Testing a New Ensemble Vegetation Classification Method Based on Deep Learning and Machine Learning Methods Using Aerial Photogrammetric Images

Abstract

The objective of this research is to report results from a new ensemble method for vegetation classification that uses deep learning (DL) and machine learning (ML) techniques. Deep learning and machine learning architectures have recently been used in methods for vegetation classification, proving their efficacy in several scientific investigations. However, some limitations have been highlighted in the literature, such as insufficient model variance and restricted generalization capabilities. Ensemble DL and ML models has often been recommended as a feasible method to overcome these constraints. A considerable increase in classification accuracy for vegetation classification was achieved by growing an ensemble of decision trees and allowing them to vote for the most popular class. An ensemble DL and ML architecture is presented in this study to increase the prediction capability of individual DL and ML models. Three DL and ML models, namely Convolutional Neural Network (CNN), Random Forest (RF), and biased Support vector machine (B-SVM), are used to classify vegetation in the Eastern part of Serbia, together with their ensemble form (CNN-RF-BSVM). The suggested DL and ML ensemble architecture achieved the best modeling results with overall accuracy values (0.93), followed by CNN (0.90), RF (0.91), and B-SVM (0.88). The results showed that the suggested ensemble model outperformed the DL and ML models in terms of overall accuracy by up to 5%, which was validated by the Wilcoxon signed-rank test. According to this research, RF classifiers require fewer and easier-to-define user-defined parameters than B-SVMs and CNN methods. According to overall accuracy analysis, the proposed ensemble technique CNN-RF-BSVM also significantly improved classification accuracy (by 4%).

Full text

Testing a New Ensemble Vegetation
Classification Method Based on Deep
Learning and Machine Learning
Methods Using Aerial
Photogrammetric Images
Siniša Drobnjak
1
,
2
*, Marko Stojanović
1
,
2
, Dejan Djordjević
1
,
2
,Saša Bakrač
1
,
2
,
Jasmina Jovanović
3
and Aleksandar Djordjević
3
1
Military Geographical Institute, Belgrade, Serbia,
2
Military Academy, University of Defense, Belgrade, Serbia,
3
Geography
Faculty, University of Belgrade, Belgrade, Serbia
The objective of this research is to report results from a new ensemble method for
vegetation classification that uses deep learning (DL) and machine learning (ML)
techniques. Deep learning and machine learning architectures have recently been used
in methods for vegetation classification, proving their efficacy in several scientific
investigations. However, some limitations have been highlighted in the literature, such
as insufficient model variance and restricted generalization capabilities. Ensemble DL and
ML models has often been recommended as a feasible method to overcome these
constraints. A considerable increase in classification accuracy for vegetation classification
was achieved by growing an ensemble of decision trees and allowing them to vote for the
most popular class. An ensemble DL and ML architecture is presented in this study to
increase the prediction capability of individual DL and ML models. Three DL and ML
models, namely Convolutional Neural Network (CNN), Random Forest (RF), and biased
Support vector machine (B-SVM), are used to classify vegetation in the Eastern part of
Serbia, together with their ensemble form (CNN-RF-BSVM). The suggested DL and ML
ensemble architecture achieved the best modeling results with overall accuracy values
(0.93), followed by CNN (0.90), RF (0.91), and B-SVM (0.88). The results showed that the
suggested ensemble model outperformed the DL and ML models in terms of overall
accuracy by up to 5%, which was validated by the Wilcoxon signed-rank test. According to
this research, RF classifiers require fewer and easier-to-define user-defined parameters
than B-SVMs and CNN methods. According to overall accuracy analysis, the proposed
ensemble technique CNN-RF-BSVM also significantly improved classification
accuracy (by 4%).
Keywords: ensemble method, machine learning, deep learning, vegetation classification, satellite and aerial images
Edited by:
Jelena Golijanin,
University of East Sarajevo, Bosnia
and Herzegovina
Reviewed by:
Luís Pádua,
Centre for the Research and
Technology of Agro-Environmental
and Biological Sciences (CITAB),
Portugal
Ke-Seng Cheng,
National Taiwan University, Taiwan
*Correspondence:
Siniša Drobnjak
sinisadrobnjak@vs.rs
Specialty section:
This article was submitted to
Environmental Informatics and Remote
Sensing,
a section of the journal
Frontiers in Environmental Science
Received: 14 March 2022
Accepted: 06 May 2022
Published: 25 May 2022
Citation:
Drobnjak S, StojanovićM, DjordjevićD,
BakračS, JovanovićJ and DjordjevićA
(2022) Testing a New Ensemble
Vegetation Classification Method
Based on Deep Learning and Machine
Learning Methods Using Aerial
Photogrammetric Images.
Front. Environ. Sci. 10:896158.
doi: 10.3389/fenvs.2022.896158
Frontiers in Environmental Science | www.frontiersin.org May 2022 | Volume 10 | Article 8961581
ORIGINAL RESEARCH
published: 25 May 2022
doi: 10.3389/fenvs.2022.896158

1 INTRODUCTION
Forests are a valuable natural resource in many countries, with
wood and forestry products serving as the primary export
cheeses. They’re also crucial in water management, tourism
and recreation, wildlife protection, and soil erosion control.
The process of photosynthesis allows plants to play a critical
role in all major planetary cycles, including water circulation in
nature, energy exchange, oxygen, carbon dioxide, and other
elements between biotic and abiotic regions (Drobnjak et al.,
2018;Wang et al., 2021).
Satellite and aerial images are effective instruments for
monitoring and studying forests and other vegetation. Satellite
images are useful equipment for forest monitoring, and remote
sensing research has become a very effective method. Satellite
images can be used to explore the borders between different types
of vegetation, the degree of vegetation development, vegetation
morphology, forest health, tree canopy humidity, diverse textures,
biomass, and a variety of other parameters (Drobnjak et al., 2013;
Bakračet al., 2018;Drobnjak et al., 2018).
Only radiometric, spatial, and spectrally enhanced images are
suitable for further digital analysis to collect the data required for
vegetation classification. Classification is the process of grouping
pixels into thematic groups or classes using statistical methods
and detecting the association between their digital values. It is one
of the most difficult processes in computer image processing in
terms of operator knowledge. In practice, classification methods
entail assessing the image’s content and grouping pixels into the
proper data categories (Running et al., 1995;Yu et al., 2006;Xie
et al., 2008). The unification is carried out according to a
predetermined numerical analysis decision rule (application of
the corresponding key). This is accomplished by statistically
categorizing pixels into thematic groups based on their digital
values, as well as the relationship between the contents of the
entities, referred to as “class”(Running et al., 1995).
The use of a combination of many classifiers to achieve a single
classification has been documented in the remote sensing
literature several times in recent years (Yu et al., 2006;Xie
et al., 2008;Engler et al., 2013;Kussul et al., 2017;Meng et al.,
2017;Amini et al., 2018;Drobnjak et al., 2018;Ayhan et al., 2020).
The ensemble classifier that results is often found to be more
accurate than any of the individual classifiers that make up the
ensemble. To categorize unknown causes, an ensemble classifier
employs weighted or unweighted voting to integrate the decisions
of a group of classifiers (Dietterich, 2000;Engler et al., 2013). For
vegetation classification, studies that used boosting with a
decision tree as the base classifier indicated a considerable
increase in classification accuracy (Chan and Paelinckx, 2008;
Xie et al., 2008). In the past, the random forest (RF) algorithm has
proved successful in producing realistic vegetation maps
(Ghimire et al., 2010). RF has been successfully utilized to
extract physiological plant features (Doktor et al., 2014),
estimate plant biomass (Adam et al., 2014), and map plant
species in studies using multispectral data for forest sciences
(Burai et al., 2015).
SVM is frequently cited as the best method for dealing with
difficult classification issues such as tree species discrimination,
with RF coming in second (Ghosh et al., 2014). Ghosh et al.
(2014) used information from a broader electromagnetic
spectrum (450–2,500 nm) to employ SVM and RF on
multispectral data to categorize five tree species in managed
woods in central Germany.
The purpose of this paper is to discuss the findings obtained
utilizing a combination of Random Forest, a biased Support
vector machine, and a Convolutional Neural Network
classifier. All mentioned classifiers use a bootstrapped sample
of the training data to select a random set of features and create a
classifier. This generates a large number of trees (classifiers), and
then unweighted voting is used to assign an unknown pixel to a
class (Shaheen and Verma, 2016;Sothe et al., 2020;Gašparović
and Dobrinić, 2020;Zhang et al., 2020;Fei et al., 2022). The new
ensemble classifier’s performance is also compared to that of
single classifiers in terms of classification accuracy, training time,
and user-defined parameters (Meng et al., 2017).
Machine learning algorithms define computer-based tools that
allow for exploratory data and statistical analysis to uncover
unknown patterns and relationships in dataset values ahead of
time. The current study used supervised and flexible machine
learning algorithms, deep learning algorithms, and their
ensemble to categorize vegetation areas in the eastern part of
Republic Serbia’s Suva Planina Mountain.
2 MATERIALS AND METHODS
2.1 Study Area and Remote Sensed Data
Acquisition
Forest area in Republic Serbia covered 27,200 km
2
which is
approximately 31.1% of the country area. The study area
includes parts of Mountain Suva Planina near NišCity,
between latitudes of 43°15′15″–43°19′45″N, and longitudes of
22°20′15″–22°30′00″E. The area covered by the test area is
109.7 km
2
. The minimum altitude of the test area is 326.4 m,
the maximum altitude is 1,154.8 m, and the average altitude of the
test area is 680.9 m. It is located in the eastern part of the Republic
of Serbia (Figure 1).
Data from the digital sensors of the satellite system Sentinel-
2A and the digital aerial photogrammetric camera Leica ADS80
were used to create the combination of aerial photogrammetric
and satellite images (Running et al., 1995;Amarsaikhan and
Douglas, 2004).
Sentinel-2A is the first optical Earth observation sensor
developed and built by Airbus (Airbus Defense and
Space—ADS) for the European Space Agency’s (ESA) needs as
part of the European Copernicus program (Table 1). Sentinel-2A
is the first civil optical Earth observation satellite with sensors in
four “Red Edge”wavelengths, which provides critical data on
vegetation on the planet’s surface (Fernández-Manso et al., 2016;
Mallinis et al., 2018).
In addition, the aerial Photogrammetric Acquisition System of
the Military Geographic Institute consists of airplane Piper
Seneca V and digital aerial photogrammetric camera Leica
ADS80 (Figure 2): The system provides a modern approach
in the field of collecting and analyzing geospatial data for the
Frontiers in Environmental Science | www.frontiersin.org May 2022 | Volume 10 | Article 8961582
Drobnjak et al. New Ensemble Vegetation Classification Method

needs of the defense system entities and other users in the country
(Drobnjak et al., 2018).
In this study, we used data obtained from a multispectral
sensor (panchromatic, RGB, and infrared bands)—digital camera
Leica ADS80 (Drobnjak et al., 2018), which has a line sensor with
a resolution of 6.5 μm, with 12,000 pixels per line or 24,000 pixels
when using HiRes Mode, with Lens focus 62.7 mm. The above
aerial photogrammetric images were downscaled with satellite
images of the Sentinel 2A mission.
During the field research in 2020 and 2021, samples for
training and testing datasets were collected. Localization of
selected tree species was achieved during data collecting. Only
regions currently occupied by living trees above 5 m height were
deemed acceptable location sources during field data collecting.
The chosen sampling sites are required to have a minimum of five
trees of the same species within a 3-m radius of the GPS receiver.
In this study, we used the Trimble T10 tablet GPS device which is
a powerful, rugged device created for survey fieldwork, mapping,
and GIS data collection and at the same time supports demanding
desktop applications. Trimble T10 has Windows 10 Enterprise
operating system, with a 10.1″screen size, Intel i7 processor,
internal GPS with SBAS, 8 GB memory, and 256 GB data storage.
FIGURE 1 | Location of the study area.
FIGURE 2 | Aerial photogrammetric recording system.
Frontiers in Environmental Science | www.frontiersin.org May 2022 | Volume 10 | Article 8961583
Drobnjak et al. New Ensemble Vegetation Classification Method

Only measurements with a localization error of less than 1.5 m
were chosen. The coordinates of polygon corners were recorded
for larger areas and then used for pixel extraction. Areas that were
definitely in shade and pixels that were uncertain were eliminated.
2.2 Methods
The Leica ADS80 multispectral dataset was then utilized to
extract training and testing samples from these locations. Leica
ADS80 capabilities include perfectly co-registered multispectral
bands and true stereo image collection. The spatial resolution of
the multispectral (RGB and Infrared bands) aerial
photogrammetric images used in the paper was 40 cm. The
flight altitude of the plane during the aerial photogrammetric
scanning was 4,000 m. Using a combination of aerial
photogrammetric images and satellite images, the spatial
resolution was downscaled to 2.5 m. Machine and Deep
learning classification methods were used on such images to
create a thematic layer of vegetation.
Supervised vegetation classification consists of a training stage
and an evaluation performance stage, and a confusion matrix is
constructed and used for accuracy assessment. In this study, we
used collected reference test samples with different NDVI indexes
and different vegetation textures and shapes. Using a GIS
program, we categorized the different forest types data as
training and testing samples for our experimental setup. The
labeled data was collected in the field, alongside additional high-
resolution imagery from other datasets and imaging (both
satellite and aerial). We defined a total of eight vegetation
classes based on the different types of forest vegetation found
in the test region and included them in the analysis. Test samples
were directly mapped from aerial photogrammetric images as
polygons of different dimensions and thus stored in the reference
test sample database.
A total of 398 forest-type vegetation features (polygons) and
225 non-forest vegetation features (e.g., water, soil, grass, and
other land coverings) were annotated on a combination of aerial
and satellite photos, resulting in 623 various sizes polygons.
Although the proximity of polygons makes it appear like some
of them are present in both subsets, this is not the case. This
happened only when small polygons were represented in the
figure size because the training and testing sets had completely
distinct features. We used the bootstrap technique to define the
training and testing datasets to explore the performance of the
machine and deep learning algorithms in the classification of
forest vegetation (polygon features).
The sample size and quality of training data have generally had
a large impact on the classification accuracy. In this regard, we
divided the dataset while ensuring that both training and testing
sets contained similar sampling patterns, being representatives of
all conditions observed in the area during labeling. Using a large
number of reference samples the uncertainty of the estimator can
be evaluated.
Because the majority of supervised classifiers are sensitive to
the data used for training, classification results will vary based on
the training dataset. Furthermore, in order to exclude human bias
from classification results, we chose to use a technique that
included a random selection of training and testing datasets
that belong to the already mentioned test sample polygons.
We chose the 0.632 bootstrap strategy for producing the test
and training datasets based on the work of (Ghosh et al., 2014;
Neto and Dougherty, 2015).
Bootstrapping is a statistical technique for producing random
samples and estimating the distribution of a population estimator
using a random sample or a model estimated from a random
sample (Ghosh and Prajneshu, 2011). It entails examining the
data as if it were a population in order to assess the distribution of
interest. When determining the asymptotic distribution of an
estimator or statistic is challenging, bootstrapping can be used to
replace computation with mathematical analysis.
The entire method was d divided into several iterations. Each
cycle involves a random split of all samples into test and training
datasets, with 63.2% of samples going to the training dataset and
the rest going to the test dataset, which is not used in the classifier
training process and belongs to the already mentioned test sample
polygons.
Table 2 shows the exact amount of samples/pixels assigned to
each class. Following this, classification was performed using the
given training samples and classification method.
Figure 3 depicts the flowchart of the method utilized in the
study. The dataset construction is demonstrated in the first step,
where all data is entered into the database, including a
combination of satellite and aerial photogrammetry photos as
well as vector data of test samples. Models of biased Support
vector machines, Random Forests, and Convolutional Neural
Networks, as well as their ensemble classification methods, were
used in the following. For this study, machine learning and deep
learning classification algorithms with their ensemble classifier
were evaluated through R software. Then, the precision, total
accuracy, and kappa coefficient were used to validate the built
models. Finally, we used the Wilcoxon Signed-Rank significance
test to statistically test the proposed techniques.
Biased Support vector machines, Random Forests,
Convolutional Neural Networks, and their ensemble
classification algorithms all used the same training data and
were tested on the same test data, ensuring that the findings
were comparable.
The next stage was to compare classifiers by analysing the
differences in producer and user classification accuracy for
classes, as well as the overall accuracy and kappa coefficient
variability. The best (most accurate) iteration for each
classification method was chosen based on the results. The
final categorization images were created using the optimal
iteration parameters. With the help of an NDVI-based mask,
non-forested areas were masked out from the final images. To
avoid the classification of bushes and young tree stands,
vegetation smaller than 2 m was concealed. We achieved this
by mapping and field testing test samples containing lower trees
and low vegetation. Pixels having an NDVI value less than 0.25
were also masked to remove buildings and manufactured
elements.
2.3 Performance Evaluation
The proportion of the total number of correctly categorized pixels
across all classes and the total number of pixels in the confusion
Frontiers in Environmental Science | www.frontiersin.org May 2022 | Volume 10 | Article 8961584
Drobnjak et al. New Ensemble Vegetation Classification Method

TABLE 1 | Characteristics of Sentinel-2A images.
Sentinel -2A bands Central wavelength (µm) Bandwidth (nm) Spatial resolution (m)
Band 1—Coastal aerosol 0.443 21 60
Band 2—Blue 0.492 66 10
Band 3—Green 0.560 36 10
Band 4—Red 0.665 31 10
Band 5—Vegetation red edge 0.704 15 20
Band 6—Vegetation red edge 0.740 15 20
Band 7—Vegetation red edge 0.783 20 20
Band 8—Near-infrared 0.833 106 10
Band 8A—Vegetation red edge 0.865 21 20
Band 9—Water vapour 0.945 20 60
Band 10—Short-wave infrared—Cirrus 1.374 31 60
Band 11—Short-wave infrared 1.614 91 20
Band 12—Short-wave infrared 2.202 175 20
TABLE 2 | Training and testing sample sizes (in pixels) used for vegetation classifications.
Vegetation
classes
Class
1
Class
2
Class
3
Class
4
Class
5
Class
6
Class
7
Class
8
Training
samples
2,154 1,145 1,874 1,054 1,745 987 875 1,987
Testing
samples
1,361 724 1,184 666 1,102 624 553 1,256
Class 1.—Coniferous vegetation over 5 m; Class 2.—Deciduous vegetation over 5 m; Class 3.—Mixed vegetation over 5 m; Class 4.—Plantation forest over 5 m; Class 5.—Shrubs and
low vegetation; Class 6.—Orchards; Class 7.—Vineyards; Class 8.—Non-vegetation areas.
FIGURE 3 | Flowchart of the used methodology in the study.
Frontiers in Environmental Science | www.frontiersin.org May 2022 | Volume 10 | Article 8961585
Drobnjak et al. New Ensemble Vegetation Classification Method

matrix is referred to as overall accuracy (the total sum of pixels
divided by the sum of diagonal elements of the matrix). The
errors associated with individual classes are described by User
and Producer accuracies. The likelihood of a reference pixel being
correctly categorized is measured by the producer’s accuracy
(total number of pixels in that category determined from
reference data divided by the total number of pixels in that
category). The likelihood that the predicted sample class
matches the reference class is the user’s accuracy (the total
number of correct classifications for a particular class and
dividing it by the row total).
Overall accuracy k
i1nii
n*100 (1)
User’s accuracy nii
ni+
(2)
Producer’s accuracy nii
n+i
(3)
With the usage of the confusion matrix, we get a coefficient of
kappa statistics which is a good indicator of the choice of
classification method consistency taking their randomness into
account. Kappa coefficient (κ) is a coefficient that quantifies the
degree of compatibility between assigned classes when
misclassification is removed.
In general, the kappa coefficient is being reduced with
enlargement of the number of classes, i.e., the better classes
are selected the greater possibility of an error in classification.
Kappa coefficient is κ= 0 for the clear compatibility between the
two total coincidental classifications and it reaches κ= 1 for
complete harmonization between the classification and data. For
unexpectedly accurate class agreement, kappa statistics are
utilized as a measure of classification accuracy.
Kappacoefficient nk
i1nii −k
i1ni+n+i
n2−k
i1ni+n+i
(4)
With a random distribution of pixels in the classes, the registered
value indicates the overall classification accuracy and consistency
between the image and the reference grid. According to Landis
and Koch (Landis and Koch, 1977), values of Kappa coefficient
greater than 0.8 indicate perfect agreement, values between 0.6
and 0.8 indicate substantial agreement, values between 0.4 and 0.6
indicate moderate agreement, and values between 0.2 and 0.4
indicate fair agreement, and values below 0.2 indicate poor
agreement. Furthermore, to compare the classification
performances of the ML, DL, and their ensemble models, a
statistical significance test (Wilcoxon signed-rank test) is used
(Woolson, 2008). The Wilcoxon signed-ranked test, a
nonparametric hypothesis test, is used to statistically evaluate
the efficacy of the models developed. The test has been widely
used to determine the statistical significance of performance
differences between models and to compare them pair-wise
(Woolson, 2008). The Wilcoxon signed-rank test’s null
hypothesis is that there is no statistical difference between the
models at a 95% confidence range. By using Wilcoxon signed-
rank test we calculate how far each value of the producer’s
accuracy, user’s accuracy, and the overall accuracy of
individual classes is from the hypothetical median. Wilcoxon
signed-rank test p-values of the producer’s accuracy, user’s
accuracy, and the overall accuracy of individual classes were
greater than 0.05 which proves there is no statistical difference
between the models at a 95% confidence range.
3 MACHINE AND DEEP LEARNING
APPLICATIONS
3.1 Machine Learning Classification
Machine learning technique emerged as a response to the rigidity
of many computer programs in comparison to the unlimited
variability of the environment. One of the most difficult aspects of
feature detection from remote sensing images has been accurately
distinguishing real-world objects from a vast number of pixels.
Machine learning is a branch of computer science that studies
algorithms that learn from examples. Classification is a task that
necessitates the application of machine learning algorithms to
learn how to assign a class label to problem domain instances. In
machine learning, there are many distinct sorts of classification
tasks to be encountered and specialized modeling approaches to
be employed for each.
3.1.1 Biased Support Vector Machine Techniques
The support vector machine (SVM) is a commonly used
statistical machine learning technique that works on the
premise of risk minimization. The support vector machine
approach divides the classes using a final surface (referred to
as an ideal hyper-plane) that maximizes the margin between the
classes in the dataset. In the same way that a regular binary SVM
determines the best separation between two classes in feature
space, a biased SVM does the same. The acquired training data
from the focal class, on the other hand, is compared against
samples taken at random from the data pool (in this case, the
vegetation pixels from the entire island), which are referred to as
“pseudo-outliers”in this context (Chan and King;Hartono et al.,
2018). Because the pseudo-outlier data has no known identity and
will comprise samples from the focus class, errors in the pseudo-
outlier class are penalized less severely than errors in the
focal class.
Furthermore, the standard SVM approach makes two
assumptions: the positive and negative training samples are of
equal size, and the cost of misclassification for samples belonging
to various classes is essentially the same. For positive and negative
samples, the Biased-SVM method is used to apply various penalty
coefficients C. In this algorithm, the minority samples are given
higher penalty factors, while the majority samples are given lower
penalty factors. As a result, the SVM classifier can concentrate on
the minority class’s misclassification rate.
Assuming that D{(xi,y
i)}(1≤i≤n)is the training set,
where xidenotes the feature vector of Piand yi∈{−1,1}is
the label of Pi, the first verified parts are m−1 and they are
positive examples labeled as yi1(1≤i≤m−1), while the
rest are unlabeled part whose labels are set to
yi−1(m≤i≤n). Furthermore, two soft margin
parameters, C1and C2, are included to highlight the differing
Frontiers in Environmental Science | www.frontiersin.org May 2022 | Volume 10 | Article 8961586
Drobnjak et al. New Ensemble Vegetation Classification Method

tolerances on the training mistakes induced by positive and
unlabeled octapeptides, respectively (Foody and Mathur, 2004;
Hartono et al., 2018;Li et al., 2021). These two factors can likewise
be used to learn from a noisy unlabeled collection with cleaved
sections. The two L1-norm soft margins biased formulation of
SVM is described by Eq. 5:
Minimize:1
2ωtω+C1
m−1
i1
ξi+C2
m−1n
im
ξi
s.t.yiωtxi+b≥1−ξi,ξi≥0,i1,2,....,n (5)
where are:
•ωis the hyperplane’s normal vector separating positive and
unlabeled sections,
•ξ
i
refers to the slack variable for each part that is used to
calculate the mistake cost, and bsignifies the offset of
hyperplane from the origin along ω.
The B-SVM model is utilized in the vegetation classification
model using the radial basis function (RBF) kernel in this study.
Because the kernel width (γ), regularization constants (C1,C
2),
and bias ball affect the performance of the B-SVM model, these
parameters should be carefully monitored. For biased Support
vector machine modeling, the R open-source software “e1071”
package was utilized, and optimal settings were specified.
Parameters of B-SVM applied for forest vegetation
classification are:
•SVM type applied for model: Radial Basis function.
•Hyper-parameter: sigma = 0.054
•Number of Support Vectors: 33,368
•Objective Function Value: −93.072 and training error: 0.160
B-SVM parameterization is also done on the training dataset
using cross-validation. We discovered that this criterion worked
well for optimizing biased SVMs and outperformed an alternate
optimization criterion in this study regarding biased SVM
optimization for vegetation mapping. We also discovered that
cross-validation performed at the crown level worked well (i.e., by
splitting crowns rather than pixels into the cross-validation
groups).
The difficulty with SVM based on structural risk reduction in
classification for their balanced data is that the classification
weight will be biased towards the majority class, causing the
classification hyperplane to be close to the minority class, making
it simple to misclassify minority samples.
3.1.2 Random Forest Classification
Breiman (2001) created the Random Forests algorithm, which
consists of a collection of tree-structured classifiers
{h(x, Θk),k1,...}where the {Θk}are independent
identically distributed random vectors and each tree casts a
unit vote for the most frequent class to the input vector (x).
Instead of using the best variables, a Random Forest (RF)
classification divides each node using a random subset of
input characteristics or predictive factors, which decreases
generalization error.
During the training period, the RF algorithm builds numerous
classification trees, and the ultimate output of the model creation
process is the average value of all classification tree outputs.
In order to run the RF model, two main parameters of the
random forest model must be defined a priori: The square root of
the number of factors (mtry)and the number of trees to run the
model (ntree). The above parameters should be optimized to
minimize the generalization error. In general, the model chooses
the most accurate parameters available.
Additionally, the Random Forest training algorithm employs
the standard technique of bagging or boot-strap aggregation for
tree learners. The Gini Index is used by the RF technique to
determine the best split selection by measuring the impurity of a
particular element in relation to the other classes. The Gini index
is a measure of a distribution’s inequality (Breiman, 1996;
Breiman, 2001;Breiman and Cutler, 2007). The Gini index
can be computed by summing the probability Piof a single
class with label ibeing chosen multiplied by the probability

k≠i
pk1−piof a mistake in categorizing that class i. The
Gini Index can be expressed as the following equation for a
given training dataset T with j classes Eq. 6:
ITp
j
i1
pi
k≠i
pk1−
j
i1
p2
i(6)
where, i∈{1,2,...,j}. Therefore, a decision tree is made to
grow to its maximum depth by using a given combination of
features.
During the classification process, RF also provides an estimate
of the relative value of the various features or variables. The RF
swaps one of the input random variables while keeping the rest
constant to assess the relevance of each satellite and aerial
photogrammetry images bands, and it assesses the loss in
accuracy through error estimation and Gini Index decline
(Liaw and Wiener, 2002;Biau, 2012).
In addition, in this study, the number of trees (m
tree
) in RF was
fixed to 650 after a preliminary analysis and the number m of
variables sampled at each node was selected to be one. No
calibration set is needed to tune the parameters.
3.2 Deep Learning Classification
3.2.1 Convolutional Neural Network
Several CNN-based methods for assigning a label to each pixel of
a classified image have been presented in recent years. Aerial
images are being used to classify land cover, land use, and
different type of vegetation using deep learning approaches for
semantic segmentation (Kussul et al., 2017). We employ a
strategy that combines classification results from manually
derived and CNN features in this study. Initially, an image
patch was used to create two sets of features (Sothe et al.,
2020;Zhang et al., 2020;Emily and Sudha, 2022):
(a) NDVI, edges, saturation, and
(b) CNN features.
Frontiers in Environmental Science | www.frontiersin.org May 2022 | Volume 10 | Article 8961587
Drobnjak et al. New Ensemble Vegetation Classification Method

The traditional manual method for effectively predicting and
classifying images takes time, and inaccurate classification results
are another major difficulty. The convolutional neural network is
a better and more scalable solution for satellite and aerial images.
The CNN employs a computational method that involves linear
algebra and matrix multiplications in order to recognize images.
The CNN beat other networks in applications such as image
processing and speech recognition. There are three layers to the
CNN: convolutional, pooling, and fully connected (Nijhawan
et al., 2018;Kattenborn et al., 2021).
The principal calculation happens to be the vegetation block
among the three in the convolutional section, which comprises
the data, filter, and feature area. The pooling layer is in charge of
downsampling, also known as data sample dimension reduction.
In the pooling layers, there is also a filter that moves over the
input but has no weight. The pooling is separated into two parts: a
Max pool and an Average pool, each of which determines the
maximum and average value. The output layers are all connected
by a node to the previous layer, and classification tasks are done
using the feature collected from the previous layer (Ayhan et al.,
2020).
In this study, the hyper-parameter of CNN model applied for
forest vegetation classification are:
•Number of filters 1,000
•Number of units in fully connected layer 150
•Dropout rate 0.5
•Learning rate 0.001
•Number of epochs 10
•Batch size 50
3.3 Ensemble Machine and Deep Learning
Ensemble learning is a general meta-approach to machine
learning that seeks the best prediction performance by
combining many methods to get the highest accuracy.
Different machine learning algorithms may not be able to
produce the best results on their own, therefore combining
them will bring out the model’s full potential and improve
accuracy (Kavzoglu et al., 2015). It has been proven that
employing an ensemble learning methodology for the
prediction and classification of a combination of satellite and
aerial images yields better results than using a single classifier
(Shaheen and Verma, 2016;Dixit, 2019;Abdi, 2020;Fei et al.,
2022). Stacking using Random Forest and biased Support vector
machine algorithms, as well as deep learning convolutional neural
networks method, were the most commonly used classifiers for
vegetation (Engler et al., 2013;Kavzoglu et al., 2015;Kussul et al.,
2017;Abdi, 2020;Ayhan et al., 2020). The use of Ensemble
methods in satellite imaging may be studied with confidence, as
the accuracy obtained is significantly greater than that of single
classifiers or classical methods (Gigovićet al., 2019b).
Ensemble learning is divided into three categories: bagging,
stacking, and boosting. Bagging is concerned with making
multiple decisions on a different sample of the same dataset
and calculating the average forecast, whereas stacking is
concerned with fitting many different types of models on the
same data and learning the combined predictions using another
type of model (Dietterich, 2000;Engler et al., 2013). The boosting
process entails sequentially adding ensemble members to correct
the previous forecast made by the other models, and then taking
the average of the predictions.
In this study, we use Bayesian averaging and efficient feature
selection to create an ensemble model that addresses these
difficulties and mitigates their effects on defect classification
performance. For each data point, Bayesian averaging makes
many different classifications (Raftery et al., 2005;
Montgomery et al., 2012). We utilize the average of all the
models’classifications to produce the final classified map
within this method. In regression problems, Bayesian
averaging can be used to make classifications, and it can be
used to compute probabilities. A new ensemble learning
technique is suggested to give robustness to data imbalance
and feature redundancy, in addition to efficient feature
selection (Vrugt and Robinson, 2007).
4 RESULTS
Biased Support vector machines, Random Forests, Convolutional
Neural Networks, and their ensemble classification algorithms all
used the same training data and were tested on the same test data,
ensuring that the findings were comparable.
Figure 4 shows the obtained results of vegetation classification
in the test area using machine learning and deep learning
methods, as well as their ensemble methods. The lines of the
vegetation contours are shown in different colors (as shown in the
legend) in order to identify the obtained classification results.
As shown in Figure 5, the classification results produce
roughly identical vegetation contours, especially in locations
where the vegetation boundary is well separated in the images
in comparison to other content. Smaller, but very significant
differences are observed in the parts of the test area where the
boundaries of vegetation are not clearly visible on the
combination of satellite and aerial images. These minor
deviations mostly affected the accuracy of the applied
classification methods.
For machine and deep learning classification accuracy testing,
the confusion (error) matrix is widely utilized. A confusion
matrix is a basic cross-tabulation of the predicted class label
against the reference data for a sample of cases at certain
locations, and it serves as a foundation for defining
classification accuracy and characterizing errors. Many
measures of classification accuracy can be derived from a
confusion matrix: kappa coefficient, overall, user’s and
producer’s accuracy. A confusions matrix are presented in the
following tables: for biased Support vector machine classification
in Table 3, for Random forest classification in Table 4, for
Convolutional Neural Networks in Table 5, and finally for
ensemble BSVM-RF-CNN in Table 6.
All four approaches achieved high overall accuracies. In other
circumstances, however, the suggested ensemble CNN-RF-BSVM
approach outperformed the others. As shown in Table 3,
reducing the number of satellite bands by deleting the less
relevant ones does not result in a significant drop in
Frontiers in Environmental Science | www.frontiersin.org May 2022 | Volume 10 | Article 8961588
Drobnjak et al. New Ensemble Vegetation Classification Method

classification accuracy. In the case of B-SVM, there is a significant
increase in classification accuracy. This could be related to the
requirement to simplify the vector space in order to build hyper-
planes.
The values of the Kappa coefficients for vegetation
classification from satellite pictures range from 0.864 for
Biased Support vector machine classification to 0.923 for
ensemble CNN-RF-BSVM classification (Tables 3–6).
In terms of the classification method utilized, it’s clear that
combining machine learning with deep learning techniques for
digital satellite and aerial image classification provides the
potential for vegetation mapping and analyzing environmental
changes. The use of a suitable machine learning or deep learning
technique aids in the selection of an appropriate classification
threshold as well as analysis bands. This reduces the need for trial
and error procedures, which are frequently utilized when
classifying data with a high degree of dimensionality.
5 DISCUSSION
According to the achieved results, the biased Support vector
machine has the lowest accuracy in relation to other
techniques used. Before the classification stage, biased SVM
and Random Forest algorithms usually include a feature
generation and selection step. We discovered that the
proposed criterion worked well for optimizing biased SVMs
and outperformed an alternate optimization criterion in
studying biased SVM for vegetation mapping. We also
discovered that cross-validation performed at the crown level
worked well (i.e., by splitting crowns rather than pixels into the
cross-validation groups).
One of the B-SVM model’s biggest advantages is its non-linear
categorization. A parametric model might thus have different
intercepts and coefficient values for each class of discrete
covariates. Furthermore, the B-SVM model is resistant to
overfitting and is not overly impacted by noisy data. The
B-SVM model benefits from complicated, non-linear
interactions and is noise-resistant. The B-SVM method’s major
flaw, on the other hand, is that it requires identifying the optimal
model after testing multiple kernel combinations and model
parameters. Meanwhile, because the results are part of a
complicated black box model, they are extremely difficult to
understand (Chan and King;Hartono et al., 2018).
Furthermore, for balanced data, the difficulty with biased SVM
based on structural risk reduction in classification is that the
classification weight will be biased towards the majority class,
causing the classification hyperplane to be close to the minority
class, making minority samples easy to misclassify (Chan and
King;Hartono et al., 2018). Reducing the number of features also
FIGURE 4 | Results of vegetation classification.
Frontiers in Environmental Science | www.frontiersin.org May 2022 | Volume 10 | Article 8961589
Drobnjak et al. New Ensemble Vegetation Classification Method

reduces overfitting concerns in remote sensing image
classification, where high-dimensional data is available but
ground truth data is scarce.
Random Forests are gradually becoming one of the most
popular machine learning algorithms due to their power,
diversity, and ease of use. The capacity to run on big datasets
with a large number of predictors and its ability to handle
thousands of input variables without variable deletion may
explain why the RF performed better than the B-SVM and
deep learning CNN models in this study (Cutler et al., 2007;
Peters et al., 2007;Biau, 2012;Amini et al., 2018). The Random
Forest model employs regression trees to estimate the dependent
FIGURE 5 | Proposed ensemble classification method with collected test samples.
TABLE 3 | Confusion (error) matrix for biased support vector machine (B-SVM) classification.
Class Method
B-SVM
1 2 3 4 5 6 7 8 Sum User’s
accuracy
(%)
1 1,975 20 25 12 23 27 17 25 2,124 92.98
2 33 1,014 28 19 27 28 15 14 1,178 86.08
3 35 17 1,674 22 35 23 14 31 1,816 92.18
4 25 23 14 887 22 32 25 25 1,028 86.28
5 33 27 37 45 1,547 18 27 25 1,726 89.63
6 22 12 33 17 33 800 31 36 962 83.16
7 14 18 37 23 17 35 734 17 881 83.31
8 17 14 26 29 41 24 12 1814 1,960 92.55
Sum 2,154 1,145 1,874 1,054 1,745 987 875 1987 11,821
Producer’s accuracy (%) 91.69 88.56 89.33 84.16 88.65 81.05 83.89 91.29
Overall accuracy (%) = 88.36
Kappa coefficient = 0.864
Frontiers in Environmental Science | www.frontiersin.org May 2022 | Volume 10 | Article 89615810
Drobnjak et al. New Ensemble Vegetation Classification Method

TABLE 4 | Confusion (error) matrix for random forest (RF) classification.
Class Method
RF
1 2 3 4 5 6 7 8 Sum User’s
accuracy
(%)
1 2,024 19 25 14 18 14 17 13 2,144 94.40
2 27 1,031 23 19 17 18 15 14 1,164 88.57
3 13 17 1,724 22 25 23 14 22 1847 93.34
4 28 17 14 917 22 22 8 28 1,028 89.20
5 12 17 15 25 1,612 18 17 25 1,729 93.23
6 11 12 33 17 13 854 21 24 974 87.68
7 14 18 14 21 17 24 771 23 888 86.82
8 25 14 26 19 21 14 12 1838 1,944 94.55
Sum 2,154 1,145 1,874 1,054 1,745 987 875 1987 11,821
Producer’s accuracy (%) 93.96 90.04 92.00 87.00 92.38 86.52 88.11 92.50
Overall accuracy (%) = 91.12
Kappa coefficient = 0.918
TABLE 5 | Confusion (error) matrix for convolution neural network (CNN) classification.
Class Method
CNN
1 2 3 4 5 6 7 8 Sum User’s
accuracy
(%)
1 1,994 13 15 17 22 11 19 24 2,115 94.28
2 35 985 32 17 14 22 12 15 1,132 87.01
3 22 34 1,725 24 19 32 24 15 1,873 92.10
4 28 25 14 909 22 22 28 25 1,045 86.99
5 15 17 15 24 1,618 18 27 14 1,733 93.36
6 21 22 33 17 12 839 17 18 958 87.58
7 14 18 14 21 17 24 731 17 842 86.82
8 25 31 26 25 21 19 17 1,859 1,998 93.04
Sum 2,154 1,145 1,874 1,054 1,745 987 875 1,987 11,821
Producer’s accuracy (%) 92.57 86.03 92.05 86.24 92.72 85.01 83.54 93.56
Overall accuracy (%) = 90.18
Kappa coefficient = 0.904
TABLE 6 | Confusion (error) matrix for ensemble BSVM-RF-CNN classification.
Class Method
Ensemble BSVM-RF-CNN
1 2 3 4 5 6 7 8 Sum User’s
accuracy
(%)
1 2,036 14 12 14 17 14 9 18 2,134 95.41
2 27 1,044 24 14 18 14 15 11 1,167 89.46
3 27 22 1,750 25 17 22 14 17 1,867 93.73
4 24 8 9 929 14 24 17 12 1,013 91.71
5 7 17 24 14 1,638 12 17 17 1,739 94.19
6 17 9 22 17 14 870 11 10 953 91.29
7 9 15 12 17 10 13 778 8 853 91.21
8 7 16 21 24 17 18 14 1,894 2,004 94.51
Sum 2,154 1,145 1,874 1,054 1,745 987 875 1,987 11,821
Producer’s accuracy (%) 94.52 91.18 93.38 88.14 93.87 88.15 88.91 95.32
Overall accuracy (%) = 92.54
Kappa coefficient = 0.923
Frontiers in Environmental Science | www.frontiersin.org May 2022 | Volume 10 | Article 89615811
Drobnjak et al. New Ensemble Vegetation Classification Method

variable’s average as the final prediction, resulting in an internally
unbiased calculation of the classification error. In comparison to
other machine learning algorithms, the RF algorithm has
significant advantages. Firstly, the RF technique can cope with
noisy or missing data as well as categorical or continuous features;
second, it does not require assumptions about the distribution of
explanatory variables; and third, it can manage interactions and
non-linearities between efficient components (Linardatos et al.,
2020). These are significant advantages that reduce the
production of outliers, especially when working with terrain
variables that have a high frequency of missing data (Amini
et al., 2018).
The Random Forests approach works by creating multiple
classification trees throughout the training period, taking
advantage of the considerable variation between individual
trees. Furthermore, by randomly modifying the predictive
variable sets and resampling the data with replacement over
the many tree stages of induction, the Random Forests
approach increases variation amongst the classification trees.
Because the average results of all trees are the result of the
model generation process, cross validation is not required for
this method (Oliveira et al., 2012;Amini et al., 2018;Gigovićet al.,
2019a). The major flaw of the RF model, on the other hand, is
that, unlike a decision tree, it is difficult to interpret. Furthermore,
the proper use of the RF model may necessitate some effort to
fine-tune the model for the data.
Convolutional neural networks can improve the likelihood of
successful classifications if big enough data sets (hundreds to
thousands of measurements, depending on the complexity of the
topic under study) are available to describe the problem. The
results show that CNN achieved high precision in the vast
majority of the cases in which it was utilized, outperforming
other common image-processing approaches (Kussul et al.,
2017). Their key is their capacity to efficiently mimic
exceedingly complicated problems and the fact that no prior
experiments are required. It’s important to remember that visual
classification and field research are only useful for obtaining
reference data if the target species or type of vegetation can be
easily identified in the imagery. This will be determined not only
by the image quality (e.g., spatial resolution), but also by the
uniqueness of the vegetation of interest’s morphological
characteristics. In any event, CNN-based vegetation species
identification is only useful if these morphological features are
present in the plant canopy.
Because different machine and deep learning algorithms may
not be capable of producing the best results on their own,
integrating them will maximize the model’s potential and
increase accuracy. It has been demonstrated that using an
ensemble learning methodology to predict and classify a
combination of satellite and aerial images produces better
results than using a single classifier.
6 CONCLUSION
The performance of ensemble approaches for vegetation
classification, which consists of three ML and DL
algorithms, was investigated in this article. Two of these
methods rely on machine learning, while the third is a deep
learning approach. We use Bayesian averaging and efficient
feature selection to create an ensemble model that addresses
these difficulties and mitigates their effects on defect
classification performance. The ensemble approach that
utilized the RGB and NIR wavelengths worked reasonably
well in tests. The results showed that the suggested
ensemble model outperformed the DL and ML models in
terms of overall accuracy by up to 7%, which was validated
by the Wilcoxon signed-rank test. Overall accuracy (OA)
analysis revealed that the suggested ensemble technique
CNN-RF-BSVM greatly enhanced classification
accuracy (by 4%).
Even though the proposed ensemble method can detect
vegetation with a reasonable level of accuracy, one future
research direction would be to use augmentation techniques
with deep learning methods to diversify the training data so
that more robust responses can be obtained when the test data
characteristics differ significantly from the training data.
According to the results of the studies, the use of a
combination of low spatial resolution satellite images and
high spatial resolution aerial photogrammetry imagery for
vegetation categorization mapping is practical, even though
there is still room for improvement. Advanced radiometric
image calibration techniques will be developed in the future to
increase the quality of the images. Experimenting with better
spectral resolution multispectral satellite images in
combination with aerial photogrammetry images, which are
becoming more cost-effective and possible, is also advised.
DATA AVAILABILITY STATEMENT
The raw data supporting the conclusion of this article will be
made available by the authors, without undue reservation.
AUTHOR CONTRIBUTIONS
SD, MS DD, and SB prepared the data layers, figures, and tables;
SD and MS performed the experiments and analyses. JJ and AD
supervised the research, finished the first draft of the manuscript,
edited and reviewed the manuscript, and contributed to the
model construction and verification.
FUNDING
This work supported research project 1.1.107/2018 “Possibilities
of automatic extraction of vegetation data by a combination of
satellite and aerial photogrammetric images”by the Ministry of
Defense of the Republic of Serbia and research project 1.21/2021
“Model for using MGI digital topographic maps in field
conditions with portable devices”by the Ministry of Defense
of the Republic of Serbia.
Frontiers in Environmental Science | www.frontiersin.org May 2022 | Volume 10 | Article 89615812
Drobnjak et al. New Ensemble Vegetation Classification Method

REFERENCES
Abdi, A. M. (2020). Land Cover and Land Use Classification Performance of
Machine Learning Algorithms in a Boreal Landscape Using Sentinel-2 Data.
GIScience Remote Sens. 57, 1–20. doi:10.1080/15481603.2019.1650447
Adam, E., Mutanga, O., Abdel-Rahman, E. M., and Ismail, R. (2014). Estimating
Standing Biomass in Papyrus (Cyperus Papyrus L.) Swamp: Exploratory of In
Situ Hyperspectral Indices and Random Forest Regression. Int. J. Remote Sens.
35 (2), 693–714. doi:10.1080/01431161.2013.870676
Amarsaikhan, D., and Douglas, T. (2004). Data Fusion and Multisource Image
Classification. Int. J. Remote Sens. 25, 3529–3539. doi:10.1080/
0143116031000115111
Amini, S., Homayouni, S., Safari, A., and Darvishsefat, A. A. (2018). Object-based
Classification of Hyperspectral Data Using Random Forest Algorithm. Geo-
Spatial Inf. Sci. 21, 127–138. doi:10.1080/10095020.2017.1399674
Ayhan, B., Kwan, C., Budavari, B., Kwan, L., Lu, Y., Perez, D., et al. (2020).
Vegetation Detection Using Deep Learning and Conventional Methods. Remote
Sens. 202012, 2502. doi:10.3390/RS12152502
Bakrač,S.,Drobnjak,S.,Stanković,S.,Vučićević,A.,andStamenković,N.(2018).
“Preparation of Photogrammetric Archive Documentation for ScientificandOther
Research,”in Sinteza 2018 - International ScientificConferenceonInformation
Technology and Data Related Research. Belgrade, Serbia: Singidunum University.
doi:10.15308/sinteza-2018-17-22
Biau, G. (2012). Analysis of a Random Forests Model. J. Mach. Learn. Res. 13,
1063–1095.
Breiman, L. (1996). Bagging Predictors. Mach. Learn 24, 123–140. doi:10.1007/
BF00058655
Breiman, L., and Cutler, A. (2007). Random Forests —Classification Description:
Random Forests. http://stat-www.berkeley.edu/users/breiman/RandomForests/cc_
home.htm
Breiman, L. (2001). Random Forests. Mach. Learn. 45, 5–32. doi:10.1023/A:
1010933404324
Burai, P., Deák, B., Valkó, O., and Tomor, T. (2015). Classification of Herbaceous
Vegetation Using Airborne Hyperspectral Imagery. Remote Sens. 7 (2),
2046–2066. doi:10.3390/rs70202046
Chan, C.-H., and King, I. (2009). “Using Biased Support Vector Machine to
Improve Retrieval Result in Image Retrieval with Self-Organizing Map,”in
International Conference on Neural Information Processing. (Berlin, HDB:
Springer), 714–719.
Chan, J. C.-W., and Paelinckx, D. (2008). Evaluation of Random Forest and
Adaboost Tree-Based Ensemble Classification and Spectral Band Selection for
Ecotope Mapping Using Airborne Hyperspectral Imagery. Remote Sens.
Environ. 112 (6), 2999–3011. doi:10.1016/J.RSE.2008.02.011
Cutler, D. R., Edwards, T. C., Jr, Beard, K. H., Cutler, A., Hess, K. T., Gibson, J., et al.
(2007). Random Forests for Classification in Ecology. Ecology 88, 2783–2792.
doi:10.1890/07-0539.1
Dietterich, T. G. (2000). “Ensemble Methods in Machine Learning,”in
International Workshop on Multiple Classifier Systems (Cagliari, Italy:
Springer), 1–15. doi:10.1007/3-540-45014-9_1
Dixit, A. (2019). Ensemble Classifier Based Multiclass Vegetation Classification
System. ICTACT Journal on Image and Video Processing 10, 2076–2082. doi:10.
21917/ijivp.2019.0295
Doktor, D., Lausch, A., Spengler, D., and Thurner, M. (2014). Extraction of Plant
Physiological Status from Hyperspectral Signatures Using Machine Learning
Methods. Remote Sens. 6 (12), 12247–12274. doi:10.3390/rs61212247
Drobnjak, S., Ćirović,G.,Sekulović,D.,andRegodić,M.(2013).Object-oriented
Classification of Multispectral Landsat 7 Satellite Images. Metal. Int. 18, 206
Available att: http://www.scopus.com/inward/record.url?eid=2-s2.0-
84874251767&partnerID=MN8TOARS.
Drobnjak, S., Marković, V., Kričković, Z., and Vučičević, A. (2018). Vegetation
Extraction from Satellite and Aerial Photogrammetric Images Using Machine
Learning Algorithms”in 8th International Scientific Conference on Defensive
Technologies, Serbia, October 11–12, 2018. Belgrade: MTI
Emily, J. A., and Sudha, N. (2022). Case Studies: Deep Learning in Remote Sensing.
Fundam. Methods Mach. Deep Learn.,425–437. doi:10.1002/9781119821908.CH18
Engler, R., Waser, L. T., Zimmermann, N. E., Schaub, M., Berdos, S., Ginzler, C.,
et al. (2013). Combining Ensemble Modeling and Remote Sensing for Mapping
Individual Tree Species at High Spatial Resolution. For. Ecol. Manag. 310,
64–73. doi:10.1016/J.FORECO.2013.07.059
Fei, S., Li, L., Han, Z., Chen, Z., and Xiao, Y. (2022). A Novel Ensemble Method for
Predicting Wheat Yield Using Feature Selection-Based Deep Learning and
Hyperspectral Vegetation Indices. Res. Sq.. doi:10.21203/rs.3.rs-1392054/v1
Fernández-Manso, A., Fernández-Manso, O., and Quintano, C. (2016).
SENTINEL-2A Red-Edge Spectral Indices Suitability for Discriminating
Burn Severity. Int. J. Appl. Earth Observation Geoinformation 50, 170–175.
doi:10.1016/J.JAG.2016.03.005
Foody, G. M., and Mathur, A. (2004). A Relative Evaluation of Multiclass Image
Classification by Support Vector Machines. IEEE Trans. Geosci. Remote Sens.
42, 1335–1343. doi:10.1109/tgrs.2004.827257
Gašparović, M., and Dobrinić, D. (2020). Comparative Assessment of Machine
Learning Methods for Urban Vegetation Mapping Using Multitemporal
Sentinel-1 Imagery. Remote Sens. 202012, 1952. doi:10.3390/RS12121952
Ghimire, B., Rogan, J., and Miller, J. (2010). Contextual Land-Cover Classification:
Incorporating Spatial Dependence in Land-Cover Classification Models Using
Random Forests and the Getis Statistic. Remote Sens. Lett. 1(1),45–54. doi:10.
1080/01431160903252327
Ghosh, A., Fassnacht, F. E., Joshi, P. K., and Koch, B. (2014). A Framework for
Mapping Tree Species Combining Hyperspectral and LiDAR Data: Role of
Selected Classifiers and Sensor across Three Spatial Scales. Int. J. Appl. Earth
Observation Geoinformation 26, 49–63. doi:10.1016/j.jag.2013.05.017
Ghosh, H., and Prajneshu, M. A. (2011). Bootstrap Study of Parameter Estimates
for Nonlinear Richards Growth Model through Genetic Algorithm. J. Appl.
Statistics 38 (3), 491–500. doi:10.1080/02664760903521401
Gigović, L., Pourghasemi, H. R., Drobnjak, S., Bai, S., Gigović, L., Pourghasemi, H.
R., et al. (2019b). Testing a New Ensemble Model Based on SVM and Random
Forest in Forest Fire Susceptibility Assessment and its Mapping in Serbia’s Tara
National Park. Forests 10, 408. doi:10.3390/F10050408
Gigović, L., Pourghasemi, H. R., Drobnjak, S., and Bai, S. (2019a). Testing a New
Ensemble Model Based on SVM and Random Forest in Forest Fire
Susceptibility Assessment and its Mapping in Serbia’s Tara National Park.
Forests 10, 408. doi:10.3390/F10050408
Hartono, H., Sitompul, O. S., Tulus, T., and Nababan, E. B. (2018). Biased Support
Vector Machine and Weighted-SMOTE in Handling Class Imbalance Problem.
Int. J. Adv. Intell. Inf. 4, 21–27. doi:10.26555/IJAIN.V4I1.146
Kattenborn, T., Leitloff, J., Schiefer, F., and Hinz, S. (2021). Review on Convolutional
Neural Networks (CNN) in Vegetation Remote Sensing. ISPRS J. Photogrammetry
Remote Sens. 173, 24–49. doi:10.1016/J.ISPRSJPRS.2020.12.010
Kavzoglu, T., Colkesen, I., and Yomralioglu, T. (2015). Object-based Classification
with Rotation Forest Ensemble Learning Algorithm Using Very-High-
Resolution WorldView-2 Image. Remote Sensing Letters 6, 834–843.
Kussul, N., Lavreniuk, M., Skakun, S., and Shelestov, A. (2017). Deep Learning
Classification of Land Cover and Crop Types Using Remote Sensing Data. IEEE
Geosci. Remote Sens. Lett. 14, 778–782. doi:10.1109/LGRS.2017.2681128
Landis, J. R., and Koch, G. G. (1977). An Application of Hierarchical Kappa-type
Statistics in the Assessment of Majority Agreement Among Multiple Observers.
Biometrics 33, 363. doi:10.2307/2529786
Li, Z., Hu, L., Tang, Z., and Zhao, C. (2021). Predicting HIV-1 Protease Cleavage
Sites with Positive-Unlabeled Learning. Front. Genet. 12, 456. doi:10.3389/
FGENE.2021.658078/BIBTEX
Liaw, A., and Wiener, M. (2002). Classification and Regression by randomForest.
R. News 2 (3), 18
Linardatos, P., Papastefanopoulos, V., and Kotsiantis, S. (2020). Explainable Ai: A
Review of Machine Learning Interpretability Methods. Entropy 23 (1), 18.
doi:10.3390/e23010018
Mallinis, G., Mitsopoulos, I., and Chrysafi, I. (2018). Evaluating and Comparing
Sentinel 2A and Landsat-8 Operational Land Imager (OLI) Spectral Indices for
Estimating Fire Severity in a Mediterranean Pine Ecosystem of Greece. GIsci.
Remote. Sens. 55, 1–18. doi:10.1080/15481603.2017.1354803
Meng, X., Shang, N., Zhang, X., Li, C., Zhao, K., Qiu, X., et al. (2017). Photogrammetric
UAV Mapping of Terrain under Dense Coastal Vegetation: An Object-Oriented
Classification Ensemble Algorithm for Classification and Terrain Correction. Remote
Sens. 9, 1187. doi:10.3390/RS9111187
Montgomery, J. M., Hollenbach, F. M., and Ward, M. D. (2012). Improving
Predictions Using Ensemble Bayesian Model Averaging. Polit. Anal. 20,
271–291. doi:10.1093/PAN/MPS002
Frontiers in Environmental Science | www.frontiersin.org May 2022 | Volume 10 | Article 89615813
Drobnjak et al. New Ensemble Vegetation Classification Method

Neto, U. M. B., and Dougherty, E. R. (2015). Error Estimation for Pattern
Recognition. Hoboken, NY, United States: John Wiley & Sons.
Nijhawan, R., Sharma, H., Sahni, H., and Batra, A. (2017). “ADeepLearning
Hybrid CNN Framework Approach for Vegetation Cover Mapping Using
Deep Features,”in Proceedings - 13th International Conference on Signal-
Image Technology and Internet-Based Systems, SITIS 2017 2018-January,
Jaipur,India,December4–7, 2017, 192. doi:10.1109/SITIS.2017.41
Oliveira, S., Oehler, F., San-Miguel-Ayanz, J., Camia, A., and Pereira, J. M. C.
(2012). Modeling Spatial Patterns of Fire Occurrence in Mediterranean Europe
Using Multiple Regression and Random Forest. For. Ecol. Manag. 275, 117–129.
doi:10.1016/j.foreco.2012.03.003
Peters, J., Baets, B. D., Verhoest, N. E. C., Samson, R., Degroeve, S., Becker, P.
D., et al. (2007). Random Forests as a Tool for Ecohydrological
Distribution Modelling. Ecol. Model. 207, 304–318. doi:10.1016/j.
ecolmodel.2007.05.011
Raftery, A. E., Gneiting, T., Balabdaoui, F., and Polakowski, M. (2005). Using
Bayesian Model Averaging to Calibrate Forecast Ensembles. Mon. Weather Rev.
133, 1155–1174. doi:10.1175/MWR2906.1
Running,S.W.,Loveland,T.R.,Pierce,L.L.,Nemani,R.R.,andHunt,E.R.
(1995). A Remote Sensing Based Vegetation Classification Logic for Global
Land Cover Analysis. Remote Sens. Environ. 51, 39–48. doi:10.1016/0034-
4257(94)00063-S
Shaheen, F., and Verma, B. (2016). “An Ensemble of Deep Learning Architectures
for Automatic Feature Extraction,”in 2016 IEEE Symposium Series on
Computational Intelligence,1–5. (Athens, Greece: SSCI). doi:10.1109/SSCI.
2016.7850047
Sothe, C., de Almeida, C. M., Schimalski, M. B., Liesenberg, V., la Rosa, L. E. C.,
Castro, J. D. B., et al. (2020). A Comparison of Machine and Deep-Learning
Algorithms Applied to Multisource Data for a Subtropical Forest Area
Classification. Int. J. Remote. Sens. 41, 1943–1969. doi:10.1080/01431161.
2019.1681600
Vrugt, J. A., and Robinson, B. A. (2007). Treatment of Uncertainty Using
Ensemble Methods: Comparison of Sequential Data Assimilation and
Bayesian Model Averaging. Water Resour. Res. 43, 1411. doi:10.1029/
2005WR004838
Wang,H.,Lv,G.,Cai,Y.,Zhang,X.,Jiang,L.,andYang,X.(2021).Determiningthe
Effects of Biotic and Abiotic Factors on the Ecosystem Multifunctionality in a
Desert-Oasis Ecotone. Ecol. Indic. 128, 107830. doi:10.1016/J.ECOLIND.2021.
107830
Woolson, R. F. (2008). Wilcoxon Signed-Rank Test,1–3. Wiley Encyclopedia of
Clinical Trialsdoi:10.1002/9780471462422.EOCT979 (Accessed September 18,
2008)
Xie, Y., Sha, Z., and Yu, M. (2008). Remote Sensing Imagery in Vegetation
Mapping: a Review. J. Plant Ecol. 1, 9–23. doi:10.1093/JPE/RTM005
Yu, Q., Gong, P., Clinton, N., Biging, G., Kelly, M., and Schirokauer, D. (2006).
Object-based Detailed Vegetation Classification with Airborne High Spatial
Resolution Remote Sensing Imagery. Photogramm. Eng. remote Sens. 72,
799–811. doi:10.14358/PERS.72.7.799
Zhang, X., Han, L., Han, L., and Zhu, L. (2020). How Well Do Deep Learning-Based
Methods for Land Cover Classification and Object Detection Perform on High
Resolution Remote Sensing Imagery? Remote Sens. 12, 417. doi:10.3390/
RS12030417
Conflict of Interest: The authors declare that the research was conducted in the
absence of any commercial or financial relationships that could be construed as a
potential conflict of interest.
Publisher’s Note: All claims expressed in this article are solely those of the authors
and do not necessarily represent those of their affiliated organizations, or those of
the publisher, the editors and the reviewers. Any product that may be evaluated in
this article, or claim that may be made by its manufacturer, is not guaranteed or
endorsed by the publisher.
Copyright © 2022 Drobnjak, Stojanović, Djordjević, Bakrač, Jovanovićand
Djordjević. This is an open-access article distributed under the terms of the
Creative Commons Attribution License (CC BY). The use, distribution or
reproduction in other forums is permitted, provided the original author(s) and
the copyright owner(s) are credited and that the original publication in this journal is
cited, in accordance with accepted academic practice. No use, distribution or
reproduction is permitted which does not comply with these terms.
Frontiers in Environmental Science | www.frontiersin.org May 2022 | Volume 10 | Article 89615814
Drobnjak et al. New Ensemble Vegetation Classification Method