sensors

Article
Urban Vegetation Mapping from Aerial Imagery Using
Explainable AI (XAI)
Abolfazl Abdollahi 1 and Biswajeet Pradhan 1,2, *

1 Centre for Advanced Modelling and Geospatial Information Systems (CAMGIS),

Faculty of Engineering and IT, University of Technology Sydney, Ultimo, NSW 2007, Australia;
abolfazl.abdollahi@student.uts.edu.au
2 Earth Observation Center, Institute of Climate Change, University Kebangsaan Malaysia,
Bangi 43600 UKM, Selangor, Malaysia
* Correspondence: Biswajeet.Pradhan@uts.edu.au or biswajeet24@gmail.com

Abstract: Urban vegetation mapping is critical in many applications, i.e., preserving biodiversity,
maintaining ecological balance, and minimizing the urban heat island effect. It is still challenging to
extract accurate vegetation covers from aerial imagery using traditional classification approaches,
because urban vegetation categories have complex spatial structures and similar spectral properties.
Deep neural networks (DNNs) have shown a significant improvement in remote sensing image
classification outcomes during the last few years. These methods are promising in this domain, yet
unreliable for various reasons, such as the use of irrelevant descriptor features in the building of the
models and lack of quality in the labeled image. Explainable AI (XAI) can help us gain insight into
these limits and, as a result, adjust the training dataset and model as needed. Thus, in this work,
we explain how an explanation model called Shapley additive explanations (SHAP) can be utilized
for interpreting the output of the DNN model that is designed for classifying vegetation covers.



We want to not only produce high-quality vegetation maps, but also rank the input parameters

 and select appropriate features for classification. Therefore, we test our method on vegetation
Citation: Abdollahi, A.; Pradhan, B. mapping from aerial imagery based on spectral and textural features. Texture features can help
Urban Vegetation Mapping from overcome the limitations of poor spectral resolution in aerial imagery for vegetation mapping. The
Aerial Imagery Using Explainable AI model was capable of obtaining an overall accuracy (OA) of 94.44% for vegetation cover mapping.
(XAI). Sensors 2021, 21, 4738. https:// The conclusions derived from SHAP plots demonstrate the high contribution of features, such as
doi.org/10.3390/s21144738
Hue, Brightness, GLCM_Dissimilarity, GLCM_Homogeneity, and GLCM_Mean to the output of the
proposed model for vegetation mapping. Therefore, the study indicates that existing vegetation
Academic Editor: Sietse Los
mapping strategies based only on spectral characteristics are insufficient to appropriately classify
vegetation covers.
Received: 18 June 2021
Accepted: 9 July 2021
Published: 11 July 2021
Keywords: XAI; deep neural network; remote sensing; SHAP; vegetation mapping

Publisher’s Note: MDPI stays neutral

with regard to jurisdictional claims in
published maps and institutional affil- 1. Introduction
iations. In urban areas, vegetation plays a significant role for various purposes, including
conserving biodiversity, preserving ecological balance, and reducing the urban heat island
effect [1]. Producing these vegetation maps is critical to assist planners in optimizing
climate change adaptation and urban ecosystem services [2]. Remote sensing data have
Copyright: © 2021 by the authors. been extensively employed for vegetation mapping in various environments due to its
Licensee MDPI, Basel, Switzerland. ability to discriminate broad scales of land cover types. For vegetation monitoring in urban
This article is an open access article and rural regions, aerial photography and satellite imaging have been used [3–6]. Urban
distributed under the terms and vegetation cover is much more fragmented than natural vegetation (i.e., forest, rangeland),
conditions of the Creative Commons making accurate extraction of vegetation cover more complex and challenging. Thus, it
Attribution (CC BY) license (https:// has been a hot topic in the remote sensing field to effectively extract and map land use
creativecommons.org/licenses/by/ types over urban vegetation cover [7]. Satellite imagery of ultra-high resolution (UHR),
4.0/).

Sensors 2021, 21, 4738. https://doi.org/10.3390/s21144738 https://www.mdpi.com/journal/sensors

Sensors 2021, 21, 4738 2 of 16

unmanned aerial vehicle (UAV) imagery, very-high resolution (VHR), and moderate-to-
high-resolution (HR) are employed in urban vegetation mapping [1]. For the high and
very high-resolution satellite imagery, QuickBird and IKONOS satellite data have been
effectively employed for mapping the urban vegetation cover [7,8]. Moreover, aerial
imagery has been broadly utilized in vegetation mapping because the images have better
resolution [9]. On the other hand, the fundamental disadvantage of multispectral satellite
data with moderate resolution such as Enhanced Thematic Mapper Plus (ETM+) and
Landsat Thematic Mapper (TM) is that they are unable to capture sufficient details of
ground objects because of its coarse spatial resolution (30 m) [10]. However, using spectral
mixture analysis, they still can be utilized to measure vegetation abundance in suburban
and urban environments. To track the spatial–temporal changes of urban vegetation,
a fraction map can be created [11]. The classification techniques can be classified into
two types. First, they can be categorized by object-based image analysis (OBIA) and
pixel-based classification methods, which are extensively employed in HR remote sensing
image classification and vegetation mapping [12,13]. The former argues that single pixels
are autonomous, and they are processed without considering their spatial relationships
with other adjacent pixels [14]. However, a “salt and pepper” effect is always apparent
in the classification results because the individual pixels no longer take the properties
of classification purposes in HR images [15]. OBIA represents the spatial neighborhood
attributes rather than an individual pixel, as opposed to pixel-based classification. OBIA’s
main characteristic is that it uses multi-scale image segmentation to combine a variety of
spatial, spectral, and textural data in the classification process, which greatly improves
accuracy [16]. In the processing of VHR images, OBIA categorization approaches have
been thoroughly explored, and numerous approaches have been developed to classify
objects [17–19]. However, during the OBIA process, two important issues remain: selecting
appropriate features for image classification and determining the right scale for image
segmentation [20]. The OBIA’s performance is mostly dependent on the researchers’
past knowledge and expertise because there is no precise process for optimizing the scale
parameter and feature [21]. Furthermore, OBIA necessitates a variety of input variables and
image segmentation, making it tough and complex to use, particularly for beginners [3].
Broadly, the classification methods can be classified into unsupervised [22] and super-
vised [23] classification techniques. Machine learning techniques are commonly used in
supervised classification. For instance, Réjichi and Chaâbane [24] created spatial-temporal
vegetation maps using the support vector machine (SVM) method, and Wei, et al. [25]
used the decision tree to map the vegetation cover. The vegetation map was obtained from
Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) data using neural network (NN)
by [26]. Tigges, et al. [27] performed an SVM approach to classify urban vegetation using
various temporal and spectral band mixtures of five RapidEye imagery for Berlin, Ger-
many. Tooke, Coops, Goodwin, and Voogt [2] applied spectral unmixing and decision tree
approach for classifying urban vegetation characteristics from QuickBird images for Van-
couver city, Canada. The Iterative Self-Organizing Data Analysis (IOSDATA) technique [28]
and K-Means method [29] are examples of unsupervised classification algorithms, which
have been popularly used for vegetation cover classification from SPOT-4 and aerial im-
agery. In [30], two totally unsupervised methods called Jenks Natural Breaks classification
method and Agglomerative Hierarchical clustering approach were used for vegetation
mapping from IKONOS images for the city of Bamenda, Cameroon. In another work [31],
Probabilistic Neural Network (PNN) approach was introduced to classify vegetation covers
from SPOT and LANDSAT data for the Marrakech city, Morocco. However, these tradi-
tional classification algorithms still need a large amount of reference data, and they are more
complex or necessitate more manual involvement for producing accurate classification
maps [32]. Moreover, due to the same spectral properties and a complex spatial structure
amongst urban vegetation groups, precise vegetation mapping based on the traditional
pixel-based classification methods from VHR satellite imagery is still challenging [33].
These approaches, which rely solely on spectral information, had limited effectiveness
Sensors 2021, 21, 4738 3 of 16

in urban set-ups [33]. In contrast to pixel-based approaches, the features employed for
classification in the object-oriented technique are the objects’ attributes. There are some
features reduction methods, which can be used to overcome the limitation of OBIA. For
instance, for data reduction and unsupervised feature extraction, principal components
analysis (PCA) [34] has been frequently employed. However, it has three major flaws: a
significant memory need, a high computational cost, and poor performance when process-
ing massive data sets. In [35], a logistic regression (LR) method was performed for both
feature selection based on their relative importance and classification of remote sensing
imagery. The assumption of linearity between the dependent and independent variables is
a key constraint of LR as well as non-linear problems cannot be solved by LR [35]. Thus,
developing an accurate strategy for picking the features that are important for a specific
classification task and alleviating the above-mentioned drawbacks is necessary.
Deep neural networks (DNN) have become extremely popular techniques in remote
sensing applications, which recognize features in several levels of representation [36]. This
is due to the fact that these approaches may easily encode spectral and spatial information
from raw images without the need for any preprocessing [37]. They are also a hierarchical
structure of deep neural networks and include a number of interconnected layers that
may learn a hierarchical feature representation from the data and extract the deep features
of the input data [37]. Therefore, we implemented a DNN model in the current work to
address the limitation of conventional classification approaches in vegetation mapping
from high-resolution remotely sensed imagery. When building a classification model, it is
critical to understand why each classification is made, except achieving high accuracy. Due
to the complicated architecture of DNNs [38], interpreting the output of these models is
typically thought to be extremely difficult [38].
In this work, we illustrate how to interpret the outputs of DNN model designed
for classification based on an explanation method. Thus, we present a methodology and
evaluation for feature selection based on an explanatory model, which is generated by a
method termed Shapley Additive explanations (SHAP) [39,40] to interpret the output of
DNN and improve the OBIA limitation in selecting appropriate features for classification.
For every feature of every data point, the technique allocates SHAP values, which are
contribution values for the output of the model. We utilize the contribution information
of every feature to arrange and encode the features based on their importance using
these SHAP values. In this scenario, choosing a set of features according to the SHAP
values entails sorting the features based on their contributions to the method’s output
and then selecting the initial features. We test our method on vegetation mapping from
high-resolution remote sensing imagery based on the spectral and textural features. In
summary, the following are the main contributions of this study: (1) applying a DNN for
producing high-quality vegetation maps from remote sensing data. (2) Implementing the
Sharp method as a feature selection mechanism to select appropriate features among the
spectral and textural features and see the contribution information of every feature on
classification and mapping the vegetation cover. This study is organized in the following
way: we give a brief overview of SHAP for selecting features and DNN structure in
Section 2. In Sections 3 and 4, we provide the experiments and discussion on the results,
respectively. In Section 5, a concluding remark is presented.

2. Methodology
The method for finding important variables that influence the classification model
is depicted in Figure 1. A random division of the data into training and test sets is used
to build the DNN model at first. The training and test dataset are utilized to create
the classification model and assess the model’s performance. SHAP is applied to the
classification network (in this case, a DNN network) to create additive attributes, which
are then utilized to ascertain significance variables and determine the effect of input factors
to the classification results. In the following sections, the technique for mapping vegetation
covers using DNN and SHAP is described in depth.
Sensors 2021, 21, x FOR PEER REVIEW 4 of 17

Sensors 2021, 21, 4738 4 of 16

classification results. In the following sections, the technique for mapping vegetation co-
vers using DNN and SHAP is described in depth.

Figure 1. The SHAP workflow for an interpretable DNN model.

Figure 1. The SHAP workflow for an interpretable DNN model.

2.1. Dataset
2.1. Dataset
We selected three images from the Aerial Imagery for Roof Segmentation (AIRS) [41]
We selected three images from the Aerial Imagery for Roof Segmentation (AIRS) [41]
dataset with a spatial resolution of 7.5 cm and a dimension of 2500 × 3000 for classification
dataset with a spatial resolution of 7.5 cm and a dimension of 2500 × 3000 for classification
purpose. The test area is a designated area of Christchurch city with 457 km22 coverage,
purpose. The test area is a designated area of Christchurch city with 457 km coverage,
which is situated on the South Island of New Zealand (Figure 2) and the images were
which is situated on the South Island of New Zealand (Figure 2) and the images were taken
taken during the flying season of 2016 [41]. Since the aerial images used in this study are
during the flying season of 2016 [41]. Since the aerial images used in this study are RGB
RGB images, we used the HSV (hue, saturation, and value) color space of the images in-
images, we used the HSV (hue, saturation, and value) color space of the images instead of
stead
RGB. ofTheRGB. The fundamental
fundamental distinctiondistinction is the
is the color color representation
representation betweenbetween
both colorbothspaces.
color
spaces.
The HSV color space uses three parameters named Hue (H), Saturation (S), and Valueand
The HSV color space uses three parameters named Hue (H), Saturation (S), (V)
Value (V) the
to depict to depict thethe
color of color of the
object. Theobject. The H
H shows theshows thecolor,
object’s object’s color,
while thewhile
S andthe S and
V values
V values represent
represent the
the object’s object’s
color color illuminance
illuminance [42]. This description
[42]. This description enables the enables
color to the color
be distin-
to
guished from the illumination and prevents the influence of illumination variations onvar-
be distinguished from the illumination and prevents the influence of illumination the
iations
color ofon the
the color Consequently,
object. of the object. Consequently,
the suggested the suggested
method method for
for vegetation vegetation
mapping usedmap-
HSV
ping
color used
space.HSV color space.
In addition, In addition,
to help to help withprocess
with the classification the classification
and improve process and im-
the vegetation
prove the vegetation classification method [43], we used textural factors
classification method [43], we used textural factors based on the gray-level-co-occurrence- based on the
gray-level-co-occurrence-matrix (GLCM) [44].
matrix (GLCM) [44]. The texture descriptor, GLCM Theistexture descriptor,
a statistical matrixGLCM
of joint is a statisti-
conditional
cal matrix of
possibility joint conditional
distributions between possibility
the greydistributions
levels of image between
pixel the
pairs grey
with levels of image
a specific dis-
pixel
tance pairs with a specific
and direction withindistance and direction
a given frame. within
We utilized a given frame.
eCognition DeveloperWe utilized eCog-
64 software to
nition Developer 64 software to obtain the features including
obtain the features including GLCM_Contrast, GLCM_Correlation, GLCM_Dissimilarity, GLCM_Contrast,
GLCM_Correlation, GLCM_Dissimilarity,
GLCM_Entropy, GLCM_Homogeneity, GLCM_Mean, GLCM_Entropy,
and GLCM_Std GLCM_Homogeneity,
values in this study.
GLCM_Mean, and GLCM_Std
The image homogeneity valuesby
is measured inGLCM
this study. The image
homogeneity, homogeneity
which presentedisthe measured
similar-
ity in gray level between nearby pixels. GLCM Contrast is a metric to show the number
of regional differences in the image opposite GLCM homogeneity. GLCM dissimilarity
receives high values when the local areas have high contrast, and it rises linearly opposite
Sensors 2021, 21, 4738 5 of 16

to Contrast. GLCM entropy is negatively correlated with energy, and it measures the
disorder of an image [45]. The frequency of occurrence of a pixel value in combination with
a specific adjacent average pixel value is used to measure GLCM Mean. GLCM standard
deviation is not similar to the simple standard deviation; it is based on the GLCM, and
deals mainly with the mixtures of neighbor and reference pixels, and GLCM correlation
estimates the linear relationship between surrounding pixels’ grey levels. These features,
along with the spectral features, are fed into the model to classify the image accurately. To
obtain samples for the training dataset and create polygons for each land cover class, we
manually analyzed the main images in ArcGIS 10.6. Table 1 shows the total number of
samples obtained in the region of research for training and test. A total of 81,831 samples
were chosen, and 60%, 15%, and 25% of them were randomly split into training, validation
and testing sets, respectively. We then employed a data normalization method to speed
up the gradient descent optimization and activation function process [46]. Thus, to avoid
anomalous gradients, we performed a z-score normalization method to the pixel values of
the aerial imagery (Equation (1)).

( X/max) − µ
X0 = (1)
σ
where, the highest pixel value of the image is defined as max, the standard deviation and
Sensors 2021, 21, x FOR PEER REVIEW 6 of 17
mean of X/max are defined as σ and µ, respectively, and the normalized data is denoted
as X 0 .

Figure 2.
Figure 2. The
The geolocation
geolocation of
of the
the used
used AIRS dataset. The
AIRS dataset. The AOI
AOI is
is indicated
indicated by the red
by the red polygon.
polygon.

2.2. DNN Architecture

Our DNN model is a series of dense layers (fully connected) that take every spectral
and textural factor’s attribute as input and classify them to build the classification map, as
illustrated in Figure 3. The output values are computed successively along with the net-
work layers when the input features are fed into the DNN. The weighted sum is produced
at every layer by multiplication of the weight vector for every unit in the current hidden
 (1)).
(1)).
max) − μ
( X // max)
X = ((X
X ''' = X / max) −− μ
μ (1)
X = σ
σ
(1)
(1)
σ
where, the
where, the highest pixel
pixel value of
of the image
image is defined
defined as m a x ,, the
the standard
standard deviation
deviation and
and
where, the highest
highest pixel value
value of the
the image is
isμdefined as
of X // max
max are σ
σ
as m
m aaxx , the standard deviation and
and μ ,, respectively,
σ
and μ , respectively, and
mean of are defined
defined as respectively, and the the normalized data is
mean of X X / max
Sensors 2021,21,
Sensors2021, 21,4738 mean 67of
of16
x FOR PEER REVIEW
' are defined asas and and the normalized
normalized data
17
data is
is
denoted as X '.
' .
denoted as X
denoted as X .
Table 1.
Table 1. The
The size
size of
of training
training and
and testing
testing sets
sets employed
employed inin this
this work
work counted
counted at
at polygon
polygon and
and pixel-level.
pixel-level.
TableTable
1. The1.size
Theofsize ofBecause
training
training andthe
of testing
and testing lack
sets sets employed
of gradient
employed in this counted
workimpact,
in thisvanishing
work counted at polygon
great
at polygon and pixel-level.
computing
and efficiency, and
pixel-level.
sparsityPolygons
property during Training Samples Validation
back-propagation
Training Samples Validation Sam-
training, Sam- Testing
ReLUTesting Samples
is a popular
Samples option ofLegend
activa-
Class
Class Polygons
Training Training Validation
Samples Samples Validation
Samples Sam- Testing
Testing SamplesSamples Legend
Class ClassPolygons Polygons
tion function in deep learning (Pixels)
[47],[48].
(Pixels) We ples (Pixels)
utilized the Softmax (Pixels)
function for theLegend
Legendoutput
(Pixels) (Pixels) (Pixels)plesples (Pixels)
(Pixels) (Pixels) (Pixels)
(Pixels)
Background layer Lout to
254 predict class possibilities
34,209 after calculating
5684 a H for the last hidden layer.
16,240
Background
Background 254
Background
254
254 34,209 34,209
34,209 5684
5684
5684 16,240
16,240
16,240
Moreover, to train the suggested network for classification, we used the categorical cross-
entropy (CCE) loss function [49]. The loss function for CCE is calculated as follows:
High-Vegetation 198
High-Vegetation
High-Vegetation
198 11,611
198 11,611
11,611 2049
2049
2049 5854
5854 5854
High-Vegetation 198 11,611 S P C2049 5854
LCCE ( g , f (I ), δ1 ) =
i = 1 j =1 c =1

−1( g ij = C ) log l c (I i j ) (4)
Low-Vegetation
Low-Vegetation
Low-Vegetation 160 160
160 3085 3085
3085 544 544
544 2555
2555 2555
Low-Vegetation 160 3085 544 2555
j
where, the number of classes is C , the ground truth label is gi , the output of the last

dense
2.2. layer
DNN at the pixel
Architecture I ij is f (Iij ) , the network parameters are δ1 , the size of the batch
is S Our
, the DNN jthmodel is athe
pixel in seriesith patch
of dense I i j , the(fully
is layers number connected) that
of pixels intake
eachevery is P ,
patchspectral
and textural factor’s attribute as input and classify them to build the classification map, as
j j
every classinpossibility
illustrated Figure 3. Theof the I i is are
pixelvalues
output lc (computed
I i ) that issuccessively
denoted asalong with the network
layers when the input features are fed into the DNN. jThe weighted sum is produced at
exp(for f c (every
I i )) unit in the current hidden layer
every layer by multiplication of the I i j ) = vector
l c (weight c
.

by the input vector, which includes every unit’s outputj values from the prior layer. An (5)
exp ( f l ( I i ))
input layer Lin , an output layer Lout , and H hidden layers Lh (h ∈ {1, 2, . . . , H }) between
l =1
input and output layers generate the basic DNN architecture. Every hidden layer Lh
Theofsuggested method’s architecture
that can beisstructuredbased on aasmultilayer a vector afeed-forward neural
| Lh | , with
consists a collection of units h ∈ R | Lh |
network that is organized as below: (1) the input layer has
denoting the number of units in Lh . After that, an activation function f (·), a bias vector several features representing
bthe class
h ∈ R
| Lh |descriptors.
, and a weight (2) There
matrixare Whseven∈ R| Lhidden layers.
h1 |∗| L h | can then Four layerstoare
be used dense layers,
parameterize and
every
three regularization
hidden layer Lh . The and unitsdropout
in Lh can layers were inserted
be computed between them
more completely by to avoid the over-
fitting problem. (3) To influence the collection of obtained features from the final hidden
T
layer to the appropriate class, the aSoftmax h = f (Wclassifier
h a h−1 + bwas h ) applied in the last layer. (4) The (2)
input layer was regulated based on the kernel regularization method, and the ReLU acti-
where, the unitswas
vation function a0 inalso input layer L0tocome
theimplemented avoidfrom a compound’s
saturation during thefeatures
learning vector and
process.
= 1, 2, . . .to
hMoreover, , H. The rectified linear unit (ReLU) function was employed
maintain as many features as feasible, we deleted the pooling operation. This as the activation
function
is due toin thethis
factstudy,
that and it is calculated
we want to captureasmore contextual information and identify fine
details [50]. We used Python and the Keras with TensorFlow as a backend to apply the
f ( x ) = max(0, x ) (3)
entire process of the suggested model for vegetation mapping from aerial data.

Figure 3.
Figure 3. The
The structure
structure of
of the
the proposed
proposedDNN
DNNnetwork.
network.
Sensors 2021, 21, 4738 7 of 16

Because of the lack of gradient vanishing impact, great computing efficiency, and
sparsity property during back-propagation training, ReLU is a popular option of activation
function in deep learning [47,48]. We utilized the Softmax function for the output layer
Lout to predict class possibilities after calculating a H for the last hidden layer. Moreover, to
train the suggested network for classification, we used the categorical cross-entropy (CCE)
loss function [49]. The loss function for CCE is calculated as follows:
S P C
∑ ∑ ∑ −1( gi = C) log lc ( Ii )
j j
LCCE ( g, f ( I ), δ1 ) = (4)
i =1 j =1 c =1

j
where, the number of classes is C, the ground truth label is gi , the output of the last dense
j j
layer at the pixel Ii is f ( Ii ), the network parameters are δ1 , the size of the batch is S, the jth
j
pixel in the ith patch is Ii , the number of pixels in each patch is P, every class possibility of
j j
the pixel Ii is lc ( Ii ) that is denoted as

j
j exp( f c ( Ii ))
lc ( Ii ) = c . (5)
j
∑ exp( f l ( Ii ))
l =1

The suggested method’s architecture is based on a multilayer feed-forward neural

network that is organized as below: (1) the input layer has several features representing the
class descriptors. (2) There are seven hidden layers. Four layers are dense layers, and three
regularization and dropout layers were inserted between them to avoid the over-fitting
problem. (3) To influence the collection of obtained features from the final hidden layer
to the appropriate class, the Softmax classifier was applied in the last layer. (4) The input
layer was regulated based on the kernel regularization method, and the ReLU activation
function was also implemented to avoid saturation during the learning process. Moreover,
to maintain as many features as feasible, we deleted the pooling operation. This is due to
the fact that we want to capture more contextual information and identify fine details [50].
We used Python and the Keras with TensorFlow as a backend to apply the entire process of
the suggested model for vegetation mapping from aerial data.

2.3. SHAP Approach

SHAP is a game-theoretic way to describing the model’s performance [51]. SHAP
employs an additive feature imputation approach, in which an output model is specified
as a linear addition of input parameters to build an interpretable method. To explain
predictions from the models, a few strategies have been proposed. Under the domain of
additive feature attribution approaches, the SHAP structure [39] collects earlier suggested
explanation techniques like DeepLIFT [52] and LIME [53]. For an original method f ( x ), the
explanation approach g( x 0 ) with streamlined input x 0 can be defined as f ( x ) = g( x 0 ) =
M
φ0 + ∑ φi x 0 i if we presume a method with input parameters x = ( x1 , x2 , . . . , x p ), where
i =1
the number of input parameters is defined as p, the consistent value when all inputs are
missing is defined as φ0 , and the number of input features is denoted as M. SHAP has a
strong theoretical foundation, which is advantageous in supervised scenarios. By allocating
a significance value (SHAP value) to any variable that meets the following criteria, it explains
a particular prediction using Shapley values. The requirements [54] include (1) consistency—
if we change a model to make it more dependent on a particular component, the relevance
of that component should not diminish, notwithstanding the relevance of other variables;
(2) missingness—components that aren’t present in the primary input must be ignored;
(3) local accuracy—the explanation method has to at least in accordance with the main
model’s output. Therefore, SHAP can describe both global and local methods effectively.
To develop an interpretable model, which takes into consideration the closeness to the
Sensors 2021, 21, 4738 8 of 16

instance, a local model employs a minimal set of background from the data [53]. There are
various approaches like Tree SHAP, Deep SHAP, and Kernel SHAP to approximate SHAP
values [39]. Kernel SHAP, a model-agnostic estimation, constructs a local explanatory
model using Shapley values and linear LIME [53]. Thus, in this work, Kernel SHAP was
used because it gives better precise estimations with lower model assessments compared
to the other sampling-based estimates.

2.4. Assessment Measures

The proposed method’s efficacy was tested using measurement variables such as
precision, recall, F1 score, and overall accuracy (OA), as indicated in Equations (6)–(9).
The recall is defined as the percentage of accurately detected pixels among all actual
pixels. The proportion of accurately categorized pixels among all predicted pixels is known
as precision. The F1 score comprises recall and precision, and overall accuracy (OA) is
measured as adding the ratio of correctly identified pixels and dividing by the total number
of pixels [50].
TP
Recall = (6)
TP + FN
TP
Precision = (7)
TP + FP
2 × Precision × Recall
F1 score = (8)
Precision + Recall
TP + TN
OA = (9)
N
where, FP denotes false-positive pixels, FN denotes false-negative pixels, TP denotes
true-positive pixels, TN denotes true-negative pixels, and N denotes the number of pixels.

2.5. DDN Implementation

To analyze the influence of low and high-contributed features on vegetation classifi-
cation, we first fed all the variables into the model and obtained the classification results.
Subsequently, we just fed the low-contributed variables and high-contributed variables
obtained by the SHAP method into the model and achieved the classification results. In
the following, the features explanations achieved by the SHAP approach are firstly pre-
sented and then both quantitative and qualitative results achieved by the DNN model for
vegetation mapping based on all variables and high-contributed variables is discussed.

3. Results
In this section, we graphically presented the features explanations achieved by the
SHAP approach. A bar plot, which illustrates the global relevance of the features, or a
partial dependence plot, which shows the impact of shifting an individual characteristic, is
commonly used to depict the influence of the features in a model [55]. However, other forms
of visualizations are possible since SHAP values are outcomes of individualized attributes
that are specific to every prediction. Because SHAP dependence graphs reflect the impacts
of feature relationships better than partial dependence plots, they are a good alternative
to partial dependence plots. To highlight the global significance, SHAP summary graphs
replace usual bar plots, and local explanations are displayed using force plots [55]. For
example, Figure 4 shows an explanation of a prediction (vegetation class) obtained with
a DNN model based on a force plot. The explanation demonstrates how multiple factors
work together to push the output of the method from the “base” to the “predicted” value
for the initial prediction. The prediction process begins with the baseline. The average of
all predictions is the baseline for Shapley values. Each Shapley value is represented by
an arrow in the plot, indicating whether the prediction should be increased or decreased.
Features that cause a high classification value are represented in red, while those that cause
a lower value are indicated in blue. For example, for the vegetation class, the sample has a
 obtained with a DNN model based on a force plot. The explanation demonstrates how
multiple factors work together to push the output of the method from the “base” to the
“predicted” value for the initial prediction. The prediction process begins with the base-
line. The average of all predictions is the baseline for Shapley values. Each Shapley value
is represented by an arrow in the plot, indicating whether the prediction should be in-
Sensors 2021, 21, 4738 9 of 16
creased or decreased. Features that cause a high classification value are represented in
red, while those that cause a lower value are indicated in blue. For example, for the vege-
tation class, the sample has a prediction of 0.97, while the baseline is 0.6. The Hue,
GLCM_Std,ofand
prediction 0.97,GLCM_Mean can improve
while the baseline the Hue,
is 0.6. The final output prediction
GLCM_Std, while Brightness
and GLCM_Mean can
can reduce
improve thethe prediction.
final output prediction while Brightness can reduce the prediction.

Figure 4. Local
Figure 4. Local interpretation
interpretationusing
usingSHAP
SHAPforce
forceplot
plotfor
forthe
theinitial
initialprediction
prediction produced
producedbybythethe
DNN
DNN model. The
model. outcome
The outcomeis
is pushed
pushed lowerlower by blue
by blue feature
feature attributions,
attributions, whereas
whereas theare
the results results
pushedarehigher
pushed higher
than thanvalue”
the “base the “base value”
by red byattributions.
feature red feature
attributions.
In addition, we used a summary plot to integrate the feature’s significance with the
feature’s impacts.we
In addition, A Shapley value for aplot
used a summary feature and a sample
to integrate is represented
the feature’s by each
significance withpoint
the
on the summary
feature’s impacts.plot. Figure value
A Shapley 5a,b, present the summary
for a feature plotsisfor
and a sample the non-vegetation
represented and
by each point
vegetation classesplot.
on the summary for different labels,
Figure 5a,b, respectively.
present The x-axis
the summary determines
plots the Shapley value,
for the non-vegetation and
and the y-axis
vegetation shows
classes for the features.
different The respectively.
labels, color denotesThe the x-axis
feature’s value, which
determines ranges
the Shapley
from
value,lowandtothe
high. We shows
y-axis can getthe a sense of Shapley
features. values
The color dispensation
denotes per feature
the feature’s using
value, which
overlapping points that are jittered in the direction of the y-axis. The
ranges from low to high. We can get a sense of Shapley values dispensation per feature features are ranked
in order
using of significance
overlapping points [53].
thatFor
are example,
jittered inFigure 5a shows
the direction thaty-axis.
of the a highThe
degree of Hue,
features are
GLCM_Mean
ranked in order and ofGLCM_Homogeneity
significance [53]. For contents
example,have
Figurea significant
5a shows and thatbeneficial impact
a high degree of
on
Hue,classification.
GLCM_Mean Theand
red GLCM_Homogeneity
color represents a high,contents x-axisarepresents
and the have significanta positive value.
and beneficial
On the opposite,
impact the feature
on classification. Thesaturation is negatively
red color represents related
a high, andtothe
thex-axis
targetrepresents
factor. Moreover,
a posi-
for the normal bar graph, as shown in Figure 6, we could take
tive value. On the opposite, the feature saturation is negatively related to the the average absolute value
target factor.
of SHAP values to every feature. The SHAP values on the x-axis
Moreover, for the normal bar graph, as shown in Figure 6, we could take the average represent the magnitude
of the difference in log-odds. In Figure 5, we can see how SHAP values are allocated to the
features globally. All features are continual in this scenario, and their average influence
on classification is vertically arranged. We can observe how GLCM_Mean has the most
significant power on the classification of Class 1 that is High-Vegetation class, while Hue
showed a high effect on the Class 0 (Background) classification. In the same manner,
high values of GLCM_Homogeniety and GLCM_Dissimilarity are similarly crucial in the
classification of non-vegetation and vegetation classes.
When the value of the features is constant, the SHAP dependence graphs indicate
the expected results of a method. In the dependence plots, we can demonstrate how the
model is dependent on a feature by illustrating how the model outputs shift when the
features vary. For instance, in Figure 7, we plotted the interactions between variables for
the non-vegetation class (Background (7a)) and vegetation class (7b) concerning SHAP
values to gain a better understanding of their relationships. This is illustrated for Hue and
GLCM_Mean in Figure 7, where the colors reflect the SHAP values given to one of them,
and x and y-axis represent the magnitude of these variables. In fact, the effect of Hue is
depicted for variation of GLCM_Mean from −1.5 to 1.5. The high values of the variable
GLCM_Mean are represented by red, while the low values are represented by blue. In
Figure 7a, the SHAP values for Hue are negative when Hue is less than 0; the SHAP values
are extremely low for low Hue and GLCM_Mean, which results in a lower possibility of
Background class.
 the magnitude of the difference in log-odds. In Figure 5, we can see how SHAP values are
allocated to the features globally. All features are continual in this scenario, and their av-
erage influence on classification is vertically arranged. We can observe how GLCM_Mean
has the most significant power on the classification of Class 1 that is High-Vegetation class,
Sensors 2021, 21, 4738 while Hue showed a high effect on the Class 0 (Background) classification. In10 the of 16 same
manner, high values of GLCM_Homogeniety and GLCM_Dissimilarity are similarly cru-
cial in the classification of non-vegetation and vegetation classes.

FigureFigure
5. The
5. DNN
The DNNmodel’s SHAP
model’s SHAPsummary
summaryplot for: (a)
plot for: (a)non-vegetation;
non-vegetation;andand (b) vegetation
(b) vegetation classes.
classes. The classification
The higher higher classifica-
tion isisrelated
related to the
the higher
higherSHAP
SHAP value
value of aoffeature.
a feature.
rs 2021, 21, Sensors 2021, 21,
x FOR PEER 4738
REVIEW 11 of 17 11 of 16

Figure 6. Taking the average absolute value of the SHAP values to measure the relative signifi-
cance of every feature for every class (e.g., Class 0 is Background, Class 1 is High-Vegetation, and
Class 2 is Low-Vegetation).

When the value of the features is constant, the SHAP dependence graphs indicate the
expected results of a method. In the dependence plots, we can demonstrate how the model
is dependent on a feature by illustrating how the model outputs shift when the features
vary. For instance, in Figure 7, we plotted the interactions between variables for the non-
vegetation class (Background (7a)) and vegetation class (7b) concerning SHAP values to
gain a better understanding of their relationships. This is illustrated for Hue and
GLCM_Mean in Figure 7, where the colors reflect the SHAP values given to one of them,
and x and y-axis represent the magnitude of these variables. In fact, the effect of Hue is
depicted for variation of GLCM_Mean from −1.5 to 1.5. The high values of the variable
GLCM_Mean are represented by red, while the low values are represented by blue. In
Figure 6. Taking the6.average
Figure TakingFigure 7a, value
absolute theabsolute
the average SHAP
of thevalues
SHAP
value offor Hue
values
the toare
SHAP negative
measure
values the when Hue
relative
to measure is less than
significance
the relative 0; thefeature
of every
signifi- SHAPforvalues
cance
every class (e.g., of every
Class are extremely low
feature for every
0 is Background, for
Classclass low Hue and GLCM_Mean,
(e.g., Class 0 is Background,
1 is High-Vegetation, and Class 2 Class which results in a lower
1 is High-Vegetation, and
is Low-Vegetation). possibility of
Background class.
Class 2 is Low-Vegetation).

When the value of the features is constant, the SHAP dependence graphs indicate the
expected results of a method. In the dependence plots, we can demonstrate how the model
is dependent on a feature by illustrating how the model outputs shift when the features
vary. For instance, in Figure 7, we plotted the interactions between variables for the non-
vegetation class (Background (7a)) and vegetation class (7b) concerning SHAP values to
gain a better understanding of their relationships. This is illustrated for Hue and
GLCM_Mean in Figure 7, where the colors reflect the SHAP values given to one of them,
and x and y-axis represent the magnitude of these variables. In fact, the effect of Hue is
depicted for variation of GLCM_Mean from −1.5 to 1.5. The high values of the variable
GLCM_Mean are represented by red, while the low values are represented by blue. In
Figure 7a, the SHAP values for Hue are negative when Hue is less than 0; the SHAP values
are extremely low for low Hue and GLCM_Mean, which results in a lower possibility of
Background class.
Figure 7. SHAP
Figure 7. SHAP dependence
dependence plots
plots for
for the
the DNN
DNN method
method regarding
regarding to Hue and
to Hue and GLCM_Mean
GLCM_Mean for
for non-vegetation
non-vegetation (a)
(a) and
and
vegetation (b) classes.
vegetation (b) classes.

4. Discussion
Several images with complex backgrounds were used to verify the proposed DNN
model for classifying vegetation cover and SHAP method for selecting the appropriate
spectral and textural features to see the contribution information of each feature on classifi-
cation. Based on allocating the SHAP values to every feature and interpreting the effect of
the features on the classification in Section 3, we figured out that the features such as Hue,
Brightness, GLCM_Dissimilarity, GLCM_Homogeneity, and GLCM_Mean showed high
contribution information to the output of the proposed model for classification, while the
other features showed low contribution. We showed the qualitative results obtained by the
DNN network for the vegetation mapping based on high-ranked features in Figure 8, while
the qualitative results for the model with all the input features is depicted in Figure 9. Two
Figure 7. SHAP dependence plots for the DNN method regarding to Hue and GLCM_Mean for non-vegetation (a) and
rows and three columns separate the sub-figures. The first row shows the main images,
vegetation (b) classes.
while the second row shows the classification results by the DNN method. As seen in
Figure 9, the model could not produce high-quality vegetation maps when we use all the
features with low contribution features (e.g., GLCM_Correlation, GLCM_Entropy, Value,
 8, while the qualitative results for the model with all the input features is depicted in F
ure 9. Two rows and three columns separate the sub-figures. The first row shows the ma
images, while the second row shows the classification results by the DNN method.
Sensors 2021, 21, 4738
seen in Figure 9, the model could not produce high-quality vegetation maps when we u
12 of 16
all the features with low contribution features (e.g., GLCM_Correlation, GLCM_Entrop
Value, GLCM_Contrast etc.); the method misclassified pixels (predicted more FPs) wh
yielded low classification accuracy. For example, in complex backgrounds where obs
GLCM_Contrast etc.); the method misclassified pixels (predicted more FPs) which yielded
cles such as buildings,
low classification shadows,
accuracy. and other
For example, elements
in complex obscure vegetation
backgrounds pixels,
where obstacles suchthe mod
misclassified non-vegetation
as buildings, pixels
shadows, and other as vegetation
elements pixels. In
obscure vegetation contrast,
pixels, using
the model the high-co
misclas-
sified non-vegetation pixels as vegetation pixels. In contrast, using the
tributed input features, the suggested DNN technique could get better results and genhigh-contributed
input features, the suggested DNN technique could get better results and generate high-
ate high-resolution maps (shown in Figure 8). By looking at the figures, we can see t
resolution maps (shown in Figure 8). By looking at the figures, we can see the effect of the
effecthigh-contributed
of the high-contributed features
features on the on theofperformance
performance the DNN model. ofThese
the DNN model.
features helpedThese f
tures the
helped
modelthe modelato
to identify identify
smaller a smaller
number number
of FP pixels for theof FP pixels
classes for the
and classify theclasses
images and cl
more accurately.
sify the images more accurately.

Figure 8. Qualitative outcomes

Figure 8. Qualitative of theofpresented
outcomes DNN
the presented model
DNN modelfor vegetation
for vegetation mapping
mapping withwith high-contributed
high-contributed input features.
input features.

We also generated the assessment metrics to measure the quantitative outputs achieved
by our proposed methodology. Tables 2 and 3 show the average quantitative results for
the imagery, with all the features and with high-contributed input features, respectively.
As demonstrated in Table 2, the proposed approach achieved classification accuracy of an
average OA of 93.11% when we utilized all the input features, while the model achieved
an average OA of 94.44% when we used the high-ranked input features for classification
(Table 3). In other words, adding high-contributed features to the input improves classifica-
tion map accuracy, and the method could increase accuracy by 1.33% compared to results
obtained with all inputs. In addition to increasing the model’s OA, the combination of
high-contributed features also enhanced the accuracy for other metrics of each class label.
 Sensors 2021, 21, 4738 13 of 16

For instance, the average accuracy of F1 score was attained at 90.66% when the suggested
network was applied to images with all features. In contrast, after using the features with
high contribution, the model was able to achieve an accuracy of 92.66% for the F1 score. In
fact, the proposed network could increase quantitative outcomes by 2%, allowing for more
Sensors 2021, 21, x FOR PEER REVIEW accurate identification of vegetation pixels, even in the presence of complicated areas and 13 of 17
large areas of occlusion. The reason is that the proposed DNN model extract deep features
from input data in a hierarchical manner and thus improves the classification results.

Figure 9. Qualitative
Figure outcomes
9. Qualitative of the
outcomes presented
of the presentedDNN
DNN model forvegetation
model for vegetation mapping
mapping withwith
all theall the features.
input input features.

We 2.also
Table generated
Quantitative resultsthe assessment
attained metrics
by the model to input
with all the measure the
features. quantitative outputs
achieved by our proposed methodology. Tables 2 and 3 show the average quantitative
Precision Recall F1_score OA
results for the imagery, with all the features and with high-contributed input features,
Background 0.94 0.98 0.96
respectively. As demonstrated
High-vegetation 0.93
in Table 2,0.83
the proposed approach
0.88
achieved
93.11
classification
accuracy of an average OA
Low-vegetation 0.90of 93.11% when 0.87 we utilized 0.88
all the input features, while the
model achieved an average OA of 94.44% when we used the high-ranked input features
for classification (Table
Table 3. Quantitative 3). attained
results In other bywords,
the modeladding high-contributed
with high-contributed features.features to the input
improves classification map accuracy, and the method could increase accuracy by 1.33%
Precision Recall F1_score OA
compared to results obtained with all inputs. In addition to increasing the model’s OA,
Background 0.95 0.99 0.97
the combination of high-contributed features also enhanced the accuracy for other metrics
High-vegetation 0.95 0.84 0.89 94.44
of each class label. For instance,
Low-vegetation 0.90 the average
0.94 accuracy of 0.92
F1 score was attained at 90.66%
when the suggested network was applied to images with all features. In contrast, after
using the features with high contribution, the model was able to achieve an accuracy of
92.66% for the F1 score. In fact, the proposed network could increase quantitative out-
comes by 2%, allowing for more accurate identification of vegetation pixels, even in the
presence of complicated areas and large areas of occlusion. The reason is that the proposed
Sensors 2021, 21, 4738 14 of 16

Moreover, we compared our results with other comparative methods to show the
effectiveness of the proposed technique in urban vegetation mapping. For example, Zhang,
Feng, and Jiang [21] applied an object-oriented approach that can incorporate spatial and
spectral information of objects in the classification procedure for vegetation mapping from
IKONOS images. They achieved an average OA of 84.18%, which has a 10.26% difference
compared to our method. In another work, Feng, Liu, and Gong [1] implemented a
hybrid approach that combined RF and texture analysis to distinguish land covers in urban
vegetated regions from UAV imagery. An OA of 88.4% was achieved by their method, while
it is less than the OA obtained by our method. In [4], XGBoost method was applied for
vegetation mapping from MODIS and LANDSAT data for two various cases (New Zealand
and the Dzungarian Basin in China). They obtained an average OA of 92.65% that is 1.79%
less than the results achieved by our method. By comparing the results achieved with other
methods and our proposed method, it is shown that the DNN technique demonstrated
efficiency in vegetation mapping from aerial images.

5. Conclusions
We used aerial imagery to categorize land-use types and generate urban vegetation
maps using a deep neural network. However, to (1) comprehend model decisions, (2) grasp
complicated inherent non-linear relations, and (3) determine the model’s suitability for
more monitoring and evaluation, it is critical to comprehend why a data-driven method
makes any classification based on specific input data. Thus, we also used an explanation
model called SHAP to interpret the outputs of a DNN model designed for classification
and analyze the feature importance. In fact, we applied the SHAP technique to select
appropriate features for classification based on allocating SHAP values for every feature
of every data point that contributes to the model’s output. We utilized some spectral and
textural features as an input for the model to classify the imagery and generate vegetation
maps. Based on interpreting the model’s output using SHAP, we realized features such
as GLCM_Homogeneity, GLCM_Mean, GLCM_Dissimilarity, Hue, GLCM_Homogeneity,
GLCM_Mean, and Brightness presented high contribution information to the output of the
proposed model, while GLCM_Correlation, GLCM_Entropy, Value, and GLCM_Contrast
showed low contribution. Therefore, we tried to classify the images with those low-ranked
and high-ranked features to find the influence of these attributes on the model’s perfor-
mance. We achieved both visualization and quantitative outcomes, which confirmed that
the presented DNN model could enhance the OA accuracy and generate high-resolution
vegetation maps using the high-contributed features. This study also showed that the
SHAP yielded the features’ importance values and enabled the interpretation of the model
and its predictions, which confirm that the method is applicable and appropriate for inter-
preting the machine learning models. To end this, SHAP allows us to conduct in-depth
analysis of the data and leads us in the selecting appropriate features and AI model for veg-
etation classification. However, the DNN algorithm disregarded the importance between
features, which implies we may lose the association between features. Secondly, when
training a large dataset, the kernel SHAP approach is slow. Finally, the work’s interpreters
aided us in breaking open the black box, but further research is needed to compare and
evaluate existing interpretation approaches.

Author Contributions: A.A. carried out the investigations, analyzed the data, and drafted the
article; B.P. conceptualized, supervised, investigated, and visualized the article, conducted project
administration, resource allocation, writing—review and editing, and reorganized the article; B.P.
effectively enhanced the article and acquired the finance. All authors have read and agreed to the
published version of the manuscript.
Funding: The Centre for Advanced Modelling and Geospatial Information Systems, Faculty of Engineer-
ing and IT, University of Technology Sydney, funded this research.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Sensors 2021, 21, 4738 15 of 16

Data Availability Statement: The link to download the Massachusetts dataset can be found in the
online version, at https://www.airs-dataset.com/ (accessed on 18 June 2021).
Conflicts of Interest: There are no conflicts of interest reported by the authors.

References
1. Feng, Q.; Liu, J.; Gong, J. UAV remote sensing for urban vegetation mapping using random forest and texture analysis. Remote
Sens. 2015, 7, 1074–1094. [CrossRef]
2. Tooke, T.R.; Coops, N.C.; Goodwin, N.R.; Voogt, J.A. Extracting urban vegetation characteristics using spectral mixture analysis
and decision tree classifications. Remote Sens Environ. 2009, 113, 398–407. [CrossRef]
3. Li, X.; Shao, G. Object-based urban vegetation mapping with high-resolution aerial photography as a single data source. Int. J.
Remote Sens. 2013, 34, 771–789. [CrossRef]
4. Zhang, H.; Eziz, A.; Xiao, J.; Tao, S.; Wang, S.; Tang, Z.; Zhu, J.; Fang, J. High-resolution vegetation mapping using eXtreme
gradient boosting based on extensive features. Remote Sens. 2019, 11, 1505. [CrossRef]
5. Guan, H.; Su, Y.; Hu, T.; Chen, J.; Guo, Q. An object-based strategy for improving the accuracy of spatiotemporal satellite imagery
fusion for vegetation-mapping applications. Remote Sens. 2019, 11, 2927. [CrossRef]
6. Sharma, R.C.; Hara, K.; Tateishi, R. High-resolution vegetation mapping in japan by combining sentinel-2 and landsat 8 based
multi-temporal datasets through machine learning and cross-validation approach. Land Degrad. 2017, 6, 50. [CrossRef]
7. Johansen, K.; Coops, N.C.; Gergel, S.E.; Stange, Y. Application of high spatial resolution satellite imagery for riparian and forest
ecosystem classification. Remote Sens. Environ. 2007, 110, 29–44. [CrossRef]
8. Lu, D.; Hetrick, S.; Moran, E. Land cover classification in a complex urban-rural landscape with QuickBird imagery. Photogramm.
Eng. Remote Sens. 2010, 76, 1159–1168. [CrossRef]
9. Honkavaara, E.; Saari, H.; Kaivosoja, J.; Pölönen, I.; Hakala, T.; Litkey, P.; Mäkynen, J.; Pesonen, L. Processing and assessment of
spectrometric, stereoscopic imagery collected using a lightweight UAV spectral camera for precision agriculture. Remote Sens.
2013, 5, 5006–5039. [CrossRef]
10. Laliberte, A.S.; Goforth, M.A.; Steele, C.M.; Rango, A. Multispectral remote sensing from unmanned aircraft: Image processing
workflows and applications for rangeland environments. Remote Sens. 2011, 3, 2529–2551. [CrossRef]
11. Powell, R.L.; Roberts, D.A.; Dennison, P.E.; Hess, L.L. Sub-pixel mapping of urban land cover using multiple endmember spectral
mixture analysis: Manaus, Brazil. Remote Sens. Environ. 2007, 106, 253–267. [CrossRef]
12. Agüera, F.; Aguilar, F.J.; Aguilar, M.A. Using texture analysis to improve per-pixel classification of very high resolution images
for mapping plastic greenhouses. ISPRS J. Photogramm. 2008, 63, 635–646. [CrossRef]
13. Szantoi, Z.; Escobedo, F.; Abd-Elrahman, A.; Smith, S.; Pearlstine, L. Analyzing fine-scale wetland composition using high
resolution imagery and texture features. Int. J. Appl. Earth Obs. Geoinf. 2013, 23, 204–212. [CrossRef]
14. Cleve, C.; Kelly, M.; Kearns, F.R.; Moritz, M. Classification of the wildland–urban interface: A comparison of pixel-and object-
based classifications using high-resolution aerial photography. Comput. Environ. Urban. Syst. 2008, 32, 317–326. [CrossRef]
15. Yu, Q.; Gong, P.; Clinton, N.; Biging, G.; Kelly, M.; Schirokauer, D. Object-based detailed vegetation classification with airborne
high spatial resolution remote sensing imagery. Photogramm. Eng. Remote Sens. 2006, 72, 799–811. [CrossRef]
16. Abdollahi, A.; Pradhan, B. Integrated technique of segmentation and classification methods with connected components analysis
for road extraction from orthophoto images. Expert Syst. Appl. 2021, 176, 114908. [CrossRef]
17. Walter, V. Object-based classification of remote sensing data for change detection. ISPRS J. Photogramm. 2004, 58, 225–238.
[CrossRef]
18. Blaschke, T. Object based image analysis for remote sensing. ISPRS J. Photogramm. 2010, 65, 2–16. [CrossRef]
19. Mathieu, R.; Freeman, C.; Aryal, J. Mapping private gardens in urban areas using object-oriented techniques and very high-
resolution satellite imagery. Landsc. Urban. Plan. 2007, 81, 179–192. [CrossRef]
20. Puissant, A.; Rougier, S.; Stumpf, A. Object-oriented mapping of urban trees using Random Forest classifiers. Int. J. Appl. Earth
Obs. Geoinf. 2014, 26, 235–245. [CrossRef]
21. Zhang, X.; Feng, X.; Jiang, H. Object-oriented method for urban vegetation mapping using IKONOS imagery. Int. J. Remote Sens.
2010, 31, 177–196. [CrossRef]
22. Wang, H.; Zhao, X.; Zhang, X.; Wu, D.; Du, X. Long time series land cover classification in China from 1982 to 2015 based on
Bi-LSTM deep learning. Remote Sens. 2019, 11, 1639. [CrossRef]
23. Otukei, J.R.; Blaschke, T. Land cover change assessment using decision trees, support vector machines and maximum likelihood
classification algorithms. Int. J. Appl. Earth Obs. Geoinf. 2010, 12, S27–S31. [CrossRef]
24. Réjichi, S.; Chaâbane, F. SVM spatio-temporal vegetation classification using HR satellite images. In Sensors, Systems, and
Next-Generation Satellites; SPIE: Bellingham, WA, USA, 2011; p. 817620.
25. Wei, W.; Polap, D.; Li, X.; Woźniak, M.; Liu, J. Study on remote sensing image vegetation classification method based on decision
tree classifier. In IEEE Symposium Series on Computational Intelligence (SSCI); IEEE: Piscataway, NJ, USA, 2018; pp. 2292–2297.
26. Zhang, C.; Xie, Z. Combining object-based texture measures with a neural network for vegetation mapping in the Everglades
from hyperspectral imagery. Remote Sens. Environ. 2012, 124, 310–320. [CrossRef]
27. Tigges, J.; Lakes, T.; Hostert, P. Urban vegetation classification: Benefits of multitemporal RapidEye satellite data. Remote Sens.
Environ. 2013, 136, 66–75. [CrossRef]
Sensors 2021, 21, 4738 16 of 16

28. Boles, S.H.; Xiao, X.; Liu, J.; Zhang, Q.; Munkhtuya, S.; Chen, S.; Ojima, D. Land cover characterization of Temperate East Asia
using multi-temporal VEGETATION sensor data. Remote Sens. Environ. 2004, 90, 477–489. [CrossRef]
29. Lv, Z.; Hu, Y.; Zhong, H.; Wu, J.; Li, B.; Zhao, H. Parallel k-means clustering of remote sensing images based on mapreduce.
In Proceedings of the International Conference on Web Information Systems and Mining, Sanya, China, 23–24 October 2010;
pp. 162–170.
30. Anchang, J.Y.; Ananga, E.O.; Pu, R. An efficient unsupervised index based approach for mapping urban vegetation from IKONOS
imagery. Int. J. Appl. Earth Obs. Geoinf. 2016, 50, 211–220. [CrossRef]
31. Iounousse, J.; Er-Raki, S.; El Motassadeq, A.; Chehouani, H. Using an unsupervised approach of Probabilistic Neural Network
(PNN) for land use classification from multitemporal satellite images. Appl. Soft Comput. 2015, 30, 1–13. [CrossRef]
32. Qiu, J.; Wu, Q.; Ding, G.; Xu, Y.; Feng, S. A survey of machine learning for big data processing. EURASIP J. Adv. Signal. Process.
2016, 2016, 1–16.
33. Schiewe, J.; Tufte, L.; Ehlers, M. Potential and problems of multi-scale segmentation methods in remote sensing. GeoBIT GIS 2001,
6, 34–39.
34. Jia, X.; Kuo, B.-C.; Crawford, M.M. Feature mining for hyperspectral image classification. Proc. IEEE. 2013, 101, 676–697.
[CrossRef]
35. Cheng, Q.; Varshney, P.K.; Arora, M.K. Logistic regression for feature selection and soft classification of remote sensing data. IEEE
Geosci. Remote. Sens. Lett. 2006, 3, 491–494. [CrossRef]
36. Liu, Z.; Li, X.; Luo, P.; Loy, C.-C.; Tang, X. Semantic image segmentation via deep parsing network. In Proceedings of the IEEE
International Conference on Computer Vision, Santiago, Chile, 7–13 December 2015; pp. 1377–1385.
37. Abdollahi, A.; Pradhan, B.; Alamri, A.M. An ensemble architecture of deep convolutional segnet and unet networks for building
semantic segmentation from high-resolution aerial images. Geocarto Int. 2020, 1–13. [CrossRef]
38. Antwarg, L.; Miller, R.M.; Shapira, B.; Rokach, L. Explaining Anomalies Detected by Autoencoders Using SHAP. arXiv 2019,
arXiv:1903.02407.
39. Lundberg, S.; Lee, S.-I. A unified approach to interpreting model predictions. arXiv 2017, arXiv:1705.07874.
40. Matin, S.S.; Pradhan, B. Earthquake-Induced Building-Damage Mapping Using Explainable AI (XAI). Sensors 2021, 21, 4489.
[CrossRef] [PubMed]
41. Chen, Q.; Wang, L.; Wu, Y.; Wu, G.; Guo, Z.; Waslander, S.L. Aerial imagery for roof segmentation: A large-scale dataset towards
automatic mapping of buildings. ISPRS J. Photogramm. Remote Sens. 2018, 147, 42–55. [CrossRef]
42. Hassanein, M.; Lari, Z.; El-Sheimy, N. A new vegetation segmentation approach for cropped fields based on threshold detection
from hue histograms. Sensors 2018, 18, 1253. [CrossRef]
43. Lin, F.; Zhang, D.; Huang, Y.; Wang, X.; Chen, X. Detection of corn and weed species by the combination of spectral, shape and
textural features. Sustainability 2017, 9, 1335. [CrossRef]
44. Lan, Z.; Liu, Y. Study on multi-scale window determination for GLCM texture description in high-resolution remote sensing
image geo-analysis supported by GIS and domain knowledge. ISPRS Int. J. Geo Inf. 2018, 7, 175. [CrossRef]
45. Zhou, Y.; Zhang, R.; Wang, S.; Wang, F. Feature selection method based on high-resolution remote sensing images and the effect
of sensitive features on classification accuracy. Sensors 2018, 18, 2013. [CrossRef] [PubMed]
46. Sameen, M.I.; Pradhan, B.; Aziz, O.S. Classification of Very High Resolution Aerial Photos Using Spectral-Spatial Convolutional
Neural Networks. J. Sens. 2018, 2018, 1–12. [CrossRef]
47. Glorot, X.; Bordes, A.; Bengio, Y. Deep sparse rectifier neural networks. In Proceedings of the Fourteenth International Conference
on Artificial Intelligence and Statistics, Fort Lauderdale, FL, USA, 11–13 April 2011; pp. 315–323.
48. Abdollahi, A.; Pradhan, B.; Sharma, G.; Maulud, K.N.A.; Alamri, A. Improving road semantic segmentation using generative
adversarial network. IEEE Access 2021, 9, 64381–64392. [CrossRef]
49. Murugan, P. Implementation of Deep Convolutional Neural Network in Multi-Class Categorical Image Classification. arXiv 2018,
arXiv:1801.01397.
50. Abdollahi, A.; Pradhan, B.; Alamri, A. VNet: An end-to-end fully convolutional neural network for road extraction from
high-resolution remote sensing data. IEEE Access 2020, 8, 179424–179436. [CrossRef]
51. Chen, S. Interpretation of multi-label classification models using shapley values. arXiv 2021, arXiv:2104.10505.
52. Shrikumar, A.; Greenside, P.; Kundaje, A. Learning important features through propagating activation differences. In Proceedings
of the International Conference on Machine Learning, Sydney, Australia, 6–11 August 2017; pp. 3145–3153.
53. Ribeiro, M.T.; Singh, S.; Guestrin, C. Why should i trust you? Explaining the predictions of any classifier. In Proceedings
of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA,
13–17 August 2016; pp. 1135–1144.
54. Mangalathu, S.; Hwang, S.-H.; Jeon, J.-S. Failure mode and effects analysis of RC members based on machine-learning-based
SHapley Additive exPlanations (SHAP) approach. Eng. Struct. 2020, 219, 110927. [CrossRef]
55. García, M.V.; Aznarte, J.L. Shapley additive explanations for NO2 forecasting. Ecol. Inform. 2020, 56, 101039. [CrossRef]