International Journal of Applied Earth Observation and Geoinformation 52 (2016) 422–432

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation and

Geoinformation
journal homepage: www.elsevier.com/locate/jag

Enhancing the performance of regional land cover mapping

Weicheng Wu a,∗ , Claudio Zucca b , Fadi Karam c , Guangping Liu d
a
State-Key Lab of Nuclear Resources and Environment, East China Institute of Technology (ECIT), Nanchang, 330013 Jiangxi, China
b
ICARDA (International Center for Agricultural Research Center in the Dry Areas), Amman, Jordan
c
Litani River Authority, Beirut, Lebanon
d
Faculty of Sciences, East China Institute of Technology (ECIT), 330013, Nanchang, Jiangxi, China

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

Article history: Different pixel-based, object-based and subpixel-based methods such as time-series analysis, decision-
Received 12 May 2016 tree, and different supervised approaches have been proposed to conduct land use/cover classification.
Received in revised form 19 July 2016 However, despite their proven advantages in small dataset tests, their performance is variable and less sat-
Accepted 20 July 2016
isfactory while dealing with large datasets, particularly, for regional-scale mapping with high resolution
data due to the complexity and diversity in landscapes and land cover patterns, and the unacceptably long
Keywords:
processing time. The objective of this paper is to demonstrate the comparatively highest performance
Multisource data integration
of an operational approach based on integration of multisource information ensuring high mapping
Phenological contrast
Topographic features
accuracy in large areas with acceptable processing time. The information used includes phenologically
Separability contrasted multiseasonal and multispectral bands, vegetation index, land surface temperature, and topo-
Accuracy graphic features. The performance of different conventional and machine learning classifiers namely
Malahanobis Distance (MD), Maximum Likelihood (ML), Artificial Neural Networks (ANNs), Support Vec-
tor Machines (SVMs) and Random Forests (RFs) was compared using the same datasets in the same IDL
(Interactive Data Language) environment. An Eastern Mediterranean area with complex landscape and
steep climate gradients was selected to test and develop the operational approach. The results showed
that SVMs and RFs classifiers produced most accurate mapping at local-scale (up to 96.85% in Overall
Accuracy), but were very time-consuming in whole-scene classification (more than five days per scene)
whereas ML fulfilled the task rapidly (about 10 min per scene) with satisfying accuracy (94.2–96.4%).
Thus, the approach composed of integration of seasonally contrasted multisource data and sampling
at subclass level followed by a ML classification is a suitable candidate to become an operational and
effective regional land cover mapping method.
© 2016 Elsevier B.V. All rights reserved.

1. Introduction
Land cover (LC) and land use (LU) data are fundamental inputs
for a wide range of environmental planning, management and
research applications. Nowadays, LC mapping mostly relies on
remote sensing building on more than 40 years of scientific
research and technological developments from local to global scale
(Atkinson and Tatnall, 1997; Chen et al., 2015; DeFries et al., 1998;
Friedl et al., 2002; Gong et al., 1992, 2013; Hansen et al., 2000;
Haralick et al., 1973; Wu and Zhang, 2003; Wu et al., 2013a). How-
ever, accuracy and reliability may become a challenge when using
high resolution data for regional and global mapping. For example,
as reported by Gong et al. (2013) concerning their global LC map-
∗ Corresponding author.
E-mail addresses: Wuwc123@gmail.com, wchwu@ecit.cn (W. Wu).
ping using Landsat data, the Overall Accuracies (OAs) were below
70% for all continental-scale and below 75% for most national-scale
maps except for some countries like Algeria, Saudi Arabia, Libya
where LC patterns are simple.
The conventional classification approaches adopted pattern
recognition techniques including both supervised and unsuper-
vised algorithms, assuming that the study area is composed of a
number of unique internally homogenous classes that are mutu-
ally exclusive (Townshend, 1984). However, such assumption is not
applicable to most natural or semi-natural areas where there are
mixed pixels (Adams et al., 1995; Atkinson, 2005; Hill and Schutt,
2000; Van Der Meer, 1995), and especially LC types exist as continua
rather than as a mosaic of discrete classes (Foody et al., 1992; Kent
et al., 1997; Wu and Zhang, 2003). As a result, the classes intergrade
showing a low degree of separability, and cannot be distinguished
by means of sharp boundaries (Foody et al., 1992). The separability
of classes can be evaluated by the Jeffreys-Matusita Distance (JMD)

http://dx.doi.org/10.1016/j.jag.2016.07.014
0303-2434/© 2016 Elsevier B.V. All rights reserved.
 W. Wu et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 422–432 423

according to Swain and King (1973) and Richards and Jia (2006). other widely adopted supervised approaches. The specific objective
For the pair of classes i and j, this distance can be expressed as: is to quantify the achieved improvement in terms of separability of
classes, accuracy of the classification, and processing time by inte-
JMDij = 2(1 − e−˛ ) (1)
gration of multisource high resolution data for area with complex
where landscape.
1 T  −1  
˛= i − j (Ci + Cj )/2 i − j
8 2. Data and methods
1 1
  
+ ln ( |Ci + Cj |)/ |Ci | × |Cj | (2) 2.1. Study area
2 2
Ci —the covariance matrix of class i; i —the mean vector of class i;
The study area is located in the Eastern Mediterranean Region
ln—the natural logarithm function; T—the transposition function;
and coincides with the area covered by Landsat scenes with
and |Ci |—the determinant of Ci ; the same meanings for the counter-
path/row numbers of 174/35–174/37 (Fig. 1). This area was cho-
parts of class j. JMD ranges from 0 to 2.0; when it is below 1.0, two
sen because it is a dryland characterized by steep climatic gradients
classes (of a class-pair) are not separable; when it is between 1.0
with various landforms and complex LC patterns, thus a challenging
and 1.5, two classes are separable but with confusion, and when it
site for remote sensing-based LC mapping. Two subset sites with
is between 1.5 and 1.9, two classes are clearly separable; only when
contrasting LC and LU characteristics were also defined (Fig. 1) for
JMD is above 1.9 the class-pair is completely separable.
experimental purposes as explained below.
For poorly separable classes the accuracy of classification is the
In the study area rainfall is mostly concentrated between
major problem in LC mapping. For this purpose, a number of authors
November and April and ranges from around 650 mm on the
have explored the possibility to improve mapping accuracy by tak-
western coastal slopes to less than 100 mm in the eastern dry
ing into account the texture (Gong et al., 1992; Haralick et al., 1973;
rangelands and deserts. Three main landforms are respectively,
Zhang, 2001) or by object-based segmentation (Mao and Jain, 1992;
from the west to the east, the coastal plains and piedmont, the
Blaschke, 2010; Pu et al., 2011) or by combining both pixel- and
mountain–valley–mountain sequence of the north-south stretch-
object-based approaches (Huth et al., 2012; Chen et al., 2015). In
ing coastal ranges, and the eastern plateau. Natural vegetation
addition to the traditional unsupervised (e.g., IsoData, K-Means)
cover mainly consists of coniferous and broadleaf forests in the
and supervised algorithms, e.g., Mahalanobis Distance (MD) and
highlands, shrublands and maquis in the mountain slopes, and
Maximum Likelihood (ML), a number of authors have introduced
herbaceous rangelands in the eastern hills and plateau (Wu, 2014).
machine learning algorithms that can capture the non-parametric
Irrigation is mainly concentrated in the Aleppo Plain, Orontes
signatures of classes such as Artificial Neural Networks (ANNs,
and Litani watersheds and Jordan River valley. The main spring
Atkinson and Tatnall, 1997; Benediktsson et al., 1990; Kavzoglu
crops are irrigated wheat and vegetables, and rainfed barley,
and Mather, 2003), Support Vector Machines (SVMs, Foody and
whereas summer crops are irrigated cotton, maize, sunflower,
Mathur, 2004; Huang et al., 2002; Kavzoglu and Colkesen, 2009;
sesame, water melon and vegetables. Olive is widespread in rainfed
Vapnik and Lerner, 1963) and Random Forests (RFs, Breiman, 2001;
areas, interleaved with fig and pistachio. Orchards including citrus,
Rodriguez-Galiano et al., 2012; Waske et al., 2012).
apple, cherry, peach, etc., are mainly distributed in the western
For mixed pixels, various subpixel processing techniques have
coastal plains and slopes. Date, banana and vineyards are mostly
been proposed to decompose land cover fraction or to improve
present in the Bekaa and Jordan River valleys. The major land
LC mapping accuracy, e.g., linear spectral unmixing (Adams et al.,
use/cover classes of the study area are summarized in Table 1.
1986; Foody and Cox, 1994; Hill and Schutt, 2000; Lu and Weng,
In Table 1 the category “Conifers” does not only include
2004; Smith et al., 1990; Van Der Meer, 1995), linear optimization
monospecific pine and/or cedar stands, but also mixed formations
(Verhoeye and De Wulf, 2002), Hopfield neural network (Tatem
including broadleaved species. The distinction between forests
et al., 2002), pixel-swapping (Atkinson, 2005), subpixel/pixel
(Conifers and Broadleaf) and “Woodland” or “Woody Shrubland”,
attraction (Mertens et al., 2006), etc.
is based on the FAO Land Cover Classification System (LCCS, Di
Some authors have also integrated a set of single or time-series
Gregorio and Jansen, 2000): forests have tree canopy cover (CC)
vegetation indices (VIs) such as NDVI (Normalized Difference Vege-
above 60%, whereas CC is between 20 and 60% for woodlands and
tation Index) or EVI (Enhanced Vegetation Index) and land surface
less than 20% for sparse woodlands (Wu et al., 2013b). Since sparse
temperature (LST) to undertake LC mapping (Friedl et al., 2002;
woody formations are generally used as grazing land in the study
Loveland et al., 2000; Lu et al., 2014). Furthermore, topographic fea-
area, this class was considered as part of the “Rangelands”.
tures have been employed in LC classification to improve accuracy
(Benediktsson et al., 1990; Rodriguez-Galiano et al., 2012), particu-
larly by the ESA-funded DesertWatch project (Pace et al., 2006; ESA, 2.2. Data
2008), based on the assumption that landscape features restrain to
a certain extent land use or land cover. For example, irrigated land Landsat 5 TM (Thematic Mapper) spring (01 May 2007) and
generally occurs in flat to gently sloping land. Phenological pat- summer (21 August 2007) images were acquired for the scene
terns and features (Zhu and Wan, 1963) have also played a role in 174/35. Landsat 8 OLI (Operational Land Imager) and TIRS (Ther-
LC mapping (Friedl et al., 2002; Jia et al., 2014; Lu et al., 2014). The mal Infrared Sensor) data were obtained for the scenes 174/36 (02
above mentioned DesertWatch project and Rodriguez-Galiano et al. April 2014 and 24 August 2014) and 174/37 (18 April 2014 and
(2012) used paired season-contrasted spring and summer images 24 August 2014). The two dates represent respectively the spring
instead of time-series data to enhance LC classification. vegetative maximum and the summer minimum, they are thus
The goal of this research is to demonstrate the performance highly contrasted. Extensive ground-truthing work was conducted
of a LC mapping procedure based on the integrated use of the in the period 2007–2011 in Syria and in 2013–2014 in Lebanon and
phenology-contrasted information including multispectral (MS) Jordan (GPS locations in Fig. 1). Google Earth was used as a comple-
bands of images, GDVI (Generalized Difference Vegetation Index) mentary source for areas not covered by field work. SRTM (Shuttle
which is more sensitive than other VIs for dryland characterization Radar Topography Mission) DEM data (90 m in resolution) were
(Wu, 2014), LST, and topographic features extracted from a Digi- obtained and used to generate elevation (E), slope (S), and aspect
tal Elevation Model (DEM), and to compare it with that of some (A) information.
424 W. Wu et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 422–432

Fig. 1. Location of the study area (defined by Landsat scenes 174/35–174/37) including the two subsets, and the field observation points.

2.3. Methods and Cubic Convolution for DEM data, to minimize the defect of
bad pixels in SRTM data.
2.3.1. Dataset preparation 3) Spring and summer GDVI with power number of 2 was derived
The following major processing steps were undertaken for the from the multispectral bands; LST was calculated in terms of
scene 174/35 for developing and testing the approaches, while Chander et al. (2009) and USGS (2015):
the other two scenes were used for regional-scale application as
explained in section 2.3.5: LST = K 2 /ln((K 1 /L ␭ ) + 1) (3)

where L —spectral radiance, K1 and K2 —conversion constants. For

1) Atmospheric correction on Landsat images was performed by
means of the COST model (Chavez, 1996): DOS (Dark-object Sub-
traction) technique (Chavez, 1988) was used to determine the
haze of each band, then spectral radiance in digital number was
converted into reflectance.
2) Spring and summer Landsat images and SRTM data were resized
to the same image dimension (7156 × 6858) and pixel size
(30 m). Nearest Neighbor resampling was used for the images
TM thermal band, K1 and K2 are respectively 607.76 and 1260.56
(Chander et al., 2009); for Landsat 8 TIRS, K1 and K2 are respectively
774.89 and 1321.08 for band 10, and 480.89 and 1201.84 for band
11 (USGS, 2015). It is worthy of mention that there is a difference of
about 0.15–1.8 K of LST from the Landsat 8 bands 10 and 11; hence,
the below mentioned LST is actually their average.
4) E, S, and A derived from SRTM data.
 W. Wu et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 422–432 425

Table 1
Land use/cover classes defined in the study area based on the field observation with reference to the European CORINE classification scheme (CEC, 1994).

Land Use/Cover Classes Description

Artificial areas
Built-Up Mainly urban (industrial residential, etc.), roads, infrastructures
Mining or Construction Sites Mining areas or sites under construction

Agricultural areas
Spring Irrigation Wheat and vegetables, locally barley
Summer Irrigation Cotton, maize, melon, sesame, and vegetables
Rotated Irrigation Perennial crops irrigated in both spring and summer
Spring Rainfed Mainly barley, locally wheat, cultivated for harvesting
Orchards Citrus, apple, cherry, pistachio, date palm, banana, etc. Can include olive groves.
Terraced Rainfed with Olives In the mountainous terraces, rainfed crops (barley and wheat) cultivated under or mixed with olive groves or fruit trees
Vineyards Vineyards
Olives Olive groves
Greenhouse Greenhouses
Fallows Cropland not being cultivated at the date.
Rainfed Pastures Mainly barley, locally wheat, cultivated for grazing, not for harvesting; managed grassland in Golan Heights

Natural and semi-natural areas

Broadleaf Forests Deciduous and sclerophyll tree formations, and maquis, with tree CC > 60%
Conifers Mainly pine, locally cedar, with tree CC > 60%
Hylophytes Shrub formations in salt marsh in Jordan Valley with CC usually >60%
Woodlands Woody shrublands including maquis, with tree CC between 20% and 60%
Rangelands Unmanaged grassland locally including sparse trees or shrubs, with tree CC < 20%
Bare Lands Bare soil, bare rocks, and deserts with vegetation cover generally <5%
Saline Land Salt marshes
Beaches Coastal sand deposits
Snow Snow cover

Other areas
Water Bodies Sea, lakes, artificial water bodies
Burnt Scars Burnt areas

5) Multisource and multifactor datasets were compiled as follows
to test the effects of different degrees of integration of spectral,
phenological and topographical information on the classifica-
tion performance:
a. 3-band datasets: Three uncorrelated bands (MS1, 4 and 7) in
reflectance from both the spring and summer images were
compiled into two 3-band datasets as bands 1, 2 and 3 in
visible spectral region of Landsat TM and ETM+ data are mutu-
ally correlated (R2 = 0.982 − 0.984), and so are bands 5 and 7
in shortwave infrared spectral region; and the combination
of MS741 as RGB constitutes the best pseudo-natural color
composite after atmospheric correction (Wu, 2003). These
uncorrelated bands were thus taken to avoid redundancy of
spectral information and saving running time;
b. 6-band dataset: containing the two 3-band MS147
(spring + summer) datasets;
c. 8-band dataset: containing the 6-band dataset + two bands of
GDVI (spring + summer);
d. 10-band dataset: the 8-band dataset + two bands of LST
(spring + summer);
e. 12-band dataset: the 10-band dataset + E + S
f. 13-band dataset: the 12-band dataset + A.
6) Subsetting
For a number of classes such as Olives, Rangelands, Woodlands,
Conifers, and so on, multiple subclasses were also defined and sam-
pled to deal with the observed variability of classes with landscape.
For example, “Rangeland 1” for herbaceous rangeland, “Rangeland
2” for woody herbaceous rangeland, and “Rangeland 3” where the
latter mixed cover was observed on dark volcanic soils, etc. “Olive 1”
for olive groves on brown soils, “Olive 2” for those on light-colored,
whitish soils containing lime and/or gypsum, “Olive 3” for those
cultivated on dark volcanic soils, “Olive 4”, for young groves with
low canopy cover (e.g., CC < 5–10%, close to bare soil), and “Olive 5”
for mature olive groves with CC > 45–50%, and so on.
In total, 13, 19 and 44 classes and subclasses (including clouds
and shadows) were recognized respectively for the Subset 1, Sub-
set 2 and whole-scene datasets of 174/35, and as many as possible
ROIs were selected to cover the whole datasets for each class
and subclass. In total, the training samples (ROIs-1) accounted for
respectively 11.72%, 6.02%, and 6.79% of the total pixels to be clas-
sified in Subsets 1, 2 and whole scene datasets in agreement with
the 5–10% rate recommended by Zhuang et al. (1994).
An independent set of samples (ROIs-2) was selected in Subsets
1, 2 and whole-scene composites following the same principle as for
ROIs-1 for cross-validation purposes, covering respectively 12.82%,
6.52% and 11.08% of the total pixels.

The performance of different supervised classifiers (listed 2.3.3. Separability of classes

below) was tested at different dataset sizes, i.e., at both subset
and whole-scene levels. Hence, multisource datasets, Subset 1
(1384 × 1211 pixels) and Subset 2 (1943 × 1776 pixels), which are
different in both location and land use/cover types, were prepared.
2.3.2. Ground-truth sampling in subsets and whole scene
Based on the field observation and visual interpretation of the
acquired satellite images with reference to Google Earth, ground-
truth samples represented by regions of interests (ROIs) were
defined on the colorful composites for each land use/cover class
in Subset 1, Subset 2 and whole-scene datasets.
The separability of the classes was investigated on the whole-
scene datasets using JMD to fully consider the complexity in
landforms and variability in land use/cover. The impacts of differ-
ent degrees of multisource data integration (from 3 bands, 6 bands
up to 13 bands) on the separability of the problematic class-pairs
were quantified.
2.3.4. Performance of different classifiers
The conventional and machine learning supervised classifiers
namely MD, ML, ANNs, SVMs and RFs were tested on Subset 1, Sub-
set 2 and whole-scene datasets for evaluating their performances
426 W. Wu et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 422–432

Fig. 2. Land use/cover maps of Subset 1 produced from different supervised classifiers.
(a) Pseudo natural color composite of MS 7, 4 and 1 of the spring images as RGB; (b) Result of ANNs (6-band dataset, though its OA reached 95.61%, Woody shrublands were
missed and Olives misclassified as Orchards); (c) Result of RFs (8-band dataset; OA: 96.84%), (d) Result of ML (13-band dataset; OA: 96.08%), and (e) Result of SVMs (13-band
dataset; OA: 96.74%)

using the Overall Accuracy (OA), Kappa Coefficient (KC) and pro-
cessing time.
MD is a direction-sensitive distance classifier that uses statistics
for each class assuming all class covariances are equal, while ML
assumes that the statistics for each class in each band are normally
distributed and calculates the probability of a given pixel belonging
to a specific class (Richards and Jia, 2006). These two classifica-
tions were conducted without setting distance error or probability
threshold so that all pixels were classified.
ANNs are a layered feed-forward neural network classification
technique in which the multilayer perceptron (MLP) backpropaga-
tion algorithm is commonly used (Mas and Flores, 2008). In our test,
the logistic (or sigmoid) function was selected as activation func-
tion with hidden layer number (representing non-linear degree)
of 1 as suggested; and training iterations were respectively set
to 100 (Subset1), 150 (Subset2) and 1000 (Whole-scene) avoiding
overfitting. The training threshold contribution (␪), denoting the
contribution of the internal weight with respect to the activation
level of the node, ranges from 0 to 1.0; the training rate (␩), the
magnitude of the adjustment of the weights, also comes between
0 and 1.0 (a higher rate will speed up the training but will also
increase the risk of oscillations of the training result). The training
momentum rate (␣), varies from 0 to 1.0, and a higher ␣ will con-
duct training with larger steps than a lower one. The default values
of ␪, ␩ and ␣ are respectively 0.9, 0.2 and 0.9 within ENVI (ENvi-
ronment for Visualizing Images) 5.2 package. Kavzoglu and Mather
(2003) noted that selecting 0.1–0.2 for ␩ and 0.5–0.6 for ␣, a bet-
ter classification accuracy can be reached. Apart from the default
setting, we tested also the classification accuracy by tuning these
parameters as indicated.
 W. Wu et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 422–432 427

Table 2
Performance of the conventional supervised classifiers at Subsets 1 and 2. Time duration includes both training (TRN) and classification (CLS).

Datasets MD ML

OA KC Time (TRN+CLS) (s) OA KC Time (TRN+CLS) (s)

Subset 1 (1384×1211) 3-Band (spring) 55.05% 0.4191 8 (2+6) 88.75% 0.8145 13 (4+9)
6-Band 80.01% 0.6763 14 (2+12) 93.99% 0.8992 14 (4+10)
8-Band 90.40% 0.8404 12 (3+9) 93.85% 0.8969 13 (4+9)
10-Band 91.14% 0.8527 12 (3+9) 94.74% 0.9118 16 (4+12)
12-Band 91.06% 0.8515 33 (8+25) 96.01% 0.9329 17 (4+13)
13-Band 91.74% 0.8627 24 (8+26) 96.08% 0.9342 18 (4+14)

Subset 2 (1943×1776) 3-Band (spring) 71.25% 0.6666 13 (4+9) 80.53% 0.7709 14 (3+11)
6-Band 75.97% 0.7216 25 (5+20) 88.50% 0.8641 17 (3+14)
8-Band 80.27% 0.77 31 (7+24) 90.45% 0.8869 18 (4+14)
10-Band 82.99% 0.7995 33 (8+25) 93.34% 0.9208 19 (4+15)
12-Band 84.54% 0.8176 37 (10+27) 94.31% 0.9323 30 (4+26)
13-Band 84.87% 0.8215 37 (8+29) 94.41% 0.9334 31 (4+27)

Table 3
Performance of the machine learning classifiers at Subset 1 and Subset 2.

Datasets ANNs RFs SVMs

OA KC Time (TRN+CLA) OA KC Time (TRN+CLA) OA KC Time (TRN+CLA)

(min) (min) (min)

Subset 1 (1384×1211) 3-Band (spring) 93.18% 0.8928 27 (26+1) 96.39% 0.9436 19 (16+3) 95.59% 0.9307 19 (16+3)
6-Band 95.61% 0.9315 29 (22+1) 96.85% 0.9507 17 (14+3) 96.81% 0.9501 45 (22+23)
8-Band 90.49% 0.8674 17 (15+2) 96.84% 0.9507 25 (22+3) 96.94% 0.9523 47 (23+24)
10-Band 85.87% 0.772 29 (28+1) 96.38% 0.9436 22 (19+3) 96.59% 0.9469 53 (31+22)
12-Band 88.00% 0.8131 26 (25+1) 95.93% 0.9368 22 (20+2) 96.76% 0.9496 42 (20+ 22)
13-Band 87.49% 0.8057 32 (30+2) 95.23% 0.9528 24 (21+3) 96.74% 0.9494 56 (26+30)

Subset 2 (1943×1774) 3-Band (spring) 74.01% 0.6825 30 (28+2) 83.56% 0.8008 543
6-Band 89.18% 0.8701 31 (29+2) 88.91% 0.8664 532
8-Band 90.41% 0.885 31 (29+2) 95.71% 0.9487 59 (50+9) 95.71% 0.9466 397
10-Band 64.04% 0.5467 34 (32+2) 96.42% 0.9572 49 (41+8) 96.09% 0.9485 486
12-Band 70.49% 0.6369 24 (23+1) 96.21% 0.9547 185
13-Band 77.46% 0.7271 97 (95+2) 95.66% 0.9481 60 (54+ 6) 96.41% 0.9571 178

Table 4
Separability improvement by integrating multisource data.

Class-Pairs JMD

3-band 3-band 6-band 8-band 10-band 13-band

3MS147 (spring) 3MS147 (summer) 6MS147 6MS147+2GDVI 6MS147+2GDVI + 2T 6MS147 + 2GDVI
+2T + E + S + A

Built-Up ∧ Bare Lands 1.1544 1.5386 1.6163 1.8669 1.9928 1.9966

Built-Up ∧ Olives 1.2112 1.4819 1.5806 1.6496 1.8113 1.9234
Built-Up ∧ Rangelands 1.2689 1.4297 1.7165 1.8215 1.9249 1.9742
Built-Up ∧ Beach 1.2927 1.3691 1.7524 1.8488 1.954 1.9989
Bare Lands ∧ Olives 0.9967 1.1824 1.3753 1.9009 1.9598 1.9903
Bare Lands ∧ Rangelands 1.2077 1.3989 1.5868 1.7978 1.8851 1.9470
Rangelands ∧ Olives 0.5848 1.0687 1.3989 1.6231 1.6813 1.7866
Rangelands ∧ Woodland 1.1078 1.0346 1.3487 1.5461 1.7261 1.8405
Woodlands ∧ Olives 1.0892 0.7737 1.3762 1.8548 1.9525 1.9721
JMD Mean 1.1012 1.2531 1.5280 1.7677 1.8764 1.9366

For SVMs that use support vectors to maximize the margin and
find the optimal hyperplanes among the clusters (Huang et al.,
2002; Kavzoglu and Colkesen, 2009), different kernel type functions
namely linear, polynomial, radial basis and sigmoid were tested.
RFs are a combination of decision-tree classifiers such that each
tree depends on the values of a random vector sampled indepen-
dently and with the same distribution for all trees in the forest
(Breiman, 2001). For this classification, tree number in training was
set to 100 and number of features was determined by square root of
all features, and Gini Index was used to define the impurity (Waske
et al., 2012).
After the classification, the LC subclasses (e.g., the Olive and
Rangeland subclasses) were respectively merged together, and
checked against the ground-truth data (ROIs-2) to calculate the
final OA and KC to evaluate the performance of different classifiers.
The subpixel mapping as mentioned in Section 1 would also be
promising. However, linear spectral unmixing required selection
of endmember which is not straightforward in landscape complex
areas, and the endmember number may not exceed the band num-
ber, e.g., 3–4 (Adams et al., 1995; Lu and Weng, 2004), which can
hardly reflect the full spectrum of LC diversity in the study area. It
would be time-consuming as well if the class boundary was defined
by thresholding on the endmember components or on the end-
member ternary diagram (Adams et al., 1995; Lu and Weng, 2004).
Other soft algorithms were successfully tested in small sites with
simple land cover, e.g., 2–3 classes (Verhoeye and De Wulf, 2002;
Tatem et al., 2002; Atkinson, 2005; Mertens et al., 2006). It was not
sure whether these techniques were applicable in complex LC areas
(>20 classes). We hence decided not to test these techniques in this
study.
428 W. Wu et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 422–432

Table 5
Performance of different classifiers on the whole-scene datasets of the scene 174/35.

Supervised Classifiers and OA KC Time (TRN + CLA) Remark

Multisource Datasets (min)

Conventional MD (12-band) 84.80% 0.7882 6 Result largely acceptable but there is confusion
Classification MD (13-band) 83.53% 0.7704 6 Same as the above
ML (3-band, 75.70% 0.7193 8 Result with strong confusion
spring)
ML (6-band) 91.03% 0.8742 8 The majority of classes is clearly classified
ML (8-band) 93.41% 0.9074 8 Result is rather good with minor confusion
ML (10-band) 94.67% 0.9223 9 Similar to the above
ML (12-band) 95.15% 0.9319 10 Results very good with very minor confusion
between Built-Up and Beaches, and between
Orchards and Olives, etc.
ML (13-band) 95.26% 0.9334 10 Same as the above

Machine Learning ANNs (6-band) 91.44% 0.8798 441 (433 + 8) Despite of rather high OA, there was strong
Classification confusion between Built-Up and Beaches,
Rainfed Pastures and Rangelands, and so on
ANNs (8-band) 81.87% 0.7480 463 (450 + 13) Result is spurious
RFs N/A N/A N/A Not finished during 5-day test period
SVMs N/A N/A N/A Not finished during 5-day test period

Table 6
Accuracies of the whole-scene land use/cover categorization in the study area.
The results were obtained from the 13-band datasets with ML classification. PA—Producer’s Accuracy, UA—User’s Accuracy

Whole-scene Accuracy 174/35 174/36 174/37

OA% KC OA% KC OA% KC

95.26 0.9334 96.44 0.9228 94.20 0.9171

Class Accuracy PA (%) UA (%) PA (%) UA (%) PA (%) UA (%)

Artificial areas
Built-Up 95.25 98.52 94.74 99.64 94.81 99.38
Mining or Construction Sites 97.40 100.00 98.30 42.79 72.98 70.61

Agricultural areas
Spring Irrigation 97.00 99.35 94.77 92.94 90.81 92.55
Summer Irrigation 98.54 97.98 99.05 77.14 97.79 83.63
Rotated Irrigation 95.54 61.02 93.55 54.92 91.46 63.63
Spring Rainfed 97.37 98.70 94.05 96.18 76.63 93.12
Rainfed for Grazing (Pastures) 83.34 88.97 97.94 38.26 87.33 69.11
Orchards 94.07 98.05 91.99 54.63 46.27 76.21
Terraced Rainfed with Olives / / 96.99 94.36 / /
Vineyards / / / / 86.43 74.32
Olives 98.73 95.16 91.11 81.01 86.91 79.96
Greenhouse / / 98.22 21.86 100.00 1.45
Fallow 95.43 25.78 87.65 72.05 92.90 78.77

Natural and semi-natural areas

Broadleaf Forests 94.98 53.67 97.80 50.95 94.38 27.42
Conifers 90.97 96.19 90.85 95.28 85.52 94.53
Hylophytes / / / / 99.44 36.65
Woodlands 91.15 83.01 93.75 70.45 65.97 93.76
Rangelands 91.04 94.48 77.43 97.44 88.20 78.66
Bare Lands 36.80 88.43 89.79 86.69 93.84 95.73
Saline Land 99.91 12.67 99.77 100.00 99.96 100.00
Beaches 99.13 99.47 99.15 100.00 99.23 38.92
Snow / / 99.95 100.00 100.00 100.00

Other areas
Water Bodies 99.87 99.98 100.00 100.00 99.99 99.99
Burnt Scars / / 100.00 73.94 99.47 100.00

2.3.5. Regional-scale mapping with EnMap-Box (Waske et al., 2012; Van der Linden et al., 2015)
Based on the tests in the above Sections 2.3.2–2.3.4, the best installed on a PC equipped with 16 GB of RAM and Intel(R) Core
performed classifier together with the phenologically contrasted i7-4510 CPU (4 processors). All tests were undertaken in IDL (Inter-
multisource data integration and sampling scheme was proposed active Data Language) environment.
and applied to the other two scenes (174/36 and 174/37) for
regional-scale mapping. The ROIs-1 and ROIs-2 took up respectively
3. Results and discussion
8.90% and 5.82% of pixels, and 7.40% and 6.64% of pixels for the two
scenes.
3.1. Subset scales
For performance evaluation purposes the conventional and
machine learning classifications by ANNs and SVMs were con-
Among the conventional classifiers, MD and ML, especially ML,
ducted with software ENVI 5.2, and RFs classification was realized
could produce reasonable maps with high accuracy and reliabil-
 W. Wu et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 422–432 429

Fig. 3. Land use/cover maps of Subset 2 produced by different supervised classifiers.

(a) Pseudo natural color composite of the spring images MS741 as RGB; (b) Result of ANN (8-band dataset; OA: 90.41%); (c) Result of RFs (10-band; OA: 96.42%); (d) Result
of ML (13-band; OA: 94.41%); and (e) Result of SVMs (13-band; OA: 96.41%)

ity in both Subsets (Table 2). With the increase in band number
(higher degree of multisource data integration) the OA and KC
also increased, with best accuracy for 13-band integration (e.g.,
94.4–96.1% for products by ML classifier at both subset sites).
Concerning the machine learning algorithms, the ANNs classi-
fier produced maps with a high accuracy when applied to 6-band
datasets at Subsets 1 (95.61%) but not at Subset 2 (89.18–90.4%,
Table 3). It failed to generate satisfying results (low OA and KC)
with high band number datasets (e.g. >8). Additional tests using
different combinations of the model parameters, for example, ␪,
␩ and ␣ respectively set to 0.3–0.6, 0.5–0.8, 0.2–0.5, did not show
better results.
The RFs classifier showed consistently high accuracy in all cases,
from 3-band to 13-band datasets, in both sites. With the increase in
band number by adding T, E, S and A, no substantial improvement in
OA and KC was observed, implying that integration of phenology-
contrasted information is sufficient to enhance the performance of
RFs (Table 3).
SVMs performed very well at both subsets especially when
radial basis was selected as kernel function. Linear kernel type
saved about 20–26% of time for each running but had lower OA
by about 0.5–4.0% than radial basis. The increase in band number
also led to a further improvement in OA and KC similar to the case
of ML.
430 W. Wu et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 422–432

Fig. 4. Regional-scale land use/cover map of an Eastern Mediterranean area.

Note: OAs are respectively 95.26%, 96.44% and 94.20% for scenes 174/35, 174/36 and 174/37. White color represents clouds and shadows in spring and/or summer images.

In summary, the conventional classifier, ML, performed quite
well in categorizing land use/cover at both Subsets 1 and 2; its OA
increased with the integration degree of multisource information,
and it reached the best at 13-band datasets. The popularly applied
ANNs did not ideally perform as expected in the complex LC area
(e.g., Subset 2), whereas RFs and SVMs allowed to achieve highly
reliable mapping with OAs from 95.66% to 96.94%. The disadvan-
tage was, however, their long processing time, particularly SVMs.
With an increase of 105% in pixel number from Subset 1 to Sub-
set 2, the time used by ML increased from 17–18 s to 30–31 s only
(Table 2), yet, the increase was dramatic for SVMs, from 19–56 min
to 178–543 min. The maps produced by different classifiers (ML,
ANNs, RFs and SVMs) for Subsets 1 and 2 are respectively presented
in Figs. 2 and 3.
3.2. Whole-scene scale
3.2.1. Separability of classes
The separability of the LC classes was analyzed at whole-
scene scale (7156 × 6858 pixels), for different levels of multisource
data integration (from 3 to 13 bands). The results show that the
improvement in the separability of the easily confused class-pairs,
 W. Wu et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 422–432
namely Bare Lands and Olives, Rangelands and Olives, Woodlands
and Olives, Rangelands and Woodlands, and Built-Up and Bare
Lands, etc., is remarkable (Table 4), from a JMD mean of 1.10–1.25
(separable but with strong confusion) of the 3-band datasets to
1.94 (completely separable) of the 13-band dataset. Phenology-
contrasted multispectral (MS) and biophysical information (GDVI
and LST) accounted for 90–92% of the improvement in separability,
while topographic features contributed by 8–10% (Table 4).
3.2.2. Classification accuracy
The test on the performance of different classifiers on the whole-
scene datasets (7156 × 6858 pixels, 3–13 bands) using the same
whole-scene ROIs-1 and ROIs-2 revealed that MD ran fast and the
maps generated were largely reasonable (83.52–84.8% of OA for
both 12- and 13-band datasets) except for the confusion between
Olives and Built-Up, and between Beaches and Built-Up; yet, ML
showed much better performance, rapidly yielding LC maps with
OAs of 94.67–95.26% for datasets of more than 8 bands (Table 5).
Among the machine learning algorithms, ANNs classifier accom-
plished the whole-scene classification in about 7–8 h with an OA
reaching 91.44% for the 6-band dataset (Table 5); however, it
misclassified Olives into Rangelands, Barelands and Clouds, and
Beaches into Built-Up, among others. The best performed classifiers
at subset scales, RFs and SVMs, could not complete the whole-scene
classification during the 5-day test period.
Therefore, ML performed best at whole-scene scale; its map
derived from the 13-band dataset is presented in Fig. 4 and the
accuracies of different LC types are shown in Table 6.
3.3. Regional-scale mapping
To perform large scale (regional to global) LC mapping using high
resolution data such as Landsat, the capacity of the classification
approach to process the whole-scene datasets and produce reliable
maps of high accuracy within acceptable time duration are critical
factors.
Both SVMs and RFs algorithms have a strong capacity for group-
ing clusters and can hence deliver accurate classification results.
However, they are time-consuming and more suited to local scale
LC mapping with small datasets viewing that the very powerful
processing facilities used by Gong et al. (2013), for example, are
not available to most analysts and users. Large datasets can be
processed by tiling but one may face problems linked to abrupt
connections among different tiles after mosaicking and time-
consuming work to amend them.
The ML classifier is the most widely known and employed con-
ventional classifier thanks to its robustness (Huang et al., 2002;
Gong et al., 2013) though questioned for its parametric assumption.
As our tests revealed, after integrating the phenology-contrasted
spectral and biophysical information and topographic features, the
ML classification was able to sort out the low separability problem;
and sampling to the subclass level helped to resolve the problem
related to non-parametric signatures of classes.
Thus, the proposed integration procedure followed with a
subclass-level sampling and ML classification was considered as
a relevant approach and applied to the adjacent scenes 174/36
and 174/37 for regional-scale mapping. The results were satisfac-
tory, of which OAs are respectively 96.4% and 94.2%. The accuracies
of each LC class are illustrated in Table 6 and their LC maps are
demonstrated in Fig. 4.
4. Conclusions
This research developed an operational approach to map LC
in large areas using high resolution data. The effectiveness of the
method, based on the integration of both spectral and ancillary
431
information, was assessed by comparing the performance of the
most popular classification algorithms. The innovative aspect of
this research lies in the assessment of the gains deriving from
different levels of progressive integration of different pieces of
information that were individually reported as relevant by var-
ious publications. The experiments revealed that integration of
phenology-contrasted multisource data (MS, GDVI, LST) and topo-
graphic features can significantly improve the separability of the
problematic LC classes and the overall mapping accuracy. After such
integration, the multispectral space (e.g., 3-dimension) becomes
a high-dimension space (13-dimension), and the non-separable
clusters become separable. The ML classifier yielded the best per-
formance at whole-scene scale.
The tests were conducted in the Mediterranean region, the pro-
posed approach, however, can be applied to other climate-type
areas by taking the local phenological pattern into consideration.
In humid areas, NDVI or EVI should be used instead of GDVI which
gets easily saturated in the densely vegetated areas (Wu, 2014).
ASTER imagery could also be used. Datasets not including the ther-
mal band such as SPOT, CBERS, and RapidEye may yield a lower
accuracy, about 0.9–2.7% of degradation (see the difference in accu-
racy between 8-band and 10-band in Tables 2 and 5), nevertheless,
an OA of >92% could still be achievable.
Acknowledgements
The field observation was conducted while the first author
was working with ICARDA (International Center for Agricultural
Research in the Dry Areas). The study was supported by the research
fund of the State-Key Lab of Nuclear Resources and Environment,
ECIT (No. NRE1501) for Weicheng Wu, and by the CGIAR Research
Program in Dryland Systems (CRP-DS) fund for Claudio Zucca.
We thank Dr Claudia Kunzer for her useful discussion and sug-
gestions about the study. Landsat images were acquired from
the USGS data server (http://glovis.usgs.gov/); SRTM data were
obtained from the CGIAR-CSI (http://srtm.csi.cgiar.org/); and coun-
try borderline shapefiles were from the Natural Earth (http://www.
naturalearthdata.com/).
References
Adams, J.B., Smith, M.O., Johnson, P.E., 1986. Spectral mixture modeling: a new
analysis of rock and soil types at the Viking Lander 1 site. J. Geophys. Res. 91
(B8), 8098–8112.
Adams, J.B., Sabol, D.E., Kapos, V., Filho, R.A., Roberts, D.A., Smith, M.O., Gillespie,
A.R., 1995. Classification of multispectral images based on fractions of
endmembers: application to land-cover change in the Brazilian Amazon.
Remote Sens. Environ. 52, 137–154.
Atkinson, P.M., Tatnall, A.R., 1997. Neural networks in remote sensing. Int. J.
Remote Sens. 18, 699–709.
Atkinson, P.M., 2005. Sub-pixel target mapping from soft-classified remotely
sensed imagery. Photogramm. Eng. Remote Sens. 71, 839–846.
Benediktsson, J.A., Swain, P.H., Ersoy, O.K., 1990. Neural network approaches
versus statistical methods in classification of multisource remote sensing data.
IEEE Trans. Geosci. Remote Sens. 28, 540–552.
Blaschke, T., 2010. Object based image analysis for remote sensing. ISPRS J.
Photogram. Remote Sens. 65, 2–16.
Breiman, L., 2001. Random forests. Mach. Learn. 45 (1), 5–32.
CEC (Commission of the European Communities), 1994. CORINE Land cover, Part 1:
Methodology. http://www.eea.europa.eu/publications/COR0-landcover
(accessed 18.09.15.).
Chander, G., Markham, B.L., Helder, D.L., 2009. Summary of current radiometric
calibration coefficients for Landsat MSS, TM, ETM+, and EO-1 ALI sensors.
Remote Sens. Environ. 113 (5), 893–903.
Chavez Jr., P.S., 1988. An improved dark-object subtraction technique for
atmospheric scattering correction of multispectral data. Remote Sens. Environ.
24, 459–479.
Chavez Jr., P.S., 1996. Image-based atmospheric corrections—revisited and
improved. Photogramm. Eng. Remote Sens. 62 (9), 1025–1036.
Chen, J., Chen, J., Liao, A., Cao, X., Chen, L., Chen, X., He, C., Han, G., Peng, S., Lu, M.,
Zhang, W., Tong, X., Mills, J., 2015. Global land cover mapping at 30 m
resolution: a POK-based operational approach. ISPRS J. Photogramm. Remote
Sens. 103, 7–27, http://dx.doi.org/10.1016/j.isprsjprs.2014.09.002.
432 W. Wu et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 422–432
DeFries, R.S., Hansen, M., Townshend, J.R.G., Sohlberg, R., 1998. Global land cover
classification at 8 km spatial resolution: use of training data derived from
Landsat imagery in decision tree classifiers. Int. J. Remote Sens. 19, 3141–3168.
Di Gregorio A., Jansen L.J.M., 2000. Land Cover Classification System (LCCS):
Classification Concepts and User Manual FOR SOFTWARE VERSION 1.0. FAO,
Rome. http://www.fao.org/docrep/003/x0596e/x0596e00.htm (accessed
02.02.09.).
ESA, 2008. DesertWatch Final Technical Report. http://dup.esrin.esa.int/files/131-
176-149-30 2009430103852.pdf (accessed 12.06.15.).
Foody, G.M., Cox, D.P., 1994. Sub-pixel land cover composition estimation using a
linear mixture model and fuzzy membership functions. Int. J. Remote Sens. 15,
619–631.
Foody, G.M., Mathur, A., 2004. A relative evaluation of multiclass image
classification by support vector machines. IEEE Trans. Geosci. Remote Sens. 42,
1335–1343.
Foody, G.M., Campbell, N.A., Trodd, N.M., Wood, T.F., 1992. Derivation and
applications of probabilistic measures of class membership from the
maximum-likelihood classification. Photogramm. Eng. Remote Sens. 58 (9),
1335–1341.
Friedl, M.A., McIver, D.K., Hodges, J.C.F., Zhang, X.Y., Muchoney, D., Strahler, A.H.,
Woodcock, C.E., Gopal, S., Schneider, A., Cooper, A., Baccini, A., Gao, F., Schaaf,
C., 2002. Global land cover mapping from MODIS: algorithms and early results.
Remote Sens. Environ. 83, 287–302.
Gong, P., Marceau, D.J., Howarth, P.J., 1992. A comparison of spatial feature
extraction algorithms for land use classification with SPOT HRV data. Remote
Sens. Environ. 40, 137–151.
Gong, P., Wang, J., Yu, L., Zhao, Y., Zhao, Y., Liang, L., et al., 2013. Finer resolution
observation and monitoring of global land cover: first mapping results with
Landsat TM and ETM+ data. Int. J. Remote Sens. 34 (7), 2607–2654, http://dx.
doi.org/10.1080/01431161.2012.748992.
Hansen, M.C., DeFries, R.S., Townshend, J.R.G., Sohlberg, R., 2000. Global land cover
classification at 1 km spatial resolution using a classification tree approach. Int.
J. Remote Sens. 21, 1331–1364.
Haralick, R.M., Shanmugam, K., Dinstein, I., 1973. Textural features for image
classification. IEEE Trans. Syst. Man Cybern. SMC-3 (6), 610–621.
Hill, J., Schutt, B., 2000. Mapping complex patterns of erosion and stability in dry
Mediterranean ecosystems. Remote Sens. Environ. 74, 557–569.
Huang, C., Davis, L.S., Townshend, J.R.G., 2002. An assessment of support vector
machines for land cover classification. Int. J. Remote Sens. 23 (4), 725–749.
Huth, J., Kuenzer, C., Wehrmann, T., Gebhardt, S., Tuan, V.Q., Dech, S., 2012. Land
cover and land use classification with TWOPAC: towards automated processing
for pixel- and object-based image classification. Remote Sens. 4, 2530–2553.
Jia, K., Liang, S., Wei, X., Yao, Y., Su, Y., Jiang, B., Wang, X., 2014. Land cover
classification of Landsat data with phenological features extracted from time
series MODIS NDVI data. Remote Sens. 6, 11518–11532, http://dx.doi.org/10.
3390/rs61111518.
Kavzoglu, T., Colkesen, I., 2009. A kernel functions analysis for support vector
machines for land cover classification. Int. J. Appl. Earth Obs. Geoinf. 11,
352–359.
Kavzoglu, T., Mather, P.M., 2003. The use of backpropagating artificial neural
networks in land cover classification. Int. J. Remote Sens. 24 (23), 4907–4938.
Kent, M., Gill, W.J., Weaver, R.E., Armitage, R., 1997. Landscape and plant
community boundaries in biogeography. Prog. Phys. Geogr. 21 (3), 315–354.
Loveland, T.R., Reed, B.C., Brown, J.F., Ohlen, D.O., Zhu, Z., Yang, L., Merchant, J.W.,
2000. Development of a global land cover characteristics database and IGBP
DISCover from 1 km AVHRR data. Int. J. Remote Sens. 21 (6–7), 1303–1330.
Lu, D., Weng, Q., 2004. Spectral mixture analysis of the urban landscape in
Indianapolis with Landsat ETM+ imagery. Photogramm. Eng. Remote Sens. 70
(9), 1053–1062.
Lu, L., Kuenzer, C., Guo, H., Li, Q., Long, T., Li, X., 2014. A novel land cover
classification map based on a MODIS time-series in Xinjiang, China. Remote
Sens. 6, 3387–3408.
Mao, J., Jain, A., 1992. Texture classification and segmentation using
multiresolution simultaneous autoregressive models. Pattern Recognit. 25 (2),
173–188.
Mas, J.F., Flores, J.J., 2008. The application of artificial neural networks to the
analysis of remotely sensed data. Int. J. Remote Sens. 29 (3), 617–663, http://
dx.doi.org/10.1080/01431160701352154.
Mertens, K.C., De Baets, B., Verbeke, L.P.C., De Wulf, R.R., 2006. A sub-pixel
mapping algorithm based on sub-pixel/pixel spatial attraction models. Int. J.
Remote Sens. 27, 3293–3310.
Pace, G., Zucca, C., Wu, W., 2006. Sviluppo di un algoritmo di classificazione
automatico per la generazione di mappe di land cover finalizzate al
monitoraggio della desertificazione. In: Atti della 10a Conferenza Nazionale
ASITA, vol. II, pp. 1505–1510, Bolzano, Nov. 14–17.
Pu, R., Landry, S., Yu, Q., 2011. Object-based urban detailed land cover classification
with high spatial resolution IKONOS imagery. Int. J. Remote Sens. 32 (12),
3285–3308.
Richards, J.A., Jia, X., 2006. Remote Sensing Digital Image Analysis: An Introduction,
4th ed. Springer-Verlag, Berlin.
Rodriguez-Galiano, V.F., Ghimire, B., Rogan, J., Chica-Olmo, M., Rigol-Sanchez, J.P.,
2012. An assessment of the effectiveness of a random forest classifier for
land-cover classification. ISPRS J. Photogramm. Remote Sens. 67, 93–104,
http://dx.doi.org/10.1016/j.isprsjprs.2011.11.002.
Smith, M.O., Ustin, S.L., Adams, J.B., Gillespie, A.R., 1990. Vegetation in deserts: I. A
regional measure of abundance from multispectral images. Remote Sens.
Environ. 31, 1–26.
Swain, P.H., King, R.C., 1973. Two effective feature selection criteria for
multispectral remote sensing (LARS Technical Note 042673). In: Proceedings of
The International Joint Conference on Pattern Recognition, Washington, D.C.,
November, pp. 536–540.
Tatem, A.J., Lewis, H.G., Atkinson, P.M., Nixon, M.S., 2002. Super-resolution land
cover pattern prediction using a Hopfield neural network. Remote Sens.
Environ. 79 (1), 1–14.
Townshend, J.R.G., 1984. Agricultural land cover discrimination using thematic
mapper spectral bands. Int. J. Remote Sens. 6, 681–698.
USGS, 2015. Landsat 8 (L8) data users handbook (V1.0). http://landsat.usgs.gov/
documents/Landsat8DataUsersHandbook.pdf (accessed 30.06.15.).
Van Der Meer, F., 1995. Spectral unmixing of Landsat Thematic Mapper data. Int. J.
Remote Sens. 16 (16), 3189–3194.
Van der Linden, S., Rabe, A., Held, M., Jakimow, B., Leitão, P.J., Okujeni, A.,
Schwieder, M., Suess, S., Hostert, P., 2015. The EnMAP-Box—a toolbox and
application programming interface for EnMAP data processing. Remote Sens.
7, 11249–11266.
Vapnik, V., Lerner, A., 1963. Pattern recognition using generalized portrait method.
Autom. Remote Control 24, 774–780.
Verhoeye, J., De Wulf, R.R., 2002. Land cover mapping at sub-pixel scales using
linear optimization techniques. Remote Sens. Environ. 79 (1), 96–104.
Waske, B., van der Linden, S., Oldenburg, C., Jakimow, B., Rabe, A., Hostert, P., 2012.
imageRF—a user-oriented implementation for remote sensing image analysis
with Random Forests. Environ. Model. Softw. 35, 192–193, http://dx.doi.org/
10.1016/j.envsoft.2012.01.014.
Wu, W., Zhang, W., 2003. Present land use and cover patterns and their
development potential in North Ningxia, China. J. Geogr. Sci. 13 (1), 54–62.
Wu, W., De Pauw, E., Zucca, C., 2013a. Use remote sensing to assess impacts of land
management policies in the Ordos rangelands in China. Int. J. Digit. Earth 6
(Supplement 2), 81–102, http://dx.doi.org/10.1080/17538947.2013.825656.
Wu, W., De Pauw, E., Hellden, U., 2013b. Assessing woody biomass in African
tropical savannas by multiscale remote sensing. Int. J. Remote Sens. 34 (13),
4525–4549, http://dx.doi.org/10.1080/01431161.2013.777487.
Wu, W., 2003. Application de la géomatique au suivi de la dynamique
environnementale en zones arides, PhD thesis. Université de Paris 1, Paris.
Wu, W., 2014. The generalized difference vegetation index (GDVI) for dryland
characterization. Remote Sens. 6 (2), 1211–1233, http://dx.doi.org/10.3390/
rs6021211.
Zhang, Y., 2001. Texture-integrated classification of urban treed areas in
high-resolution color-infrared imagery. Photogramm. Eng. Remote Sens. 67
(12), 1359–1365.
Zhu, K.Z., Wan, W.M., 1963. Phenology. Science Press, Beijing (in Chinese).
Zhuang, X., Engel, B.A., Lozano-Garcia, D.F., Fernandez, R.N., Johannsen, C.J., 1994.
Optimization of training data required for neuro-classification. Int. J. Remote
Sens. 15, 3271–3277.