ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32

Contents lists available at ScienceDirect

ISPRS Journal of Photogrammetry and Remote Sensing

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

Investigating the robustness of the new Landsat-8 Operational Land

Imager derived texture metrics in estimating plantation forest
aboveground biomass in resource constrained areas
Timothy Dube ⇑, Onisimo Mutanga
Discipline of Geography, School of Agricultural, Earth and Environmental Sciences, University of KwaZulu-Natal, P/Bag X01, Scottsville, Pietermaritzburg 3209, South Africa

a r t i c l e i n f o a b s t r a c t

Article history: The successful launch of the 30-m Landsat-8 Operational Land Imager (OLI) pushbroom sensor offers a
Received 21 November 2014 new primary data source necessary for aboveground biomass (AGB) estimation, especially in
Received in revised form 8 May 2015 resource-limited environments. In this work, the strength and performance of Landsat-8 OLI image
Accepted 3 June 2015
derived texture metrics (i.e. texture measures and texture ratios) in estimating plantation forest species
AGB was investigated. It was hypothesized that the sensor’s pushbroom design, coupled with the pres-
ence of refined spectral properties, enhanced radiometric resolution (i.e. from 8 bits to 12 bits) and
Keywords:
improved signal-to-noise ratio have the potential to provide detailed spectral information necessary
Estimation accuracy
Landsat-8 OLI texture metrics
for significantly strengthening AGB estimation in medium-density forest canopies. The relationship
Pushbroom sensor between image texture metrics and measurements of forest attributes can be used to help characterize
Signal-to-noise ratio complex forests, and enhance fine vegetation biophysical properties, a difficult challenge when using
Regression ensemble spectral vegetation indices especially in closed canopies. This study examines the prospects of using
Swath-width Landsat-8 OLI sensor derived texture metrics for estimating AGB for three medium-density plantation
Band texture ratios forest species in KwaZulu Natal, South Africa. In order to achieve this objective, three unique data
pre-processing techniques were tested (analysis I: Landsat-8 OLI raw spectral-bands vs. raw texture
bands; analysis II: Landsat-8 OLI raw spectral-band ratios vs. texture band ratios and analysis III:
Landsat-8 OLI derived vegetation indices vs. texture band ratios). The landsat-8 OLI derived texture
parameters were examined for robustness in estimating AGB using linear regression,
stepwise-multiple linear regression and stochastic gradient boosting regression models. The results of
this study demonstrated that all texture parameters particularly band texture ratios calculated using a
3  3 window size, could enhance AGB estimation when compared to simple spectral reflectance, simple
band ratios and the most popular spectral vegetation indices. For instance, the use of combined texture
ratios yielded the highest R2 values of 0.76 (RMSE = 9.55 t ha1 (18.07%) and CV-RMSE of 0.18); 0.74
(RMSE = 12.81 t ha1 (17.72%) and CV-RMSE of 0.08); 0.74 (RMSE = 12.67 t ha1 (06.15%) and CV-RMSE
of 0.06) and 0.53 (RMSE = 20.15 t ha1 (14.40%) and CV-RMSE of 0.15) overall for Eucalyptus dunii,
Eucalyptus grandis, Pinus taeda individually and all species, respectively. Overall, the findings of this study
provide the necessary insight and motivation to the remote sensing community, particularly in resource
constrained regions, to shift towards embracing various texture metrics obtained from the
readily-available and cheap multispectral Landsat-8 OLI sensor.
Ó 2015 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier
B.V. All rights reserved.

1. Introduction
Aboveground biomass (AGB) is an important indicator of
bio-physical processes related to forest dynamics. Information on
AGB is valuable in understanding and monitoring ecosystem
⇑ Corresponding author. Tel.: +27 735470648.
E-mail address: dube.timoth@gmail.com (T. Dube).
response and its contribution to the global carbon cycle and cli-
mate change for accurate greenhouse gas inventorying, and terres-
trial carbon accounting (Chinembiri et al., 2013; Gara et al., 2014;
Güneralp et al., 2014; Lu, 2006). Accurate and regularly repeated
AGB estimation at regional or local scales is critical in reducing
the uncertainty in estimating carbon sequestration and emissions
(Güneralp et al., 2014; Lu, 2006). Information on AGB is important
in greenhouse gas life cycle assessments and global climate change

http://dx.doi.org/10.1016/j.isprsjprs.2015.06.002
0924-2716/Ó 2015 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V. All rights reserved.
 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32
mitigation strategies. Moreover, accurate and repeated monitoring
of forest ecosystems status can also help in introducing appropri-
ate planning and monitoring conservation efforts. Therefore iden-
tifying affordable, timely and readily-available remote sensing
datasets, together with robust image processing techniques, is
essential to improve forest AGB estimates.
Although the application of remote sensing in forest AGB esti-
mation remains one of the most favourable and invaluable poten-
tial approaches, the issue of image resolution (i.e. spectral and
spatial properties) plays a major role in the accurate retrieval of
AGB estimates. Spectral vegetation indices normally rely on the
relationship between the red and near infra-red bands of the elec-
tromagnetic spectrum to augment the spectral influence from the
green vegetation, concurrently reducing contribution from the soil
background, sun zenith angle, sensor viewing angle, senesced veg-
etation and the atmosphere (Bannari et al., 1995; Carlson and
Ripley, 1997; Choudhury, 1987; Elvidge and Chen, 1995; Foody
et al., 2003; Huete, 1988; Lu, 2006; Tarpley et al., 1984; Teillet
et al., 1997; Tucker, 1979; Wiegand et al., 1991). However, previ-
ous studies show that the use of spectral vegetation indices com-
puted from medium-to-coarse spatial resolution multispectral
sensors in places with moderate or high canopy closure produce
poor results primarily due to saturation challenges and the existing
problem of multiple layering (Dube et al., 2014b; Godsmark, 2010;
Ingram et al., 2005; Lu, 2006; Mutanga et al., 2012; Mutanga and
Skidmore, 2004; Nichol and Sarker, 2011). Recent studies on
aboveground biomass estimation advocate for the use of texture
parameters instead of spectral vegetation indices (Kelsey and
Neff, 2014). This is because texture parameters correlate much bet-
ter with field datasets, since they allow for finer distinction of veg-
etation structural details (Eckert, 2012; Fuchs et al., 2009; Nichol
and Sarker, 2011; Sarker and Nichol, 2011).
Image-texture analysis technique can be applied in identifying
spectrally unique objects, based on a function of local variance in
the image and is also related to the spatial resolution and the size
of the dominant scene objects (Haralick et al., 1973). Image-texture
is an important source of information, capable of identifying differ-
ent aspects of forest stand structure including density, age, leaf
area index in medium-to-high spatial resolution images
(Champion et al., 2008). Some of the major strengths of applying
image-texture measures include: (1) the ability to simplify and
define complex forest canopy structures such as multiple layering
and even closed canopies; and (2) enhancing detection of fine veg-
etation biophysical properties; a complex challenge when using
spectral vegetation indices especially in closed canopies
(Champion et al., 2008). Moreover, earlier work (i.e. using high spa-
tial resolution sensors) demonstrates that texture measures have
the capability to enhance vegetation discrimination (Dekker,
2003; Podest and Saatchi, 2002; Shimada et al., 2014). However,
the relationship between medium spatial resolution sensors and
forest AGB has not yet been fully examined, especially when com-
pared to the use of other image properties, such as raw spectral
band information and vegetation indices. This is perhaps due to
the difficulty of identifying and selecting proper texture parame-
ters together with the optimal window-size and offset.
So far, the available studies have mostly demonstrated the
application of texture measures derived from high resolution sen-
sors such as synthetic aperture radar, Worldview-2, Advanced
Visible and Near Infrared Radiometer type-2 (AVNIR-2) and
SPOT-5 in estimating various tree-structural attributes namely:
stand age, leaf area index, stand density and biomass (Eckert,
2012; Nichol and Sarker, 2011; Pandey et al., 2010; Pinto et al.,
2012; Sarker and Nichol, 2011) and partly from medium resolution
multispectral sensors, such as the Landsat products (Cutler et al.,
2012; Kelsey and Neff, 2014). The majority of these studies have
shown that image-texture measures have the potential to
13
accurately improve the detection of different forest stand charac-
teristics. More importantly, image-texture measures have the
capability to enhance the discrimination of spatial information
and simultaneously enhancing AGB detection levels by increasing
the saturation levels that could not be measured with spectral veg-
etation indices (Eckert, 2012; Kuplich et al., 2005; Santos et al.,
2003; Sarker and Nichol, 2011; Vashum and Jayakumar, 2012; Xu
et al., 2011). For instance, Sarker and Nichol (2011) estimated
AGB with a plausible adjusted r-square value of 0.88 using texture
measures derived from the Advanced Land Observation Satellite,
AVNIR-2. On the other hand, Eckert (2012) obtained high carbon
estimates (R2-value 0.84 and relative RMSE of 6.8%) for degraded
forest, using WorldView-2 derived texture measures.
Thus, from the afore-mentioned studies, it is clear that image
textural measures have the potential to provide an attractive
opportunity for monitoring tree-structural attributes (i.e. AGB, leaf
area index and chlorophyll content etc.) in areas with moderate or
high canopy closure. However, the problem with the use of high
resolution images in resource-constrained regions of sub-Saharan
Africa, south-east Asia and South America is the associated cost,
limited availability and related-technical challenges. Currently,
the limited number of AGB studies in these regions, using high res-
olution images for instance, is probably an indication of the limita-
tions associated with these datasets (Adelabu and Dube, 2014;
Dube et al., 2014a; Dube and Mutanga, 2015; Dube et al., 2014b;
Dube et al., in-press; Koch, 2010). Nevertheless, considering the
inevitability of forest AGB estimation at regional to global scales,
the prospects of investigating the performance of image texture
parameters obtained from the multispectral sensors with
medium-spatial resolution, a large swath-width and a repeated
coverage is necessary in resource constrained regions.
The adoption of the free-and-readily available remotely sensed
datasets is critical in such resource-constrained regions. The
newly-launched pushbroom Landsat-8 OLI sensor with 30-m spa-
tial resolution is one such dataset. So far, the rich information con-
tained in this sensor has not yet been fully exploited in
understanding the distribution of AGB. This mainly due to the fact
that this sensor has just recently been launched and possibly due
to the purported challenges of saturation and the presence of a
few spectral bands, which can be used to compute simple spectral
vegetation indices. Among the different types of the
readily-available multispectral remote sensing sensors, a deriva-
tion of texture measures from this remotely sensed dataset (i.e.
pushbroom Landsat-8 OLI sensor) can improve the estimation of
regional to local scale forest AGB. The newly-launched Landsat-8
OLI sensor is hypothesized to display a great potential in estimat-
ing AGB across different scales especially in data-scarce areas. This
is due to the enhanced sensor’s sensitivity to different vegetation
properties (e.g. chlorophyll, leaf area index and AGB), the presence
of a large swath-width (i.e. 185-km) coupled with improved
signal-to-noise ratios (SNR). These sensor improvements enhance
the radiometric sensitivity which in turn improves spectral
strength to detect the most important forest structural properties
by thereby minimizing saturation problems. These saturation
problems were normally common with the prior Landsat 7 ETM+
products. Although the new Landsat-8 OLI sensor presents a more
attractive potential in vegetation studies than its counterparts (i.e.
MODIS, earlier Landsat products etc.), previous work using spectral
vegetation indices computed from the Landsat-8 OLI, obtained rel-
atively low AGB estimation accuracies (i.e. an R2 of 0.69 between
predicted and observed biomass and a moderately high average
root mean square error (RMSE) of 14.91 t ha1) (Dube and
Mutanga, 2015). This shows that accurate forest AGB estimation
is not only dependent on the Landsat-8 OLI dataset but also
requires advanced and robust image processing techniques such
as texture metrics. Therefore, texture metrics are perceived to
14 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32

improve the estimation accuracy of AGB when compared to the use
of vegetation indices (Dube and Mutanga, 2015; Nichol and Sarker,
2011; Sarker and Nichol, 2011). The main aim of this paper was
therefore to investigate the performance of texture metrics derived
from the Landsat-8 OLI in estimating AGB in the Clan Sappi Forests
in KwaZulu-Natal Province, South Africa.
2. Materials and methods
2.1. Study area
The study site was conducted at the sappi Clan plantation forest
(a paper and pulp company) in the midlands region of KwaZulu
Natal province, South Africa (Fig. 1). The Clan Sappi Forests area
is situated between Latitudes 29°240 46.7400 S, 29°170 45.9400 S and
Longitude 30°180 32.8900 E, 30°280 28.2100 E. The area is characterized
by moderately steep and undulating topography ranging from
644 to 1266 m above sea-level with predominantly deep-fertile
lithic soils such as the leptosols, cambisols, acrisols and lixisols
(Sappi, 1993; Scott and Lesch, 1997). The area experiences
sub-tropical climatic conditions, with the rainy season occurring
during the summer months (i.e. October–February) with annual
rainfall ranging from 730 to 1500 mm and an average annual tem-
perature of 21.7 C. Different species of eucalyptus and pine are
present within the area and these trees are mainly grown for pulp-
wood production.

Fig. 1. Study area location with the distribution and position of sampled plots.
 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32
Table 1
Measured aboveground biomass descriptive statistics (t ha1).
Species type # Of samples Min. Max. Avg.
E. dunii 63.00 33.24 96.49 52.86
E. grandis 65.00 106.03 225.07 170.30
P. taeda 53.00 137.11 298.04 206.07
2.2. Field-data inventorying
Field data collection was done from the 30th of July to 22nd
August, 2013 coinciding with Sappi annual routine-field surveys.
A grid-based systematic sampling technique, with ten meter radius
circular sample plots, distributed systematically (100 m) within
the stand was applied during the field data collection period
(Wessels and Kassier, 1985). Diameter-at-breast height (DBH)
and tree height (Ht) were measured from the three dominant
Clan Sappi Forests which are Eucalyptus dunii (ED), Eucalyptus gran-
dis (EG) and Pinus taeda (PT). During data collection, trees with
DBH P 5 cm were considered. The vector maps of the study, cour-
tesy of Sappi forest together with the aid of an expert forest spe-
cialist were used to navigate and identify the selected plots of
interest within the area. During the field campaign, plot centre
GPS locations (n = 181 plots) were also recorded with sub-meter
accuracy using the Trimble GeoXH 6000 series handheld Global
Position System (GPS). The measurement of DBH and Ht were done
using the Haglof Digitech Calliper and Vertex IV laser instrument
respectively.
To derive in situ AGB for individual 181 sampled plots, two
unique methods were chosen. Aboveground biomass was calcu-
lated using volume with species specific biomass expansion factors
(BEF) as well as the general functional group equations for the
three plantation species. The use of different techniques in com-
puting AGB in this study was driven by the prevailing unique struc-
tural and taxonomical properties of E. dunii, E. grandis and P. taeda
(Atta-Boateng and William, 1998; Henry et al., 2011; Lu, 2005;
Penman et al., 2003). Specifically, the general functional group
equation was developed in an area with rainfall of approx. 800–
1500 mm and a temperature range between 21 and 34 °C which
is comparable to climate conditions of the study area.
Furthermore this equation was developed based on a DBH range
of 0.6–56 cm, a characteristic almost identical to the midlands
region of KwaZulu Natal, South Africa. Specifically, AGB for E. dunii
and E. grandis species were derived using volume (m3/ha) together
with species specific BEF available in literature (Dovey, 2009). Tree
volume was computed based on mathematical function described
in Bredenkamp (2000). For P. taeda, individual tree DBH measure-
ments were converted to AGB via a general functional group equa-
tion (Penman et al., 2003) and subsequently summed across the
entire plot. Table 1 indicates descriptive statistics for individual
species AGB, with P. taeda having high biomass (298.04 t ha1)
when compared to the other species.
2.3. Landsat-8 OLI image acquisition and pre-processing
The Landsat-8 OLI sensor was launched on the 11th of February
2013 by the National Aeronautics and Space Administration and
the United States Geological Survey (NASA–USGS). Officially
Landsat-8 OLI sensor began normal operations on the 30th of
May 2013, and presents a number of key improvements in design
and spectral configuration. The novel and attractive spectral prop-
erties associated with the medium-resolution multispectral
Landsat-8 OLI sensor include a refined spectral range for certain
bands, (i.e. critical for improving the vegetation spectral responses
across the near-infrared) enhanced radiometric resolution (e.g.
15
from 8 to 12 bits: which significantly allows the characterization
of different vegetation structural characteristics) (El-Askary et al.,
Std dev. 2014; Pahlevan and Schott, 2013). Moreover, the changes in the
16.39 Landsat-8 OLI sensor design have also seen great advances in
29.94 signal-to-noise ratios (SNR), almost twice as good as Landsat 7
42.83 Enhanced Thematic Mapper plus: ETM+ (Irons et al., 2012). It is
because of these appealing unique sensor features or properties
that the utility of texture metrics derived from Landsat-8 OLI
image was tested in estimating plantation forest species AGB.
The Landsat-8 OLI image used in the study was acquired on the
3rd of August 2013 using the pushbroom or ‘‘along track’’
Operational Land Imager sensor. The Clan Sappi Forests area is cov-
ered by one Landsat-8 OLI tile with path/row: 168/80 and obtained
from the USGS Earth Resources Observation and Science (EROS)
Centre archive (http://earthexplorer.usgs/). The image was
acquired during a clear and sunny sky condition with sun azimuth
angle of 39.78° and sun elevation angle of 33.14°. The image spatial
resolution was about 30-m with seven spectral bands ranging from
0.43 to 2.29 lm (VNIR-SWIR: Visible Near-Infrared and Short
Wave Infrared) and bandwidths ranging between 0.3 to 18 lm.
The seven image wavebands correspond to band 1 – coastal/aero-
sol (0.43–0.45 lm), band – 2 blue (0.45–0.51 lm), band 3 – green
(0.53–0.59 lm), band 4 – red (0.64–0.67 lm), band 5 – near IR
(0.85–0.88 lm), band 6 – swir (1.57–1.65 lm) and band 7 – swir
(2.11–2.29 lm). The Landsat-8 OLI sensor has a temporal resolu-
tion of 16 days.
The Landsat-8 OLI image bands were converted from digital
number format (DN) to reflectance using ENVI 5.1 software, fol-
lowing the methodology summarized on the USGS website
(http://landsat.usgs.gov). Subsequently, the acquired Landsat-8
OLI image was atmospherically corrected using the MODTRAN
based on the Fast Line-of-sight Atmospheric Analysis of Spectral
Hypercube (FLAASH) radiative transfer algorithm (Matthew et al.,
2000; Perkins et al., 2005) which is embedded in the ENVI soft-
ware. The Landsat-8 OLI image was then geometrically corrected
(GCS Hartebeesthoek 1994) using 23 ground control points
(GCPs) of major features (e.g. roads and rock outcrops). The first
order polynomial function was used and a nearest-neighbour
resampling protocol was applied. The total transformation root
mean square error (RMSE) of less than a pixel was attained.
2.4. Landsat-8 OLI derived variables
Four simple spectral reflectances (i.e. blue, green, red, and nir)
as well as six simple spectral band ratios (i.e. blue/green, blue/red,
blue/nir, green/red green/nir, and red/nir) mostly applied in litera-
ture (Nichol and Sarker, 2011), were derived from the Landsat-8
OLI imagery. Five vegetation indices were also computed
(Table 2) based on the previous findings, which highlighted their
successful application as predictors for AGB estimation (Dube
and Mutanga, 2015; Güneralp et al., 2014). The texture parameters
used here have been explored in Eckert (2012), Kelsey and Neff
(2014) and Cutler et al. (2012). The spectral indices and texture
information applied in this study were extracted at each location,
and the pixel containing a sample plot was identified from the
Landsat-8 OLI imagery for analyses.
2.5. Landsat-8 OLI sensor texture metrics derivation
For this particular study, the Landsat-8 OLI image texture met-
rics were statistically derived using popular Gray Level
Co-occurrence Matrix (GLCM) texture algorithms calculated from
a relative displacement vector (d, h), which explains the spatial dis-
tribution of the level pairs separated by d in direction h (Haralick
et al., 1973). Although numerous image-texture metrics can be
generated from the most popular Gray Level Co-occurrence
16 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32

Table 2 compared with those derived using selected spectral vegetation

Selected spectral vegetation indices and image texture measures derived from indices (Table 2).
Landsat-8 OLI applied in the estimation of aboveground biomass.

Parameters References or equation 2.6. Statistical analysis

Single band reflectance
Band 2-blue, band 3-green, band 4-red, band 5-near-infrared For this study, AGB was estimated using linear regression, step-
Computed vegetation indices wise multiple regression and stochastic gradient boosting (SGB)
1. Moisture Stress Index (MSI) SWIR/NIR Rock et al. (1986) predictive methods. The choice and application of the SGB method
2. Normalized Difference Vegetation Index Rouse et al. (1974), Jordan in this study was based mainly on the algorithm’s technical abili-
(NDVI) (NIR  Red)/(NIR + Red) (1969)
3. Corrected Normalized Difference Vegetation Nemani et al. (1993)
ties over traditional parametric regression approaches which
  include (i) the non-parametric nature of the regressor, (ii) the
SWIRSWIRmin
Index (NDVIc) NIRRed
NIRþRed
1  SWIR max SWIRmin
potential to screen the most important predictors (i.e. capability
4. Normalized difference infrared index (NDII) Hardisky et al. (1983),
(NIR  SWIR 1)/(NIR + SWIR 1) Kimes et al. (1981)
to identify and select suitable predictor variables from the entire
5. Soil adjusted vegetation index (SAVI) Huete (1988) dataset); (iii) the ability to handle non-linear and hierarchical rela-
ð1þLÞðNIRþRedÞ
NIRþRedþL
; L = 0.5 tionships between predictor variables; (iv) acceptance of the miss-
Grey level co-occurrence matrix texture measures ing values and the ability to improve final model accuracy through
1. Angular Second moment (ASM) PN1 2
i;j¼0 iP i;j handling small perturbations in the training dataset; and (v) the
PN1
2. Contrast (CO)
i;j¼0 iP i;j ði  jÞ2 algorithm is resilient to overfitting (Carreiras et al., 2012; Dube
 
3. Correlation (CR) PN1 ðili Þðilj Þ et al., 2014b; Güneralp et al., 2014; Moisen et al., 2006). In addi-
i;j¼0 P i;j ðr2i Þðr2j Þ
tion, the stochastic gradient boosting algorithm, unlike other
4. Data Range (DR) maxfXg  minfXg; approaches, utilises the combined strengths of bagging (a machine
Where X ¼ x1 ; x2 ; . . . ; xn
5. Dissimilarity (DI) PN1 learning algorithm which has the capability to improve model sta-
i;j¼0 iP i;j ji  jj
PN1 bility and the final predictive accuracy) and boosting (an adaptive
6. Entropy (EN)
i;j¼0 iP i;j ð ln P i;j Þ
PN1 P i;j
method for integrating multiple simple models to provide
7. Homogeneity (HO)
i;j¼0 i 1þðijÞ2 improved predictive performance) methods (De’ath, 2007; Elith
PN1
8. Mean (ME) et al., 2008; Friedman, 2002; Leathwick et al., 2006; Moisen
i;j¼0 iP i;j
pffiffiffiffiffiffi
9. Standard deviation (std.dev) VA; et al., 2006). On the other hand, stepwise multiple linear regression
where was also applied in modeling aboveground biomass and the
P 2
VA ¼ N1 i;j¼0 iP ij ði  MEÞ obtained results were compared to those obtained using SGB
P
10. Variance (r2 )
2
i;j
ð xi;j l Þ regression algorithm. Also, statistical parameters such as tolerance
n1
(Tol.) and variable inflator factor (VIF) were applied to test for mul-
For Eqs. (3) and (4) the SWIR represents the SWIR I band of Landsat 8 OLI. ticollinearity effects. To demonstrate multicollinearity problems, a
value less than 0.10 for tolerance (Belsley, 1991) and a value
greater than 10 for VIF were considered as determinants (Sarker
Matrix, in this study ten were selected and these include mean, and Nichol, 2011).
variance, homogeneity, contrast, dissimilarity, entropy, second
moment, correlation, standard deviation and data range 2.7. Landsat-8 OLI data preparation for AGB estimation
(Baeza-Yates and Ribeiro-Neto, 1999; Blaschke et al., 2014;
Haralick, 1979; Haralick et al., 1973; Sarker and Nichol, 2011) Models of aboveground biomass were developed using the lin-
(Table 2). The choice of these texture metrics was prompted by ear regression, stepwise multiple linear regression and stochastic
their strength and successful application displayed in previous gradient boosting regression models with spectral/texture infor-
AGB estimation studies conducted in dense tropical forests using mation as the predictor variables and aboveground biomass as
high resolution image datasets (Bastin et al., 2014; Cutler et al., the response variable. Specifically, the remotely sensed predictor
2012; Eckert, 2012; Sarker and Nichol, 2011; Sarker et al., 2013). variables were (i) raw-spectral bands, (ii) raw-texture measures,
Although the selected texture metrics have demonstrated suc- (iii) spectral band ratios, (iv) vegetation indices and (v) texture
cessful application in most AGB studies, literature indicates that band ratios. Models of estimating aboveground biomass were con-
the identification and selection of the most appropriate texture ducted following three analytic steps:
measures require additional selection of the most suitable moving
window sizes (Bastin et al., 2014; Cutler et al., 2012; Eckert, 2012; i. Raw spectral-bands and spectral-band ratios derived from
Sarker and Nichol, 2011; Sarker et al., 2013). Thus, in order to the Landsat-8 OLI sensor were used to model AGB individu-
define the optimal window size required for accurate AGB estima- ally in linear regression. The raw spectral-bands and
tion, the selected texture measures were computed from four spectral-band ratios were combined and collectively used
Landsat-8 OLI spectral bands (i.e. band – 2 blue, band 3 – green, in modeling AGB using stepwise multiple linear regression
band 4 – red, and band 5 – near IR) using four window sizes (i.e. as well as using stochastic gradient boosting regression
3  3, 5  5, 7  7 and 9  9) based on four offsets ([1, 0], [1, 1], models. We then compared the strength of
[0, 1] and [1, 1]). Literature indicates that texture measures stepwise-multiple linear regression and stochastic gradient
derived from a small window size are sensitive to fine scale varia- boosting algorithm (using R2, RMSE, CV-RMSE and RMSE%)
tions in pixel brightness when compared to those derived using in estimating AGB and the best performing algorithm was
large window sizes (Kelsey and Neff, 2014). Thus, to determine a then applied on texture metrics.
suitable window size for accurate AGB retrieval, a window with a ii. Raw band texture measures (i.e. 10 texture metrics on 4
strong relationship between estimated and measured AGB was window sizes on four bands with some combination of off-
considered as indicated in literature (Kelsey and Neff, 2014). sets) from a single Landsat-8 OLI spectral band were individ-
Furthermore, to evaluate the strength of image texture measures ually tested in modeling AGB, using a stochastic gradient
(i.e. best performing texture measures from the texture metrics boosting model. Also, texture metrics were then combined
on four window sizes on four bands with a combination of differ- and collectively used in modeling AGB, using stochastic gra-
ent offsets) and band texture ratios, the results obtained were dient boosting regression model.
 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32
iii. Vegetation indices and band-texture ratios (b2/b3, b2/b4,
b2/b5, b3/b4, b3/b5, and b4/b5) generated from the
Landsat-8 OLI sensor were tested individually tested in mod-
eling AGB, using a stochastic gradient boosting model.
Similarly, texture metrics were then combined and collec-
tively used in modeling AGB, using stochastic gradient
boosting regression model.
In this study, field measured forest AGB values were used to
train and validate the model. Thus, the field measured forest AGB
dataset was randomly split into 70% (two-thirds of field data)
and 30% (one-third of field data) for model building and testing
(Kohavi, 1995; Riggins et al., 2009). The SGB model was parameter-
ized using three most important user-defined hyper-parameters
namely: (i) the learning rate (lr), critical for determining the contri-
bution of each tree to the final model; (ii) the tree complexity (tc),
which represents the number of independent variables interacting
to determine each split; and (iii) the number of regression trees
(nt): (models that relate a response to their predictors by recursive
binary splits) (Elith et al., 2008; Hastie et al., 2001; Leathwick et al.,
2006). For this work, the shrinkage rate (which determines the
contribution of each tree to the growing model) was set between
0.0001 and 0.1, whereas the tree complexity (controls whether
interactions are fitted) was set between 1 and 5. The bagging frac-
tion, which determines a fraction of the training dataset selected
randomly for computing each tree was set at 0.3 and 0.5 during
data analysis. To assess the tested regression model accuracies,
the coefficient of determination (R2), root mean square error
(RMSE), percentage root mean square error (RMSE%) and coeffi-
cient of variance of the root mean square error (RMSE-CV) were
computed using the test dataset (Eqs. ()()()1–3).
rffiffiffi n
1X
RMSE ¼ ^ i Þ2
ðy  y
n i¼1 i
where RMSE is the Root Mean Square Error; n is the number of
^i is the estimated values,
observed values; yi is the observed value, y
 mean of the observed values.
y estimating AGB for individual forest species and all species inves-
qffiffiP
1 n 2
n i¼1 ðyi
^i Þ
y
RMSEð%Þ ¼  100

y
where RMSE% is the Percentage Root Mean Square Error; n is the
^i is the esti-
number of observed values; yi is the observed value, y
 mean of the observed values.
mated values, y
 
RMSE
RMSE  CV ¼

y
where RMSE-CV is the Coefficient of Variance of the Root Mean
 mean of the
Square Error; RMSE is the Root Mean Square Error; y
observed values.
3. Results
3.1. Analysis I: Landsat-8 OLI raw spectral-bands vs. raw texture bands
A few texture parameters derived using a 3  3 window size
and an offset [0, 1] were identified as the best variables for the
accurate estimation of AGB (Table 3). These include mean, entropy,
correlation, dissimilarity, homogeneity, and variance. Raw
Landsat-8 bands yielded weaker AGB estimates across all tree spe-
cies (i.e. ED, EG, PT and all species data (combined data)) when
compared to estimates obtained using raw Landsat-8 texture mea-
sures (Table 3 and Fig. 2). Despite poor performance from the raw
Landsat-8 bands, the near-infrared and the green bands yielded
slightly higher AGB estimates for ED, EG, PT and all species data
combined. Using texture parameters derived from the NIR band,
17
an R2 value of 0.29; RMSE of 19.23 t ha1 and CV-RMSE of 0.37
were obtained for ED. Similarly, using texture parameters derived
from the green band, an R2 value of 0.24, RMSE of 32.26 t ha1
and CV-RMSE of 0.62 were obtained for ED species. A similar per-
formance was observed for E. grandis, P. taeda and all species.
Comparatively, raw texture measures derived from Landsat-8 OLI
yielded overall better AGB estimates than any individual spectral
band. For instance, high average R2 and RMSE and CV-RSME were
attained from the use of Landsat-8 OLI raw texture measures
derived from the NIR, green and blue bands respectively. In con-
trary, the use of the step-wise multiple linear regression together
with a stochastic gradient boosting algorithm shows that the latter
performed better, yielding the highest R2 (0.51), lowest RMSE
(52.31 t ha1) and RMSE% (37.53%) for all species dataset combined
(Table 3).
Moreover, it can be observed that when all Landsat-8 OLI spec-
tral bands (i.e. combined raw Landsat-8 OLI bands) were used, AGB
estimates improved significantly when compared to the use of
individual raw Landsat-8 OLI bands. Overall, the combined
Landsat-8 OLI raw texture measures yielded plausible AGB esti-
mates for ED, EG, PT and all species (Fig. 2a–c and Table 3).
Although, combined raw Landsat-8 OLI bands produced lower
AGB estimates, their results were better when compared with
those obtained using individual bands (Fig. 2d).
3.2. Analysis II: Landsat-8 OLI raw spectral – band ratios vs. texture
band ratios
This study finds that texture based models derived using a
stochastic gradient boosting algorithm display enhanced accuracy
in AGB estimation relative to the use of spectral bands and spectral
vegetation indices (Fig. 3a–d). Although the results derived using
ð1Þ both raw spectral band ratios and texture ratios appear to be
promising, raw spectral band ratios based AGB estimates generally
yielded low accuracy (Table A1 in Appendix). For instance, weaker
R2 values were obtained from the use of raw spectral band ratios in
tigated. Fig. 3a and b show that blue/nir, green/nir and red/nir
raw spectral band ratios had slightly higher R2 values especially
ð2Þ when compared to the use of single raw spectral bands.
However, a comparison of AGB estimates between raw-spectral
band ratios and raw-texture ratios indicate improved performance
from the latter (Fig. 3a and b). Overall, Landsat-8 OLI sensor
texture-ratios derived from the following band combinations:
band 2/band3, band 3/ band 5, and band 4/ band 5 had a stronger
ð3Þ correlation between predicted and observed biomass than other
methods investigated (see Fig. 4).
The improved and higher R2 values of 0.49 (RMSE = 20.45 t ha1
and CV-RMSE of 0.39), 0.39 (RMSE = 35.60 t ha1 and CV-RMSE of
0.21), 0.52 (RMSE = 21.85 t ha1 and CV-RMSE of 0.11) were
obtained for ED, EG and PT species AGB estimates, respectively
from red/nir texture ratio. A similar performance from the red/nir
texture ratio was observed in estimating all species AGB, where an
R2 value 0.27 (RMSE = 35.50 t ha1 and CV-RMSE of 0.26) was
attained (Fig. 3a and b and Table 4). Texture based AGB estimates
demonstrate a substantial enhancement when compared with the
best findings obtained using Landsat-8 OLI raw spectral and band
ratios. A further demonstration of good performance in AGB esti-
mation was also observed from the results obtained using the com-
bined Landsat-8 OLI texture ratios (Fig. 3a and b). The use of
combined texture ratios yielded even higher R2 values of 0.76
(RMSE = 9.55 t ha1 and CV-RMSE of 0.18); 0.74 (RMSE = 1
2.81 t ha1 and CV-RMSE of 0.08); 0.74 (RMSE = 12.67 t ha1 and
CV-RMSE of 0.06) and 0.53 (RMSE = 20.15 t ha1 and CV-RMSE of
0.15) overall for ED, EG, PT and all species dataset respectively. It
is also important to note that overall, the 3  3 window size and
18 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32

Table 3
Aboveground biomass estimates derived using Landsat-8 OLI spectral bands and the best performing texture parameters from a 3  3 window size and an offset [0, 1] using linear,
stepwise-multiple linear and stochastic gradient boosting regression models.

Model Plant. species Image data Variable name R2 RMSE t ha1 CV-RMSE Tol. VIF
and window size (RMSE%)
Linear E. dunii Blue – 0.01 39.64 (74.99%) 0.75 0.99 1.01
Linear Green – 0.18 36.79 (69.60%) 0.70 0.82 1.22
Linear Red – 0.09 40.31 (76.26%) 0.76 0.91 1.10
Linear NIR – 0.15 38.97 (73.72%) 0.74 0.85 1.18
SMLR All bands – 0.19 35.12 (66.44%) 0.66 0.81 1.23
SGB All bands – 0.25 29.68 (56.15%) 0.56 0.75 1.33
Linear E. grandis Blue – 0.01 78.02 (45.81%) 0.46 1.00 1.00
Linear Green – 0.21 61.93 (36.37%) 0.36 0.79 1.27
Linear Red – 0.02 74.25 (43.60%) 0.44 0.98 1.02
Linear NIR – 0.22 37.40 (21.96%) 0.22 0.78 1.28
SMLR All bands – 0.26 46.52 (27.32%) 0.27 0.74 1.35
SGB All bands – 0.33 55.46 (32.57%) 0.33 0.67 1.49
Linear P. taeda Blue – 0.01 78.53 (38.11%) 0.38 0.99 1.01
Linear Green – 0.19 63.11 (30.63%) 0.31 0.81 1.23
Linear Red – 0.01 74.89 (36.34%) 0.36 0.99 1.01
Linear NIR – 0.13 59.47 (28.86%) 0.29 0.87 1.15
SMLR All bands – 0.20 46.52 (22.57%) 0.23 0.80 1.25
SGB All bands – 0.34 38.29 (27.37%) 0.27 0.66 1.52
Linear All species Blue – 0.01 80.73 (57.91%) 0.58 0.99 1.01
Linear Green – 0.12 71.34 (51.18%) 0.51 0.88 1.14
Linear Red – 0.02 84.13 (60.39%) 0.60 0.98 1.02
Linear NIR – 0.27 69.66 (49.79%) 0.50 0.73 1.37
SMLR All bands – 0.24 61.73 (44.29%) 0.44 0.76 1.32
SGB All bands – 0.51 52.31 (37.53%) 0.37 0.49 2.04
SGB E. dunii b2 texture 1, 5, 6 (3  3) 0.20 36.34 (68.75%) 0.69 0.80 1.25
SGB b3 texture 6, 1, 8, 3 (3  3) 0.24 32.26 (61.03%) 0.62 0.76 1.32
SGB b4 texture 6, 1, 5 (3  3) 0.18 29.37 (55.56%) 0.56 0.82 1.22
SGB b5 texture 1, 8, 6,2 (3  3) 0.29 19.23 (36.38%) 0.37 0.71 1.41
SGB All data 1, 6, 8, 5, 3, 2 (3  3) 0.68 10.29 (19.47%) 0.20 0.32 3.13
SGB E. grandis b2 texture 1, 6, 5 (3  3) 0.35 37.35 (21.93%) 0.22 0.65 1.54
SGB b3 texture 6, 1, 8, 3 (3  3) 0.25 33.42 (19.62%) 0.20 0.75 1.33
SGB b4 texture 6, 1, 5 (3  3) 0.19 31.12 (18.27%) 0.18 0.81 1.23
SGB b5 texture 8, 1, 2,6 (3  3) 0.46 29.30 (17.20%) 0.17 0.54 1.85
SGB All data 2, 6, 8, 1, 3, 5 (3  3) 0.67 12.13 (07.12%) 0.07 0.33 3.03
SGB P. taeda b2 texture 5, 6, 1 (3  3) 0.23 41.55 (20.16%) 0.21 0.77 1.30
SGB b3 texture 6, 1, 8, 3 (3  3) 0.33 35.23 (17.10%) 0.17 0.67 1.49
SGB b4 texture 6, 5, 1 (3  3) 0.27 40.58 (19.69%) 0.20 0.73 1.37
SGB b5 texture 8, 1, 6,2 (3  3) 0.43 21.19 (10.28%) 0.10 0.57 1.75
SGB All data 1, 2, 3, 5, 8, 6 (3  3) 0.65 15.77 (07.65%) 0.08 0.35 2.86
SGB All species b2 texture 1, 6, 5 (3  3) 0.23 59.94 (43.00%) 0.43 0.77 1.30
SGB b3 texture 6, 3, 8, 1 (3  3) 0.18 63.15 (45.30%) 0.46 0.82 1.22
SGB b4 texture 5, 1, 6 (3  3) 0.07 71.03 (50.96%) 0.51 0.93 1.08
SGB b5 texture 1, 2, 6, 8 (3  3) 0.42 29.03 (20.83%) 0.21 0.58 1.72
SGB All data 1, 2, 8, 5, 3, 6 (3  3) 0.55 26.33 (18.89%) 0.19 0.45 2.22

SMLR = stepwise multiple-linear regression; SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second
moment, 8-correlation, 9-standard deviation and 10-data range.

offset [0, 1], yielded the best results when compared to the results gradient boosting model determined from Landsat-8 OLI texture
obtained using other window sizes (i.e. 5  5, 7  7 and 9  9) and band ratios (mean, entropy, dissimilarity, correlation, homogene-
the other offsets which include [1, 1], [0, 1] and [1, 1] (Appendix ity, and variance) computed using the optimal window size of
Tables A2–A16). 3  3 and an offset of [0, 3].

3.3. Analysis III: Landsat-8 OLI derived vegetation indices vs. texture
band ratios
This work finds that Landsat-8 OLI derived texture ratios per-
formed better than the selected vegetation indices, in terms of
R2; RMSE and CV-RMSE (Fig. 5a). For instance, the use of combined
texture ratios resulted in high R2 values of 0.76 (RMSE = 9.55 t ha1
and CV-RMSE of 0.18) for E. dunii whereas based on the five
selected vegetation indices, weaker results were obtained, produc-
ing an R2 of 0.71 and a RMSE of 10.66 t ha1. Similar performance
was also observed for the other two species i.e. E. grandis and P.
taeda as well as for all species datasets combined. Fig. 5b shows
the AGB map obtained using the best performing stochastic
4. Discussion
The present study aimed to investigate the performance of tex-
ture metrics derived from the newly-launched Landsat-8 OLI sen-
sor in estimating plantation forest species AGB in the midlands
region of KwaZulu Natal, South Africa. Specifically, this research
investigated whether the various texture metrics obtained from
the 30-m Landsat-8 OLI sensor with refined-spectral properties
as well as the improved signal-to-noise ratio have the potential
to enhance AGB estimation when compared to other methods
investigated i.e. spectral bands, bands ratios, vegetation indices.
This research shows that texture metrics (i.e. mean, entropy,
dissimilarity, correlation, homogeneity, and variance) derived from
 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32 19

Fig. 2. A comparison between the strength of Landsat-8 OLI sensor spectral bands and the best performing texture parameters derived using a 3  3 window size and an
offset [0, 1] in aboveground biomass estimation based on the three modeling techniques (i.e. linear regression, multiple linear regression, SGB algorithm). (a) Represents E.
dunii, (b) stands for E. grandis, (c) stands for P. taeda and (d) represents all species data combined.

the 30-m Landsat-8 OLI multispectral sensor using a 3  3 window
size and an offset of [0, 1] overall enhanced AGB estimation accu-
racy when compared to the use of (1) raw-spectral bands, (2)
raw-spectral band ratios and (3) spectral vegetation indices. The
findings of this study are similar to those available in literature,
which demonstrate the above-mentioned texture parameters as
critical for AGB estimation (Kelsey and Neff, 2014; Nichol and
Sarker, 2011; Sarker and Nichol, 2011). For example, these studies
have shown that mean, entropy, dissimilarity, correlation, homo-
geneity, and variance texture parameters are capable of enhancing
aboveground biomass estimation accuracies. Also, the study by
Kelsey and Neff (2014) using Landsat data, has demonstrated that
a 3  3 window size and an offset of [0, 1] results in improved AGB
estimates when compared to other offsets and larger window
sizes. High AGB estimation accuracies in terms of R2 and RMSE
and CV-RMSE were obtained for the three different plantation for-
est species (ED, EG and PT) under study. Also, similar performances
from texture band ratios were observed in estimating all species
AGB from all the three analyses stages (i.e. analysis I, II and III).
Improved performance and estimation accuracy from the 30-m
multispectral landsat-8 OLI sensor can be associated with the
advanced image processing methodologies (e.g. texture retrieval)
adopted and the advanced machine learning algorithms applied
in data analysis (Dube and Mutanga, 2015; Güneralp et al., 2014;
Kelsey and Neff, 2014). These results are in line with the findings
from literature which reported good AGB predictive accuracies
obtained from the use of texture metrics such as entropy, mean
and correlation based on a 3  3 window size and an offset of
[0, 1] (Kelsey and Neff, 2014). Literature further shows that
machine learning algorithms, such as stochastic gradient have
the potential to significantly improve AGB estimates. (De’ath,
2007; Moisen et al., 2006). This is mainly because these algorithms
utilise the combined strengths of bagging and boosting which have
the capability to improve model stability and the final predictive
accuracy.
Furthermore, the results from this work have shown that tex-
ture metrics derived from medium multispectral sensors (10–
100 m pixel size) offers invaluable opportunities for improving
and better understanding forest AGB in (1) areas with tall forest
and complex canopies structure, (2) areas associated with limited
resources and (3) limited access to high resolution satellite, such
as hyperspectral, LiDAR and radar etc. The plausible performance
and strength of texture metrics in improving AGB estimates is in
line with the findings from other studies done elsewhere using
other optical sensors, such as synthetic aperture radar (SAR) and
Worldview-2, among others (Bastin et al., 2014; Champion et al.,
2013; Cutler et al., 2012; Eckert, 2012; Kelsey and Neff, 2014;
Ploton et al., 2013; Sarker and Nichol, 2011; Singh et al., 2014).
Credible performance displayed by the use of Landsat-8 OLI
derived texture metrics may be attributed to the presence of cru-
cial and sensitive vegetation information as well as related bio-
physical properties such as tree age, leaf area index and most
importantly biomass (Champion et al., 2008; Fuchs et al., 2009),
a complex challenge when using spectral vegetation indices.
Kelsey and Neff (2014) state that texture analysis is efficient in
addressing saturation problems associated with vegetation indices
20 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32

Fig. 3. A comparison between the strength of Landsat-8 OLI sensor spectral band ratios and texture measure band ratios in aboveground biomass estimation using a
stochastic gradient boosting algorithm. (a) Represents E. dunii, (b) stands for E. grandis, (c) stands for P. taeda and (d) represents all species data combined. The exact texture
metrics used in each ratio calculation are presented in Table 3.

Fig. 4. Scatterplots of the predicted vs. observed aboveground biomass of the best models. (a)–(d) represent E. dunii, E. grandis, P. taeda and all species datasets combined
based Landsat 8 OLI texture ratios derived using a 3  3 window size, an offset of [0, 1] and stochastic gradient boosting algorithm.
 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32 21

Table 4
Aboveground biomass estimates derived using Landsat-8 OLI band texture ratios from a 3  3 window size and an offset [0, 1] using stochastic gradient boosting regression
models.

Model Plant. species Image data Selected texture ratios R2 RMSE t ha1 (RMSE%) CV-RMSE Tol. VIF
SGB E. dunii b2/b3 1, 6, 2 0.24 32.28 (61.07%) 0.62 0.76 1.32
b2/b4 6, 5, 8 0.42 30.11 (56.96%) 0.58 0.58 1.72
b2/b5 6, 1 0.37 36.57 (69.18%) 0.70 0.63 1.59
b3/b4 8, 6, 2 0.39 35.63 (67.40%) 0.68 0.61 1.64
b3/b5 1, 6, 8 0.27 27.92 (52.82%) 0.53 0.73 1.37
b4/b5 6, 7, 5 0.49 20.45 (38.69%) 0.39 0.51 1.96
All data 4, 1, 6, 5, 8, 7, 2 0.76 09.55 (18.07%) 0.18 0.24 4.17
SGB E. grandis b2/b3 1, 2, 6 0.25 38.40 (24.27%) 0.23 0.75 1.33
b2/b4 6, 8, 5 0.15 41.34 (19.55%) 0.24 0.85 1.18
b2/b5 6, 1 0.33 33.29 (21.60%) 0.20 0.67 1.49
b3/b4 1, 8, 6 0.37 36.78 (22.45%) 0.22 0.63 1.59
b3/b5 1, 6, 8 0.26 38.23 (20.90%) 0.22 0.74 1.35
b4/b5 6, 5, 7 0.39 35.60 (07.52%) 0.21 0.61 1.64
All data 6, 4, 1, 8, 5, 2, 7 0.74 12.81 (17.72%) 0.08 0.26 3.85
SGB P. taeda b2/b3 1, 6, 2 0.30 36.52 (33.93%) 0.19 0.70 1.43
b2/b4 6, 5, 8 0.05 69.92 (33.93%) 0.36 0.95 1.05
b2/b5 6, 1 0.36 34.57 (16.78%) 0.18 0.64 1.56
b3/b4 8, 6, 2 0.24 37.69 (18.29%) 0.19 0.76 1.32
b3/b5 1, 6, 8 0.25 39.74 (19.28%) 0.20 0.75 1.33
b4/b5 6, 7, 5 0.52 21.85 (10.60%) 0.11 0.48 2.08
All data 6, 4, 1, 8, 5, 2, 7 0.74 12.67 (06.15%) 0.06 0.26 3.85
SGB All species b2/b3 1, 6, 2, 4 0.16 43.24 (30.91%) 0.31 0.84 1.19
b2/b4 6, 5, 8 0.21 47.23 (33.76%) 0.34 0.79 1.27
b2/b5 6, 1 0.12 67.96 (48.58%) 0.49 0.88 1.14
b3/b4 8, 6, 2 0.11 68.43 (48.92%) 0.50 0.89 1.12
b3/b5 1, 6, 8 0.24 51.29 (36.66%) 0.37 0.76 1.32
b4/b5 6, 7, 5 0.27 35.50 (25.38%) 0.26 0.73 1.37
All data 1, 6, 8, 5, 2, 7, 4 0.53 20.15 (14.40%) 0.15 0.47 2.13

SMLR = stepwise multiple-linear regression; SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second
moment, 8-correlation, 9-standard deviation and 10-data range.

Fig. 5. (a) A comparison between the strength of Landsat-8 OLI sensor spectral vegetation indices and band texture ratios in aboveground biomass estimation and (b) best
performing Landsat-8 OLI image texture-derived aboveground biomass map.

when mapping biomass especially in dense canopies as it corre-
lates more with AGB and carbon than spectral parameters. The
other reason for the unique performance of texture measures in
estimating AGB can be explained by their capability to detect vary-
ing forest canopy structural characteristics as well as the inherent
sensitivity to the spatial aspects of the canopy than spectral reflec-
tance or band ratios (Eckert, 2012).
Additionally, improvements in AGB estimation were obtained
from the use of simple texture ratios derived from a 30-m
Landsat-8 OLI sensor. So far, no study, to the best of our knowledge,
has used texture ratios derived from medium spatial resolution
satellite data (e.g. pixel size greater 10-m) to estimate AGB. The
greater performance from optical Landsat-8 OLI image texture
metrics when compared to spectral vegetation indices can be
attributed to a number factors and these include: (1) the simplifi-
cation of complex forest canopy structures and the stronger corre-
lation between texture measures and biophysical properties, such
as leaf area index and AGB (Barbosa et al., 2014; Eckert, 2012;
22 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32
Fuchs et al., 2009; Nichol and Sarker, 2011; Sarker and Nichol,
2011); and (2) the capability to minimize atmospheric effects
and canopy geometry, sensor view angle and sun angle by this sen-
sor (Fan et al., 2014; Franklin et al., 1996; Laurent et al., 2014;
Mousivand et al., 2014; Wulder et al., 1998). In addition, the great
performance from texture ratios in modeling AGB can be attributed
to the fact that these ratios combine the strengths of both texture
and ratios which is critical in improving AGB estimation accuracies
(Nichol and Sarker, 2011; Sarker and Nichol, 2011). Also the good
performance of texture metrics derived from Landsat-8 OLI sensor
can be attributed to its push-broom design, which is associated
with an improved signal-to-noise ratio; high radiometric resolu-
tion of 12-bits making it (1) more sensitivity and robust in detect-
ing crucial vegetation structural attributes (El-Askary et al., 2014;
Pahlevan and Schott, 2013) and (2) the narrow or refined
near-infrared spectral band that prevents the absorption effect of
water vapour (e.g. at 0.825 lm) thereby permitting accurate sur-
face spectral detection while minimizing satellite spectral satura-
tion problems (Lu, 2006).
One of the observations from this study is that when forests
have complex canopies (e.g. high biomass levels) (Basuki et al.,
2011); vegetation indices derived from medium spatial resolution
sensors become less useful in estimating AGB when compared to
the use of texture metrics (Mutanga and Skidmore, 2004;
Thenkabail et al., 2000). Despite the poor performance of vegeta-
tion indices in this study, other studies indicate their successful
application in estimating AGB in simple structured forests with
less complex canopies where AGB and LAI are low (e.g. less than
100% vegetation cover) (Darvishzadeh et al., 2011; Englhart et al.,
2012; Gallardo-Cruz et al., 2012; Goh et al., 2014; Vashum and
Jayakumar, 2012). The other possible explanation for the poor per-
formance from the spectral vegetation indices computed in this
study may be as a result of the presence of tall forests; hence sha-
dow effects might have contributed immensely to the spectral
reflectances captured by the sensors resulting in their poor perfor-
mance. Most of these vegetation indices have been designed based
on simple and less complex forest canopies hence their inability to
estimate AGB in medium-to-dense forests. Moderately weaker per-
formance by raw spectral bands and vegetation indices derived
from Landsat-8 OLI sensor observed in this study this may also
be attributed to saturation problems associated with
medium-to-coarse resolution multispectral sensors when estimat-
ing AGB in medium-to-high density vegetation canopies. Literature
shows that vegetation indices e.g. normalized difference vegeta-
tion indices derived mainly using the red and near infrared bands
of the electromagnetic spectrum suffer immensely from saturation
problems as they asymptotically reach a saturation level beyond a
certain density of AGB and/or leaf area index (Adam et al., 2014;
Gao et al., 2000; Mutanga and Skidmore, 2004; Sarker and
Nichol, 2011; Thenkabail et al., 2000, 2004). These studies have
greatly demonstrated that the use of spectral vegetation indices,
such as the normalized difference vegetation index have limited
significance in estimating AGB in areas with high vegetation cover
during the peak of periods.
Finally, improvement in AGB estimation using the 30-m
Landsat-8 OLI multispectral sensor derived texture metrics can
be linked with the strength of advanced machine learning algo-
rithms when compared to the use of stepwise multiple linear
regression methods. Previous research has precisely demonstrated
the strength and capabilities of advanced machine learning tech-
niques (i.e. stochastic gradient boosting (SGB) and the most popu-
lar random forest (RF) regression ensembles) in simplifying and
enhancing AGB estimation accuracy when compared to the use
of simple multiple linear methods (Adam et al., 2014; Carreiras
et al., 2012; Dube et al., 2014b; Güneralp et al., 2014). One of the
most important strengths of SGB regression ensemble is its robust-
ness in handling inaccurate training data, outliers, missing and
unbalanced datasets (Dube et al., 2014b; Elith et al., 2008), and
the capacity to handle, ascertain and choose the most crucial pre-
dictor variables from a huge amount of predictors (Carreiras et al.,
2012; Dube et al., 2014b).
5. Conclusion
This research investigated the performance and strength of tex-
ture metrics calculated from the newly-launched pushbroom
designed Landsat-8 OLI sensor in estimating medium-density
AGB for plantation forest species in the midlands region of
KwaZulu Natal, South Africa.
The findings of this work have demonstrated that:
1. Texture metrics yielded more accurate AGB estimates when
compared to the use of Landsat-8 OLI sensor derived spectral
vegetation indices.
2. AGB for medium-density E. dunii, E. grandis, P. taeda and all spe-
cies data combined can be accurately estimated from using the
newly-launched 30-m Landsat-8 OLI multispectral sensor
derived texture ratios. So far, no study to the best of our knowl-
edge has examined the strength of texture ratios derived from
optical remote sensing sensors with a pixel size greater than
10-m for AGB estimation.
3. Texture parameters derived from the newly-launched
Landsat-8 OLI dataset provides an important tool for the cre-
ation, mapping and continuous updating AGB maps which is
critical for well-informed land management purposes.
4. Stochastic gradient boosting regression ensemble has proven
useful and more dependable in enhancing AGB for
medium-density plantation forest species based on medium
spatial resolution multispectral Landsat-8 OLI sensor derived
texture metrics.
Overall, this study presents a successful and effective applica-
tion of texture metrics derived from the newly-launched
push-broom Landsat-8 OLI multispectral data and stochastic gradi-
ent boosting algorithm in estimating AGB in resource constrained
regions. The results of this study can possibly aid in understanding
the contribution of forest ecosystems in regulating atmospheric
carbon. Moreover, the results from this study can also provide
the required insight and inspiration to forest managers, environ-
mental and remote sensing experts to focus their attention in
terms of research, in coming up with possible, reliable and less
costly approaches for deriving accurate estimates of AGB particu-
larly in resource-constrained areas from the readily available med-
ium resolution sensors.
Acknowledgements
The authors are thankful to National Aeronautics and Space
Administration (NASA) for providing the newly-launched
Landsat-8 OLI images of the study area. Our word of thanks is also
extended to ACCESS-SA: Applied Center for Climate and Earth
Systems Science in South Africa, for funding this research under
the theme ‘‘Land Use and Land Cover Change’’. Sappi Forest
Company is thanked for providing field datasets. We would like
to thank Dr Riyad Ismail in helping with the implementation of
SGB and RF algorithms in R. Authors would also like to thank
Khoboso E. Seutloali and Lucky S. Nkomo for helping with field
work data collection.
 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32 23

Appendix A

See Tables A1–A16

Table A1
Aboveground biomass estimates derived using Landsat-8 OLI spectral band ratios based on a stochastic gradient boosting algorithm.

Model Plant. species Band ratio R2 RMSE t ha1 CV-RMSE Tol. VIF
SGB E. dunii b2/b3 0.23 39.42 (74.57%) 0.75 0.77 1.30
b2/b4 0.06 40.23 (76.11%) 0.76 0.94 1.06
b2/b5 0.24 39.70 (75.10%) 0.75 0.76 1.32
b3/b4 0.20 37.58 (71.09%) 0.71 0.80 1.25
b3/b5 0.25 34.16 (64.62%) 0.65 0.75 1.33
b4/b5 0.23 33.32 (63.03%) 0.63 0.77 1.30
All data 0.36 29.17 (55.18%) 0.55 0.64 1.56
SGB E. grandis b2/b3 0.19 52.33 (30.73%) 0.31 0.81 1.23
b2/b4 0.06 58.45 (34.32%) 0.34 0.94 1.06
b2/b5 0.25 49.36 (28.98%) 0.29 0.75 1.33
b3/b4 0.20 45.90 (26.95%) 0.27 0.80 1.25
b3/b5 0.24 44.00 (25.84%) 0.26 0.76 1.32
b4/b5 0.33 43.97 (25.82%) 0.26 0.67 1.49
All data 0.53 39.88 (23.42%) 0.23 0.47 2.13
SGB P. taeda b2/b3 0.15 52.61 (25.53%) 0.26 0.85 1.18
b2/b4 0.01 59.83 (29.03%) 0.29 0.98 1.02
b2/b5 0.22 51.28 (24.88%) 0.25 0.78 1.28
b3/b4 0.23 50.11 (24.32%) 0.24 0.77 1.30
b3/b5 0.22 44.57 (21.63%) 0.22 0.78 1.28
b4/b5 0.20 44.28 (21.49%) 0.21 0.80 1.25
All data 0.36 32.55 (15.80%) 0.16 0.64 1.56
SGB All species b2/b3 0.12 73.21 (52.33%) 0.52 0.88 1.14
b2/b4 0.01 71.63 (51.20%) 0.51 0.98 1.02
b2/b5 0.03 77.29 (55.25%) 0.55 0.97 1.03
b3/b4 0.07 75.43 (53.92%) 0.54 0.93 1.08
b3/b5 0.17 63.57 (45.44%) 0.45 0.83 1.20
b4/b5 0.18 63.10 (45.11%) 0.45 0.82 1.22
All data 0.24 50.66 (36.21%) 0.36 0.76 1.32

Table A2
Aboveground biomass estimates derived using Landsat-8 OLI band texture ratios from a 3  3 window size and an offset [1, 1] using stochastic gradient boosting regression
models.

Model Plant. species Image data Selected texture ratios R2 RMSE t ha1 CV-RMSE Tol. VIF
SGB E. dunii b2/b3 5, 2, 1 0.20 35.79 (67.71%) 0.68 0.80 1.25
b2/b4 3, 6, 4 0.36 34.28 (64.85%) 0.65 0.64 1.56
b2/b5 6, 1, 3 0.35 40.31 (76.26%) 0.76 0.65 1.54
b3/b4 8, 6, 3 0.34 37.53 (71.00%) 0.71 0.66 1.52
b3/b5 1, 6, 8 0.24 30.62 (57.93%) 0.58 0.76 1.32
b4/b5 8, 7, 5 0.40 23.78 (44.99%) 0.45 0.60 1.67
All data 1, 4, 6, 8, 7, 10, 2 0.68 15.47 (29.27%) 0.29 0.32 4.13
SGB E. grandis b2/b3 1, 3, 6 0.22 43.85 (25.75%) 0.26 0.78 1.28
b2/b4 6, 2, 5 0.15 45.29 (26.59%) 0.27 0.85 1.18
b2/b5 6, 1, 4 0.30 36.41 (21.38%) 0.21 0.70 1.43
b3/b4 2, 8, 6 0.33 37.65 (22.11%) 0.22 0.67 1.49
b3/b5 1, 6, 8 0.22 40.14 (23.57%) 0.24 0.78 1.28
b4/b5 6, 5, 7 0.34 38.25 (22.46%) 0.22 0.66 1.52
All data 6, 8, 8,1, 4, 7, 2 0.65 14.37 (08.44%) 0.08 0.35 2.86
SGB P. taeda b2/b3 2, 6, 1 0.22 38.64 (18.75%) 0.19 0.78 1.28
b2/b4 4, 6, 8 0.02 73.55 (35.69%) 0.36 0.98 1.02
b2/b5 6, 1, 2 0.30 38.30 (18.59%) 0.19 0.70 1.43
b3/b4 2, 6, 8 0.22 40.43 (19.62%) 0.20 0.78 1.28
b3/b5 1, 6, 4 0.21 43.71 (21.21%) 0.21 0.79 1.27
b4/b5 7, 4, 6 0.49 25.16 (12.21%) 0.12 0.51 1.96
All data 8, 1, 4, 2, 7, 5, 4 0.69 15.14 (07.35%) 0.07 0.31 3.23
SGB All species b2/b3 2, 4, 1 0.14 45.68 (32.77%) 0.33 0.86 1.16
b2/b4 8, 4, 6 0.18 50.19 (36.01%) 0.36 0.82 1.22
b2/b5 1, 6, 3 0.10 71.31 (51.16%) 0.51 0.90 1.11
b3/b4 6, 8, 2 0.09 71.25 (51.12%) 0.51 0.91 1.10
b3/b5 6, 1, 2 0.20 56.37 (40.44%) 0.40 0.80 1.25
b4/b5 7, 8, 4 0.24 39.64 (28.44%) 0.28 0.76 1.32
All data 8, 1, 4, 2, 7, 4, 6 0.49 29.89 (21.44%) 0.21 0.51 1.96

SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second moment, 8-correlation, 9-standard deviation and 10-
data range.
24 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32

Table A3
Aboveground biomass estimates derived using Landsat-8 OLI band texture ratios from a 3  3 window size and an offset [1, 0] using stochastic gradient boosting regression
models.

Model Plant. species Image data Selected texture ratios R2 RMSE t ha1 CV-RMSE Tol. VIF
SGB E. dunii b2/b3 6, 2, 1 0.17 38.12 (72.12%) 0.72 0.83 1.20
b2/b4 4, 8, 6 0.31 37.45 (70.85%) 0.71 0.69 1.45
b2/b5 1, 6, 3 0.29 45.34 (85.77%) 0.86 0.71 1.41
b3/b4 6, 8, 4 0.26 42.25 (79.93%) 0.80 0.74 1.35
b3/b5 8, 1, 6 0.20 35.14 (66.48%) 0.66 0.80 1.25
b4/b5 9, 6, 3 0.34 27.72 (52.44%) 0.52 0.66 1.52
All data 1, 4, 2, 9, 8, 4, 6 0.57 20.46 (38.71%) 0.39 0.43 2.33
SGB E. grandis b2/b3 6, 1, 2 0.17 46.90 (27.54%) 0.28 0.83 1.20
b2/b4 4, 7, 6 0.13 49.85 (29.27%) 0.29 0.87 1.15
b2/b5 4, 6, 1 0.21 39.82 (23.38%) 0.23 0.79 1.27
b3/b4 6, 7, 2 0.27 41.26 (24.23%) 0.24 0.73 1.37
b3/b5 8, 6, 1 0.19 45.33 (26.62%) 0.27 0.81 1.23
b4/b5 4, 6, 7 0.28 42.10 (24.72%) 0.25 0.72 1.39
All data 2, 6, 7, 1, 7, 4, 8 0.59 18.36 (10.78%) 0.11 0.41 2.44
SGB P. taeda b2/b3 6, 1, 2 0.18 43.69 (21.20%) 0.21 0.82 1.22
b2/b4 8, 5, 7 0.01 79.34 (38.50%) 0.39 0.99 1.01
b2/b5 1, 6, 2 0.29 43.21 (20.97%) 0.21 0.71 1.41
b3/b4 6, 2, 8 0.20 44.84 (21.76%) 0.22 0.80 1.25
b3/b5 1, 6, 8 0.21 46.29 (22.46%) 0.22 0.79 1.27
b4/b5 6, 7, 5 0.42 31.39 (15.23%) 0.15 0.58 1.72
All data 1, 5, 6, 4, 8, 2, 7 0.60 24.68 (11.98%) 0.12 0.40 2.50
SGB All species b2/b3 6, 2, 4 0.12 49.36 (35.41%) 0.35 0.88 1.15
b2/b4 4, 6, 5 0.13 53.97 (38.72%) 0.39 0.87 1.25
b2/b5 1, 5, 6 0.11 76.34 (54.77%) 0.55 0.89 1.12
b3/b4 8, 6, 2 0.10 77.30 (55.46%) 0.55 0.90 1.11
b3/b5 1, 8, 6 0.17 61.30 (43.98%) 0.44 0.83 1.20
b4/b5 6, 7, 3 0.22 43.82 (31.44%) 0.31 0.78 1.28
All data 6, 1, 7, 8, 4, 5, 2 0.42 36.75 (26.36%) 0.26 0.58 1.72

SMLR = stepwise multiple-linear regression; SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second
moment, 8-correlation, 9-standard deviation and 10-data range.

Table A4
Aboveground biomass estimates derived using Landsat-8 OLI band texture ratios from a 3  3 window size and an offset [1, 1] using stochastic gradient boosting regression
models.

Model Plant. species Image data Selected texture ratios R2 RMSE t ha1 CV-RMSE Tol. VIF
SGB E. dunii b2/b3 6, 4, 1 0.15 49.76 (94.17%) 0.72 0.85 1.18
b2/b4 6, 4, 3 0.29 50.34 (95.23%) 0.71 0.71 1.41
b2/b5 1, 6, 2 0.27 52.36 (99.05%) 0.86 0.73 1.37
b3/b4 6, 2, 4 0.25 53.17 (98.69%) 0.80 0.75 1.33
b3/b5 1, 6, 2 0.21 49.48 (93.61%) 0.66 0.79 1.27
b4/b5 7, 6, 4 0.32 43.74 (82.75%) 0.52 0.68 1.47
All data 1, 6, 4, 8, 2, 7 0.55 37.91 (71.72%) 0.39 0.45 2.22
SGB E. grandis b2/b3 1, 2, 6 0.15 50.37 (29.58%) 0.28 0.85 1.17
b2/b4 4, 6, 8 0.11 58.31 (34.24%) 0.29 0.89 1.12
b2/b5 1, 6, 4 0.19 39.82 (23.38%) 0.23 0.81 1.23
b3/b4 2, 8, 6 0.25 53.24 (31.26%) 0.24 0.75 1.33
b3/b5 6, 1, 8 0.20 52.12 (30.60%) 0.27 0.80 1.25
b4/b5 7, 5, 4 0.27 49.92 (29.31%) 0.25 0.73 1.37
All data 6, 4, 8, 1, 2, 5, 7 0.55 40.22 (23.62%) 0.11 0.45 2.22
SGB P. taeda b2/b3 2, 6, 1 0.16 50.63 (24.57%) 0.21 0.84 1.19
b2/b4 1, 6, 8 0.02 87.42 (42.42%) 0.39 0.98 1.02
b2/b5 6, 1, 2 0.27 53.47 (25.95%) 0.21 0.73 1.37
b3/b4 8, 6, 1 0.15 52.84 (25.64%) 0.22 0.85 1.18
b3/b5 1, 6, 8 0.19 49.73 (24.13%) 0.22 0.81 1.23
b4/b5 7, 6, 4 0.43 44.67 (21.68%) 0.15 0.57 1.75
All data 6, 1, 8, 4, 7, 5, 2 0.54 35.18 (17.07%) 0.12 0.46 2.17
SGB All species b2/b3 6, 1, 2, 4 0.11 55.31 (39.68%) 0.35 0.89 1.12
b2/b4 4, 3, 1 0.14 57.45 (41.22%) 0.39 0.86 1.11
b2/b5 1, 6, 2 0.10 81.29 (58.32%) 0.55 0.90 1.11
b3/b4 8, 6, 1 0.09 86.15 (61.81%) 0.55 0.91 1.10
b3/b5 1, 6, 8 0.16 73.24 (52.54%) 0.44 0.84 1.19
b4/b5 6, 7, 4 0.23 56.22 (40.33%) 0.31 0.77 1.30
All data 6, 1, 8, 4, 2, 7, 4 0.39 43.78 (31.41%) 0.26 0.61 1.64

SMLR = stepwise multiple-linear regression; SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second
moment, 8-correlation, 9-standard deviation and 10-data range.
 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32 25

Table A5
Aboveground biomass estimates derived using Landsat-8 OLI band texture ratios from a 5  5 window size and an offset [0, 1] using stochastic gradient boosting regression
models.

Model Plant. species Image data Selected texture ratios R2 RMSE t ha1 CV-RMSE Tol. VIF
SGB E. dunii b2/b3 5, 4, 3 0.22 34.00 (64.32%) 0.64 0.78 1.28
b2/b4 6, 3, 1 0.39 32.58 (61.63%) 0.62 0.61 1.64
b2/b5 6, 1, 5 0.33 37.40 (70.75%) 0.71 0.67 1.49
b3/b4 8, 8, 2 0.36 35.73 (67.59%) 0.68 0.64 1.56
b3/b5 1, 6, 8 0.23 28.01 (52.99%) 0.53 0.77 1.30
b4/b5 6, 7, 4 0.45 21.39 (40.47%) 0.40 0.55 1.82
All data 1, 4, 6, 8, 7, 3 0.70 11.68 (22.10%) 0.22 0.30 3.33
SGB E. grandis b2/b3 1, 2, 6 0.22 39.18 (23.01%) 0.23 0.78 1.28
b2/b4 7, 8, 6 0.16 42.35 (24.87%) 0.25 0.84 1.19
b2/b5 1, 6, 5 0.31 33.92 (21.68%) 0.22 0.69 1.45
b3/b4 8, 1, 6 0.32 37.45 (21.99%) 0.22 0.68 1.47
b3/b5 1, 6, 8 0.27 39.66 (23.29%) 0.23 0.73 1.37
b4/b5 4, 6, 7 0.36 37.41 (21.97%) 0.22 0.64 1.56
All data 1, 8, 6, 4, 4, 2, 7 0.69 17.70 (10.39%) 0.09 0.31 3.23
SGB P. taeda b2/b3 5, 6, 2 0.24 40.00 (19.41%) 0.18 0.76 1.31
b2/b4 3, 6, 4 0.04 70.28 (34.10%) 0.34 0.96 1.04
b2/b5 6, 4, 2 0.30 35.33 (17.14%) 0.17 0.70 1.43
b3/b4 6, 8, 2 0.21 37.91 (18.40%) 0.18 0.79 1.27
b3/b5 6, 1, 3 0.25 40.00 (19.41%) 0.19 0.75 1.33
b4/b5 7, 6, 4 0.49 22.18 (10.76%) 0.11 0.51 1.96
All data 5, 2, 7, 6, 1, 8, 4 0.69 13.43 (06.52%) 0.07 0.31 3.23
SGB All species b2/b3 2, 1, 4, 5 0.16 44.55 (31.96%) 0.32 0.84 1.19
b2/b4 6, 3, 2 0.18 48.00 (34.44%) 0.34 0.82 1.22
b2/b5 6, 4, 7 0.11 69.24 (49.67%) 0.50 0.89 1.12
b3/b4 8, 6, 2 0.08 70.38 (50.49%) 0.50 0.92 1.09
b3/b5 1, 6, 8 0.23 53.74 (38.55%) 0.39 0.77 1.30
b4/b5 7, 6, 4 0.25 36.89 (26.47%) 0.26 0.75 1.33
All data 6, 2, 1, 7, 3, 8, 4 0.50 22.69 (16.28%) 0.16 0.50 2.00

SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second moment, 8-correlation, 9-standard deviation and 10-
data range.

Table A6
Aboveground biomass estimates derived using Landsat-8 OLI band texture ratios from a 5  5 window size and an offset [1, 1] using stochastic gradient boosting regression
models.

Model Plant. species Image data Selected texture ratios R2 RMSE t ha1 CV-RMSE Tol. VIF
SGB E. dunii b2/b3 1, 6, 2 0.18 38.12 (72.12%) 0.72 0.88 1.22
b2/b4 6, 4, 8 0.34 37.08 (70.15%) 0.70 0.66 1.52
b2/b5 6, 3, 1 0.33 41.03 (77.62%) 0.78 0.67 1.49
b3/b4 8, 6, 2 0.32 38.07 (72.02%) 0.72 0.68 1.47
b3/b5 1, 6, 8 0.20 33.62 (63.60%) 0.64 0.80 1.25
b4/b5 6, 7, 4 0.38 27.31 (51.66%) 0.52 0.62 1.61
All data 4, 5, 2, 1, 8, 7, 6 0.65 16.75 (31.69%) 0.32 0.35 2.86
SGB E. grandis b2/b3 6, 2, 3, 4 0.20 44.82 (26.32%) 0.26 0.80 1.25
b2/b4 8, 6, 4 0.13 46.79 (27.48%) 0.27 0.87 1.15
b2/b5 1, 6, 7 0.27 36.50 (21.43%) 0.21 0.73 1.37
b3/b4 8, 2, 6 0.30 37.47 (22.00%) 0.22 0.70 1.43
b3/b5 1, 6, 8 0.19 43.33 (25.44%) 0.25 0.81 1.23
b4/b5 6, 5, 7 0.33 38.25 (22.46%) 0.22 0.67 1.49
All data 6, 4, 2, 7, 8, 5, 1 0.63 15.73 (09.24%) 0.09 0.37 2.70
SGB P. taeda b2/b3 1, 6, 2 0.18 38.99 (18.92%) 0.19 0.82 1.22
b2/b4 6, 4, 8 0.03 76.18 (36.97%) 0.37 0.97 1.03
b2/b5 6, 1, 7 0.27 40.73 (19.77%) 0.20 0.73 1.36
b3/b4 8, 7, 6 0.20 45.17 (21.92%) 0.22 0.80 1.25
b3/b5 8, 6, 1 0.18 42.06 (20.41%) 0.20 0.82 1.21
b4/b5 7, 6, 4 0.44 27.58 (13.38%) 0.13 0.56 1.79
All data 2, 7, 6, 4, 8, 1, 4 0.67 19.43 (09.43%) 0.09 0.33 3.03
SGB All species b2/b3 1, 2, 4, 6 0.13 46.33 (33.24%) 0.33 0.87 1.15
b2/b4 6, 4, 8 0.15 50.99 (36.58%) 0.37 0.85 1.18
b2/b5 1, 5, 6 0.09 74.96 (53.78%) 0.54 0.91 1.10
b3/b4 8, 6, 2 0.07 74.98 (53.79%) 0.54 0.93 1.07
b3/b5 8, 1, 6 0.18 57.23 (41.06%) 0.41 0.82 1.21
b4/b5 7, 8, 6 0.21 40.82 (29.28%) 0.29 0.79 1.26
All data 1, 8, 5, 2, 7, 4, 6 0.45 33.77 (24.23%) 0.24 0.55 1.81

SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second moment, 8-correlation, 9-standard deviation and 10-
data range.
26 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32

Table A7
Aboveground biomass estimates derived using Landsat-8 OLI band texture ratios from a 5  5 window size and an offset [1, 0] using stochastic gradient boosting regression
models.

Model Plant. species Image data Selected texture ratios R2 RMSE t ha1 CV-RMSE Tol. VIF
SGB E. dunii b2/b3 3, 2, 6 0.14 43.17 (81.67%) 0.82 0.86 1.16
b2/b4 6, 4, 5 0.26 41.26 (78.06%) 0.78 0.74 1.35
b2/b5 6, 1, 3 0.23 48.30 (91.37%) 0.91 0.77 1.30
b3/b4 2, 8, 6 0.21 43.00 (81.35%) 0.81 0.79 1.27
b3/b5 6, 8, 1 0.19 35.99 (68.09%) 0.68 0.81 1.23
b4/b5 6, 7, 5 0.30 27.87 (52.72%) 0.53 0.70 1.43
All data 4, 8, 7, 3, 1, 6, 5 0.53 22.46 (42.49%) 0.42 0.47 2.13
SGB E. grandis b2/b3 1, 2, 6 0.15 47.66 (27.99%) 0.28 0.85 1.17
b2/b4 6, 4, 8 0.10 50.23 (29.50%) 0.29 0.90 1.11
b2/b5 1, 3, 6 0.19 51.10 (30.01%) 0.30 0.81 1.23
b3/b4 2, 8, 6 0.25 41.89 (24.60%) 0.25 0.75 1.33
b3/b5 1, 6, 3 0.20 46.78 (27.47%) 0.27 0.80 1.25
b4/b5 6, 1, 4 0.25 43.00 (25.25%) 0.25 0.75 1.33
All data 6, 5, 2, 7, 4, 1, 8 0.55 18.85 (11.07%) 0.11 0.45 2.22
SGB P. taeda b2/b3 1, 6, 2 0.20 44.01 (21.36%) 0.21 0.80 1.25
b2/b4 6, 7, 8 0.01 79.70 (38.68%) 0.39 0.99 1.01
b2/b5 6, 3, 1 0.25 44.78 (21.73%) 0.22 0.75 1.33
b3/b4 2, 6, 8 0.17 45.32 (21.99%) 0.22 0.83 1.20
b3/b5 1, 6, 8, 3, 2 0.19 46.50 (22.57%) 0.23 0.81 1.23
b4/b5 6, 7, 5 0.39 32.73 (15.88%) 0.16 0.61 1.64
All data 6, 5, 4, 3, 8, 2, 7 0.55 26.40 (12.81%) 0.13 0.45 2.22
SGB All species b2/b3 2, 4, 1, 6 0.11 50.32 (36.10%) 0.36 0.89 1.12
b2/b4 5, 6, 8 0.09 53.98 (38.73%) 0.39 0.91 1.09
b2/b5 6, 3, 2 0.12 78.34 (56.20%) 0.56 0.88 1.14
b3/b4 8, 6, 7 0.13 79.13 (56.77%) 0.57 0.87 1.15
b3/b5 1, 2, 6, 8 0.14 63.11 (45.28%) 0.45 0.86 1.16
b4/b5 6, 7, 5 0.18 45.28 (32.48%) 0.32 0.82 1.23
All data 5, 6, 8, 2, 7, 4, 3 0.38 36.77 (26.38%) 0.26 0.62 1.61

SMLR = stepwise multiple-linear regression; SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second
moment, 8-correlation, 9-standard deviation and 10-data range.

Table A8
Aboveground biomass estimates derived using Landsat-8 OLI band texture ratios from a 5  5 window size and an offset [1, -1] using stochastic gradient boosting regression
models.

Model Plant. species Image data Selected texture ratios R2 RMSE t ha1 CV-RMSE Tol. VIF
SGB E. dunii b2/b3 1, 6, 2 0.14 51.37 (97.18%) 0.97 0.86 1.16
b2/b4 6, 5, 8 0.22 49.58 (93.79%) 0.94 0.78 1.28
b2/b5 6, 1, 3 0.26 49.00 (92.70%) 0.93 0.74 1.35
b3/b4 8, 6, 2 0.24 49.30 (93.27%) 0.93 0.76 1.35
b3/b5 1, 6, 8 0.20 50.73 (95.97%) 0.96 0.80 1.25
b4/b5 6, 7, 5 0.30 46.93 (88.78%) 0.89 0.70 1.43
All data 3, 8, 6, 5, 2, 7, 4 0.52 43.24 (81.80%) 0.82 0.48 2.08
SGB E. grandis b2/b3 1, 2, 6 0.13 53.61 (31.48%) 0.31 0.87 1.15
b2/b4 6, 8, 5 0.10 61.73 (36.25%) 0.36 0.90 1.11
b2/b5 6, 1, 3 0.17 43.28 (25.41%) 0.25 0.83 1.20
b3/b4 2, 8, 6 0.21 56.07 (32.92%) 0.33 0.79 1.27
b3/b5 1, 6, 8 0.16 54.25 (31.86%) 0.32 0.84 1.19
b4/b5 6, 7, 5 0.23 54.44 (31.97%) 0.32 0.77 1.29
All data 6, 4, 2, 3, 7, 1, 8, 5 0.55 43.63 (25.62%) 0.26 0.45 2.22
SGB P. taeda b2/b3 1, 6, 2 0.15 55.35 (26.86%) 0.27 0.85 1.18
b2/b4 6, 5, 8 0.01 88.13 (42.77%) 0.43 0.99 1.01
b2/b5 6, 1, 2 0.26 56.38 (27.36%) 0.27 0.74 1.35
b3/b4 2, 8, 6 0.14 57.00 (27.66%) 0.28 0.86 1.16
b3/b5 1, 6, 8 0.17 53.16 (25.80%) 0.26 0.83 1.20
b4/b5 7, 5, 6 0.41 47.28 (22.94%) 0.23 0.59 1.69
All data 6, 5, 2, 7, 4, 1, 8 0.51 38.94 (18.90%) 0.19 0.49 2.04
SGB All species b2/b3 1, 6, 2, 3 0.11 55.38 (39.73%) 0.40 0.89 1.12
b2/b4 5, 6, 8 0.14 58.02 (41.62%) 0.42 0.86 1.16
b2/b5 1, 6, 2 0.11 83.46 (59.88%) 0.60 0.89 1.12
b3/b4 8, 2, 6 0.08 86.77 (62.25%) 0.62 0.92 1.09
b3/b5 8, 1, 6 0.14 73.83 (52.97%) 0.53 0.86 1.16
b4/b5 6, 7, 5 0.20 56.24 (40.35%) 0.40 0.80 1.25
All data 1, 8, 6, 5, 4, 2, 7 0.35 44.16 (31.68%) 0.32 0.65 1.54

SMLR = stepwise multiple-linear regression; SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second
moment, 8-correlation, 9-standard deviation and 10-data range.
 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32 27

Table A9
Aboveground biomass estimates derived using Landsat-8 OLI band texture ratios from a 7  7 window size and an offset [0, 1] using stochastic gradient boosting regression
models.

Model Plant. species Image data Selected texture ratios R2 RMSE t ha1 CV-RMSE Tol. VIF
SGB E. dunii b2/b3 1, 6, 2 0.19 35.01 (66.23%) 0.66 0.81 1.23
b2/b4 6, 5, 8 0.38 33.70 (63.75%) 0.64 0.62 1.61
b2/b5 6, 3, 1 0.31 39.55 (74.82%) 0.75 0.69 1.45
b3/b4 8, 6, 2 0.36 35.98 (68.07%) 0.68 0.64 1.56
b3/b5 1, 6, 8 0.20 30.09 (56.92%) 0.57 0.80 1.25
b4/b5 6, 5, 7 0.43 23.44 (44.34%) 0.44 0.57 1.75
All data 7, 2, 4, 1, 6, 5, 8 0.65 33.47 (63.32%) 0.63 0.35 2.85
SGB E. grandis b2/b3 1, 2, 6 0.17 40.73 (23.92%) 0.24 0.83 1.20
b2/b4 8, 5, 6 0.16 42.82 (25.14%) 0.25 0.84 1.19
b2/b5 1, 6, 3 0.30 34.51 (20.26%) 0.20 0.70 1.43
b3/b4 2, 8, 6 0.28 37.85 (22.23%) 0.22 0.72 1.39
b3/b5 8, 1, 6 0.25 40.22 (23.62%) 0.24 0.75 1.33
b4/b5 7, 6, 5 0.33 39.66 (23.29%) 0.23 0.67 1.49
All data 5, 2, 7, 6, 4, 1, 8 0.63 19.14 (11.24%) 0.11 0.37 2.70
SGB P. taeda b2/b3 1, 6, 2 0.21 42.79 (20.76%) 0.21 0.79 1.27
b2/b4 6, 5, 8 0.03 70.88 (34.40%) 0.34 0.97 1.03
b2/b5 6, 1, 3 0.25 36.99 (17.95%) 0.18 0.75 1.33
b3/b4 8, 2, 6 0.18 36.99 (17.95%) 0.18 0.82 1.21
b3/b5 1, 6, 3, 8 0.24 40.73 (19.77%) 0.20 0.76 1.31
b4/b5 7, 5, 6 0.43 23.67 (11.49%) 0.11 0.57 1.75
All data 6, 4, 1, 8, 5, 2, 7 0.68 13.58 (06.59%) 0.07 0.32 3.12
SGB All species b2/b3 1, 2, 6 0.15 45.69 (32.78%) 0.33 0.85 1.18
b2/b4 6, 5, 8 0.14 48.83 (35.03%) 0.35 0.86 1.16
b2/b5 3, 6, 1 0.10 70.41 (50.51%) 0.51 0.90 1.11
b3/b4 8, 6, 2 0.05 73.87 (53.00%) 0.53 0.95 1.05
b3/b5 1, 8, 6 0.19 53.72 (38.54%) 0.39 0.81 1.23
b4/b5 6, 7, 5 0.24 36.55 (26.22%) 0.26 0.76 1.32
All data 1, 6, 8, 5, 2, 7, 4 0.45 24.39 (17.50%) 0.17 0.55 1.82

SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second moment, 8-correlation, 9-standard deviation and 10-
data range.

Table A10
Aboveground biomass estimates derived using Landsat-8 OLI band texture ratios from a 7  7 window size and an offset [1, 1] using stochastic gradient boosting regression
models.

Model Plant. species Image data Selected texture ratios R2 RMSE t ha1 CV-RMSE Tol. VIF
SGB E. dunii b2/b3 1, 6, 2 0.15 40.32 (76.28%) 0.76 0.85 1.18
b2/b4 6, 8, 5 0.30 40.67 (76.94%) 0.77 0.70 1.43
b2/b5 6, 1, 3 0.32 41.98 (79.42%) 0.79 0.68 1.47
b3/b4 8, 2, 6 0.31 38.99 (73.06%) 0.74 0.69 1.45
b3/b5 1, 6, 8 0.20 34.00 (64.32%) 0.64 0.80 1.25
b4/b5 7, 6, 5 0.35 28.33 (53.59%) 0.54 0.65 1.53
All data 4, 1, 6, 7, 2, 8 0.63 17.21 (32.56%) 0.33 0.37 2.70
SGB E. grandis b2/b3 1, 2, 6 0.16 45.87 (26.93%) 0.27 0.84 1.19
b2/b4 6, 5, 8 0.12 46.95 (27.57%) 0.28 0.88 1.14
b2/b5 6, 3, 1 0.25 38.28 (22.48%) 0.22 0.75 1.33
b3/b4 2, 8, 6 0.29 37.84 (22.22%) 0.22 0.71 1.41
b3/b5 1, 8, 6, 0.15 45.19 (26.54%) 0.27 0.85 1.18
b4/b5 6, 5, 7 0.30 40.76 (23.93%) 0.24 0.70 1.42
All data 6, 4, 1, 8, 5, 2, 7 0.61 17.94 (10.53%) 0.11 0.39 2.56
SGB P. taeda b2/b3 1, 6, 2 0.18 45.62 (22.14%) 0.22 0.82 1.21
b2/b4 6, 5, 8 0.02 76.69 (37.22%) 0.37 0.98 1.02
b2/b5 1, 6, 2 0.23 41.80 (20.28%) 0.20 0.77 1.30
b3/b4 8, 6, 2 0.20 45.96 (22.30%) 0.22 0.80 1.25
b3/b5 1, 6, 8 0.18 42.70 (20.72%) 0.21 0.82 1.22
b4/b5 7, 6, 5 0.41 30.38 (14.74%) 0.15 0.59 1.69
All data 6. 1, 8, 5, 2, 7 0.63 23.05 (11.19%) 0.11 0.37 2.70
SGB All species b2/b3 1, 6, 4 0.10 46.87 (33.63%) 0.34 0.90 1.11
b2/b4 8, 6, 5 0.12 53.91 (38.68%) 0.39 0.88 1.14
b2/b5 6, 1, 3 0.07 75.06 (53.85%) 0.54 0.93 1.08
b3/b4 8, 6, 2 0.05 76.28 (54.72%) 0.55 0.95 1.05
b3/b5 1, 6, 8 0.15 60.37 (43.31%) 0.43 0.85 1.18
b4/b5 6, 7, 5 0.16 44.38 (31.84%) 0.32 0.84 1.19
All data 1, 6, 8, 5, 2, 7, 4 0.40 35.20 (25.25%) 0.25 0.60 1.67

SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second moment, 8-correlation, 9-standard deviation and 10-
data range.
28 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32

Table A11
Aboveground biomass estimates derived using Landsat-8 OLI band texture ratios from a 7  7 window size and an offset [1, 0] using stochastic gradient boosting regression
models.

Model Plant. species Image data Selected texture ratios R2 RMSE t ha1 CV-RMSE Tol. VIF
SGB E. dunii b2/b3 1, 6, 2 0.14 43.17 (81.67%) 0.82 0.86 1.16
b2/b4 6, 5, 8 0.26 41.26 (78.06%) 0.78 0.74 1.35
b2/b5 6, 1, 2 0.23 48.30 (91.37%) 0.91 0.77 1.30
b3/b4 8, 6, 2 0.21 43.00 (81.35%) 0.81 0.79 1.27
b3/b5 1, 6, 8 0.19 35.99 (68.09%) 0.68 0.81 1.23
b4/b5 6, 7, 5 0.30 27.87 (52.72%) 0.53 0.70 1.43
All data 4, 1, 6, 5, 8, 7, 2 0.53 22.46 (42.49%) 0.42 0.47 2.13
SGB E. grandis b2/b3 1, 2, 6 0.15 47.66 (27.99%) 0.28 0.85 1.17
b2/b4 6, 8, 5 0.10 50.23 (29.50%) 0.29 0.90 1.11
b2/b5 6, 1, 3 0.19 51.10 (30.01%) 0.30 0.81 1.23
b3/b4 2, 8, 6 0.25 41.89 (24.60%) 0.25 0.75 1.33
b3/b5 1, 6, 8 0.20 46.78 (27.47%) 0.27 0.80 1.25
b4/b5 6, 5, 7 0.25 43.00 (25.25%) 0.25 0.75 1.33
All data 6, 4, 1, 8, 5, 2, 7 0.55 18.85 (11.07%) 0.11 0.45 2.22
SGB P. taeda b2/b3 1, 6, 2 0.20 44.01 (21.36%) 0.21 0.80 1.25
b2/b4 6, 5, 8 0.01 79.70 (38.68%) 0.39 0.99 1.01
b2/b5 6, 1, 3 0.25 44.78 (21.73%) 0.22 0.75 1.33
b3/b4 8, 6, 2 0.17 45.32 (21.99%) 0.22 0.83 1.20
b3/b5 1, 6, 8 0.19 46.50 (22.57%) 0.23 0.81 1.23
b4/b5 6, 7, 5 0.39 32.73 (15.88%) 0.16 0.61 1.64
All data 6, 4, 1, 5, 2, 7, 8 0.55 26.40 (12.81%) 0.13 0.45 2.22
SGB All species b2/b3 1, 6, 2, 4 0.11 50.32 (36.10%) 0.36 0.89 1.12
b2/b4 6, 5, 8 0.09 53.98 (38.73%) 0.39 0.91 1.09
b2/b5 6, 1 0.12 78.34 (56.20%) 0.56 0.88 1.14
b3/b4 8, 6, 2 0.13 79.13 (56.77%) 0.57 0.87 1.15
b3/b5 1, 6, 8 0.14 63.11 (45.28%) 0.45 0.86 1.16
b4/b5 6, 7, 5 0.18 45.28 (32.48%) 0.32 0.82 1.23
All data 1, 6, 2, 7, 4, 8, 5 0.38 36.77 (26.38%) 0.26 0.62 1.61

SMLR = stepwise multiple-linear regression; SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second
moment, 8-correlation, 9-standard deviation and 10-data range.

Table A12
Aboveground biomass estimates derived using Landsat-8 OLI band texture ratios from a 7  7 window size and an offset [1, -1] using stochastic gradient boosting regression
models.

Model Plant. species Image data Selected texture ratios R2 RMSE t ha1 CV-RMSE Tol. VIF
SGB E. dunii b2/b3 1, 2, 6 0.13 51.45 (97.33%) 0.97 0.87 1.15
b2/b4 6, 5, 8 0.19 49.68 (93.98%) 0.94 0.81 1.23
b2/b5 6, 1, 3 0.21 49.55 (93.74%) 0.94 0.79 1.27
b3/b4 2, 8, 6 0.20 49.77 (94.15%) 0.94 0.80 1.25
b3/b5 1, 8, 6 0.19 50.78 (96.07%) 0.96 0.81 1.23
b4/b5 6, 7, 5 0.28 46.99 (88.90%) 0.89 0.72 1.39
All data 4, 1, 6, 5, 8, 7, 3 0.47 43.90 (83.05%) 0.83 0.53 1.88
SGB E. grandis b2/b3 1, 6, 2 0.11 53.70 (31.53%) 0.32 0.89 1.12
b2/b4 6, 5, 8 0.10 62.00 (36.41%) 0.36 0.90 1.11
b2/b5 1, 6, 2 0.15 43.79 (25.71%) 0.26 0.85 1.18
b3/b4 8, 2, 6 0.18 56.87 (33.39%) 0.33 0.82 1.22
b3/b5 1, 6, 8 0.15 55.10 (32.35%) 0.32 0.85 1.18
b4/b5 6, 5, 7 0.20 54.90 (32.24%) 0.32 0.80 1.25
All data 6, 4, 1, 8, 5, 2, 7 0.50 44.00 (25.84%) 0.26 0.50 2.00
SGB P. taeda b2/b3 1, 6, 2 0.13 57.60 (27.95%) 0.28 0.87 1.15
b2/b4 6, 8, 5 0.01 88.93 (43.16%) 0.43 0.99 1.01
b2/b5 1, 6, 3 0.22 56.86 (27.59%) 0.28 0.78 1.28
b3/b4 8, 6, 2 0.11 57.69 (28.00%) 0.28 0.89 1.12
b3/b5 1, 6, 8 0.16 53.70 (26.06%) 0.26 0.84 1.19
b4/b5 7, 6, 5 0.38 48.02 (23.30%) 0.23 0.62 1.61
All data 6, 5, 2, 7,4, 1, 8 0.46 39.50 (19.17%) 0.19 0.54 1.85
SGB All species b2/b3 1, 6, 3, 5 0.10 56.66 (40.65%) 0.41 0.90 1.11
b2/b4 5, 6, 8 0.13 58.92 (42.27%) 0.42 0.87 1.15
b2/b5 3, 1, 6 0.11 83.91 (60.20%) 0.60 0.89 1.12
b3/b4 8, 6, 2 0.07 87.21 (62.57%) 0.63 0.93 1.08
b3/b5 1, 8, 8 0.15 74.00 (53.09%) 0.53 0.85 1.18
b4/b5 6, 5, 7 0.19 56.79 (40.74%) 0.41 0.81 1.23
All data 1, 6, 7, 4, 8, 5, 2 0.32 48.18 (34.56%) 0.35 0.68 1.47

SMLR = stepwise multiple-linear regression; SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second
moment, 8-correlation, 9-standard deviation and 10-data range.
 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32 29

Table A13
Aboveground biomass estimates derived using Landsat-8 OLI band texture ratios from a 9  9 window size and an offset [0, 1] using stochastic gradient boosting regression
models.

Model Plant. species Image data Selected texture ratios R2 RMSE t ha1 CV-RMSE Tol. VIF
SGB E. dunii b2/b3 1, 6, 3 0.15 36.54 (69.13%) 0.69 0.85 1.18
b2/b4 6, 5, 8 0.35 34.45 (65.17%) 0.65 0.65 1.54
b2/b5 6, 1, 5 0.29 41.28 (78.09%) 0.78 0.71 1.41
b3/b4 8, 6, 3 0.33 37.73 (71.38%) 0.71 0.67 1.49
b3/b5 8, 1, 6 0.17 32.16 (60.84%) 0.61 0.83 1.20
b4/b5 6, 7, 5 0.40 24.00 (45.40%) 0.45 0.60 1.67
All data 1, 4, 6, 8, 5, 7, 3 0.62 35.97 (68.05%) 0.68 0.38 2.63
SGB E. grandis b2/b3 2, 1, 6 0.15 41.60 (24.43%) 0.24 0.85 1.18
b2/b4 6, 8, 5 0.14 42.99 (25.24%) 0.25 0.86 1.16
b2/b5 6, 1, 2 0.27 35.51 (20.85%) 0.21 0.73 1.37
b3/b4 2, 8, 6, 3 0.25 38.23 (22.45%) 0.22 0.75 1.33
b3/b5 1, 6, 8 0.22 41.27 (24.23%) 0.24 0.78 1.28
b4/b5 6, 5, 7 0.30 40.39 (23.72%) 0.24 0.70 1.43
All data 6, 4, 1, 8, 5, 7, 2 0.59 23.46 (13.78%) 0.14 0.41 2.44
SGB P. taeda b2/b3 1, 6, 2 0.18 43.00 (20.87%) 0.21 0.82 1.22
b2/b4 6, 5, 8 0.02 71.76 (34.82%) 0.35 0.98 1.02
b2/b5 6, 1, 3 0.25 37.33 (18.12%) 0.18 0.75 1.33
b3/b4 8, 6, 2, 3 0.15 37.69 (19.29%) 0.18 0.85 1.18
b3/b5 1, 6, 8 0.21 40.99 (19.89%) 0.20 0.79 1.27
b4/b5 6, 7, 5 0.39 23.83 (11.56%) 0.12 0.61 1.64
All data 6, 4, 1, 8, 5, 2, 7 0.65 15.40 (07.47%) 0.07 0.35 2.86
SGB All species b2/b3 1, 6, 2 0.13 47.02 (33.73%) 0.34 0.87 1.15
b2/b4 6, 5, 8 0.12 49.35 (35.40%) 0.35 0.88 1.14
b2/b5 6, 2, 1 0.09 72.48 (52.00%) 0.52 0.91 1.09
b3/b4 8, 2, 6 0.06 74.90 (53.73%) 0.54 0.94 1.06
b3/b5 1, 6, 8 0.15 53.81 (38.60%) 0.39 0.85 1.18
b4/b5 6, 7, 5 0.21 37.59 (26.97%) 0.27 0.79 1.27
All data 1, 7, 4, 6, 5, 2 0.43 26.18 (18.78%) 0.19 0.57 1.75

SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second moment, 8-correlation, 9-standard deviation and 10-
data range.

Table A14
Aboveground biomass estimates derived using Landsat-8 OLI band texture ratios from a 9  9 window size and an offset [1, 1] using stochastic gradient boosting regression
models.

Model Plant. species Image data Selected texture ratios R2 RMSE t ha1 CV-RMSE Tol. VIF
SGB E. dunii b2/b3 1, 6, 2 0.13 41.50 (78.51%) 0.79 0.87 1.15
b2/b4 6, 5, 8 0.25 41.62 (78.74%) 0.79 0.75 1.33
b2/b5 6, 1, 3 0.26 42.00 (79.46%) 0.79 0.74 1.35
b3/b4 8, 6, 2 0.29 39.17 (74.10%) 0.74 0.71 1.41
b3/b5 1, 3, 6 0.17 34.89 (66.00%) 0.66 0.83 1.20
b4/b5 6, 7, 5 0.34 29.75 (56.28%) 0.56 0.66 1.51
All data 4, 1, 6, 3, 8, 7, 5 0.57 18.56 (35.11%) 0.35 0.43 2.33
SGB E. grandis b2/b3 1, 2, 6 0.14 46.11 (22.97%) 0.27 0.86 1.16
b2/b4 6, 5, 8 0.10 47.23 (27.93%) 0.28 0.90 1.11
b2/b5 6, 1, 3 0.23 39.85 (23.40%) 0.23 0.77 1.30
b3/b4 2, 8, 6 0.25 38.65 (22.70%) 0.23 0.75 1.33
b3/b5 1, 6, 8 0.14 47.30 (27.77%) 0.28 0.86 1.16
b4/b5 6, 5, 7 0.29 40.99 (24.07%) 0.24 0.71 1.41
All data 6, 4, 5, 2, 7, 1, 8 0.57 20.18 (11.85%) 0.12 0.43 2.32
SGB P. taeda b2/b3 1, 6, 2 0.17 47.34 (22.97%) 0.23 0.83 1.20
b2/b4 6, 5, 8 0.02 78.16 (37.93%) 0.38 0.98 1.02
b2/b5 6, 1, 3 0.19 43.73 (21.22%) 0.21 0.81 1.23
b3/b4 8, 6, 2 0.19 46.89 (22.75%) 0.23 0.81 1.23
b3/b5 1, 6, 8 0.16 43.71 (21.21%) 0.21 0.84 1.19
b4/b5 7, 6, 5 0.38 32.58 (15.81%) 0.16 0.62 1.61
All data 6, 1, 8, 4, 5, 2, 7 0.60 24.99 (12.13%) 0.12 0.40 2.50
SGB All species b2/b3 1, 6, 2, 4 0.11 47.22 (33.88%) 0.34 0.89 1.12
b2/b4 6, 5, 8 0.10 54.87 (39.36%) 0.39 0.90 1.11
b2/b5 6, 1, 3 0.05 75.96 (54.49%) 0.54 0.95 1.05
b3/b4 8, 6, 2, 3 0.06 76.89 (55.16%) 0.55 0.94 1.06
b3/b5 6, 1, 8 0.13 62.73 (45.00%) 0.45 0.87 1.15
b4/b5 6, 7, 5 0.14 46.17 (33.12%) 0.33 0.86 1.16
All data 1, 6, 8, 5, 2, 3, 4, 7 0.38 36.93 (26.49%) 0.26 0.62 1.61

SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second moment, 8-correlation, 9-standard deviation and 10-
data range.
30 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32

Table A15
Aboveground biomass estimates derived using Landsat-8 OLI band texture ratios from a 9  9 window size and an offset [1, 0] using stochastic gradient boosting regression
models.

Model Plant. species Image data Selected texture ratios R2 RMSE t ha1 CV-RMSE Tol. VIF
SGB E. dunii b2/b3 1, 6, 2 0.13 43.79 (82.86%) 0.83 0.87 1.15
b2/b4 6, 5, 8 0.22 41.80 (79.08%) 0.79 0.78 1.28
b2/b5 6, 1, 3 0.20 49.11 (92.91%) 0.93 0.80 1.25
b3/b4 8, 6, 2 0.19 43.93 (83.11%) 0.83 0.81 1.23
b3/b5 1, 6, 8 0.19 36.73 (69.49%) 0.69 0.81 1.23
b4/b5 6, 7, 5 0.27 28.38 (53.69%) 0.54 0.73 1.37
All data 4, 1, 6, 5, 8, 7, 2 0.49 23.67 (44.78%) 0.45 0.51 1.96
SGB E. grandis b2/b3 1, 2, 6 0.13 48.80 (28.66%) 0.29 0.87 1.15
b2/b4 6, 8, 5 0.07 50.95 (29.92%) 0.30 0.93 1.08
b2/b5 6, 1, 3 0.16 52.73 (30.96%) 0.31 0.84 1.19
b3/b4 2, 8, 6 0.22 42.46 (24.93%) 0.25 0.78 1.28
b3/b5 1, 6, 8 0.20 47.30 (27.77%) 0.28 0.80 1.25
b4/b5 5, 7, 6 0.23 44.31 (26.02%) 0.26 0.77 1.30
All data 6, 4, 1, 8, 5, 2, 7 0.54 19.20 (11.27%) 0.11 0.46 2.17
SGB P. taeda b2/b3 1, 6, 2 0.17 46.28 (22.46%) 0.22 0.83 1.20
b2/b4 6, 5, 8 0.02 79.93 (38.79%) 0.39 0.98 1.02
b2/b5 6, 1, 3 0.26 46.27 (22.45%) 0.22 0.74 1.35
b3/b4 8, 6, 2 0.16 46.38 (22.51%) 0.23 0.84 1.19
b3/b5 1, 6, 8 0.17 47.83 (23.21%) 0.23 0.83 1.20
b4/b5 6, 7, 5 0.38 33.67 (16.34%) 0.16 0.62 1.61
All data 5, 4, 1, 8, 5, 2, 3, 7 0.53 28.03 (13.60%) 0.14 0.47 2.13
SGB All species b2/b3 1, 6, 2, 4 0.10 53.07 (38.07%) 0.38 0.90 1.11
b2/b4 6, 5, 8 0.07 54.38 (39.01%) 0.39 0.93 1.08
b2/b5 6, 1, 3 0.13 79.46 (57.01%) 0.57 0.87 1.15
b3/b4 8, 6, 2 0.11 80.17 (57.51%) 0.58 0.89 1.12
b3/b5 6, 8, 1 0.12 64.58 (46.33%) 0.46 0.88 1.14
b4/b5 6, 5, 7 0.15 46.40 (33.29%) 0.33 0.85 1.18
All data 1, 5, 2, 6, 4, 6, 8, 3 0.35 37.66 (27.02%) 0.27 0.65 1.54

SMLR = stepwise multiple-linear regression; SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second
moment, 8-correlation, 9-standard deviation and 10-data range.

Table A16
Aboveground biomass estimates derived using Landsat-8 OLI band texture ratios from a 9  9 window size and an offset [1, 1] using stochastic gradient boosting regression
models.

Model Plant. species Image data Selected texture ratios R2 RMSE t ha1 CV-RMSE Tol. VIF
SGB E. dunii b2/b3 1, 6, 2 0.09 52.00 (98.37%) 0.98 0.91 1.09
b2/b4 6, 5, 8 0.12 51.10 (96.67%) 0.97 0.88 1.14
b2/b5 6, 1, 3 0.16 50.69 (95.89%) 0.96 0.84 1.19
b3/b4 8, 6, 2 0.15 50.33 (95.21%) 0.95 0.85 1.18
b3/b5 1, 6, 8 0.13 52.37 (99.07%) 0.99 0.87 1.15
b4/b5 8, 7, 5 0.23 49.16 (93.00%) 0.93 0.77 1.30
All data 4, 1, 6, 8, 7, 2, 5 0.44 45.28 (85.66%) 0.86 0.56 1.10
SGB E. grandis b2/b3 1, 2, 6 0.09 55.27 (32.45%) 0.32 0.91 1.15
b2/b4 6, 8, 5 0.11 65.28 (38.33%) 0.38 0.89 1.18
b2/b5 1, 3, 6 0.13 46.29 (27.18%) 0.27 0.87 1.18
b3/b4 8, 6, 2 0.15 58.29 (34.23%) 0.34 0.85 1.22
b3/b5 1, 6, 8 0.15 56.97 (33.45%) 0.33 0.85 1.18
b4/b5 6, 5, 7 0.18 55.75 (32.74%) 0.33 0.82 1.22
All data 6, 4, 5, 2, 7, 1, 8 0.46 48.24 (28.33%) 0.28 0.54 1.85
SGB P. taeda b2/b3 1, 6, 2 0.11 60.82 (29.51%) 0.30 0.89 1.12
b2/b4 5, 8, 6 0.01 90.76 (44.04%) 0.44 0.99 1.01
b2/b5 3, 6, 1 0.20 59.24 (28.75%) 0.29 0.80 1.25
b3/b4 8, 6, 3 0.10 59.38 (28.82%) 0.27 0.90 1.11
b3/b5 1, 6, 8 0.13 56.54 (27.44%) 0.26 0.87 1.15
b4/b5 6, 7, 5 0.35 51.43 (24.96%) 0.25 0.65 1.54
All data 6, 8, 5, 2, 7, 4, 1 0.41 43.61 (21.16%) 0.21 0.59 1.69
SGB All species b2/b3 1, 2, 4, 6 0.05 58.24 (41.78%) 0.43 0.95 1.05
b2/b4 5, 8, 6 0.09 60.35(43.30%) 0.43 0.91 1.10
b2/b5 6, 1, 3 0.07 85.46 (61.31%) 0.61 0.93 1.08
b3/b4 8, 8, 2 0.04 89.76 (64.39%) 0.64 0.96 1.04
b3/b5 1, 6, 8 0.12 77.28 (55.44%) 0.55 0.88 1.14
b4/b5 7, 5, 6 0.15 60.39 (43.32%) 0.43 0.85 1.18
All data 1, 6, 2, 7, 4, 3, 8, 5 0.27 52.60 (37.74%) 0.38 0.73 1.37

SMLR = stepwise multiple-linear regression; SGB = stochastic gradient boosting. 1-mean, 2-variance, 3-homogeneity, 4-contrast, 5-dissimilarity, 6-entropy, 7-second
moment, 8-correlation, 9-standard deviation and 10-data range.
 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32 31

References Fan, Y., Koukal, T., Weisberg, P.J., 2014. A sun-crown-sensor model and adapted C-
correction logic for topographic correction of high resolution forest imagery.
ISPRS J. Photogramm. Remote Sens. 96, 94–105.
Adam, E., Mutanga, O., Abdel-Rahman, E.M., Ismail, R., 2014. Estimating standing
Foody, G.M., Boyd, D.S., Cutler, M.E.J., 2003. Predictive relations of tropical forest
biomass in papyrus (Cyperus papyrus L.) swamp: exploratory of in situ
biomass from Landsat TM data and their transferability between regions.
hyperspectral indices and random forest regression. Int. J. Remote Sens. 35,
Remote Sens. Environ. 85, 463–474.
693–714.
Franklin, S.E., Wulder, M.A., Lavigne, M.B., 1996. Automated derivation of
Adelabu, S., Dube, T., 2014. Employing ground and satellite-based QuickBird data
geographic window sizes for use in remote sensing digital image texture
and random forest to discriminate five tree species in a Southern African
analysis. Comput. Geosci. 22, 665–673.
Woodland. Geocarto Int., 1–15
Friedman, J.H., 2002. Stochastic gradient boosting. Comput. Stat. Data Anal. 38, 367–
Atta-Boateng, J., William, M.J., 1998. A method for classifying commercial tree
378.
species of an uneven-aged mixed species tropical forest for growth and yield
Fuchs, H., Magdon, P., Kleinn, C., Flessa, H., 2009. Estimating aboveground carbon in
model construction. For. Ecol. Manage. 104, 89–99.
a catchment of the Siberian forest tundra: combining satellite imagery and field
Baeza-Yates, R., Ribeiro-Neto, B., 1999. Modern Information Retrieval. ACM press,
inventory. Remote Sens. Environ. 113, 518–531.
New York.
Gallardo-Cruz, J.A., Meave, J.A., González, E.J., Lebrija-Trejos, E.E., Romero-Romero,
Bannari, A., Morin, D., Bonn, F., Huete, A., 1995. A review of vegetation indices.
M.A., Pérez-García, E.A., Gallardo-Cruz, R., Hernández-Stefanoni, J.L., Martorell,
Remote Sens. Rev. 13, 95–120.
C., 2012. Predicting tropical dry forest successional attributes from space: is the
Barbosa, J., Broadbent, E., Bitencourt, M., 2014. Remote sensing of aboveground
key hidden in image texture? PLoS ONE 7, e30506.
biomass in tropical secondary forests: a review. Int. J. Forestry Res. 2014.
Gao, X., Huete, A.R., Ni, W., Miura, T., 2000. Optical-biophysical relationships of
Bastin, J.-F., Barbier, N., Couteron, P., Adams, B., Shapiro, A., Bogaert, J., De Cannière,
vegetation spectra without background contamination. Remote Sens. Environ.
C., 2014. Aboveground biomass mapping of African forest mosaics using canopy
74, 609–620.
texture analysis: towards a regional approach. Ecol. Appl.
Gara, T., Murwira, A., Chivhenge, E., Dube, T., Bangira, T., 2014. Estimating wood
Basuki, T.M., Skidmore, A.K., van Laake, P.E., van Duren, I., Hussin, Y.A., 2011. The
volume from canopy area in deciduous woodlands of Zimbabwe. Southern
potential of spectral mixture analysis to improve the estimation accuracy of
Forests: a J. Forest Sci. 76, 237–244.
tropical forest biomass. Geocarto Int. 27, 329–345.
Godsmark, R., 2010. The South African forestry and forest products industry.
Belsley, D.A., 1991. Conditioning Diagnostics. Wiley Online Library.
Forestry South Africa.
Blaschke, T., Hay, G.J., Kelly, M., Lang, S., Hofmann, P., Addink, E., Queiroz Feitosa, R.,
Goh, J., Miettinen, J., Chia, A.S., Chew, P.T., Liew, S.C., 2014. Biomass estimation in
van der Meer, F., van der Werff, H., van Coillie, F., 2014. Geographic object-based
humid tropical forest using a combination of ALOS PALSAR and SPOT 5 satellite
image analysis-towards a new paradigm. ISPRS J. Photogramm. Remote Sens.
imagery. Asian J. Geoinform. 13.
87, 180–191. _ Filippi, A.M., Randall, J., 2014. Estimation of floodplain aboveground
Güneralp, I.,
Bredenkamp, B., 2000. Volume and mass of logs and standing trees. Southern
biomass using multispectral remote sensing and nonparametric modeling. Int. J.
African Institute of Forestry, Menlo Park.
Appl. Earth Obs. Geoinf. 33, 119–126.
Carlson, T.N., Ripley, D.A., 1997. On the relation between NDVI, fractional vegetation
Haralick, R.M., 1979. Statistical and structural approaches to texture. Proc. IEEE 67,
cover, and leaf area index. Remote Sens. Environ. 62, 241–252.
786–804.
Carreiras, J.M.B., Vasconcelos, Maria J., Lucas, R.M., 2012. Understanding the
Haralick, R.M., Shanmugam, K., Dinstein, I.H., 1973. Textural features for image
relationship between aboveground biomass and ALOS PALSAR data in the
classification. Syst., Man Cybernet., IEEE Trans SMC-3, 610–621.
forests of Guinea-Bissau (West Africa). Remote Sens. Environ. 121, 426–442.
Hardisky, M., Klemas, V., Smart, M., 1983. The influence of soil salinity, growth form,
Champion, I., Da Costa, J.P., Godineau, A., Villard, L., Dubois-Fernandez, P., Le Toan,
and leaf moisture on the spectral radiance of. Spartina alterniflora, 77–83.
T., 2013. Canopy structure effect on SAR image texture versus forest biomass
Hastie, T., Tibshirani, R., Friedman, J., 2001. The Elements of Statistical Learning Data
relationships. EARSeL eProceedings 12, 25–32.
Mining, Inference, and Prediction, 2nd ed. Springer, New York.
Champion, I., Dubois-Fernandez, P., Guyon, D., Cottrel, M., 2008. Radar image
Henry, M., Picard, N., Trotta, C., Manlay, R.J., Valentini, R., Bernoux, M., Saint-André,
texture as a function of forest stand age. Int. J. Remote Sens. 29, 1795–1800.
L., 2011. Estimating tree biomass of sub-Saharan African forests: a review of
Chinembiri, T.S., Bronsveld, M.C., Rossiter, D.G., Dube, T., 2013. The precision of C
available allometric equations. Silva Fennica 45, 477–569.
stock estimation in the Ludhikola watershed using model-based and design-
Huete, A.R., 1988. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25,
based approaches. Nat. Resour. Res. 22, 297–309.
295–309.
Choudhury, B.J., 1987. Relationships between vegetation indices, radiation
Ingram, J.C., Dawson, T.P., Whittaker, R.J., 2005. Mapping tropical forest structure in
absorption, and net photosynthesis evaluated by a sensitivity analysis.
southeastern Madagascar using remote sensing and artificial neural networks.
Remote Sens. Environ. 22, 209–233.
Remote Sens. Environ. 94, 491–507.
Cutler, M., Boyd, D., Foody, G., Vetrivel, A., 2012. Estimating tropical forest biomass
Irons, J.R., Dwyer, J.L., Barsi, J.A., 2012. The next Landsat satellite: the Landsat data
with a combination of SAR image texture and Landsat TM data: an assessment
continuity mission. Remote Sens. Environ. 122, 11–21.
of predictions between regions. ISPRS J. Photogramm. Remote Sens. 70, 66–77.
Jordan, C.F., 1969. Derivation of leaf-area index from quality of light on the forest
Darvishzadeh, R., Matkan, A., Eskandari, N., 2011. Evaluation of ALOS-AVNIR2
floor. Ecology, 663–666.
spectral indices for prediction of rice biomass. J. Geograph. Landscape.
Kelsey, K.C., Neff, J.C., 2014. Estimates of aboveground biomass from texture
De’ath, G., 2007. Boosted trees for ecological modeling and prediction. Ecology 88,
analysis of landsat imagery. Remote Sens. 6, 6407–6422.
243–251.
Kimes, D., Markham, B., Tucker, C., McMurtrey III, J., 1981. Temporal relationships
Dekker, R.J., 2003. Texture analysis and classification of ERS SAR images for map
between spectral response and agronomic variables of a corn canopy. Remote
updating of urban areas in The Netherlands. Geosci. Remote Sens., IEEE Trans.
Sens. Environ. 11, 401–411.
41, 1950–1958.
Koch, B., 2010. Status and future of laser scanning, synthetic aperture radar and
Dovey, S.B., 2009. Estimating biomass and macronutrient content of some
hyperspectral remote sensing data for forest biomass assessment. ISPRS J.
commercially important plantation species in South Africa. Southern Forests
Photogramm. Remote Sens. 65, 581–590.
71, 245–251.
Kohavi, R., 1995. A study of cross-validation and bootstrap for accuracy estimation
Dube, T., Gumindoga, W., Chawira, M., 2014a. Detection of land cover changes
and model selection. IJCAI, 1137–1145.
around Lake Mutirikwi, Zimbabwe, based on traditional remote sensing image
Kuplich, T., Curran, P., Atkinson, P., 2005. Relating SAR image texture to the biomass
classification techniques. African J. Aquatic Sci. 39, 89–95.
of regenerating tropical forests. Int. J. Remote Sens. 26, 4829–4854.
Dube, T., Mutanga, O., 2015. Evaluating the utility of the medium-spatial resolution
Laurent, V.C., Schaepman, M.E., Verhoef, W., Weyermann, J., Chávez, R.O., 2014.
Landsat 8 multispectral sensor in quantifying aboveground biomass in uMgeni
Bayesian object-based estimation of LAI and chlorophyll from a simulated
catchment, South Africa. ISPRS J. Photogramm. Remote Sens. 101, 36–46.
Sentinel-2 top-of-atmosphere radiance image. Remote Sens. Environ. 140, 318–
Dube, T., Mutanga, O., Elhadi, A., Ismail, R., 2014b. Intra-and-inter species biomass
329.
prediction in a plantation forest: testing the utility of high spatial resolution
Leathwick, J.R., Elith, J., Francis, M.P., Hastie, T., Taylor, P., 2006. Variation in
Spaceborne multispectral RapidEye sensor and advanced machine learning
demersal fish species richness in the oceans surrounding New Zealand: an
algorithms. Sensors 14, 15348–15370.
analysis using boosted regression trees. Mar. Ecol. Prog. Ser. 321, 267–281.
Dube, T., Mutanga, O., Ismail, R., in-press. Predicting Eucalyptus stand volumes in
Lu, D., 2005. Aboveground biomass estimation using Landsat TM data in the
African environments: an analysis using stochastic gradient boosting with
Brazilian Amazon. Int. J. Remote Sens. 26, 2509–2525.
multisource spatial data. Int. J. Remote Sens.
Lu, D., 2006. The potential and challenge of remote sensing-based biomass
Eckert, S., 2012. Improved forest biomass and carbon estimations using texture
estimation. Int. J. Remote Sens. 27, 1297–1328.
measures from WorldView-2 satellite data. Remote Sens. 4, 810–829.
Matthew, M.W., Adler-Golden, S.M., Berk, A., Richtsmeier, S.C., Levine, R.Y.,
El-Askary, H., Abd El-Mawla, S.H., Li, J., El-Hattab, M.M., El-Raey, M., 2014. Change
Bernstein, L.S., Acharya, P.K., Anderson, G.P., Felde, G.W., Hoke, M.L., 2000.
detection of coral reef habitat using Landsat-5 TM, Landsat 7 ETM+ and Landsat
Status of atmospheric correction using a MODTRAN4-based algorithm,
8 OLI data in the Red Sea (Hurghada, Egypt). Int. J. Remote Sens. 35, 2327–2346.
AeroSense 2000. Int. Soc. Opt. Photon., 199–207
Elith, J., Leathwick, J.R., Hastie, T., 2008. A working guide to boosted regression trees.
Moisen, G.G., Freeman, E.A., Blackard, J.A., Frescino, T.S., Zimmermann, N.E.,
J. Anim. Ecol. 77, 802–813.
Edwards Jr, T.C., 2006. Predicting tree species presence and basal area in
Elvidge, C.D., Chen, Z., 1995. Comparison of broad-band and narrow-band red and
Utah: a comparison of stochastic gradient boosting, generalized additive
near-infrared vegetation indices. Remote Sens. Environ. 54, 38–48.
models, and tree-based methods. Ecol. Model. 199, 176–187.
Englhart, S., Keuck, V., Siegert, F., 2012. Modeling aboveground biomass in tropical
Mousivand, A., Menenti, M., Gorte, B., Verhoef, W., 2014. Global sensitivity analysis
forests using multi-frequency SAR data – A comparison of methods. Select. Top.
of the spectral radiance of a soil-vegetation system. Remote Sens. Environ. 145,
Appl. Earth Observat. Remote Sens., IEEE J. 5, 298–306.
131–144.
32 T. Dube, O. Mutanga / ISPRS Journal of Photogrammetry and Remote Sensing 108 (2015) 12–32
Mutanga, O., Adam, E., Cho, M.A., 2012. High density biomass estimation for
wetland vegetation using WorldView-2 imagery and random forest regression
algorithm. Int. J. Appl. Earth Obs. Geoinf. 18, 399–406.
Mutanga, O., Skidmore, A.K., 2004. Narrow band vegetation indices overcome the
saturation problem in biomass estimation. Int. J. Remote Sens. 25, 3999–4014.
Nemani, R., Pierce, L., Running, S., Band, L., 1993. Forest ecosystem processes at the
watershed scale: sensitivity to remotely-sensed Leaf Area Index estimates. Int. J.
Remote Sens. 14, 2519–2534.
Nichol, J.E., Sarker, M.R., 2011. Improved biomass estimation using the texture
parameters of two high-resolution optical sensors. Geosci. Remote Sens., IEEE
Trans. 49, 930–948.
Pahlevan, N., Schott, J.R., 2013. Leveraging EO-1 to evaluate capability of new
generation of landsat sensors for coastal/inland water studies. Select. Top. Appl.
Earth Observat. Remote Sens., IEEE J. 6, 360–374.
Pandey, U., Kushwaha, S., Kachhwaha, T., Kunwar, P., Dadhwal, V., 2010. Potential of
Envisat ASAR data for woody biomass assessment. Tropical Ecol. 51, 117.
Penman, J., Gytarsky, M., Hiraishi, T., Krug, T., Kruger, D., Pipatti, R., Buendia, L.,
Miwa, K., Ngara, T., Tanabe, K., 2003. Good practice guidance for land use, land-
use change and forestry. Intergovernmental Panel on Climate Change: Institute
for Global Environmental Strategies, Kanagawa, Japan. <http://www.ipcc-nggip.
iges.or.jp/public/gpglulucf/gpglulucf_files/GPG_LULUCF_FULL.pdf>.
Perkins, T., Adler-Golden, S., Matthew, M., Berk, A., Anderson, G., Gardner, J., Felde,
G., 2005. Retrieval of atmospheric properties from hyper and multispectral
imagery with the FLAASH atmospheric correction algorithm. Remote Sens. Int.
Soc. Opt. Photon., pp 59790E–59790E-59711.
Pinto, N., Simard, M., Dubayah, R., 2012. Using InSAR coherence to map stand age in
a boreal forest. Remote Sens. 5, 42–56.
Ploton, P., Pélissier, R., Barbier, N., Proisy, C., Ramesh, B., Couteron, P., 2013. Canopy
Texture Analysis for Large-scale Assessments of Tropical Forest Stand Structure
and Biomass, Treetops at Risk. Springer, pp. 237–245.
Podest, E., Saatchi, S., 2002. Application of multiscale texture in classifying JERS-1
radar data over tropical vegetation. Int. J. Remote Sens. 23,
1487–1506.
Riggins, J.J., Tullis, J.A., Stephen, F.M., 2009. Per-segment aboveground forest
biomass estimation using LIDAR-derived height percentile statistics. GISci.
Remote Sens. 46, 232–248.
Rock, B., Vogelmann, J., Williams, D., Vogelmann, A., Hoshizaki, T., 1986. Remote
detection of forest damage. BioScience-Am. Instit. Biol. Sci. 36.
Rouse, J.W., Haas, R.H., Schell, J.A., Deering, D.W., Harlan, J.C., 1974. Monitoring the
Vernal Advancements and Retrogradation (Greenwave Effect) of Nature
Vegetation; NASA/GSFC Final Report. NASA, Greenbelt, MD, USA.
Santos, J.R., Freitas, C.C., Araujo, L.S., Dutra, L.V., Mura, J.C., Gama, F.F., Soler, L.S.,
Sant’Anna, S.J., 2003. Airborne P-band SAR applied to the aboveground biomass
studies in the Brazilian tropical rainforest. Remote Sens. Environ. 87, 482–493.
Sappi, 1993. Forest Land Types of the Natal Region Sappi Forests Research, Howick,
KwaZulu Natal, South Africa.
Sarker, L.R., Nichol, J.E., 2011. Improved forest biomass estimates using ALOS
AVNIR-2 texture indices. Remote Sens. Environ. 115, 968–977.
Sarker, M.L.R., Nichol, J., Iz, H.B., Ahmad, B.B., Rahman, A.A., 2013. Forest biomass
estimation using texture measurements of high-resolution dual-polarization C-
band SAR data. Geosci. Remote Sens., IEEE Trans. 51, 3371–3384.
Scott, D.F., Lesch, W., 1997. Streamflow responses to afforestation with Eucalyptus
grandis and Pinus patula and to felling in the Mokobulaan experimental
catchments, South Africa. J. Hydrol. 199, 360–377.
Shimada, M., Itoh, T., Motooka, T., Watanabe, M., Shiraishi, T., Thapa, R., Lucas, R.,
2014. New global forest/non-forest maps from ALOS PALSAR data (2007–2010).
Remote Sens. Environ.
Singh, M., Malhi, Y., Bhagwat, S., 2014. Biomass estimation of mixed forest
landscape using a Fourier transform texture-based approach on very-high-
resolution optical satellite imagery. Int. J. Remote Sens. 35, 3331–3349.
Tarpley, J., Schneider, S., Money, R.L., 1984. Global vegetation indices from the
NOAA-7 meteorological satellite. J. Climate Appl. Meteorol. 23, 491–494.
Teillet, P., Staenz, K., William, D., 1997. Effects of spectral, spatial, and radiometric
characteristics on remote sensing vegetation indices of forested regions.
Remote Sens. Environ. 61, 139–149.
Thenkabail, P.S., Smith, R.B., De Pauw, E., 2000. Hyperspectral vegetation indices
and their relationships with agricultural crop characteristics. Remote Sens.
Environ. 71, 158–182.
Thenkabail, P.S., Stucky, N., Griscom, B.W., Ashton, M.S., Diels, J., van der Meer, B.,
Enclona, E., 2004. Biomass estimations and carbon stock calculations in the oil
palm plantations of African derived savannas using IKONOS data. Int. J. Remote
Sens. 25, 5447–5472.
Tucker, C.J., 1979. Red and photographic infrared linear combinations for
monitoring vegetation. Remote Sens. Environ. 8, 127–150.
Vashum, K.T., Jayakumar, S., 2012. Methods to estimate above-ground biomass and
carbon stock in natural forests-a review. J. Ecosyst. Ecography 2, 1–7.
Wessels, N.O., Kassier, H.W., 1985. A computerised system for forest management
and silvicultural planning and control in even-aged plantation forestry. S Afr.
For. J. 132, 62–64.
Wiegand, C., Richardson, A., Escobar, D., Gerbermann, A., 1991. Vegetation indices in
crop assessments. Remote Sens. Environ. 35, 105–119.
Wulder, M.A., LeDrew, E.F., Franklin, S.E., Lavigne, M.B., 1998. Aerial image texture
information in the estimation of northern deciduous and mixed wood forest
leaf area index (LAI). Remote Sens. Environ. 64, 64–76.
Xu, X., Du, H., Zhou, G., Ge, H., Shi, Y., Zhou, Y., Fan, W., Fan, W., 2011. Estimation of
aboveground carbon stock of Moso bamboo (Phyllostachys heterocycla var.
pubescens) forest with a Landsat Thematic Mapper image. Int. J. Remote Sens.
32, 1431–1448.