AIMS Geosciences, 10(1): 62–79.

DOI: 10.3934/geosci.2024005
Received: 15 September 2023
Revised: 24 October 2023
Accepted: 12 November 2023
Published: 25 January 2024
http://www.aimspress.com/journal/geosciences

Research article

Machine Learning-Based Classification of Small-Sized Wetlands Using

Sentinel-2 Images

Eric Ariel L. Salas*, Sakthi Subburayalu Kumaran, Robert Bennett, Leeoria P. Willis and Kayla
Mitchell

Agricultural Research Development Program (ARDP), Central State University, Wilberforce, OH

45384, USA

* Correspondence: Email: esalas@centralstate.edu; Tel.: +01-937-376-6637.

Abstract: Wetlands are invaluable ecosystems, offering essential services such as carbon sequestration,
water purification, flood control and habitat for countless aquatic species. However, these critical
environments are under increasing threat from factors like industrialization and agricultural expansion.
In this research, we focused on small-sized wetlands, typically less than 10 acres in size, due to their
unique ecological roles and groundwater recharge contributions. To effectively protect and manage
these wetlands, precise mapping and monitoring are essential. To achieve this, we exploited the
capabilities of Sentinel-2 imagery and employ a range of machine learning algorithms, including
Random Forest (RF), Classification and Regression Tree (CART), Gradient Tree Boost (GTB), Naive
Bayes (NB), k-nearest neighbors (KNN) and Support Vector Machine (SVM). Our evaluation used
variables, such as spectral bands, indices and image texture. We also utilized Google Earth Engine
(GEE) for streamlined data processing and visualization. We found that Random Forest (RF) and
Gradient Tree Boost (GTB) outperformed other classifiers according to the performance evaluation.
The Normalized Difference Water Index (NDWI) came out to be one of the important predictors in
mapping wetlands. By exploring the synergistic potential of these algorithms, we aim to address
existing gaps and develop an optimized approach for accurate small-sized wetland mapping. Our
findings will be useful in understanding the value of small wetlands and their conservation in the face
of environmental challenges. They will also lay the framework for future wetland research and
practical uses.

Keywords: GEE; mapping wetlands; remote sensing wetlands; random forest; Google Earth Engine
 63

1. Introduction

Among important ecosystems, wetlands are unique. These ecosystems are useful for hydrological
cycle, carbon sequestration, water purification, flood control and provide habitats for thousands of
aquatic flora and fauna [1]. Wetlands are, however, threatened by industrialization and agricultural
intensification [2]. In the USA, the main cause of wetland loss is the conversion to agricultural land
use [3]. Globally, about 35% of wetlands have been lost between 1970 and 2015 [4].
Small-sized wetlands, the focus of our study, comprise a diverse array of ecologically and
hydrologically significant ecosystems. Typically occupying an area of less than 10 acres (0.04 km2),
these wetlands are situated within a variety of terrestrial and aquatic landscapes [5]. Hydrologically,
these wetlands are noticeable by their remarkable ability to retain water over extended periods,
particularly during wet seasons, thereby contributing substantially to local groundwater recharge [6].
From an ecological perspective, small-sized wetlands hold a fundamental role in supporting unique
habitats for a diverse range of wetland-dependent species [7]. These types of wetlands usually have
specific vegetation types, such as emergent aquatic plants and hydric soils, further setting them apart
from the surrounding ecosystems [8].
It is crucial to map and monitor wetlands for conservation and management strategies [9]. Remote
sensing technology, including satellite imagery, has emerged as a valuable tool for mapping and
monitoring wetlands at a regional and global scale [10,11]. The European Space Agency’s (ESA)
Sentinel satellites, specifically Sentinel-2, have proven particularly useful in providing high-resolution
optical and radar imagery to map wetlands. The multispectral Sentinel-2 imagery can identify
vegetation and water bodies, while its high-resolution capabilities enable the detection and
characterization of small-sized wetlands [12].
To enhance the precision and efficiency of wetland classification using Sentinel-2 imagery, there
has been a growing interest in the application of machine learning (ML) algorithms. Machine learning
is a subset of artificial intelligence and involves training algorithms to recognize patterns in data and
subsequently classify new data based on these learned patterns. In the context of satellite imagery, ML
algorithms offer the potential for automating the mapping process and accurately classifying various
land cover types, including wetlands [13]. Among the ML algorithms, Random Forest (RF) and
Classification and Regression Tree (CART) have emerged as prominent choices for mapping wetlands
with Sentinel-2 images [14,15]. RF, for instance, leverages the combination of multiple decision trees
to produce precise predictions through the aggregation of their outputs. Its capacity to handle complex
and high-dimensional datasets, while also enabling the assessment of variable importance, positions
RF as a robust tool for wetland mapping [16]. Similarly, CART constructs binary decision trees based
on feature attributes, facilitating the accurate classification of wetland areas [17].
In addition to RF and CART, various other ML algorithms have found application in the mapping
of small-sized wetlands using Sentinel-2 imagery. Among these, Gradient Tree Boost (GTB) stands out
as an ensemble learning method, similar to RF, which combines multiple decision trees to enhance
classification accuracy [18]. Naive Bayes (NB) is another notable algorithm, offering a probabilistic
approach that assumes independence between features, and it has demonstrated successful outcomes in
wetland mapping [19]. In a different manner, k-nearest neighbors (KNN), a non-parametric algorithm,
distinguishes itself by classifying data points based on their proximity to known data points [20]. Further,
the Support Vector Machine (SVM) also emerges as an effective ML tool, constructing hyperplanes to
delineate distinct classes within the feature space [21].

AIMS Geosciences Volume 10, Issue 1, 62–79.

 64

Several studies have been conducted regarding the effectiveness of these ML algorithms in
wetland mapping tasks. For example, Mahdianpari et al. [22] and Waleed et al. [23] employed RF and
CART algorithms with Sentinel-2 images to achieve high overall accuracies (>95%) in wetland
classification. Li et al. [24] utilized GTB for wetland classification and reported improved accuracy
compared to individual decision trees. Wu et al. [25] successfully applied NB and KNN algorithms in
wetland mapping and highlighted its ability to handle complex classification tasks. Finally, Gemechu
et al. [26] demonstrated the effectiveness of SVM algorithm in temporal wetland classification in
Guangling County, China with accuracies ranging from 86% to 98.1%.
To process and analyze the Sentinel-2 images using these ML classifiers, we take advantage of
the capabilities of Google Earth Engine (GEE), a cloud-based platform that combines a vast archive
of satellite imagery with geospatial analysis functions [27]. GEE provides a powerful computing
infrastructure and pre-built functions for image analysis and enables efficient data processing and
visualization. It offers access to Sentinel satellite imagery, including Sentinel-2 data and facilitates the
integration of ML algorithms for wetland mapping [28,29]. GEE could overcome computational
limitations and process various satellite imagery at the same time, which allows for the integration of
multiple ML algorithms and the exploration of various input features for wetland classification [27].
Despite the advancements in ML algorithms and the utilization of Sentinel-2 imagery for wetland
mapping, there is a research gap in the development of an integrated approach that combines multiple
machine learning algorithms to improve the accuracy and efficiency of small-sized wetland
classification. While studies have individually demonstrated the effectiveness of algorithms, especially
RF, there is a need to explore the potential of other algorithms in a synergistic manner to enhance
wetland mapping processes. Our specific objectives of this study include:
1. Assessing the individual performance of RF, CART, GTB, NB, kNN and SVM classifiers in
wetland classification using Sentinel-2 imagery.
2. Evaluating the impact of different input variables, such as spectral bands from Sentinel-2
images, spectral indices and image texture on the performance of each classifier.
3. Mapping of the wetland areas using the best performing classifier.
Our results would provide valuable insights into the development of an optimized and accurate
wetland mapping approach. By exploring the synergistic potential of these algorithms and variables,
our research aims to address existing gaps and develop an optimized approach for accurate small-sized
wetland mapping. Such an approach would be helpful for other researchers, conservationists, land
managers, policy-makers and environmental practitioners who are actively involved in the
preservation and management of small-sized wetlands. Finally, our results would also address the gaps
that exist in the National Wetlands Inventory (NWI) database.

2. Materials and methods

2.1. Study Area

The study area includes a number of wetland systems, including floodplain woodlands, marshes,
wet prairies, woodland pools and scrub-shrub wetlands. These small-sized wetlands are managed by
Beaver Creek Wetlands Greenway (BWG) Community Land Trust. Located in Beavercreek, Ohio,
USA, the wetlands are protected by the Ohio Environmental Protection Agency (OEPA) Clean Water
Act. Due to the variety of flora and fauna species, school groups, birdwatchers and wildflower

AIMS Geosciences Volume 10, Issue 1, 62–79.

 65

enthusiasts frequent the wetlands. The Beaver Creek stream runs through the wetlands from north to
south. The BWG boundary and the perimeter of our study area are depicted in Figure 1. According to
the NLCD land use, most of the urban areas are in the east, while cultivated crops predominate in the
west. Although wetland is one of the NLCD classes, due to their small sizes, they are not properly
mapped out.

Figure 1. The BWG study area is located in southwestern Ohio.

2.2. General Methodology

The methods for classifying wetlands using GEE and utilizing six classifiers involved several
steps.
1. Sentinel-2 images were retrieved and preprocessed using GEE. Relevant spectral bands and
features were extracted from the images, capturing important information about wetland
vegetation and water characteristics.
2. The six machine learning classifiers, including RF, CART, GTB, NB, KNN and SVM, were
implemented to classify the wetland areas.
3. These classifiers were trained on the extracted spectral indices, image texture and spectral
bands, leveraging the ability of the classifiers to recognize patterns and classify new data
based on learned patterns.

AIMS Geosciences Volume 10, Issue 1, 62–79.

 66

4. The performance of each classifier was assessed individually, evaluating their accuracy in
wetland classification using Sentinel-2 imagery.
5. The impact of different input variables on the classification performance of each approach
was also evaluated by grouping them into four sets: Set A (spectral bands, spectral indices,
image texture); set B (spectral bands, spectral indices); set C (spectral bands, image texture);
set D (spectral bands).
6. The accuracy of the multiple classifiers using the various set of variables was validated by
splitting the sampling data into training and testing, providing insights into an optimized and
accurate wetland mapping approach.

2.3. Google Earth Engine (GEE) Platform and Sentinel-2 Images

GEE (earthengine.google.com) is a cloud-based platform that provides researchers with a vast

collection of satellite imagery and geospatial datasets [30]. It offers a comprehensive and diverse range
of data sources, including the 10-meter spatial resolution Sentinel-2 which is essential for wetland
mapping. The extensive GEE data archive, coupled with its efficient data storage and processing
capabilities, allowed this research seamless access to the necessary image data for analysis. By
utilizing the power of GEE, we retrieved and preprocessed the Sentinel-2 imagery from June 1 to
October 31, 2022 (summer to fall seasons), extracted relevant spectral bands and features from the
images and performed various data manipulations required for wetland classification. The use of GEE
ensures a streamlined and efficient workflow, enabling us to leverage its vast data resources and
advanced analysis capabilities for accurate wetland mapping.
We utilized summer to fall images for wetland classification because this period captures crucial
temporal variations in wetland vegetation dynamics and water presence. During this period, wetlands
undergo distinct changes in vegetation phenology, water levels and land cover, which can provide
valuable information for accurate classification. By analyzing imagery from June 1 to October 31,
2022, we captured the seasonal transitions and phenological patterns necessary for distinguishing
wetland classes and improved the overall accuracy of wetland mapping.

2.4. Selected Classifiers RF, CART, GTB, NB, KNN and SVM

We selected RF, CART, GTB, NB, KNN and SVM as our classifiers based on their demonstrated
effectiveness in wetland classification using remote sensing data. Each classifier offers unique
advantages and has been widely used in various land cover mapping studies.
RF is an ensemble learning method that combines multiple decision trees to generate accurate
predictions by aggregating their outputs [31]. It utilizes the concept of bagging and random feature
selection to reduce overfitting and improve generalization. The RF algorithm combines the predictions
of individual decision trees using a voting mechanism or averaging method, as shown in Equation 1:

𝑦𝑅𝐹 ∑ 𝑦 𝑥 , 𝑤ℎ𝑒𝑟𝑒 𝑦 𝑥 𝜖 1,1 (1)

where 𝑦𝑅𝐹 is the predicted class label for a given sample 𝑥, 𝑦 , 𝑥 is the predicted class label of the
i-th decision tree and N is the total number of decision trees in the forest.

AIMS Geosciences Volume 10, Issue 1, 62–79.

 67

CART constructs binary decision trees based on feature attributes, enabling precise classification
of wetland areas [32]. The splitting criterion used in CART is often based on Gini impurity or
information gain, as shown in Equation 2:

𝐺𝑖𝑛𝑖 𝑝 1 ∑ 𝑝𝑘 (2)

where 𝑝𝑘 represents the proportion of samples in class k at a given node. CART recursively partitions
the feature space based on the selected splitting criterion, forming a binary tree structure.
GTB, like RF, is an ensemble learning method that combines multiple decision trees to improve
classification accuracy [32]. GTB differs from RF in the way it constructs subsequent trees. The
boosting process of GTB updates the weights of misclassified samples, allowing subsequent trees to
focus on the previously misclassified samples. The final prediction of GTB is a weighted sum of
predictions from all the trees, as shown in Equation 3:

𝑦𝐺𝑇𝐵 ∑ 𝑣𝑦 𝑥 , 𝑤ℎ𝑒𝑟𝑒 𝑦 𝑥 𝜖 1,1 (3)

where 𝑦𝐺𝑇𝐵 is the predicted class label for a given sample 𝑥, 𝑦 , 𝑥 is the predicted class label of the
i-th decision tree, N is the total number of decision trees in the ensemble and v is the learning rate.
NB is a probabilistic classifier that assumes independence between features and has been applied
successfully in wetland mapping [33]. It estimates the conditional probability of a sample belonging
to a particular class using Bayes' theorem. The predicted class label is determined by selecting the class
with the highest probability, as shown in Equation 4:

𝑦𝑁𝐵 𝑎𝑟𝑔𝑚𝑎𝑥 , ,..., 𝑃 𝑦 ∏ 𝑃 𝑥 |𝑦 (4)

where 𝑦𝑁𝐵 is the predicted class label for a given sample, 𝑃 𝑦 is the prior probability of class y,
𝑃 𝑦 is the conditional probability of feature 𝑥 given class y, n is the number of features and K is the
total number of classes.
KNN is a non-parametric algorithm that classifies data points based on their proximity to other
known data points [34]. It assigns the class label based on the majority vote of its nearest neighbors.
The class label of a sample is determined by Equation 5:
𝑦𝐾𝑁𝑁 𝑚𝑜𝑑𝑒 𝑦 |𝑥 is one of the k nearest neighbors of x}) (5)

where 𝑦𝐾𝑁𝑁 is the predicted class label for a given sample 𝑥 𝑦 is the class label of the i-th nearest
neighbor and k is the number of nearest neighbors.
SVM is a powerful algorithm that constructs a hyperplane to separate different classes in the
feature space [35]. It aims to maximize the margin between the support vectors and the decision
boundary. The class label of a sample is determined by the sign of the discriminant function, as shown
in Equation 6:
𝑦𝑆𝑉𝑀 𝑠𝑖𝑔𝑛 𝑓 𝑥 (6)

where 𝑦𝑆𝑉𝑀 is the predicted class label for a given sample 𝑥, 𝑓 𝑥 is the discriminant function and
the sign function determines the class label based on the sign of the discriminant function.

AIMS Geosciences Volume 10, Issue 1, 62–79.

 68

Using the strengths of these classifiers, we aim to improve the accuracy and efficiency of small
wetland classification.

2.5. Image Texture and Other Indices as Predictors

We added image texture measures such as entropy, along with vegetation indices like NDVI and
water indices like NDWI as they have crucial roles in wetland classification, as they can capture
important spectral and textural characteristics of wetland environments. Entropy, as a measure of
randomness or disorder within an image, provides valuable information about the spatial distribution
and heterogeneity of land cover classes within a wetland area. Texture measures have been widely
used in remote sensing studies for land cover classification and change detection [10,36].
NDVI is a widely applied vegetation index that quantifies the presence and vigor of vegetation
by leveraging the contrast between near-infrared (NIR) and red reflectance. Wetlands typically exhibit
distinct vegetation characteristics, and NDVI helps in differentiating wetland vegetation from other
land cover types [14,37]. High NDVI values indicate dense vegetation cover, which is often associated
with wetland areas due to their unique hydrological conditions and plant species composition. NDWI
is a spectral index commonly used to identify water bodies. It exploits the contrast between the NIR
and shortwave infrared (SWIR) bands to highlight the presence of water [21]. Wetlands are
characterized by the presence of water, and the inclusion of NDWI as a variable helps in accurately
delineating wetland boundaries and differentiating wetlands from non-wetland land cover classes [19].
By incorporating entropy with NDVI and NDWI, together with the Sentinel-2 bands (B2: blue;
B3: green; B4: red; B5, B6, B7, B8: VNIR; B11, B12: SWIR) into the classification process, we
effectively captured the heterogeneity of wetland environments, distinguish wetland vegetation from
other land cover types and accurately identify water bodies within wetland areas. This comprehensive
set of variables enhanced the discriminatory power of the classification algorithms and enhanced the
accuracy of wetland mapping and monitoring studies.

2.6. Sets of Variables to Assess the Impact of Each Predictor

We ran the classifiers using four different sets of variables (A, B, C and D) to understand the
impact of including different types of variables on the classification performance.
1. Set A (spectral bands, spectral indices, image texture): This set includes a comprehensive
range of variables, namely spectral bands, spectral indices (vegetation and water) and image texture.
Spectral bands capture information about the reflectance properties of different land cover classes,
while spectral indices, such as the NDVI and NDWI, provide insights into specific land cover
characteristics like vegetation and water content. Image texture measures, such as entropy, describe
the spatial arrangement and patterns within the imagery. By combining these variables, the
classifiers can leverage a diverse set of information for improved discrimination and classification
accuracy.
2. Set B (spectral bands, spectral indices): This set focuses on combining spectral bands and
spectral indices, excluding the image texture measures. It allows us to evaluate the contribution of
spectral indices in enhancing the performance of the classifiers. Spectral indices, being derived
from specific band combinations, provide valuable information about vegetation health, moisture
content and other land cover characteristics. By including these indices alongside spectral bands,

AIMS Geosciences Volume 10, Issue 1, 62–79.

 69

the classifiers can utilize additional spectral information for more accurate classification.
3. Set C (spectral bands, image texture): This set explores the impact of image texture measures
when combined with spectral bands. Image texture measures capture information related to spatial
patterns, such as the heterogeneity or smoothness of land cover classes. By incorporating image
texture alongside spectral bands, the classifiers can consider textural characteristics as
discriminative features, which may aid in distinguishing between different land cover types.
4. Set D (spectral bands): This set serves as a baseline comparison where only spectral bands are
used as input variables. By excluding spectral indices and image texture measures, the classifiers'
performance can be assessed based solely on the spectral information captured by the satellite
imagery. This set helps evaluate the extent to which additional variables (spectral indices and image
texture) contribute to improving the classification accuracy compared to using spectral bands alone.

2.7. Field Data and Accuracy Assessment

We collected 60 points samples from the field (using GPS) and interpreted additional 20 samples
from the high-resolution images using Google Earth. To assess the wetland mapping classification
accuracy of the classifiers, we employed a data splitting approach to create training and testing datasets.
We divided our dataset into 70% training set and 30% testing set. This division of dataset guarantees
an unbiased assessment of the classifiers' accuracy by evaluating their ability to generalize to new data.
During the training phase, the classifiers were trained using the labeled training dataset, where
each pixel was assigned a known wetland class label. This process involved learning the underlying
patterns and relationships between the spectral characteristics and wetland classes present in the
training data. The trained classifiers were then applied to the testing dataset, where the class labels
were withheld, and the classifiers' predictions were compared against the true class labels to evaluate
their accuracy.
We assessed the accuracy using established metrics such as overall accuracy (OA) and kappa
coefficient. These metrics provide quantitative measures of the classifier performance in correctly
identifying wetland classes and assessing the agreement between the predicted and reference class
labels [38]. By utilizing the training and testing data splitting approach, we evaluated the accuracy of
the RF, CART, GTB, NB, KNN and SVM classifiers and compared their mapping performance.

3. Results

3.1. Performance of the Six Classifiers

We ran the classifiers using four different sets. GTB demonstrated high overall training and
validation accuracies (from 0.94 to 0.97), indicating good classification performance (Table 1). These
numbers suggested a substantial level of agreement between the predicted and reference classifications.
The RF classifier exhibited equally high overall training and validation accuracies, ranging from 0.93
to 0.98, and the kappa coefficients ranging from 0.97 to 0.98.
The SVM showed relatively lower overall training and validation accuracies, ranging from 0.90
to 0.92, while kappa coefficients, ranging from 0.86 to 0.91. CART demonstrated high overall training
accuracy (0.91 to 0.95), with kappa coefficients ranging from 0.92 to 0.95. KNN showed lower overall
training and validation accuracies, both consistently at 0.67. The kappa coefficient was also low at

AIMS Geosciences Volume 10, Issue 1, 62–79.

 70

0.48. Finally, NB exhibited moderate overall training accuracy, ranging from 0.72 to 0.75. The
performance in terms of kappa coefficient ranged from 0.72 to 0.73.
Based on the provided results, the RF and GTB classifiers outperformed the other classifiers in
terms of OA and kappa coefficients. These two classifiers demonstrated higher accuracy and agreement
with the reference classifications, indicating their effectiveness in mapping the LCLU/wetland classes.
In the end, we used the RF and GTB classifiers as the final models for mapping the wetlands.

Table 1. Overall training accuracy/validation accuracy/kappa of the six ML classifiers in

mapping the LCLU classes. A= spectral bands, spectral indices, image texture; B = spectral
bands, spectral indices; C = spectral bands, image texture; and D = spectral bands.
Classifier A B C D
CART 0.95/0.91/0.95 0.91/0.90/0.92 0.90/0.90/0.92 0.91/0.90/0.92
RF 0.98/0.94/0.98 0.98/0.93/0.98 0.98/0.93/0.97 0.97/0.92/0.97
SVM 0.92/0.91/0.86 0.92/0.90/0.86 0.91/0.90/0.86 0.91/0.90/0.86
GTB 0.97/0.94 0.97/0.93 0.97/0.93 0.97/0.93
KNN 0.67/0.67/0.48 0.67/0.66/0.48 0.67/0.65/0.48 0.67/0.67/0.48
NB 0.75/0.73 0.74/0.73 0.75/0.72 0.72/0.73

When we evaluated the efficacy of the classifiers in mapping wetlands against all other classes
(that is, excluding the wetland class), the results indicated varying levels of relative OA (Table 2).
Among the classifiers tested, RF and GTB exhibited the highest relative OA values of 0.93, suggesting
their effectiveness in accurately mapping wetlands against the combined class. Both RF and GTB
outperformed other classifiers in correctly classifying the samples.

Table 2. Overall training accuracy/validation accuracy/kappa of the six ML classifiers in

mapping wetlands against all other classes. A= spectral bands, spectral indices, image
texture; B = spectral bands, spectral indices; C = spectral bands, image texture; and D =
spectral bands.
Classifier Relative Overall Accuracy (Wetlands)
CART 0.90
RF 0.93
SVM 0.90
GTB 0.93
KNN 0.66
NB 0.73

CART and SVM achieved a relative OA of 0.90, indicating similar performance in accurately
classifying the wetlands against the combined class. Although CART and SVM performed slightly
lower than RF and GTB, they demonstrated good accuracy. In contrast, KNN exhibited a lower relative
OA of 0.66, indicating a comparatively weaker performance in correctly classifying the wetlands. This
suggests that KNN may struggle to distinguish wetlands from other land cover types effectively. NB
achieved a moderate relative OA of 0.73, indicating its moderate performance. While NB performed
better than KNN, it fell behind RF, GTB, CART and SVM in accuracy.

AIMS Geosciences Volume 10, Issue 1, 62–79.

 71

Overall, the results highlighted the superior performance of RF and GTB classifiers in accurately
mapping the wetlands, suggesting their suitability for distinguishing and mapping wetlands. CART
and SVM classifiers also exhibited good accuracy, while KNN showed relatively weaker performance,
and NB achieved moderate accuracy.

3.2. Ranking the Importance of the Variables

We evaluated the importance of specific variables in the classification when using only two
classes (wetlands vs others) (Table 3). Across the different variable sets (A, B, C and D), certain
variables consistently emerged as significant contributors to the classification process. In variable set
A, which includes spectral bands, spectral indices and image texture, B4 (Red spectral band), B11
(SWIR spectral band) and NDVI (spectral index) were identified as the most important variables.
Similarly, in variable set B (spectral bands and spectral indices), B11, B8, B5 and B2 (all spectral
bands) played crucial roles in distinguishing between the two classes. Variable set C (spectral bands
and image texture) further emphasized the significance of B11, B8, B5, B2 and B12 (spectral bands)
in accurate classification. Finally, variable set D (spectral bands) confirmed the importance of B11, B8,
B3, B2, B4 and B5 (all spectral bands) for effective discrimination between the two classes. Among
all variables, entropy is absent from the list of important variables, suggesting that it may have had
limited discriminatory power in the classification. Overall, the consistent ranking of these variables
underscores their relevance in accurately classifying the land cover classes under consideration,
emphasizing the value of specific spectral bands and spectral indices, such as NDVI, in the
classification process.

Table 3. Important variables in the classification. A= spectral bands, spectral indices,

image texture; B = spectral bands, spectral indices; C = spectral bands, image texture; and
D = spectral bands.
Rank A B C D
1 B4 B11 B11 B11
2 B11 B8 B8 B8
3 NDVI B5 B5 B3
4 B2 B2 B2 B2
5 B3 B12 B12 B4
6 B12 NDWI B4 B5

3.3. Mapping the Wetlands Using the Best Performing Classifiers

We mapped the wetlands the two best-performing classifiers (GTB and RF), using Sentinel-2
imagery through the GEE environment (Figure 2).

AIMS Geosciences Volume 10, Issue 1, 62–79.

 72

Figure 2. Resulting classified map of the BWG wetlands using the (a) GTB and (b) RF
models. The wetland class is highlighted in yellow.

4. Discussion

4.1. Performance of the Six Classifiers

We conducted an assessment of the effectiveness of six popular ML classifiers in the framework

of small-sized wetland mapping. Our findings revealed a range of performance levels among the
classifiers. Notably, RF exhibited a higher degree of accuracy, followed by CART, SVM and GTB.
Conversely, KNN and NB exhibited comparatively lower levels of accuracy. These results agreed with
prior research findings that support the efficacy of decision tree-based classifiers for wetland mapping
[17,18,20].
In contrast to our results, Gemechu et al. [26] presented findings indicating that SVM classifier
outperformed RF and other classifiers in terms of accuracy, achieving levels as high as 98.1% in
wetland mapping. They particularly underscored the proficiency of SVM in handling high-dimensional
data and capturing complex relationships, which ultimately contributed to improved classification
accuracy. Furthermore, a study by Amani et al. [39] reported satisfactory results with the k-nearest
neighbors (KNN) classifier when applied to wetland mapping using remote sensing data. Their results
placed importance on the significance of optimizing KNN algorithm parameters to achieve precise
wetland classification. These divergent results underscore the importance of taking into account

AIMS Geosciences Volume 10, Issue 1, 62–79.

 73

specific contextual variables and parameters when making decisions regarding classifier selection for
wetland mapping applications. Considerations relating to the study area, dataset characteristics and
optimization techniques should be weighed cautiously to ensure dependable and accurate wetland
classification results.
Furthermore, Xing et al. [40] conducted a comprehensive investigation of ensemble classifiers,
among them RF, in the context of wetland mapping. Their investigation yielded high overall accuracy
when using RF; however, they observed a tendency for overestimating wetland areas attributed to
inherent algorithmic characteristics. In response, they recommended incorporating post-processing
techniques to mitigate this challenge and enhance the precision of wetland mapping.
Our study contributes to the existing pool of knowledge pertaining to wetland mapping by
conducting a thorough evaluation of various ML classifiers aimed at mapping small-sized wetlands.
Although our findings indicated a relatively high level of accuracy in the case of decision tree-based
classifiers, it remains imperative for future research studies to continue exploring the performance of
the classifiers within different contextual frameworks. Additionally, the prospect of integrating
complementary approaches to advance wetland mapping accuracy warrants further investigation.

4.2. Spatial Distribution of Wetlands

One of our primary objectives was the precise mapping of small-sized wetlands, a task for which
conventional approaches, illustrated by the NWI, often prove inadequate due to their inability to
accurately delineate these wetland features. To address this critical gap, we utilized higher resolution
Sentinel-2 imagery and centered our methodology on a multifaceted approach and employed a diverse
set of variables to map the spatial distribution of these wetlands.
Our findings agreed with the results of prior research studies that have underscored the limitations
of the NWI in effectively delineating small wetlands. Du et al. [41] conducted an assessment of the
NWI program and revealed the challenges faced in achieving precision in wetland mapping,
particularly when dealing with the smaller wetland patches – attributable to the NWI adoption of a
relatively larger mapping unit, approximately 0.20 hectares. Similarly, Chignell et al. [42] expressed
the limitations of the NWI, especially when applied in remote and forested terrains and advocated for
the integration of supplementary methodologies to improve the mapping of wetlands. These studies
support the compelling necessity for alternative methodologies that could overcome the limitations
inherent in traditional wetland mapping approaches.
Numerous studies have documented the challenges in utilizing remote sensing technologies for
wetland mapping, particularly in environments characterized by intricate landscapes or dense
vegetation cover. Sánchez-Espinosa and Schröder [43] reported difficulties when mapping wetlands
using Landsat imagery, a dilemma mostly related to mixed land cover types and spectral confusion.
Further, Zhang et al. [32] identified restrictions in the classification of wetlands with Sentinel-2 data
within regions categorized by dense vegetation and spectral signatures that exhibit significant overlap.
These disagreements in findings underscore the importance of carefully considering the unique
characteristics of the study area and the primary challenges associated with remote sensing methods
in wetland mapping.
In our study, we overcome the limitations posed by the NWI by utilizing a higher spatial
resolution Sentinel-2 imagery, thus enabling a more accurate mapping of small wetlands. This
approach provides invaluable insights into the spatial distribution of these ecologically pivotal

AIMS Geosciences Volume 10, Issue 1, 62–79.

 74

ecosystems, as represented by the BWG area. However, it is important to acknowledge the limitations
of our models, such as potential challenges in classifying complex landscapes and the need for further
research to integrate additional data sources into the algorithm.

4.3. Importance of Input Variables

The consistent patterns observed across the spectrum of variables used in the ML models – various
spectral bands, spectral indices and image texture attributes – have underscored the importance of
specific variables in the accurate delineation of wetlands. Particularly, the B4 and B11 bands, in
combination with the NDVI, have emerged as pivotal variables for wetland mapping. The spectral
bands B8, B5 and B2 also exhibited importance when discriminating wetlands from other land cover
classes. Our results align with the findings by Mahdianpari et al. [22] that emphasized the critical role
of SWIR bands and spectral indices in the classification of wetlands. Further, Judah and Hu [37]
emphasized the significance of NDVI, particularly when derived from images during the spring
months, as a critical variable for wetland classification. The absence of image texture as an important
variable in the classification process supports the observations presented by Liu et al. [24] when
mapping wetlands within the Tibetan Plateau. These consistencies in research results highlight the
importance of specific variables that significantly contribute to the efficacy of wetland classification
methodologies.

4.4. Implications

Our research results have considerable implications for conservation, operation and scientific
exploration of small-sized wetlands. The accuracy achieved in mapping these wetlands through the
utilization of advanced ML techniques and remote sensing data offers critical insights regarding their
spatial distribution, characteristics and their ecological significance.
Wetland mapping is a crucial component of conservation and management initiatives, particularly
when addressing small-scale wetlands due to their ecological importance as habitats for a diverse array
of plant and animal species [44]. The mapping of these wetlands not only facilitates the implementation
of targeted conservation strategies, including habitat protection and restoration planning but also
ensures the preservation of biodiversity and the continuity of essential ecosystem functions [45].
Furthermore, the mapping of small-sized wetlands proves essential for the comprehensive evaluation
and quantification of their contributions to ecosystem services. These ecosystems play an active role
in vital processes such as water purification, flood regulation and carbon sequestration [46]. The
precision achieved through accurate wetland mapping streamlines the assessment of their spatial
distribution, facilitating the precise quantification of their roles in delivering ecosystem services. This,
consecutively, encourages decision-making that is more informed and discerning.
This research could provide insights for BWG wetlands habitat assessment and restoration
initiatives, particularly given the key role of our study area as a critical habitat for numerous rare and
endangered flora and fauna species. The precise mapping of wetlands not only facilitates the
identification of habitat areas of utmost importance but also helps in the identification of tailored
restoration efforts. These efforts bolster hydrological connectivity and thereby advancing the cause of
biodiversity conservation and the rejuvenation of ecosystems [47].

AIMS Geosciences Volume 10, Issue 1, 62–79.

 75

Efforts to monitor of small-sized wetlands could greatly benefit from accurate mapping data [45].
Researchers could utilize these mapped wetlands as study sites to investigate and assess environmental
impacts, and monitor changes over time. The availability of reliable wetland maps enhances the
efficiency and accuracy of field data collection, optimizing resource allocation and permitting a more
focused investigation into specific wetland areas of interest [22,37]. Finally, our results underscore the
importance of wetland mapping in supporting policy development and land-use planning. Including
accurate wetland information in land use plans and zoning regulations ensures the protection of these
ecologically sensitive areas and minimizes potential conflicts between wetland conservation and
development activities. Reliable wetland maps aid in assessing the ecological implications of proposed
land use changes and support sustainable land management practices.

5. Conclusions

This research highlights the importance and effectiveness of utilizing advanced machine learning
techniques and remote sensing data for mapping small-sized wetlands. The findings demonstrated that
machine learning classifiers, such as RF and GTB, can accurately classify and map small-sized
wetlands. These classifiers outperformed other methods, such as SVM, CART, KNN and NB, in terms
of overall accuracy and agreement with reference classifications.
Accurately mapping small-sized wetlands has benefits for wetland conservation, management
and research. Apart from enabling targeted conservation efforts, climate change adaptation strategies
and habitat assessment and restoration, it is essential for scientific research and informed policy
support. Our results and methods could contribute to mapping small-sized wetlands, and provide
necessary maps for wetland preservation. The application of advanced machine learning techniques
using GEE in small-sized wetland mapping creates ideas for further research and application in other
geographic regions and ecosystems. Overall, this study emphasized the significance of accurate
wetland mapping using machine learning approaches. It serves as a foundation for practical
applications aimed at improving wetland management and promoting their long-term sustainability.

Use of AI tools declaration

The authors declare they have not used Artificial Intelligence (AI) tools in the creation of this
article.

Acknowledgments

The study was also supported by NIFA/USDA through Central State University Evans-Allen
Research Program Fund Number NI201445XXXXG018-0001.

Conflict of interest

All authors declare no conflicts of interest in this paper.

AIMS Geosciences Volume 10, Issue 1, 62–79.

 76

References

1. Finlayson CM, Davidson NC, Spiers AG, et al. (1999) Global wetland inventory – current status
and future priorities. Mar Freshwater Res 50: 717–727. https://doi.org/10.1071/MF99098
2. Kotze DC, Ellery WN, Macfarlane DM, et al. (2012) A rapid assessment method for coupling
anthropogenic stressors and wetland ecological condition. Ecol Indic 13: 284–293.
https://doi.org/10.1016/j.ecolind.2011.06.023
3. Dahl T, Allord G Technical Aspects of Wetlands. History of Wetlands in the Conterminous United
States. Available from: https://water.usgs.gov/nwsum/WSP2425/history.html.
4. Darrah SE, Shennan-Farpón Y, Loh J, et al. (2019) Improvements to the Wetland Extent Trends
(WET) index as a tool for monitoring natural and human-made wetlands. Ecol Indic 99: 294–298.
https://doi.org/10.1016/j.ecolind.2018.12.032
5. Davey Resource Group (2006) GIS Wetlands Inventory and Restoration Assessment, Cuyahoga
County, Ohio, Cuyahoga Soil and Water Conservation District.
6. van der Kamp G, Hayashi M (1998) The Groundwater Recharge Function of Small Wetlands in
the Semi-Arid Northern Prairies. Great Plains Res 8: 39–56.
7. Tang Z, Li Y, Gu Y, et al. (2016) Assessing Nebraska playa wetland inundation status during 1985–
2015 using Landsat data and Google Earth Engine. Environ Monit Assess 188: 654.
https://doi.org/10.1007/s10661-016-5664-x
8. Gibbs JP (1993) Importance of small wetlands for the persistence of local populations of wetland-
associated animals. Wetlands 13: 25–31. https://doi.org/10.1007/BF03160862
9. Wang W, Sun M, Li Y, et al. (2022) Multi-Level Comprehensive Assessment of Constructed
Wetland Ecosystem Health: A Case Study of Cuihu Wetland in Beijing, China. Sustainability 14:
13439. https://doi.org/10.3390/su142013439
10. Szantoi Z, Escobedo FJ, Abd-Elrahman A, et al. (2015) Classifying spatially heterogeneous
wetland communities using machine learning algorithms and spectral and textural features.
Environ Monit Assess 187: 262. https://doi.org/10.1007/s10661-015-4426-5
11. Chatziantoniou A, Psomiadis E, Petropoulos GP (2017) Co-Orbital Sentinel 1 and 2 for LULC
Mapping with Emphasis on Wetlands in a Mediterranean Setting Based on Machine Learning.
Remote Sens 9: 1259. https://doi.org/10.3390/rs9121259
12. Wei C, Guo B, Fan Y, et al. (2022) The Change Pattern and Its Dominant Driving Factors of
Wetlands in the Yellow River Delta Based on Sentinel-2 Images. Remote Sens 14: 4388.
https://doi.org/10.3390/rs14174388
13. Malinowski R, Lewiński S, Rybicki M, et al. (2020) Automated Production of a Land Cover/Use
Map of Europe Based on Sentinel-2 Imagery. Remote Sens 12: 3523.
https://doi.org/10.3390/rs12213523
14. Cai Y, Lin H, Zhang M (2019) Mapping paddy rice by the object-based random forest method
using time series Sentinel-1/Sentinel-2 data. Adv Space Res 64: 2233–2244.
https://doi.org/10.1016/j.asr.2019.08.042
15. Jamali A, Mahdianpari M, Brisco B, et al. (2021) Deep Forest classifier for wetland mapping
using the combination of Sentinel-1 and Sentinel-2 data. GIScience Remote Sens 58: 1072–1089.
https://doi.org/10.1080/15481603.2021.1965399

AIMS Geosciences Volume 10, Issue 1, 62–79.

 77

16. Millard K, Richardson M (2013) Wetland mapping with LiDAR derivatives, SAR polarimetric
decompositions, and LiDAR–SAR fusion using a random forest classifier. Can J Remote Sens 39:
290–307. https://doi.org/10.5589/m13-038
17. Onojeghuo AO, Onojeghuo AR, Cotton M, et al. (2021) Wetland mapping with multi-temporal
sentinel-1 & -2 imagery (2017–2020) and LiDAR data in the grassland natural region of alberta.
GIScience Remote Sens 58: 999–1021. https://doi.org/10.1080/15481603.2021.1952541
18. Pham H-T, Nguyen HQ, Le KP, et al. (2023) Automated Mapping of Wetland Ecosystems: A
Study Using Google Earth Engine and Machine Learning for Lotus Mapping in Central Vietnam.
Water 15: 854. https://doi.org/10.3390/w15050854
19. Bhatnagar S, Gill L, Regan S, et al. (2020) Mapping vegetation communities inside wetlands
using Sentinel-2 imagery in Ireland. Int J Appl Earth Obs Geoinformation 88: 102083.
https://doi.org/10.1016/j.jag.2020.102083
20. Ruiz LFC, Guasselli LA, Simioni JPD, et al. (2021) Object-based classification of vegetation
species in a subtropical wetland using Sentinel-1 and Sentinel-2A images. Sci Remote Sens 3:
100017. https://doi.org/10.1016/j.srs.2021.100017
21. Kaplan G, Avdan U (2019) Evaluating Sentinel-2 Red-Edge Bands for Wetland Classification.
Proceedings 18: 12. https://doi.org/10.3390/ECRS-3-06184
22. Mahdianpari M, Jafarzadeh H, Granger JE, et al. (2020) A large-scale change monitoring of
wetlands using time series Landsat imagery on Google Earth Engine: a case study in
Newfoundland. GIScience Remote Sens 57: 1102–1124.
https://doi.org/10.1080/15481603.2020.1846948
23. Waleed M, Sajjad M, Shazil MS, et al. (2023) Machine learning-based spatial-temporal
assessment and change transition analysis of wetlands: An application of Google Earth Engine in
Sylhet, Bangladesh (1985–2022). Ecol Indic 75: 102075.
https://doi.org/10.1016/j.ecoinf.2023.102075
24. Liu Q, Zhang Y, Liu L, et al. (2021) A novel Landsat-based automated mapping of marsh wetland
in the headwaters of the Brahmaputra, Ganges and Indus Rivers, southwestern Tibetan Plateau.
Int J Appl Earth Obs Geoinformation 103: 102481. https://doi.org/10.1016/j.jag.2021.102481
25. Wu N, Crusiol L, Liu G, et al. (2023) Comparing Machine Learning Algorithms for Pixel/Object-
Based Classifications of Semi-Arid Grassland in Northern China Using Multisource Medium
Resolution Imageries. Remote Sens 15: 750. https://doi.org/10.3390/rs15030750
26. Gemechu GF, Rui X, Lu H (2022) Wetland Change Mapping Using Machine Learning Algorithms,
and Their Link with Climate Variation and Economic Growth: A Case Study of Guangling County,
China. Sustainability 14: 439. https://doi.org/10.3390/su14010439
27. Ghorbanian A, Zaghian S, Asiyabi RM, et al. (2021) Mangrove Ecosystem Mapping Using
Sentinel-1 and Sentinel-2 Satellite Images and Random Forest Algorithm in Google Earth Engine.
Remote Sens13: 2565. https://doi.org/10.3390/rs13132565
28. Hemati MA, Hasanlou M, Mahdianpari M, et al. (2021) Wetland Mapping of Northern Provinces
of Iran Using Sentinel-1 and Sentinel-2 in Google Earth Engine, 2021 IEEE International
Geoscience and Remote Sensing Symposium IGARSS, 96–99.
https://doi.org/10.1109/IGARSS47720.2021.9554984
29. Gxokwe S, Dube T, Mazvimavi D (2022) Leveraging Google Earth Engine platform to
characterize and map small seasonal wetlands in the semi-arid environments of South Africa. Sci.
Total Environ 803: 150139. https://doi.org/10.1016/j.scitotenv.2021.150139

AIMS Geosciences Volume 10, Issue 1, 62–79.

 78

30. Tamiminia H, Salehi B, Mahdianpari M, et al. (2020) Google Earth Engine for geo-big data
applications: A meta-analysis and systematic review. ISPRS J Photogramm Remote Sens 164:
152–170. https://doi.org/10.1016/j.isprsjprs.2020.04.001
31. Pal M (2005) Random forest classifier for remote sensing classification. Int J Remote Sens 26:
217–222. https://doi.org/10.1080/01431160412331269698
32. Zhang L, Hu Q, Tang Z (2022) Assessing the contemporary status of Nebraska’s eastern saline
wetlands by using a machine learning algorithm on the Google Earth Engine cloud computing
platform. Environ Monit Assess 194: 193. https://doi.org/10.1007/s10661-022-09850-8
33. Domingos P, Pazzani M (1997) On the Optimality of the Simple Bayesian Classifier under Zero-
One Loss. Mach Learn 29: 103–130. https://doi.org/10.1023/A:1007413511361
34. Hall P, Park BU, Samworth RJ (2008) Choice of neighbor order in nearest-neighbor classification.
Ann Stat 36: 2135–2152. https://doi.org/10.1214/07-AOS537
35. Boser B, Guyon I, Vapnik V (1992) A Training Algorithm for Optimal. Margin Classifiers. COLT,
144–152. https://doi.org/10.1145/130385.130401
36. Fekri E, Latifi H, Amani M, et al. (2021) A Training Sample Migration Method for Wetland
Mapping and Monitoring Using Sentinel Data in Google Earth Engine. Remote Sens 13: 4169.
https://doi.org/10.3390/rs13204169
37. Judah A, Hu B (2022) The Integration of Multi-Source Remotely Sensed Data with Hierarchically
Based Classification Approaches in Support of the Classification of Wetlands. Can J Remote Sens
48: 158–181. https://doi.org/10.1080/07038992.2021.1967732
38. Zhang M, Lin H (2022) Wetland classification using parcel-level ensemble algorithm based on
Gaofen-6 multispectral imagery and Sentinel-1 dataset. J Hydrol 606: 127462.
https://doi.org/10.1016/j.jhydrol.2022.127462
39. Amani M, Salehi B, Mahdavi S, et al. (2017) Wetland classification in Newfoundland and
Labrador using multi-source SAR and optical data integration. GIScience Remote Sens 54: 779–
796. https://doi.org/10.1080/15481603.2017.1331510
40. Xing H, Niu J, Feng Y, et al. (2023) A coastal wetlands mapping approach of Yellow River Delta
with a hierarchical classification and optimal feature selection framework. CATENA 223: 106897.
https://doi.org/10.1016/j.catena.2022.106897
41. Du L, McCarty GW, Zhang X, et al. (2020) Mapping Forested Wetland Inundation in the
Delmarva Peninsula, USA Using Deep Convolutional Neural Networks. Remote Sens 12: 644.
https://doi.org/10.3390/rs12040644
42. Chignell SM, Luizza MW, Skach S, et al. (2018) An integrative modeling approach to mapping
wetlands and riparian areas in a heterogeneous Rocky Mountain watershed. Remote Sens Ecol
Conserv 4: 150–165. https://doi.org/10.1002/rse2.63
43. Sánchez-Espinosa A, Schröder C (2019) Land use and land cover mapping in wetlands one step
closer to the ground: Sentinel-2 versus landsat 8. J Environ Manage 247: 484–498.
https://doi.org/10.1016/j.jenvman.2019.06.084
44. Sebastián-González E, Green AJ (2014) Habitat Use by Waterbirds in Relation to Pond Size,
Water Depth, and Isolation: Lessons from a Restoration in Southern Spain. Restor Ecol 22: 311–
318. https://doi.org/10.1111/rec.12078
45. Gitau P, Ndiritu G, Gichuki N (2019) Ecological, recreational and educational potential of a small
artificial wetland in an urban environment. Afr J Aquat Sci 44: 329–338.
https://doi.org/10.2989/16085914.2019.1663721

AIMS Geosciences Volume 10, Issue 1, 62–79.

 79

46. Jie Y, Zhao Y (2021) Trends in Research on Wetland Restoration based on Web of Science
Database. Ecol Environ 30: 1541–1548. https://doi.org/10.16258/j.cnki.1674-5906.2021.07.023
47. Rebelo AJ, Scheunders P, Esler KJ, et al. (2017) Detecting, mapping and classifying wetland
fragments at a landscape scale. Remote Sens Appl Soc Environ 8: 212–223.
https://doi.org/10.1016/j.rsase.2017.09.005

© 2024 the Author(s), licensee AIMS Press. This is an open access

article distributed under the terms of the Creative Commons
Attribution License (http://creativecommons.org/licenses/by/4.0)

AIMS Geosciences Volume 10, Issue 1, 62–79.

View publication stats