climate Change.
John, R., Chen, J., Kim, Y., Ou-yang, Z., Xiao, J., Park, H., Shao, C., Zhang, Y., Amarjargal,
A., Batkhshig, O., Qi, J., 2015. Differentiating anthropogenic modification and pre-
cipitation-driven change on vegetation productivity on the Mongolian Plateau.
Landsc. Ecol. 31, 547–566. https://doi.org/10.1007/s10980-015-0261-x.
Jones, D.A., Wang, W., Fawcett, R., 2009. High-quality spatial climate data-sets for
Australia. Aust. Meteorol. Oceanogr. J. 58, 233.
Jones, P.W., 1999. First- and second-order conservative remapping schemes for grids in
spherical coordinates. Mon. Weather Rev. 127, 2204–2210. https://doi.org/10.1175/
1520-0493(1999) 127<2204:FASOCR>2.0.CO;2.
Kaptué, A.T., Prihodko, L., Hanan, N.P., 2015. On regreening and degradation in Sahelian
watersheds. Proc. Natl. Acad. Sci. 112, 12133–12138. https://doi.org/10.1073/pnas.
1509645112.
Keenan, T.F., Riley, W.J., 2018. Greening of the land surface in the world’s cold regions
consistent with recent warming. Nat. Clim. Change 1. https://doi.org/10.1038/
s41558-018-0258-y.
Keersmaecker, W.D., Lhermitte, S., Hill, M.J., Tits, L., Coppin, P., Somers, B., 2017.
Assessment of regional vegetation response to climate anomalies: a case study for
Australia using GIMMS NDVI time series between 1982 and 2006. Remote Sens. 9,
34. https://doi.org/10.3390/rs9010034.
Khishigbayar, J., Fernández-Giménez, M.E., Angerer, J.P., Reid, R.S., Chantsallkham, J.,
Baasandorj, Y., Zumberelmaa, D., 2015. Mongolian rangelands at a tipping point?
Biomass and cover are stable but composition shifts and richness declines after 20
years of grazing and increasing temperatures. J. Arid Environ. 115, 100–112. https://
doi.org/10.1016/j.jaridenv.2015.01.007.
King, A.D., Alexander, L.V., Donat, M.G., 2013. The efficacy of using gridded data to
examine extreme rainfall characteristics: a case study for Australia. Int. J. Climatol.
33, 2376–2387. https://doi.org/10.1002/joc.3588.
Krueger, T., Freer, J., Quinton, J.N., Macleod, C.J.A., Bilotta, G.S., Brazier, R.E., Butler, P.,
Haygarth, P.M., 2010. Ensemble evaluation of hydrological model hypotheses. Water
Resour. Res. 46. https://doi.org/10.1029/2009WR007845.
Kuenzer, C., Dech, S., Wagner, W., 2015. Remote Sensing Time Series: Revealing Land
Surface Dynamics. Springer.
Lawley, V., Lewis, M., Clarke, K., Ostendorf, B., 2016. Site-based and remote sensing
methods for monitoring indicators of vegetation condition: An Australian review.
Ecol. Indic. 60, 1273–1283. https://doi.org/10.1016/j.ecolind.2015.03.021.
Leigh, C., Bush, A., Harrison, E.T., Ho, S.S., Luke, L., Rolls, R.J., Ledger, M.E., 2015.
Ecological effects of extreme climatic events on riverine ecosystems: insights from
Australia. Freshw. Biol. 60, 2620–2638. https://doi.org/10.1111/fwb.12515.
Li, A., Wu, J., Huang, J., 2012. Distinguishing between human-induced and climate-
driven vegetation changes: a critical application of RESTREND in inner Mongolia.
Landsc. Ecol. 27, 969–982. https://doi.org/10.1007/s10980-012-9751-2.
Liu, Y.Y., Evans, J.P., McCabe, M.F., de Jeu, R.A.M., van Dijk, A.I.J.M., Dolman, A.J.,
Saizen, I., 2013. Changing climate and overgrazing are decimating mongolian
steppes. PLoS ONE 8, e57599. https://doi.org/10.1371/journal.pone.0057599.
Liu, Z., Wu, C., Liu, Y., Wang, X., Fang, B., Yuan, W., Ge, Q., 2017. Spring green-up date
derived from GIMMS3g and SPOT-VGT NDVI of winter wheat cropland in the North
China Plain. ISPRS J. Photogramm. Remote Sens. 130, 81–91. https://doi.org/10.
1016/j.isprsjprs.2017.05.015.
Ludwig, J.A., Bastin, G.N., Wallace, J.F., McVicar, T.R., 2007. Assessing landscape health
by scaling with remote sensing: when is it not enough? Landsc. Ecol. 22, 163–169.
https://doi.org/10.1007/s10980-006-9038-6.
Ludwig, J.A., Tongway, D.J., Bastin, G.N., James, C.D., 2004. Monitoring ecological in-
dicators of rangeland functional integrity and their relation to biodiversity at local to
regional scales. Austral Ecol. 29, 108–120.
Lymburner, L., Tan, P., Mueller, N., Thackway, R., Lewis, A., Thankappan, M., Randall, L.,
Islam, A., Senarath, U., 2011. The national dynamic land cover dataset. Geosci. Aust.
Symonston Aust. 10.
Maggioni, V., Meyers, P.C., Robinson, M.D., 2016. A review of merged high-resolution
satellite precipitation product accuracy during the Tropical Rainfall Measuring
Mission (TRMM) era. J. Hydrometeorol. 17, 1101–1117. https://doi.org/10.1175/
JHM-D-15-0190.1.
Marshall, M., Okuto, E., Kang, Y., Opiyo, E., Ahmed, M., 2016. Global assessment of
Vegetation Index and Phenology Lab (VIP) and Global Inventory Modeling and
Mapping Studies (GIMMS) version 3 products. Biogeosciences 13, 625–639. https://
doi.org/10.5194/bg-13-625-2016.
Matsuura, K., Willmott, C., 2007. Terrestrial Air Temperature and Precipitation:
1900–2006 Gridded Monthly Time Series, Version 1.01.
McAlpine, C., Thackway, R., Smith, A., 2014. Towards an Australian Rangeland
Biodiversity Monitoring Framework. A Discussion Paper Developed from a
Biodiversity Monitoring Workshop. Australian Collaborative Rangelands Information
System.
MEA, 2005. Ecosystems and Human Well-being: Desertification Synthesis. World
Resources Institute, Washington, DC, USA.
New, M., Hulme, M., Jones, P., 2000. Representing twentieth-century space-time climate
variability. Part II: Development of 1901–96 monthly grids of terrestrial surface
climate. J. Clim. 13, 2217–2238. https://doi.org/10.1175/1520-0442(2000)
013<2217:RTCSTC>2.0.CO;2.
Orr, B., Cowie, A., Sanchez, V.C., Chasek, P., Crossman, N., Erlewein, A., Louwagie, G.,
Maron, M., Metternicht, G., Minelli, S., 2017. Scientific Conceptual Framework for
Land Degradation Neutrality. A Report of the Science-Policy Interface. United
Nations Convention to Combat Desertification (UNCCD), Bonn, Germany.
Paget, M.J., King, E.A., 2008. MODIS Land data sets for the Australian region.
Pan, N., Feng, X., Fu, B., Wang, S., Ji, F., Pan, S., 2018. Increasing global vegetation
browning hidden in overall vegetation greening: Insights from time-varying trends.
A.L. Burrell et al.
Remote Sens. Environ. 214, 59–72. https://doi.org/10.1016/j.rse.2018.05.018.
Pickup, G., 1998. Desertification and climate change–the Australian perspective. Clim.
Res. 11, 51–63. https://doi.org/10.3354/cr011051.
Pinzon, J.E., 2017. Personal Communication. Calibration Error in GIMMS NDVI.
Pinzon, J.E., Tucker, C.J., 2014. A non-stationary 1981–2012 AVHRR NDVI3g time series.
Remote Sens. 6, 6929–6960. https://doi.org/10.3390/rs6086929.
Poulter, B., Frank, D., Ciais, P., Myneni, R.B., Andela, N., Bi, J., Broquet, G., Canadell,
J.G., Chevallier, F., Liu, Y.Y., Running, S.W., Sitch, S., van der Werf, G.R., 2014.
Contribution of semi-arid ecosystems to interannual variability of the global carbon
cycle. Nature 509, 600–603. https://doi.org/10.1038/nature13376.
Qiu, B., Chen, G., Tang, Z., Lu, D., Wang, Z., Chen, C., 2017. Assessing the three-north
shelter forest program in china by a novel framework for characterizing vegetation
changes. ISPRS J. Photogramm. Remote Sens. 133, 75–88. https://doi.org/10.1016/j.
isprsjprs.2017.10.003.
Queensland Department of Science, Information Technology and Innovation, 2015. Land
cover change in Queensland 2012–13 and 2013–14: a Statewide Landcover and Trees
Study (SLATS) report. Statewide Landcover and Trees Study (SLATS). DSITI,
Brisbane.
M. Raupach P. Briggs V. Haverd E. King M. Paget C. Trudinger Australian water avail-
ability project 2012 (AWAP).
Raupach, M., Briggs, P., Haverd, V., King, E., Paget, M., Trudinger, C., 2009. Australian
water availability project (AWAP): CSIRO marine and atmospheric research compo-
nent: final report for phase 3.
Reeves, M.C., Washington-Allen, R.A., Angerer, J., Hunt, R., Kulawardhana, R.W., Kumar,
L., Loboda, T., Loveland, T., Metternicht, G., Ramsey, D.R., 2015. Global view of
remote sensing of rangelands: evolution, applications, future pathways. Land
Resources Monitoring, Modeling, and Mapping with Remote Sensing, Remote Sensing
Handbook.
Renzullo, L., Chappell, A., Raupach, T., Dyce, P., Ming, L., Shao, Q., 2011. An assessment
of statistically blended satellite-gauge precipitation data for daily rainfall analysis in
Australia. Presented at the Proc. 34th International Symposium on Remote Sensing of
Environment (ISRSE), Sydney, Australia, April 10–15.
Rishmawi, K., Prince, S.D., 2016. Environmental and anthropogenic degradation of ve-
getation in the Sahel from 1982 to 2006. Remote Sens. 8, 948. https://doi.org/10.
3390/rs8110948.
Ruppert, J.C., Harmoney, K., Henkin, Z., Snyman, H.A., Sternberg, M., Willms, W.,
Linstädter, A., 2015. Quantifying drylands’ drought resistance and recovery: the
importance of drought intensity, dominant life history and grazing regime. Glob.
Change Biol. 21, 1258–1270. https://doi.org/10.1111/gcb.12777.
Scheftic, W., Zeng, X., Broxton, P., Brunke, M., 2014. Intercomparison of seven NDVI
products over the United States and Mexico. Remote Sens. 6, 1057–1084. https://doi.
org/10.3390/rs6021057.
Schneider, U., Becker, A., Finger, P., Meyer-Christoffer, A., Rudolf, B., Ziese, M., 2015.
GPCC Full Data Reanalysis Version 7.0 at 0.5°: Monthly Land-Surface Precipitation
from Rain-Gauges built on GTS-based and Historic Data.
Schneider, U., Becker, A., Finger, P., Meyer-Christoffer, A., Ziese, M., Rudolf, B., 2014.
GPCC’s new land surface precipitation climatology based on quality-controlled in situ
data and its role in quantifying the global water cycle. Theor. Appl. Climatol. 115,
15–40. https://doi.org/10.1007/s00704-013-0860-x.
Schneider, U., Ziese, M., Meyer-Christoffer, A., Finger, P., Rustemeier, E., Becker, A.,
2016. The new portfolio of global precipitation data products of the global pre-
cipitation climatology centre suitable to assess and quantify the global water cycle
and resources. In: Proceedings of the International Association of Hydrological
Sciences. Presented at the Water Resources Assessment and Seasonal Prediction –
International Conference Water Resources Assessment and Seasonal Prediction,
Koblenz, Germany, 13–16 October 2015 Copernicus GmbH, pp. 29–34. https://doi.
org/10.5194/piahs-374-29-2016.
Scholes, R., Montanarella, L., Brainich, A., Barger, N., ten Brink, B., Cantele, M., Erasmus,
B., Fisher, J., Gardner, T., Holland, T.G., Kohler, F., Kotiaho, J.S., Von Maltitz, G.,
Nangendo, G., Pandit, R., Parrotta, J., Potts, M.D., Prince, S., Sankaran, M., Willemen,
L., 2018. Summary for Policymakers of the Thematic Assessment Report on Land
Degradation and Restoration of the Intergovernmental Science-Policy Platform on
Biodiversity and Ecosystem Services (IPBES). IPBES Secretariat, Bonn, Germany.
State of the Environment 2011 Committee, 2011. Australia State of the Environment
2011.
Sylla, M.B., Giorgi, F., Coppola, E., Mariotti, L., 2013. Uncertainties in daily rainfall over
Africa: assessment of gridded observation products and evaluation of a regional cli-
mate model simulation. Int. J. Climatol. 33, 1805–1817. https://doi.org/10.1002/
joc.3551.
Tang, G., Arnone III, J.A., Verburg, P.S.J., Jasoni, R.L., Sun, L., 2015. Trends and climatic
sensitivities of vegetation phenology in semiarid and arid ecosystems in the US Great
Basin during 1982–2011. Biogeosciences 12, 6985–6997. https://doi.org/10.5194/
bg-12-6985-2015.
Tian, F., Brandt, M., Liu, Y.Y., Verger, A., Tagesson, T., Diouf, A.A., Rasmussen, K., Mbow,
C., Wang, Y., Fensholt, R., 2016. Remote sensing of vegetation dynamics in drylands:
evaluating vegetation optical depth (VOD) using AVHRR NDVI and in situ green
biomass data over West African Sahel. Remote Sens. Environ. 177, 265–276. https://
doi.org/10.1016/j.rse.2016.02.056.
Tian, F., Fensholt, R., Verbesselt, J., Grogan, K., Horion, S., Wang, Y., 2015. Evaluating
temporal consistency of long-term global NDVI datasets for trend analysis. Remote
Sens. Environ. 163, 326–340. https://doi.org/10.1016/j.rse.2015.03.031.
37
ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018) 22–37
Tong, S., Zhang, J., Bao, Y., Wurina, Terigele, Weilisi, Lianxiao, 2017a. Spatial and
temporal variations of vegetation cover and the relationships with climate factors in
Inner Mongolia based on GIMMS NDVI3g data. J. Arid Land 9, 394–407. https://doi.
org/10.1007/s40333-017-0016-4.
Tong, X., Wang, K., Yue, Y., Brandt, M., Liu, B., Zhang, C., Liao, C., Fensholt, R., 2017b.
Quantifying the effectiveness of ecological restoration projects on long-term vege-
tation dynamics in the karst regions of Southwest China. Int. J. Appl. Earth Obs.
Geoinformation 54, 105–113. https://doi.org/10.1016/j.jag.2016.09.013.
van Dijk, A.I.J.M., Beck, H.E., Crosbie, R.S., de Jeu, R.A.M., Liu, Y.Y., Podger, G.M.,
Timbal, B., Viney, N.R., 2013. The Millennium Drought in southeast Australia
(2001–2009): Natural and human causes and implications for water resources, eco-
systems, economy, and society. Water Resour. Res. 49, 1040–1057. https://doi.org/
10.1002/wrcr.20123.
Verbesselt, J., Hyndman, R., Newnham, G., Culvenor, D., 2010a. Detecting trend and
seasonal changes in satellite image time series. Remote Sens. Environ. 114, 106–115.
https://doi.org/10.1016/j.rse.2009.08.014.
Verbesselt, J., Hyndman, R., Zeileis, A., Culvenor, D., 2010b. Phenological change de-
tection while accounting for abrupt and gradual trends in satellite image time series.
Remote Sens. Environ. 114, 2970–2980. https://doi.org/10.1016/j.rse.2010.08.003.
Verbesselt, J., Zeileis, A., Herold, M., 2012. Near real-time disturbance detection using
satellite image time series. Remote Sens. Environ. 123, 98–108. https://doi.org/10.
1016/j.rse.2012.02.022.
Wang, L., D’Odorico, P., Evans, J.P., Eldridge, D.J., McCabe, M.F., Caylor, K.K., King,
E.G., 2012a. Dryland ecohydrology and climate change: critical issues and technical
advances. Hydrol. Earth Syst. Sci. 16, 2585–2603. https://doi.org/10.5194/hess-16-
2585-2012.
Wang, L., Tian, F., Wang, Y., Wu, Z., Schurgers, G., Fensholt, R., 2018. Acceleration of
global vegetation greenup from combined effects of climate change and human land
management. Glob. Change Biol. https://doi.org/10.1111/gcb.14369.
Wang, T., Yan, C.Z., Song, X., Xie, J.L., 2012b. Monitoring recent trends in the area of
aeolian desertified land using Landsat images in China’s Xinjiang region. ISPRS J.
Photogramm. Remote Sens. 68, 184–190. https://doi.org/10.1016/j.isprsjprs.2012.
01.001.
Watts, L.M., Laffan, S.W., 2014. Effectiveness of the BFAST algorithm for detecting ve-
getation response patterns in a semi-arid region. Remote Sens. Environ. 154,
234–245. https://doi.org/10.1016/j.rse.2014.08.023.
Wessels, K.J., Prince, S.D., Malherbe, J., Small, J., Frost, P.E., VanZyl, D., 2007. Can
human-induced land degradation be distinguished from the effects of rainfall varia-
bility? A case study in South Africa. J. Arid Environ. 68, 271–297. https://doi.org/10.
1016/j.jaridenv.2006.05.015.
Wessels, K.J., van den Bergh, F., Scholes, R.J., 2012. Limits to detectability of land de-
gradation by trend analysis of vegetation index data. Remote Sens. Environ. 125,
10–22. https://doi.org/10.1016/j.rse.2012.06.022.
Willmott, C.J., Matsuura, K., 2001. Terrestrial air temperature and precipitation: monthly
and annual time series (1950–1999).
Wu, Z., Boke-Olén, N., Fensholt, R., Ardö, J., Eklundh, L., Lehsten, V., 2018a. Effect of
climate dataset selection on simulations of terrestrial GPP: highest uncertainty for
tropical regions. PLOS ONE 13, e0199383. https://doi.org/10.1371/journal.pone.
0199383.
Wu, X., Liu, H., Li, X., Ciais, P., Babst, F., Guo, W., Zhang, C., Magliulo, V., Pavelka, M.,
Liu, S., Huang, Y., Wang, P., Shi, C., Ma, Y., 2018b. Differentiating drought legacy
effects on vegetation growth over the temperate Northern Hemisphere. Glob. Change
Biol. 24, 504–516. https://doi.org/10.1111/gcb.13920.
Livneh Yanto, B., Rajagopalan, B., 2017. Development of a gridded meteorological da-
taset over Java island, Indonesia 1985–2014. sdata201772. Sci. Data 4. https://doi.
org/10.1038/sdata.2017.72.
Yengoh, G.T., Dent, D., Olsson, L., Tengberg, A.E., Iii, C.J.T., 2015. Main global NDVI
datasets, databases, and software. In: Use of the Normalized Difference Vegetation
Index (NDVI) to Assess Land Degradation at Multiple Scales, Springer Briefs in
Environmental Science Springer International Publishing, pp. 45–50. https://doi.org/
10.1007/978-3-319-24112-8_8.
Yin, H., Donat, M.G., Alexander, L.V., Sun, Y., 2015. Multi-dataset comparison of gridded
observed temperature and precipitation extremes over China. Int. J. Climatol. 35,
2809–2827. https://doi.org/10.1002/joc.4174.
Zhang, W., Brandt, M., Tong, X., Tian, Q., Fensholt, R., 2018. Impacts of the seasonal
distribution of rainfall on vegetation productivity across the Sahel. Biogeosciences
15, 319–330. https://doi.org/10.5194/bg-15-319-2018.
Zhang, Y., Song, C., Band, L.E., Sun, G., Li, J., 2017. Reanalysis of global terrestrial ve-
getation trends from MODIS products: browning or greening? Remote Sens Environ.
191, 145–155. https://doi.org/10.1016/j.rse.2016.12.018.
Zhu, X., Zhang, M., Wang, S., Qiang, F., Zeng, T., Ren, Z., Dong, L., 2015. Comparison of
monthly precipitation derived from high-resolution gridded datasets in arid Xinjiang,
central Asia. Quat. Int. Climate Change Mitigat. Strateg. Tienshan Mount. North.
Basins 358, 160–170. https://doi.org/10.1016/j.quaint.2014.12.027.
Zhu, Z., Piao, S., Myneni, R.B., Huang, M., Zeng, Z., Canadell, J.G., Ciais, P., Sitch, S.,
Friedlingstein, P., Arneth, A., Cao, C., Cheng, L., Kato, E., Koven, C., Li, Y., Lian, X.,
Liu, Y., Liu, R., Mao, J., Pan, Y., Peng, S., Peñuelas, J., Poulter, B., Pugh, T.A.M.,
Stocker, B.D., Viovy, N., Wang, X., Wang, Y., Xiao, Z., Yang, H., Zaehle, S., Zeng, N.,
2016. Greening of the Earth and its drivers. Nat. Clim. Change 6, 791. https://doi.
org/10.1038/nclimate3004.