GeoAI-Driven Wetland Change Analysis in the Sangamon River Watershed (2000–2025): A Comparative Assessment of Machine Learning and Deep Learning Approaches

Highlights

What are the main findings?

• The Sangamon River Watershed experienced major landscape transformation over 25 years, including a loss of 54.85% of wetland extent (860.68 km²), 46.51% of forest cover (694.88 km²), and 36.8% of water bodies (66.13 km²), alongside substantial increases in agriculture/grassland/barren land 8.56% (820.62 km²) and urban/developed areas 72.92% (799.52 km²).
• A comparative GeoAI framework integrating multi–source Earth observations, object–based image analysis, ML, CNNs, and transformer–based DL models demonstrated that transformer-based models achieved superior segmentation of spatially fragmented and spectrally heterogeneous wetlands and land–cover classes.

What are the implications of the main findings?

• The findings show the potential of advanced GeoAI and multi–source Earth observation for accurate, scalable, and transferable watershed–scale wetland monitoring and landuse change analysis.
• Results provide valuable geospatial insights for wetland conservation, restoration, and climate–resilient watershed planning in agriculturally dominated watersheds.

Abstract

Wetlands monitoring is essential for sustainable watershed biodiversity conservation, and climate resilience. This study performs a spatiotemporal wetland change analysis for the Sangamon River Watershed, Illinois between 2000 and 2025, using Landsat 5 Thematic Mapper (TM), Sentinel–2 Surface Reflectance (SR), Synthetic Aperture Radar (SAR), Gray–Level Co–occurrence Matrix (GLCM) and terrain data through cloud–based processing in Google Earth Engine (GEE), Google Colab and ArcGIS Pro 3.6.0. We conducted a comparative assessment of deep learning (Deep Neural Network [DNN], U-Net, Attention U-Net, and SegFormer), and machine learning models (Random Forest [RF], Gradient Tree Boosting [GTB], and Support Vector Machines [SVM]) through pixel–based and object–based methods. National Land Cover Database (NLCD) was used for training and validation using stratified random sampling for five categories namely wetlands, forest, agriculture/grassland/barren land, urban/developed and water. A proportion of 54.85% (860.68 km²) of wetlands extent was lost to other land uses, particularly agriculture, urban and forest, along with 46.51% (694.88 km²) forest and 36.80% (66.13 km²) water bodies loss. Agriculture/grassland/barren and urban/developed witnessed increases of 8.56% (820.62 km²) and 72.92% (799.52 km²), respectively. For Landsat–based and Sentinel–based classifications, SegFormer outperformed all ML and DL classifiers with (OA = 94%, Kappa = 0.89, mean F1 = 0.80, mean IoU = 0.70 and OA = 96%, Kappa = 0.92, mean F1 = 0.95, mean IoU = 0.73, respectively) with excellent wetland delineation (PA = 0.99, UA = 0.97, F1 = 0.98, IoU = 0.97 and PA = 0.99, UA = 0.99, F1 = 0.98, IoU = 0.99, respectively). Sentinel–based classifications had improved performance than Landsat, while object–based models consistently outperformed pixel–based methods. The Digital Elevation Model (DEM) and slope were the most influential predictors for RF models, while GLCM and SAR produced negligible influence. The integrated and comparative GeoAI framework provides a robust methodology for watershed–scale wetland monitoring and supports evidence–based conservation, restoration prioritization, climate resilience, and sustainable land–use planning, while offering strong potential for application in other agricultural watersheds following regional validation.

1. Introduction

Wetlands, often referred as the “kidneys of the landscape”, occupy only about 8% of the Earth’s surface, yet they provide a diverse range of ecological, hydrological, geological, and socio-economic benefits, including water quality improvement, climate regulation, flood mitigation, soil stabilization, carbon sequestration, and biodiversity support [1,2,3,4,5]. Despite their importance, wetlands have become one of the most threatened ecosystems both globally and in the United States (US). Rapid urbanization, agricultural expansion, accelerated sea-level rise (SLR), and climate-driven stressors, such as prolonged droughts, extreme precipitation events, and storm-induced erosion, have also contributed to wetland loss and degradation at large scale [5,6,7,8,9,10]. The loss of wetlands is also associated with increased nutrient loading (nitrogen and phosphorus), elevated sediment transport, reduced flood attenuation capacity, and the decline of essential ecosystem services leading to the environmental degradation of overall ecosystem [5,11,12,13].

The Sangamon River watershed, located in central Illinois (Figure 1), encompasses approximately 13,915 km² across 18 counties and forms a major sub–basin of the Illinois River system. Illinois, one of the most agriculture-dependent states in the US, has experienced the loss of approximately 85–90% of its historical wetland area [6]. Prior to Euro–American settlement, the watershed had extensive wetland complexes linked with prairie potholes, floodplains, depressional landscapes, and riparian corridors characteristic of the Midwest glacial plains [2,6,14]. Extensive drainage, hydrologic alterations, and channel modifications to support intensive row–crop agriculture have historically and largely transformed the Sangamon River Basin wetlands, along with urban expansion and transportation infrastructure development, leading to significant wetland loss in the region over the past century [6,7,8,12,13,14]. These cumulative modifications have left remaining wetlands highly fragmented, spatially isolated, and frequently disconnected from natural flow regimes, reducing their ecological integrity and long–term resilience.

The continuing interaction of anthropogenic pressures and climate variability call for the need for multi-temporal, spatial, remote sensing (RS), and GeoAI-based assessments of wetland change/loss, while forecasting the likelihood for future change/loss, to identify areas of concern, understand scale of wetland change/loss, and support evidence–based restoration, conservation, and protection strategies at the watershed scale.

Monitoring and managing wetlands remain generally challenging tasks due to their spatial heterogeneity, dynamic hydrologic regimes, and diverse vegetation structure. Before the emergence of advanced geospatial and RS technologies, traditional wetland analysis methods were often limited by sparse spatial and temporal coverage, high implementation costs, and labor-intensive field surveys, restricting their ability to capture wetland dynamics at watershed to regional scales. Recent advances in GeoAI approaches have incorporated machine learning (ML) and deep learning (DL) with RS data, aiming to overcome these constraints by enabling scalable, repeatable, and data-driven wetland mapping, classification, and change detection across large spatial extents and multi–decadal time periods [15,16,17,18]. Various studies have examined wetland loss/change across spatial and temporal scales at national and global levels [19,20,21,22], but substantial gaps remain in understanding wetland dynamics at watershed and regional scales, particularly within highly modified, human, and agriculturally dominated landscapes.

Recent studies in wetland classification or change analysis have incorporated multi–temporal RS data, with various ML and DL models along with Synthetic Aperture Radar (SAR) imagery for improved accuracy in the detection of wetlands/mangroves [20,23,24,25,26,27]. Widely applied ML models in related studies, including classification and regression trees (CARTs), random forest (RF), gradient tree boosting (GTB), and support vector machines (SVMs) have demonstrated strong performance in complex RS data classification tasks involving mixed land covers and spectrally similar classes [25,28,29,30,31,32,33,34]. While ML models are effective in achieving high accuracy and handling non–linear relationships among predictor variables, DL models have emerged recently as a powerful alternative. DL models, such as dense neural networks (DNNs), convolutional neural networks (CNNs), U-Net architectures, artificial neural networks (ANNs), recurrent neural networks (RNNs), and hybrid CNN-based semantic segmentation, have an edge over traditional ML models as they can capture complex spatial patterns, object boundaries, and neighborhood relationships that are particularly important in the detection of fine-scale wetland features through spatial-contextual learning, reducing salt-and-pepper noise, and improving boundary delineation [26,27,35,36,37,38,39,40].

Several studies have adopted a comparative approach of different ML and DL classifiers, particularly in land use land cover (LULC) change analysis [25,28,29,30,31,32,33,34], though relatively few have focused on comparative ML/DL approaches for wetland change dynamics [17,20,21,22,23,24,26,27,35,36,37,38,39,40]. Studies using a comparative approach in multiple ML and DL algorithms demonstrate that performance can vary depending on environmental conditions and classification objectives [25,34,41]. ML algorithms utilize distinct learning mechanisms to capture non–linear relationships in RS data; RF relies on an ensemble of independently constructed decision trees, GTB develops trees sequentially to iteratively reduce prediction errors, and SVM determines optimal hyperplanes to maximize class separability in high-dimensional feature space. When combined with multi-spectral satellite imagery and spectral indices, RF, GTB and SVM models have been widely shown to outperform other classifiers in complex landscapes characterized by mixed land covers and subtle wetland signatures [16,25,28,29,30,31,32,33,34,41].

More recent studies have shown that DL models, when integrated with multi–sensor optical and SAR datasets, outperform traditional ML classifiers in wetland mapping applications using multisensory RS imagery [26,27,42]. CNN, U-Net and transformer–based semantic segmentation models perform wetland change analysis with higher accuracy while preserving spatial continuity and landscape structure. This is because they allow end-to-end pixel-wise classification. Furthermore, recent advances in DL approaches have shifted wetland mapping from CNN-based architecture toward attention–based and transformer–based semantic segmentation models. DL models, such as U-Net, U-Net++, Swin-UNet, DeepLabV3+, and SegFormer, have demonstrated improved capability for capturing multi–scale spatial context and long–range feature dependencies in high–resolution RS imagery, particularly when integrating multi–temporal and multi–sensor datasets [26,27,35,36,37,38,39,40]. Therefore, the comparative evaluation of core and advanced ML and DL classifiers provide a comprehensive framework for understanding model strengths and limitations, identifying the most informative predictor variables, and selecting the most suitable classification approach for complex wetland environments. This integrated assessment improves classification accuracy while guiding the selection of the optimal ML or DL model for reliable wetland mapping and monitoring.

Pixel–based or object-based classification methods are two of the primary image classification methods. After the launch of IKONOS in 1999, a shift was observed from pixel–based approaches, where each pixel is classified independently according to its spectral characteristics, to object-based image analysis (OBIA or GEOBIA), a method which groups neighboring pixels into meaningful image objects based on spectral, textural, and contextual characteristics [42,43]. Numerous studies consider OBIA to be an improvement due to its ability to remove the “salt-and-pepper” effect and because of its within-class spectral variability. In wetland change analysis, OBIA is particularly valuable because wetlands are highly heterogeneous environments characterized by irregular boundaries, mixed vegetation communities, varying hydrologic conditions, and transitional ecotones that are difficult to capture using individual pixels alone [42]. However, comparative research in agricultural landscapes has shown that pixel-based approaches can achieve comparable classification performance while offering simpler workflows and lower computational demands [44]. A recent coastal wetland study comparing pixel-based, object-based, and deep-learning approaches reported higher classification accuracy for the DL approach [45]. Despite these advantages, pixel-based approaches remain valuable, especially for medium-resolution (30 m) imagery, where computational efficiency and simpler workflows are important.

Google Earth Engine (GEE), a cloud-based geospatial analysis platform with JavaScript and Python APIs, has become widely used for large-scale remote-sensing applications, including LULC mapping, wetland monitoring, change detection, and multi-source Earth observation analysis [15,34,46,47,48]. GEE provides access to extensive satellite and geospatial datasets at multiple spatial and temporal resolutions, enabling efficient processing and reproducible remote-sensing workflows [15,34,46,47,48].

This study employs an integrated and comparative wetland classification and change detection approach using multi-season (spring and summer) Landsat 5 TM and Sentinel-2 Surface Reflectance (SR) indices, Sentinel-1 SAR backscatter data, terrain, Gray-Level Co-occurrence Matrix (GLCM), and land use data to quantify wetland change detection in complex, heterogeneous, and fragmented landscapes such as the Sangamon River watershed between 2000 and 2025 (Figure 2). In addition to comparing wetland classification accuracy, this study evaluates the contribution of Sentinel-1 SAR data when integrated with Sentinel-2, terrain, and GLCM variables to determine whether SAR provides additional classification benefits in agriculture–dominated landscapes where seasonal crop dynamics and spectral similarity often complicate wetland mapping. GEE and Google Colab cloud–based platforms have been used for data preprocessing, feature generation, model development, and classification while ArcGIS Pro 3.6.0 has been used for change quantification, mapping, and visualization.

We compared traditional ML classifiers, RF, GTB, and SVM with core and advanced DL semantic segmentation models, including DNN, U-Net, Attention U-Net, and SegFormer, to quantify and spatially characterize wetland gains, losses, persistence, and conversions to other land use classes such as agriculture, barren land, grassland, developed land, and water. RF, GTB, and SVM were selected as representative ensemble tree–based and kernel-based ML algorithms that are widely used for RS image classification and which provide a robust benchmark for comparison with DL models. The selected DL models represent the evolution of semantic segmentation architectures, ranging from a baseline deep neural network (DNN) classifier to convolutional neural network (CNN)-based segmentation models (U-Net and Attention U-Net) and transformer–based architecture (SegFormer) and were selected to compare the effectiveness of multilayer perceptron, convolutional, attention-enhanced convolutional, and vision transformer architectures for wetland classification.

The key contributions of this study are as follows:

• Integrating multi-season Landsat 5 TM (for years 2000–2008) and Sentinel-2 (for years 2017–2025) optical indices, Sentinel-1 SAR backscatter, GLCM, and terrain-derived variables within a comparative ML through both pixel-based and object-based approaches and core and advanced DL-based workflows to evaluate their accuracy performance in wetland mapping and change detection.
• Investigating the contribution of Sentinel-1 SAR data when integrated with Sentinel-2, terrain, and GLCM variables to determine its effectiveness for wetland classification in agriculture-dominated landscapes.
• ▪ Applying a temporally consistent classification approach using identical predictors, training data for 2001 and 2017, and classification methodology to ensure that changes reflect true land cover transitions.
• Evaluating DL models performance against ML models for Landsat 5 TM and Sentinel-2, respectively, for different spatial resolutions, and temporal periods, sensor characteristics, and model architectures.
• Quantifies and spatially characterizes wetland dynamics, including gains, losses, and conversions to other land use types, providing critical understanding into landscape-level transformations.
• Integrating cloud-based GEE and Colab platforms that provide high computational efficiency with ArcGIS-based visualization and quantification.
• ▪ Determining the “variable importance” of variables used in this study through the RF classifier algorithm.

This study contributes to the growing body of wetland monitoring research through an integrated, comparative, reproducible, and scalable GeoAI framework that is critical for regional planning, conservation and land management policies, future restoration initiatives, and prediction of the future wetland loss based on current research and allows for its application to other watersheds, contributing to broader wetland monitoring and management efforts in the US and abroad.

2. Methods

2.1. Study Area (Sangamon River Watershed)

The Sangamon River watershed, one of the three core field sites of the Intensively Managed Landscapes Critical Zone Observatory (IML-CZO) [49], is visualized in ArcGIS Pro 3.6.0 (Figure 1). It covers 18 counties and 13,915.13 square kilometers (approx. 5372.66 square miles) in central Illinois, lies within coordinates 40°26′43.12″N, 88°43′57.24″W and 40°1′21.17″N, 90°25′58.45″W. According to the 2025 National Land Cover Database (NLCD), the current land–use composition in the Sangamon River watershed is approximately 75% of row-crop agriculture (corn and soybeans) and grassland, 13% of developed/urban area, 5% of forest, 5% of wetlands, and approx. 1% of open water and other land uses. The Sangamon River watershed was delineated using the United States Geological Survey (USGS) Watershed Boundary Dataset by merging four HUC-8 sub-basins: Upper Sangamon (07130006), South Fork Sangamon (07130007), Lower Sangamon (07130008), and Salt Creek (07130009) both in Arc GIS Pro 3.6.0 and GEE (Google, Mountain View, CA, USA).

2.2. Data Sources and Data Processing

We used the high-resolution, Sentinel-1 (S1) SAR GRD: C-band Synthetic Aperture Radar Ground Range Detected (“COPERNICUS/S1_GRD”) and multispectral Sentinel-2 (S2) Level-2A SR imagery (“COPERNICUS/S2_SR_HARMONIZED”) for years 2017–2025 and Landsat 5 (L5) thematic mapper (TM) (“LANDSAT/LT05/C02/T1_L2”) downloaded in GEE for year 2000 and 2008. The other datasets include digital elevation model (DEM) data from the USGS 3DEP 10/30 m DEM, and NLCD data (2001 and 2017) to generate training and validation samples for classification.

For the deep learning component, DNN, U-Net, Attention U-Net, and SegFormer models were implemented in Python 3 (Google Colab) using TensorFlow/Keras 2.15 and PyTorch 2.9.0. Multi-band image patches were exported from GEE as TFRecord files from the Landsat and Sentinel multi–source predictor stacks and used to train semantic segmentation models for classifying all land–cover classes.

2.3. Cloud Masking, Surface Reflectance Scaling and Band Selection

All image processing and computational analysis were conducted in JavaScript 2026 (GEE) and Python 3 (Google Colab). All GIS raster and vector data were projected on “NAD_1983_StatePlane_Illinois_West_FIPS_1202” (in ArcGIS Pro 3.6.0) and “EPSG:5070” (in GEE). L5 imagery was used to conduct analysis between 2000 to 2008 while S1 and S2 imagery was used for the years 2017 to 2025 to calculate wetland change by multiplying respective pixels in ArcGIS Pro 3.6.0. The native spatial resolution of each sensor was retained to preserve the original spectral and spatial characteristics of the imagery, avoid resampling-induced artifacts, and accurately represent the fragmented and heterogeneous distribution of wetlands, which could otherwise be distorted by resampling. This approach ensured reliable temporal comparisons within each sensor, while wetland change was quantified by converting classified pixels to area and calculating gains, losses, and land–cover transitions in ArcGIS Pro 3.6.0.

S1 and S2 optical imagery were preprocessed using a standardized approach. For Sentinel-2, we used the S2Cloudless DL algorithm to mask cloud, cirrus, cloud shadow, and snow pixels. This algorithm has demonstrated improved performance over other cloud masking algorithms, including scene classification layer (SCL), in minimizing residual contamination [50]. SAR sensitivity to surface moisture, inundation conditions, and vegetation structure improved class separability and reduced misclassification within moisture-sensitive landscapes. For S1 data, GEE’s built-in SAR processing algorithms, including thermal noise removal, radiometric calibration, and terrain correction, were used. Backscatter coefficients (VV and VH polarizations) were converted to sigma naught (σ⁰) values in decibels (dB) and filtered to reduce speckle noise. Temporal compositing and multi-date aggregation were applied to enhance signal stability and reduce noise effects. For L5 quality assessment (QA), band (QA_PIXEL) masking was used.

All images were harmonized and resampled to 10 m (S2) and 30 m (L5) spatial resolution. Reflectance values were scaled by a factor of 0.0001 to convert integer digital numbers to physical reflectance units. Detailed sensor characteristics including selected spectral bands, their spatial resolutions, and revisit frequencies are summarized in

Table 1. Band description for the Landsat 5 TM and Sentinel-2 SR L2A imagery and for the Sentinel-1 SAR GRD used in the study.

Landsat 5 TM				Sentinel–2 SR L2A					SAR GRD
Band	Spectral	Wavelength (µm)	Resolution (m)	Band	Spectral	Wavelength (µm)	Resolution (m)	Band		Polarization	Frequency/Wavelength	Resolution (m)
B1	Blue	0.45–0.52	30	B2	Blue	0.45–0.52	10	VV		Vertical transmit/Vertical receive	C–band (~5.6 cm; 5.405 GHz)	10
B2	Green	0.52–0.60	30	B3	Green	0.52–0.60	10
B3	Red	0.63–0.69	30	B4	Red	0.63–0.69	10
B4	NIR (near infrared)	0.76–0.90	30	B8	NIR	0.76–0.90	10	VH		Vertical transmit/Horizontal receive	C–band (~5.6 cm; 5.405 GHz)	10
B5	SWIR1 (short wave infrared)	1.55–1.75	30	B11	SWIR1	1.55–1.75	20
B7	SWIR2	2.08–2.35	30	B12	SWIR2	2.08–2.35	20
Temporal resolution 16 day				3–5 days				6–12 days

.

2.4. Composite Generation and Normalization

Seasonal L5 composites for 2000 and 2008 and S2 composites for 2017 and 2025 were generated in GEE for the Sangamon River watershed, Illinois. L5 and S2 imagery was first filtered by seasonal date ranges and clipped to the watershed boundary. For S2, to reduce atmospheric noise, minimize cloud contamination and short–term variability, scaled reflectance images were composited using a median reducer to generate seasonal, cloud–free mosaics for the Spring (March–May) and Summer (June–September) period, which corresponds to peak vegetation growth and maximum wetland inundation in the study area, thereby enhancing the detectability of wetland extent and loss. For L5, a 25th percentile (P25) compositing approach was applied to the image collection for Spring and Summer. The percentile–based composite was selected instead of median compositing as it minimizes the influence of bright atmospheric artifacts and improves the representation of stable land surface conditions for L5. Z-score normalization was applied to standardize spectral variability as shown in Equation (1). Mean and standard deviation statistics for each spectral band were calculated from the seasonal image collection.

$${\displaystyle \frac{z=\left\{x-\mu \right\}}{\sigma }}$$

(1)

where (x) represents the pixel value, (µ) is the mean reflectance of the band, and (σ) is the standard deviation.

A small constant was added during division to avoid instability associated with near–zero standard deviation values. The normalized and denormalized seasonal composites were subsequently renamed according to season–specific band identifiers (e.g., Blue_spring, NIR_summer) and used as predictor variables for ML and DL models. Normalization is used for all classifiers for L5 imagery and SVM for S2 imagery.

2.5. DNN, Attention U-Net and SegFormer Semantic Segmentation for Landsat 5 TM (2000–2008)

DNN, Attention U-Net and SegFormer were selected for L5 imagery for years 2000 to 2008 to evaluate the applicability of both core and advanced DL architecture for a 30 m spatial resolution. DNN was selected as a baseline DL classifier because it learns complex relationships among spectral, environmental, terrain, and SAR-derived predictor variables without relying heavily on fine-scale spatial context, making it well suited for 30 m Landsat imagery. Attention U-Net and SegFormer were included to investigate whether attention mechanisms and transformer–based architectures could improve the delineation of fragmented wetlands despite the relatively coarse spatial resolution.

2.6. U-Net, Attention U-Net, and SegFormer Semantic Segmentation for Sentinel–2 and Sentinel–1 Data (2017–2025)

For the S1 and S2 datasets (2017–2025), U-Net, Attention U-Net, and SegFormer were implemented to compare CNN, attention–based CNN, and transformer–based semantic segmentation approaches. For both DNN and U-Net, spring and summer composites were generated in GEE using similar preprocessing and composite generation as stated in Section 2.3 and Section 2.4, including all of the predictor variables used for ML models. The higher spatial resolution (10 m) of the Sentinel datasets enables the models to exploit detailed spatial patterns, object boundaries, and contextual information.

Predictor stacks and land-cover labels were exported from GEE as multi-band TFRecord image patches and were used to train the DL models in Google Colab. Spring and summer Landsat composites were generated following the preprocessing workflow described in Section 2.3 using the same predictor variables employed for the ML classifiers.

2.7. Spectral Indices

Spectral indices were calculated from both L5 and S2 imagery to enhance differentiation among vegetation, wetlands, built-up areas, open water, and moisture variations (

Table 2. Spectral indices derived in the study and their interpretation.

	Index	Purpose	Range	Equations	Number	Explanation/References	Interpretation
1	Normalized Difference Vegetation Index (NDVI)	Vegetation health	−1 to +1	$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}={\displaystyle \frac{\left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{E}\mathrm{D}\right)}{\left(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}\mathrm{E}\mathrm{D}\right)}}$	(2)	Equation (2) calculates NDVI, i.e., vegetation greenness, to check vegetation density, vigor, and overall health and use to differentiate vegetated areas from other land use types (e.g., soil, water, covered areas) [51]. Healthy vegetation absorbs red and reflects NIR due to chlorophyll and cell structure.	(<0) are clouds and water bodies. (0) are barren rock, sand, or snow. (0.2 to 0.5) are shrubs, grasslands, or sparse vegetation. (0.6 to 0.9) are dense, healthy vegetation (temperate, and tropical forests).
2	Modified Normalized Difference Water Index (MNDWI)	Wetlands and shallow water	−1 to +1	$\mathrm{M}\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}={\displaystyle \frac{\left(\mathrm{G}\mathrm{R}\mathrm{E}\mathrm{E}\mathrm{N}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}\right)}{\left(\mathrm{G}\mathrm{R}\mathrm{E}\mathrm{E}\mathrm{N}+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}\right)}}$	(3)	MNDWI in Equation (3) improves NDWI by using SWIR to suppress built–up land spectral signatures, making it more effective for water mapping in urban regions [52]. It is better at delineating waterlogged areas mixed with vegetation, which are common in the Sangamon River watershed floodplains.	(>0.2) are open water. (−1 to 0) are non–water features (built–up/soil/vegetation)
3	Normalized Difference Built–up Index (NDBI)	Differentiates built–up areas	−1 to +1	$\mathrm{N}\mathrm{D}\mathrm{B}\mathrm{I}={\displaystyle \frac{\left(\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}-\mathrm{N}\mathrm{I}\mathrm{R}\right)}{\left(\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}+\mathrm{N}\mathrm{I}\mathrm{R}\right)}}$	(4)	NDBI in Equation (4) highlights and differentiates built-up areas from vegetation and soil [53].	(−1 to 0) are vegetation, water bodies, or empty, non-urban land. (0 to 1) are built-up, urban areas. (+1) are high-density, dense urban areas. (0) is a mix of built–up, urban areas and sparse, sparse vegetation.
4	Normalized Difference Moisture Index (NDMI)	Vegetation moisture/soil moisture	−1 to +1	$\mathrm{N}\mathrm{D}\mathrm{M}\mathrm{I}={\displaystyle \frac{\left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}\right)}{\left(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}\right)}}$	(5)	NDMI in Equation (5) is useful to estimate vegetation moisture content [54].	(>0.5) indicate higher vegetation water content and moist surfaces (wetlands, irrigated areas). (0.2–0.5) indicate moderate moisture conditions. (−1 to 0) indicate dry vegetation, bare soil, or built-up areas.
5	Automated Water Extraction Index (AWEI)	Improved water detection from shadows and built–up areas	−1 to +1	$\mathrm{A}\mathrm{W}\mathrm{E}\mathrm{I}=4\left(\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1\right)-0.25\left(\mathrm{N}\mathrm{I}\mathrm{R}\right)+2.75\left(\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}2\right)$	(6)	AWEI in Equation (6) has improved water mapping accuracy by reducing noise from shadows and dark surfaces than MNDWI [55].	(>0.5) indicate open water and waterlogged areas. (−1 to 0) represent non-water features.
6	Enhanced Vegetation Index (EVI)	Enhanced vegetation health	−1 to +1	$\mathrm{E}\mathrm{V}\mathrm{I}={\displaystyle \frac{2.5\left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{E}\mathrm{D}\right)}{\left\{\mathrm{N}\mathrm{I}\mathrm{R}+6\left(\mathrm{R}\mathrm{E}\mathrm{D}\right)\right\}-\left\{7.5\left(\mathrm{B}\mathrm{L}\mathrm{U}\mathrm{E}\right)+1\right\}}}$	(7)	EVI in Equation (7) is an improvement upon the NDVI by enhancing sensitivity in high biomass regions and reducing canopy background noise and atmospheric interference [56].	(<0) indicate water, clouds, or non–vegetated surfaces. (<0.2) indicate sparse vegetation or bare land. (0.2–0.5) represent moderate vegetation. (>0.5) indicate dense vegetation with high biomass.
7	Bare Soil Index (BSI)	differentiate bare soil from vegetation, water, and built–up areas	−1 to +1	$\mathrm{B}\mathrm{S}\mathrm{I}={\displaystyle \frac{\left(\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1+\mathrm{R}\mathrm{E}\mathrm{D}\right)-\left(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{B}\mathrm{L}\mathrm{U}\mathrm{E}\right)}{\left(\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1+\mathrm{R}\mathrm{E}\mathrm{D}\right)+\left(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{B}\mathrm{L}\mathrm{U}\mathrm{E}\right)}}$	(8)	BSI in Equation (8) is used to enhance bare soil features that distinguish them from other land uses such as water, vegetation, and built-up areas [57].	(<0) indicate water bodies or dense vegetation. (≈−0.2 to 0) represent vegetated surfaces with minimal exposed soil. (0 to 0.3) indicate mixed surfaces (sparse vegetation with exposed soil). (>0.3) indicate bare soil, dry riverbeds, or exposed land surfaces.
8	Modified Soil Adjusted Vegetation Index (MSAVI)	Enhanced early–season crops detection from barren land	−1 to +1	$\mathrm{M}\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}={\displaystyle \frac{2\times \mathrm{N}\mathrm{I}\mathrm{R}+1-\sqrt{\left(2\times \mathrm{N}\mathrm{I}\mathrm{R}+1\right)-8\left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{E}\mathrm{D}\right)}}{2}}$	(9)	Equation (9) illustrates MSAVI, which is useful for distinguishing early-season crops from barren land and reduces soil background impacts [58].	(<0) indicate water, clouds, or non–vegetated surfaces. (<0.2) indicate bare soil or very sparse vegetation. (0.2–0.5) represent emerging or moderately dense vegetation, especially useful in early growth stages. (>0.5) indicate dense, healthy vegetation with minimal soil influence.
9	Urban Index (UI) (Landsat 5 TM)	differentiate barren land from other land uses	−1 to +1	$\mathrm{U}\mathrm{I}={\displaystyle \frac{\left(\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}-\mathrm{N}\mathrm{I}\mathrm{R}\right)}{\left(\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}+\mathrm{N}\mathrm{I}\mathrm{R}\right)}}=1.0\times 100$	(10)	Equation (10) illustrates UI which is used specifically for Landsat TM to distinguish built–up areas from bare soil [59].	(<0) indicate water bodies or vegetated areas. (≈−0.2 to 0) represents dense vegetation or moist surfaces. (0 to 0.2) indicate mixed land cover (transitional zones). (>0.2) indicate urban/built–up areas, dry soil, or impervious surfaces.
10	Land Surface Water Index (LSWI)	detect the water content in vegetation and soil	−1 to +1	$\mathrm{L}\mathrm{S}\mathrm{W}\mathrm{I}={\displaystyle \frac{\left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}\right)}{\left(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}\right)}}$	(11)	Equation (11) illustrates LSWI which is used to monitor the water content in vegetation and soil [60].	(<0) indicate dry soil, impervious surfaces, urban/built-up areas, barren land, or sparse vegetation with low moisture content. (≈0 to 0.1) indicate dry vegetation, agricultural fields, grasslands, or mixed land-cover conditions with limited surface moisture. (≈0.1 to 0.3) indicate healthy vegetation, moist soils, wetlands, and areas with moderate water content. (>0.3) indicate saturated soils, inundated wetlands, flooded vegetation, and open water bodies with high moisture content.

). These indices, together with spectral bands, were used as predictor variables for all classifiers for both pixel–based and object–based methods. Among these predictors, MNDWI and the NDMI are considered highly useful for wetland change analysis. While MNDWI provides superior mapping for open water and for suppressing built–up noise, NDMI tracks moisture content in vegetation and soil, providing a comprehensive view of wetland degradation, drying, or shifts from aquatic to terrestrial environments.

2.8. Digital Elevation Model (DEM), Slope, and Distance–to–Water

DEM, slope (Figure 3a,b), and distance-to-water were included as predictor variables to capture topographic gradients, surface hydrological connectivity, and inundation potential. DEM data were downloaded from the USGS 3DEP 30 m/10 m DEM and clipped to the study area extent and terrain slope was derived from the DEM, resampled using bilinear interpolation, and reprojected to consistent 30 m/10 m for L5 and S2, respectively. Surface water features were extracted from L5 and S2 imagery using the MNDWI, where pixels with values greater than zero were classified as water. A Euclidean distance-to-water layer was then generated using a fast distance transform algorithm, with distances computed from the nearest water pixel and converted to metric units.

2.9. Gray-Level Co-Occurrence Matrix (GLCM)

Multi-seasonal GLCM features were incorporated for both L5 and S2 imagery to enhance land use/wetland classification accuracy, particularly with ML-based classifiers like RF and SVM [61,62]. GLCM is a statistical approach for analyzing image texture that explicitly captures the spatial relationships between pixels. It works by quantifying how often pairs of pixels with specific intensity values occur at a defined distance and orientation (e.g., one pixel to the right, above, or diagonally adjacent to another). We computed contrast, entropy, and inverse difference moment (IDM) using a 3 × 3 window size which is useful for fine details relevant to land used with fragmented wetland composition.

2.10. Training Samples

Wetland training samples were obtained from a rasterized (30 m L5 and 10 m for S2) NLCD dataset, where wetland classes (90 and 95), woody wetlands (regions that have forest or shrubland vegetation greater than 20% of vegetative cover and where the soil or substrate is periodically saturated with or covered with water), and emergent herbaceous wetlands (regions that have perennial herbaceous vegetation greater than 80% of vegetative cover and the soil or substrate is periodically saturated with or covered with water) [63] were classified as Class1, with other land cover classes classified as forest (2), agriculture/grassland/barren land (3), urban/developed (4), and water (5) (Figure 3c). A stratified random sampling approach with same classification scheme and class distribution strategy was used for both L5 and S2 datasets to ensure balanced and representative sampling across all classes. For all pixel-based models, a total of 10,000 sample points were generated at 10/30 m resolution, with class-specific sample sizes allocated approximately proportional to the areal extent of each land–cover class while ensuring the sufficient representation of minority classes. (2600, 2000, 2500, 2200, and 700 samples per class) (

Table 3. (a) Class–wise training sample distribution for wetland and non–wetland classes for ML Classifiers. (b) Sampling approach for ML and DL models compared in the study.

(a)
		Pixel–Based Sampling			Object–Based Sampling
Class Name/Class ID	Sub–Classes	Samples per Class	Training Distribution	Testing Distribution	Samples per Class	Training Distribution	Testing Distribution
Wetland (1)	NLCD 90, 95 (woody wetlands/emergent herbaceous wetlands, freshwater-forested and shrub wetland)	2600	1822	778	3500	2450	1050
Forest (2)	NLCD 41, 42, 43,	2000	1414	586	1500	1050	450
Agriculture/grassland/barren land (3)	NLCD 81, 82, 71, 31	2500	1732	768	2000	1400	600
Urban/developed (4)	NLCD 21, 22, 23, 24	2200	1530	670	1500	1050	450
Water (5)	NLCD 11, 12	700	493	207	500	350	150
Total sample size	10,000				9000
Training size (70%)	6991				6300
Testing size (30%)	3009				2700
(b)
Model Type			Sampling Approach
RF (ML) and GTB (ML)			Pixel–based, object–based
SVM (ML)			normalized pixel-based, OBIA
Dense neural network (DL)			tabular predictor samples,
U-NET, Attention U-NET, and SegFormer (DL)			TFRecord patches

).

Sampling was conducted in the upper, middle, and lower reaches of the watershed to capture spatial heterogeneity. To reduce class bias we enforced proportional representation across sampling design for all land cover classes, while per-pixel stratified sampling also minimizes label ambiguity when compared with vector–based approaches and ensures consistency with the spatial resolution of predictor variables (Table 3). Training and validation samples with missing predictor values were removed. The resulting dataset was used to train all of the classifiers for wetland change analysis. For OBIA, image objects were generated using the simple non–iterative clustering (SNIC) segmentation algorithm, where segment-level mean predictor values and majority land-cover labels were extracted from segmented image objects. These object-level samples (9000) were subsequently used to train the RF, GTB, and SVM object-based classifiers (Table 3).

For the DNN model (L5), after the stratified random sampling based on reclassified NLCD, predictor values were extracted using sampleRegions(), were normalized to reduce differences in variable ranges and to improve neural network convergence in GEE, and were exported as CSV tables for training and validation in Google Colab. For the L5, Attention U-Net, and SegFormer models, multi–band predictor stacks and corresponding land–cover labels were exported as 64 × 64 TFRecord image patches. For the S1/S2 models, U-Net and Attention U-Net were trained using 64 × 64 TFRecord image patches, whereas SegFormer was trained using 256 × 256 patches to better capture larger spatial context (Table 3).

2.11. Training ML and DL Classifiers for Classifying Maps (2000–2008–2017–2025)

All classifiers were trained using predictor variables, including spectral bands and indices, terrain variables, hydrologic proximity layers, SAR backscatter data, and GLCM, and the dataset was divided into 70% training and 30% validation/testing using a reproducible random seed to ensure consistency across classification. Hyperparameters for each classifier were optimized through iterative testing to improve classification accuracy and reduce overfitting. ML used 500 to 550 trees with 6 variables per split, taking the square root of the total number of predictors.

Image objects were generated using the simple non-iterative clustering (SNIC) segmentation algorithm in GEE and segment-level mean predictor variables and majority land-cover labels were extracted and used to train the ML object–based classifiers. Separate classification models were developed for L5 and S2/S1 datasets to account for differences in spatial resolution, spectral characteristics, and sensor properties. Training and validation were performed at 10 m/30 m resolution for the S1/S2 and L5 datasets, respectively.

For DNN, we trained the model using categorical cross-entropy loss and the Adam optimizer with 70:30 split for training and testing subsets. The DNN architecture consisted of multiple fully connected dense layers with batch normalization, dropout regularization, and ReLU activation functions to model non-linear relationships among predictor variables to reduce overfitting. Predictor variables were normalized prior to model training to improve convergence and classification performance.

For U-Net, semantic segmentation multi-band image patches (64 × 64 pixels) were exported from GEE as TFRecord datasets and trained in Google Colab using TensorFlow/Keras. U-Net effectively learns complex spatial patterns and improves segmentation of wetlands and heterogeneous landscape boundaries through an encoder-decoder architecture with skip connections that preserve fine spatial detail. Convolutional and max–pooling layers extract hierarchical spatial features during encoding, while decoder layers progressively reconstruct segmentation outputs through upsampling and skip connections. The final layer used softmax activation for multi-class semantic segmentation.

For Attention U-Net, the same TFRecord image patches were used for training in TensorFlow/Keras using the keras-unet-collection library. Attention U-Net extends the conventional U-Net by incorporating attention gates within the skip connections, enabling the network to selectively emphasize informative spatial features while suppressing irrelevant background information. This mechanism enhances the delineation of fragmented wetlands and complex land-cover boundaries. Similar to U-Net, the model employed an encoder-decoder architecture with convolutional layers, batch normalization, ReLU activation, max-pooling, bilinear upsampling, and softmax activation for five-class semantic segmentation.

For SegFormer, multi-band TFRecord image patches (64 × 64 pixels for L5 and 256 × 256 pixels for S1/S2) were trained using PyTorch. SegFormer is a lightweight transformer–based semantic segmentation model that combines a hierarchical mix transformer (MiT) encoder with an all-MLP decoder to capture both local and global contextual information without relying on positional encodings. The larger Sentinel patch size was selected to provide sufficient spatial context for the transformer’s hierarchical attention mechanism, whereas 64 × 64 patches were retained for Landsat imagery because of its coarser 30 m spatial resolution. Class imbalance in both Attention U-Net and SegFormer was addressed using a hybrid loss function combining class–weighted cross-entropy and macro-averaged dice loss, with class weights iteratively refined using validation results. SegFormer was optimized using the AdamW optimizer with weight decay and cosine annealing learning-rate scheduling. Final classified maps were generated for the years 2000, 2008, 2017, and 2025, and subsequent wetland change was calculated using raster calculator in ArcGIS Pro.

2.12. Pearsons Correlation Analysis

Pearson correlation analysis was conducted including the optical, SAR, terrain, and spectral index variables and excluding the GLCM texture variables as they represent derived texture features from NDVI and SAR imagery (Figure 4a,b). Their contribution to model performance was instead evaluated through RF variable importance. The Landsat–based and Sentinel–based correlation analysis indicated that most predictor variables exhibited low to moderate pairwise correlations, suggesting limited multicollinearity among the selected predictors.

For Landsat-based analysis, LSWI and NDMI exhibited a perfect positive correlation (r = +1.00), whereas LSWI and NDBI showed a perfect negative correlation (r = −1.00). BSI also exhibited very strong correlations with LSWI, NDMI, and NDBI (|r| ≈ 1.00), indicating substantial redundancy among these indices. Among the spectral bands, strong positive correlations were observed between blue and green (r = +0.997) and between SWIR1 and SWIR2 (r = +0.997). Vegetation indices also showed strong positive associations, particularly between EVI and MSAVI (r = +0.9996), while NDVI and MSAVI exhibited moderate to strong positive correlations. Several moisture-related indices also showed strong correlations, reflecting their shared spectral responses. In contrast, terrain variables, including DEM, slope, and distance to water, showed weak correlations with the spectral predictors, indicating that they provide complementary information for wetland classification (Figure 4a).

For Sentinel-based analysis, BSI and NDMI (r = −1.00) and BSI and LSWI (r = −1.00) exhibited perfect negative correlations, whereas LSWI and NDMI (r = +1.00) and BSI and NDBI (r = +1.00) exhibited perfect positive correlations. Adjacent Sentinel-2 spectral bands, including blue–green (r = 0.96), green–red (r = 0.96), blue–red (r = 0.93), and SWIR1–SWIR2 (r = 0.90–0.94), showed strong positive correlations. Similarly, strong positive correlations were observed among vegetation indices, particularly EVI and MSAVI (r = 0.99) and MSAVI and NDVI (r = 0.90–0.92). Terrain variables (DEM, slope, and distance to water) exhibited weak correlations with the optical and spectral indices, indicating that they provide complementary information for wetland classification (Figure 4b).

Overall, the correlation analysis showed that the predictor set integrates multiple sources of complementary information, while the highly correlated derived indices were retained because tree–based ML and DL models are comparatively robust to multicollinearity, and each index represents distinct spectral and environmental characteristics relevant to wetland mapping.

3. Results

3.1. Accuracy Assessment

Model performance was evaluated using overall accuracy (OA), Cohen’s Kappa, producer’s accuracy (PA), user’s accuracy (UA), F1-score, and intersection over union (IoU) (

Table 4. Accuracy assessment matrices used in the study.

	Accuracy Metric	Equations	Number	Explanation
1	Overall accuracy (OA)	$\mathrm{O}\mathrm{A}={\displaystyle \frac{\left(\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}\right)}{\left(\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}\right)}}$	(12)	OA measures correctly classified pixels in all cases, representing Equation (12).
		where TP is true positive, FP is false positive, TN is true negative, and FN is false negative.
2	Kappa coefficient (KC)	$\mathrm{K}\mathrm{C}={\displaystyle \frac{2\left(\mathrm{T}\mathrm{P}·\mathrm{T}\mathrm{N}-\mathrm{F}\mathrm{N}·\mathrm{F}\mathrm{P}\right)}{\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}\right)\left(\mathrm{F}\mathrm{N}+\mathrm{T}\mathrm{N}\right)+\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{F}\mathrm{P}+\mathrm{T}\mathrm{N}\right)}}$	(13)	KC measures classification agreement beyond chance, represented with Equation (13).
		where KC is the Kappa coefficient.
3	Producer accuracy (PA)	$\mathrm{P}\mathrm{A}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}\right)}}$	(14)	PA measures the ability of the classifier to correctly identify reference samples for each class (i.e., omission error) and represents Equation (14).
4	User accuracy (UA)	$\mathrm{U}\mathrm{A}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$	(15)	UA measures the reliability of classified pixels for each class and represents Equation (15).
5	Class–wise F1–score	$\mathrm{F}1\mathrm{i}=2{\displaystyle \frac{\left(\mathrm{P}\mathrm{A}\mathrm{i}\times \mathrm{U}\mathrm{A}\mathrm{i}\right)}{\left(\mathrm{P}\mathrm{A}\mathrm{i}+\mathrm{U}\mathrm{A}\mathrm{i}\right)}}$	(16)	Class–wise F1-scores calculate the harmonic mean of PA and UA (precision and recall, respectively) when evaluating classifier performance on each class (i = 1–5) and represents Equation (16).
		where F1i is F1-score for class i.
6	Intersection over union (IoU)	$\mathrm{I}\mathrm{o}\mathrm{U}={\displaystyle \frac{\mathrm{F}\mathrm{i}}{2-\mathrm{F}\mathrm{i}}}$	(17)	Intersection over union (IoU) calculates the overlap between true positive and false positive or false negative.
		where F1i is F1-score for class i.

).

3.2. Landsat 5 TM for Period 2000–2008

As

Table 5. Comparative accuracy assessment performance of machine learning (RF, GTB, SVM) and deep learning (DNN, Attention U-Net, SegFormer) models derived from Landsat 5 TM for wetland change analysis in Sangamon River watershed (2000–2008).

Confusion Matrix	Wetland (1)	Forest (2)	Agriculture/Grassland/Barren Land (3)	Urban/Developed (4)	Water (5)	Producer Accuracy	User Accuracy	Overall Accuracy (%)	F1 Score	IoU
Pixel–based
Random Forest (RF) Kappa Coefficient = 0.57, mean IoU = 0.51
Wetland (1)	1358	110	64	25	4	0.87	0.70	67%	0.78	0.64
Forest (2)	256	711	102	85	3	0.61	0.62		0.61	0.44
Agriculture/grassland/barren land (3)	132	147	1000	206	1	0.67	0.62		0.64	0.47
Urban/developed (4)	113	135	421	665	1	0.50	0.66		0.57	0.40
Water (5)	78	30	21	22	247	0.62	0.96		0.75	0.60
Gradient Tree Boosting (GTB) Kappa Coefficient = 0.57, mean IoU = 0.51
Wetland (1)	1312	111	83	40	15	0.84	0.70	67%	0.76	0.61
Forest (2)	241	670	145	97	4	0.58	0.62		0.60	0.43
Agriculture/grassland/barren land (3)	90	120	1036	239	1	0.69	0.60		0.64	0.47
Urban/developed (4)	100	122	427	685	1	0.51	0.63		0.56	0.39
Water (5)	60	22	25	25	266	0.66	0.93		0.77	0.63
Support Vector Machine (SVM) Kappa Coefficient = 0.38, mean IoU = 0.36
Wetland (1)	1121	168	216	49	8	0.71	0.55	53%	0.62	0.45
Forest (2)	336	482	263	105	6	0.40	0.49		0.44	0.28
Agriculture/grassland/barren land (3)	206	90	1092	127	9	0.71	0.48		0.57	0.40
Urban/developed (4)	236	207	635	257	8	0.19	0.46		0.27	0.16
Water (5)	113	30	52	17	253	0.54	0.89		0.67	0.50
Object–based
Random Forest (RF) Kappa Coefficient = 0.85, mean IoU = 0.77
Wetland (1)	2056	10	17	5	1	0.98	0.89	88%	0.93	0.87
Forest (2)	116	703	40	38	3	0.78	0.83		0.80	0.67
Agriculture/grassland/barren land (3)	54	63	1038	33	0	0.87	0.88		0.87	0.77
Urban/developed (4)	48	46	71	753	2	0.82	0.90		0.86	0.75
Water (5)	30	19	7	3	231	0.80	0.97		0.88	0.79
Gradient Tree Boosting (GTB) Kappa Coefficient = 0.88, mean IoU = 0.87
Wetland (1)	2090	14	12	0	0	0.99	0.92	91%	0.95	0.91
Forest (2)	48	784	36	25	47	0.87	0.90		0.88	0.79
Agriculture/grassland/barren land (3)	32	21	1147	24	30	0.93	0.91		0.92	0.85
Urban/developed (4)	17	20	35	828	1	0.92	0.92		0.92	0.85
Water (5)	8	2	4	2	300	0.95	0.99		0.97	0.94
Support Vector Machine (SVM) Kappa Coefficient = 0.84, mean IoU = 0.78
Wetland (1)	1973	23	33	21	7	0.96	0.87	88%	0.91	0.84
Forest (2)	117	723	40	35	2	0.78	0.89		0.83	0.71
Agriculture/grassland/barren land (3)	89	41	961	39	2	0.85	0.87		0.86	0.75
Urban/developed (4)	82	22	55	731	5	0.81	0.88		0.84	0.72
Water (5)	10	4	6	0	265	0.93	0.94		0.93	0.87
Deep Learning (DL)
Dense Neural Networks (DNNs) Kappa Coefficient = 0.50, mean IoU = 0.45
Wetland (1)	1294	86	144	27	9	0.83	0.64	61%	0.72	0.56
Forest (2)	262	662	170	101	5	0.55	0.59		0.57	0.40
Agriculture/grassland/barren land (3)	219	140	965	173	3	0.64	0.53		0.59	0.41
Urban/developed (4)	146	195	487	486	6	0.36	0.60		0.46	0.29
Water (5)	100	34	23	12	251	0.59	0.91		0.72	0.56
Attention U-Net Model Kappa Coefficient = 0.85, mean IoU = 0.60
Wetland (1)	9,447,824	29,364	13,663	4127	3253	0.99	0.97	92%	0.98	0.97
Forest (2)	61,601	335,905	116,801	50,720	4663	0.59	0.50		0.55	0.38
Agriculture/grassland/barren land (3)	150,864	192,679	6,783,963	345,459	5976	0.90	0.92		0.91	0.84
Urban/developed (4)	36,838	91,424	419,245	311,265	3377	0.40	0.43		0.40	0.25
Water (5)	10,592	10,492	9162	2342	51,841	0.61	0.75		0.68	0.60
SegFormer Model Kappa Coefficient = 0.89, mean IoU = 0.70
Wetland (1)	55,675,649	299,617	104,552	16,750	5984	0.99	0.97	94%	0.98	0.97
Forest (2)	62,041	1,586,300	97,874	42,095	10,956	0.88	0.40		0.55	0.38
Agriculture/grassland/barren land (3)	1,196,325	2,014,006	40,583,655	1,477,721	37,778	0.89	0.97		0.93	0.88
Urban/developed (4)	39,537	116,979	869,816	3,460,327	48,133	0.76	0.69		0.72	0.57
Water (5)	100	2023	3532	17,201	299,913	0.93	0.74		0.83	0.70

illustrates that, for pixel-based classification, RF and GTB classifiers showed strong performance for wetlands, agriculture/grassland/barren land, and water when compared with SVM and DNN models with 67% OA, 0.57 KC, and mean IoU of 0.51. RF and GTB showed improved detection of wetland areas, with PA between 0.84–0.87, F1-scores between 0.76–0.78, and IoU values between 0.61 and 0.64. Water was classified with the highest reliability, achieving UA greater than 0.90, F1-scores of 0.75–0.77, and IoU values of 0.60–0.63. However, urban/developed areas remained the most challenging class due to spectral mixing with agriculture and wetlands, resulting in relatively low PA (0.50–0.51), F1–scores (0.56–0.57), and IoU values (0.39–0.40). The pixel–based SVM produced the lowest overall performance (OA = 53%, KC = 0.38, mean IoU = 0.36), with substantial confusion among spectrally similar classes. Although agriculture/grassland/barren land achieved a relatively high PA (0.71), urban/developed areas exhibited the poorest performance (F1 = 0.27; IoU = 0.16).

As Table 5 illustrates, object–based classification substantially improved wetland mapping compared with pixel-based approaches. The object-based RF classifier achieved an OA of 88%, KC of 0.85, and mean IoU of 0.77, producing high wetland classification accuracy (PA = 0.98, UA = 0.89, F1 = 0.93, IoU = 0.87). Agriculture/grassland/barren land (F1 = 0.87; IoU = 0.77), urban/developed (F1 = 0.86; IoU =0.75), and water (UA = 0.97, F1 = 0.88, IoU = 0.79) also exhibited substantial improvements over their pixel–based counterparts. Among all ML classifiers, the object-based GTB model achieved the highest performance, with an OA of 91%, KC of 0.88, and mean IoU of 0.87. This model demonstrated excellent class separability, yielding F1-scores of 0.95 for wetlands, 0.92 for agriculture/grassland/barren land, 0.92 for urban/developed, and 0.97 for water, with corresponding IoU values ranging from 0.79 to 0.94. These results indicate that GTB effectively captured the complex non–linear relationships among spectral, SAR, terrain, and environmental variables while accurately delineating fragmented wetlands. The object-based SVM also showed marked improvement over its pixel–based counterpart, achieving 88% OA, 0.84 KC, and a mean IoU of 0.78. Wetland (F1 = 0.91; IoU = 0.84) and water (F1 = 0.93; IoU = 0.87) exhibited particularly strong performance.

For the DL models, DNN achieved moderate classification accuracy (OA = 61%, KC = 0.50, mean IoU = 0.45), exceeding the pixel–based SVM but remaining inferior to the object-based ML classifiers. Attention U-Net with substantially improved classification performance, achieving 92% OA, 0.85 KC, and a mean IoU of 0.59. It accurately delineated wetlands with PA = 0.99, UA = 0.97, F1 = 0.98, and IoU = 0.97, while agriculture/grassland/barren land also achieved strong performance (F1 = 0.91; IoU = 0.84). However, forest (F1 = 0.55; IoU = 0.38) and urban/developed (F1 = 0.40; IoU = 0.25) remained challenging because of their heterogeneous spatial distribution and spectral similarity with adjacent land-cover classes. SegFormer outperformed all other ML and DL models, achieving 94% OA, 0.89 KC, and a mean IoU of 0.70. SegFormer showed excellent wetland delineation (F1 = 0.98; IoU = 0.97) while also improving classification of agriculture/grassland/barren land (F1 = 0.93; IoU = 0.88), urban/developed (F1 = 0.72; IoU = 0.57), and water (F1 = 0.83; IoU = 0.70). These findings demonstrate that the transformer–based architecture effectively captured both local spatial patterns and broader contextual information, resulting in improved discrimination of fragmented wetland landscapes and heterogeneous land–cover classes when compared with conventional CNN–based and machine learning approaches (Table 5).

3.3. Sentinel–2 for Period 2017–2025

As

Table 6. Comparative accuracy assessment performance of machine learning (RF, GTB, SVM) and deep learning (UNet, Attention U-Net, SegFormer) models derived from Sentinel 1 and 2 for wetland change analysis in Sangamon River watershed (2017–2025).

Confusion Matrix	Wetland (1)	Forest (2)	Agriculture/Grassland/Barren land (3)	Urban/Developed (4)	Water (5)	Producer Accuracy	User Accuracy	Overall Accuracy (%)	F1 Score	IoU
Machine Learning (ML) Models-Pixel-based
Random Forest (RF) Kappa Coefficient = 0.77, mean IoU = 0.72
Wetland (1)	675	93	13	17	4	0.85	0.82	81%	0.83	0.71
Forest (2)	125	415	43	26	1	0.69	0.75		0.72	0.56
Agriculture/grassland/barren land (3)	7	23	591	65	0	0.87	0.85		0.86	0.75
Urban/developed (4)	10	22	46	566	7	0.87	0.82		0.84	0.72
Water (5)	9	0	1	10	189	0.90	0.94		0.92	0.85
Gradient Tree Boosting (GTB) Kappa Coefficient = 0.77, mean IoU = 0.69
Wetland (1)	632	84	12	14	14	0.84	0.80	80%	0.82	0.69
Forest (2)	119	457	32	38	1	0.70	0.77		0.73	0.57
Agriculture/grassland/barren land (3)	6	23	629	73	0	0.86	0.84		0.85	0.74
Urban/developed (4)	10	30	74	494	0	0.81	0.79		0.80	0.67
Water (5)	29	0	1	4	174	0.84	0.92		0.88	0.79
Support Vector Machine (SVM) Kappa Coefficient = 0.54, mean IoU = 0.47
Wetland (1)	504	112	81	36	23	0.66	0.62	61%	0.64	0.47
Forest (2)	150	339	89	63	6	0.52	0.58		0.55	0.38
Agriculture/grassland/barren land (3)	82	71	411	166	1	0.56	0.57		0.56	0.39
Urban/developed (4)	56	64	150	334	4	0.54	0.55		0.54	0.37
Water (5)	18	2	5	8	175	0.84	0.83		0.84	0.72
Machine Learning (ML) Models-Object-based
Random Forest (RF) Kappa Coefficient = 0.82, mean, IoU = 0.74
Wetland (1)	2022	73	31	17	17	0.94	0.85	87%	0.89	0.80
Forest (2)	268	566	32	27	3	0.63	0.81		0.71	0.55
Agriculture/grassland/barren land (3)	26	33	1044	80	0	0.88	0.88		0.88	0.79
Urban/developed (4)	31	24	68	766	10	0.85	0.85		0.85	0.74
Water (5)	27	0	0	8	277	0.88	0.90		0.89	0.80
Gradient Tree Boosting (GTB) Kappa Coefficient = 0.84, mean IoU = 0.75
Wetland (1)	1984	123	24	13	16	0.92	0.88	88%	0.90	0.82
Forest (2)	203	641	26	23	3	0.70	0.77		0.73	0.58
Agriculture/grassland/barren land (3)	17	38	1042	86	0	0.87	0.90		0.88	0.79
Urban/developed (4)	22	25	54	788	11	0.87	0.86		0.86	0.75
Water (5)	25	0	0	5	282	0.89	0.90		0.89	0.80
Support Vector Machine (SVM) Kappa Coefficient = 0.84, mean IoU = 0.78
Wetland (1)	2123	7	18	0	0	0.98	0.97	88%	0. 97	0.94
Forest (2)	55	810	3	0	0	0.93	0.65		0.77	0.63
Agriculture/grassland/barren land (3)	3	412	789	41	1	0.63	0.96		0.76	0.61
Urban/developed (4)	0	1	10	868	0	0.98	0.90		0.94	0.89
Water (5)	0	0	0	48	251	0.83	0.99		0.91	0.84
Deep Learning (DL) Models
U-Net Kappa Coefficient = 0.91, mean IoU = 0.70
Wetland (1)	83,675,186	472,908	632,020	42,665	12,786	0.98	0.99	95%	0.99	0.98
Forest (2)	96	781	18	4	0	0.55	0.74		0.63	0.46
Agriculture/grassland/barren land (3)	11	72	1112	36	1	0.95	0.93		0.94	0.89
Urban/developed (4)	3	4	44	873	1	0.79	0.70		0.74	0.59
Water (5)	1	0	0	12	298	0.78	0.71		0.75	0.60
Attention U-Net Kappa Coefficient = 0.90, mean IoU = 0.73
Wetland (1)	83,812,735	307,318	755,123	48,673	24,118	0.98	0.98	95%	0.99	0.98
Forest (2)	242,025	3,587,953	1,354,160	242,751	29,655	0.66	0.98		0.63	0.46
Agriculture/grassland/barren land (3)	117,948	1,791,321	63,503,964	2,651,095	65,308	0.93	0.99		0.95	0.89
Urban/developed (4)	15,013	126,999	1,090,279	6,648,437	118,485	0.83	0.99		0.75	0.60
Water (5)	2968	2968	11,151	25,276	574,449	0.93	1.00		0.80	0.67
SegFormer Kappa Coefficient = 0.92, mean IoU = 0.73
Wetland (1)	55,592,386	214,527	238,019	34,000	1492	0.99	0.99	96%	0.98	0.99
Forest (2)	4186	1,096,925	634,066	69,680	2074	0.61	0.65		0.63	0.50
Agriculture/grassland/barren land (3)	12,681	308,702	43,242,050	2,062,264	13,992	0.94	0.95		0.95	0.90
Urban/developed (4)	556	71,835	1,016,761	3,460,327	27,979	0.76	0.62		0.68	0.51
Water (5)	0	6380	3988	26,097	275,879	0.88	0.85		0.87	0.77

illustrates, among the pixel-based ML classifiers, RF and GTB showed strong performance, with 81% and 80% OA, respectively, and identical Kappa coefficients (KC = 0.77). Both classifiers showed strong wetland classification performance with PA ranging from 0.84–0.85, UA from 0.80–0.82, F1-scores of 0.82–0.83, and IoU values of 0.69–0.71. Agriculture/grassland/barren land and urban/developed classes also exhibited high classification accuracy, with F1-scores between 0.80–0.86 and IoU values ranging from 0.67–0.75, reflecting improved class separability resulting from the integration of S2 and SAR data. Water achieved the highest classification accuracy among all classes, with UA exceeding 0.92, F1-scores between 0.88–0.92, and IoU values of 0.79–0.85. In contrast, forest remained the most challenging vegetated class, producing comparatively lower F1-scores (0.72–0.73) and IoU values (0.56–0.57) because of spectral similarity with wetlands and other vegetation. The pixel-based SVM produced the lowest performance (61% OA; KC = 0.54) and exhibited substantial confusion among wetlands, forests, agriculture, and urban classes, with IoU values below 0.50 for most classes except water (IoU = 0.72).

Among the object-based ML classifiers, the GTB achieved the highest overall performance (88% OA; KC = 0.84) while maintaining the most balanced classification across all land-cover classes. Wetlands achieved PA = 0.92, UA = 0.88, F1 = 0.90, and IoU = 0.82, whereas agriculture/grassland/barren land, urban/developed, and water achieved F1-scores of 0.88, 0.86, and 0.89, corresponding to IoU values of 0.79, 0.75, and 0.80, respectively. Forest remained comparatively less separable (F1 = 0.73; IoU = 0.58) because of spectral overlap with adjacent wetland vegetation. The object-based RF model also demonstrated strong classification performance (87% OA; KC = 0.82) with a wetland F1-score of 0.89 (IoU = 0.80) and consistently high accuracies for agriculture/grassland/barren land (F1 = 0.88; IoU = 0.79), urban/developed (F1 = 0.85; IoU = 0.74), and water (F1 = 0.89; IoU = 0.80). The object-based SVM achieved 88% OA and KC = 0.84, producing the highest wetland classification among the ML models (F1 = 0.97; IoU = 0.94) together with excellent performance for urban/developed (F1 = 0.94; IoU = 0.89) and water (F1 = 0.91; IoU = 0.84). However, forest (F1 = 0.77; IoU = 0.63) and agriculture/grassland/barren land (F1 = 0.76; IoU = 0.61) remained affected by confusion among vegetated classes (Table 6).

All sentinel-based DL models showed comparable performance and outperformed all sentinel-based ML models with improved wetland classification. SegFormer achieved the highest overall classification performance with 96% OA and a KC of 0.89, outperforming the CNN-based architectures in overall accuracy. The transformer-based model effectively captured both local and global spatial context, resulting in excellent wetland classification with PA = 0.99, UA = 0.99, F1 = 0.98, and IoU = 0.99. Agriculture/grassland/barren land was also classified with the highest accuracy (PA = 0.94, UA = 0.95, F1 = 0.95, IoU = 0.90), while the water class achieved the best performance among all DL models (F1 = 0.87; IoU = 0.77).

Urban/developed and forest remained comparatively more challenging, with F1-scores of 0.68 and 0.63 and IoU values of 0.51 and 0.50, respectively, reflecting spectral similarity among vegetated classes and mixed urban boundaries. U-Net and Attention U-Net also demonstrated excellent segmentation performance, each achieving 95% OA with KC values of 0.91 and 0.90, respectively. Both models accurately delineated fragmented wetlands (F1 = 0.99; IoU = 0.98) and produced strong classification of agriculture/grassland/barren land (F1 = 0.94–0.95; IoU = 0.89). Attention U-Net slightly improved urban (F1 = 0.75; IoU = 0.60) and water (F1 = 0.80; IoU = 0.67) classification relative to the standard U-Net.

Overall, the Sentinel-based models demonstrate a clear progression in classification performance from pixel-based ML to object-based ML and to DL approaches. Incorporating object-level spatial information substantially improved classification accuracy relative to pixel-based methods, while CNN- and transformer-based semantic segmentation models achieved the highest overall performance. Overall, the transformer-based SegFormer provided the most accurate and balanced wetland classification, demonstrating the advantages of global self-attention for mapping fragmented and heterogeneous wetlands.

3.4. Variable Importance (VI) and Relative Importance (RI) in RF Models

The comparative assessment of RF model variable importance (VI) in

Table 7. Variable importance (VI) and relative importance (RI) (%) of variables used in the random forest (RF) model derived from Landsat 5 TM (2000–2008) and Sentinel-2 SR (2017–2025).

	Landsat 5 TM (2000–2008)					Sentinel-2 SR (2017–2025)
		Pixel-Based		Object-Based		Pixel-Based		Object-Based
No.	Variables	RF VI	RI (w.r.t. Slope) (%)	RF VI	RI (w.r.t. Slope) (%)	RF VI	RI (w.r.t. DEM) (%)	RF VI	RI (w.r.t. Slope) (%)
1	AWEI	1998.51	2.50	1392.28	2.50	681.78	2.77	637.84	2.57
2	Blue_spring	2171.77	2.53	1483.51	2.67	563.83	2.80	730.06	2.95
3	Blue_summer	2162.47	2.25	1388.20	2.50	480.97	2.75	709.05	2.86
4	DEM	3898.17	4.88	3572.73	6.44	690.84	5.30	1300.62	5.26
5	DistWater	139.03	0.17	43.12	0.07	677.04	3.58	765.77	3.09
6	EVI_summer	1883.86	2.36	1183.89	2.36	1305.76	2.32	590.51	2.38
7	Green_spring	2192.45	2.74	1537.69	2.74	880.79	2.75	669.41	2.70
8	Green_summer	2117.58	2.65	1350.96	2.65	572.64	2.49	655.60	2.65
9	LSWI_spring	2014.94	2.52	1277.33	2.52	677.02	2.35	605.54	2.44
10	LSWI_summer	1813.86	2.27	1108.57	2.27	613.79	1.97	509.25	2.05
11	MNDWI	2048.35	2.56	1531.15	2.56	579.81	2.80	719.04	2.90
12	NDBI	1777.19	2.22	1091.20	2.22	495.99	2.01	526.67	2.13
13	NDMI_spring	2071.63	2.59	1562.58	2.59	580.54	2.36	611.86	2.47
14	NDMI_summer	1817.91	2.28	1091.03	2.28	483.51	1.96	521.09	2.10
15	NDVI_spring	1960.67	2.45	1357.21	2.45	765.15	3.11	714.26	2.88
16	GLCM_spring_contrast:	2033.63	2.55	1411.11	2.55	743.57	3.02	719.85	2.91
17	GLCM_spring_ent	1999.56	2.50	1398.42	2.50	718.72	2.92	695.94	2.81
18	GLCM_spring_idm	1995.57	2.50	1389.26	2.50	743.08	3.02	716.18	2.89
19	NDVI_summer	1980.28	2.48	1224.30	2.48	651.94	2.65	647.28	2.61
20	GLCM_summer_contrast	1999.84	2.50	1411.39	2.50	1052.80	4.27	893.82	3.61
21	GLCM_summer_ent	1985.15	2.48	1369.51	2.48	702.47	2.85	680.04	2.75
22	GLCM_summer_idm	2019.73	2.53	1418.07	2.53	733.13	2.98	732.11	2.96
23	NIR_spring	2288.32	2.87	1409.68	2.87	664.49	2.70	707.84	2.86
24	NIR_summer	2272.23	2.85	1370.25	2.85	556.39	2.26	595.85	2.41
25	Red_spring	2156.80	2.70	1511.61	2.70	664.83	2.70	690.66	2.79
26	Red_summer	2043.31	2.56	1357.56	2.56	675.30	2.74	683.77	2.76
27	SWIR1_spring	2297.34	2.88	1458.11	2.88	739.67	3.00	774.10	3.13
28	SWIR1_summer	1966.76	2.46	1371.00	2.46	668.18	2.71	636.96	2.57
29	SWIR2_spring	2253.85	2.82	1418.44	2.82	768.42	3.12	765.47	3.09
30	SWIR2_summer	2031.24	2.54	1350.84	2.54	710.72	2.88	626.44	2.53
31	Slope	4470.74	5.60	3362.89	5.60	1363.57	5.54	1473.20	5.95
32	BSI_spring	2024.63	2.53	1312.47	2.36	563.83	2.29	611.16	2.47
33	BSI_summer	1798.20	2.25	1113.93	2.00	480.97	1.95	533.10	2.15
34	MSAVI_spring	2062.33	2.58	1307.49	2.58	687.67	2.79	1245.09	2.81
35	MSAVI_summer	1901.76	2.38	1210.56	2.38	567.14	2.30	1049.83	2.31
36	UI_spring	2092.16	2.62	1278.16	2.62
37	UI_summer	1848.08	2.31	1177.39	2.31
38	NDVI_diff	2135.98	2.67	1568.33	2.67
39	NDMI_diff:	2130.13	2.50	1562.58	2.50
40	VH					0	0.00	0	0.00
41	VV					0.15	0.00	0.15	0.00
42	VV_VH					0	0.00	0	0.00
43	VV_contrast					0.98	0.00	0.98	0.00
44	VV_ent					1.33	0.00	1.33	0.00
45	VV_idm					2.09	0.00	2.09	0.00
Out of bag error estimate:0.34				0.11		0.19		0.13

between the L5 (2000–2008) and S2 (2017–2025) datasets indicate OBIA significantly improved model stability, spatial consistency, and predictor utilization for wetland classification in heterogeneous watershed environments by reducing OOB error from 0.34 to 0.11 in the Landsat-based model and 0.19 to 0.13 in the Sentinel-based model, respectively.

3.4.1. Landsat 5 TM (2000–2008)

Table 7 and Figure 5 indicate that for both pixel-based and object-based Landsat-based RF models, slope showed the highest RI (~5.60%), followed by DEM (~4.88%). With RI (~2.4–2.9%), SWIR1, SWIR2, NIR, and NDVI being important spectral predictors because of their sensitivity to surface moisture conditions, inundation, and land-surface detection in wetlands. GLCM are also significant in distinguishing fragmented wetland systems of spatial heterogeneity, particularly in the OBIA model. AWEI, LSWI, MNDWI, NDMI, MSAVI, and BSI showed moderate but stable VI, whereas DistWater had low VI, showing that terrain and spectral characteristics provided stronger discriminatory capability than simple proximity-based hydrologic variables.

3.4.2. Sentinel-2 SR (2017–2025)

Table 7 and Figure 5 indicate that S2 imagery had an improved and balanced RI of spectral and GLCM variables. With RI ~5%, slope and DEM were the most influential variables. A comparatively higher VI was observed for GLCM summer contrast and entropy showing improved sensitivity to fine-scale spatial variability and fragmented landscape patterns. SWIR1, SWIR2, NIR, and NDVI-derived variables showed high VI, particularly for OBIA, with improved classification of wetlands, forests, agriculture, and urban categories. MNDWI, NDMI, LSWI, and AWEI signify improved VI delineation of inundated and moisture-rich environments compared with SAR variables and SAR texture metrics, which showed very low VI.

3.5. Change Detection and Wetland Loss Calculations (2000–2008–2017–2025)

Classified maps were generated for 2000, 2008, 2017, and 2025 (Figure 6) and wetland change analysis (Figure 7) was subsequently conducted in ArcGIS Pro 3.6.0. We calculated the wetland gain, wetland loss to all other classes, and stable wetland area for 2000 and 2017 in km² as shown in

Table 8. Change detection and wetland loss calculations in the Sangamon River watershed, Illinois.

Class Name	Area (km2) in 2000	% of Land Use	Area (km2) in 2008	% of Land Use	Area (km2) in 2017	% of Land Use	Area (km2) in 2025	% of Land Use	Change in Area Between 2000–2025	% of Land Use Change Based on the Year 2000	Increase/Decrease
Wetland (1)	1569.2436	11.27%	1551.9123	11.15%	771.9249	5.54%	708.5648	5.09%	860.6788	54.85%	Decrease
Forest (2)	1494.0486	10.74%	1499.7681	10.78%	729.5228	5.24%	799.1661	5.74%	694.8825	46.51%	Decrease
Agriculture/grassland/barren land (3)	9575.7489	68.81%	9563.5845	68.75%	10,246.1084	73.66%	10,396.3652	74.74	820.6163	8.56%	Increase
Urban/developed (4)	1096.3863	7.87%	1141.8759	8.20%	2051.9583	14.75%	1895.9107	13.62	799.5244	72.92%	Increase
Water (5)	179.700	1.29%	157.9896	1.13%	113.3917	0.81%	113.5674	0.81	66.1326	36.80%	Decrease
		Total watershed area: 13,915.13

and

Table 9. Wetland loss calculations to other land uses in the Sangamon River watershed, Illinois.

Class Name	Area (km2) 2000 –2008	% of Wetland Area Transition	Area (km2) 2017–2025	% of Wetland Area Transition
Stable Wetland between 2000–2008	1363.4874	86.88%
Stable Wetland between 2017–2025			460. 9723	59.71%
Wetland to Forest (2)	50.4567	3.18%	88.794	11.54%
Wetland to Agriculture/Grassland/Barren Land (3)	119.9862	7.57%	176.7847	22.91%
Wetland to Urban (4)	18.3033	1.25%	36.2248	4.70%
Wetland to Water (5)	17.01	1.07%	9.1491	1.18%
Total	1569.2436	100%	771.9249	100%
Total watershed area: 13,915.13

. The result showed major wetland loss and fragmentation in Sangamon River watershed between 2000 and 2025 as approximately 54.85% wetlands (860.68 km²) area was lost to other land uses. Other land uses that lost their spatial extent includes forest with 46.51% (694.88 km²), and water bodies with 36.80% (66.13 km²), whereas agriculture/grassland/barren covered area witnessed an increase of 8.56% (820.62 km²), and urban/developed land with 72.92% (799.52 km²) as shown in Table 8 and Table 9 based on their existing land use in year 2000.

Table 9 shows that a larger percentage, 86.88%, of stable wetlands existed between 2000–2008 than the 59.71% between the 2017–2025 period, indicating that major changes in wetland occurred during the latter period. Between 2000–2008, 3.18% of wetland converted to forest, 7.57% wetland converted to agriculture/grassland/barren land, 1.25% of wetland converted to urban, and 1.07% of wetland was lost to other water bodies. Meanwhile, for the periods between 2017–2025, 11.54% of wetland converted to forest, 22.91% to agriculture/grassland/barren land, 4.70% to urban, and 1.18% to other water bodies. The wetland loss percentage was calculated based on the total wetland area of the base year i.e., 2000 and 2017, respectively.

Overall, the results indicate that the Sangamon River watershed witnessed substantial environmental and landscape changes in the past 25 years in the form of a significant 54.85% wetland loss in addition to a proportional forest cover loss. While agricultural practices intensified continuously throughout this period, as the most dominant land-cover class, an alarming rate of urban expansion caused significant anthropogenic pressure on watershed ecosystems and hydrologically sensitive landscapes.

4. Discussion

This study represents one of the first comprehensive multi-temporal wetland and LULC change analyses conducted for the Sangamon River watershed to identify spatial patterns of wetland loss, agricultural expansion, forest cover loss, urban, climate and anthropogenic pressure in the watershed and aims to support evidence-based decision-making for wetland conservation and restoration, climate resilience planning, agricultural best management practices and sustainable land-use policy development. Most existing studies on wetland change analysis have relied on a limited set of classification methods that incorporates a comparative framework integrating ML, DL, pixel-based, and OBIA classification. This study addresses these gaps by comprehensively evaluating and comparing core and advanced DL classifiers and multiple ML across both pixel-based and object-based approaches within a single comparative framework to have an accurate wetland classification and change detection in a complex agricultural watershed environment. The results demonstrate that while both ML and DL approaches can effectively classify wetlands, object-based ML substantially improved Landsat-based classifications, whereas transformer-based SegFormer consistently achieved the highest classification performance for both Landsat- and Sentinel-based datasets, highlighting the advantages of integrating spatial context and global feature learning for wetland mapping in complex agricultural landscapes.

For this study we strategically divide the analysis period into two temporal phases (2000–2008 and 2017–2025) based on the availability of the satellite datasets. For 2000 and 2008 we incorporated L5 imagery, as it provides one of the longest continuous Earth observation archives for long-term analysis and S2 imagery is not available for years prior to 2017. For 2017 and 2025 we used S2 integrated with S1 SAR data to exploit higher spatial resolution and complementary radar information. To preserve the native characteristics of each sensor, the original spatial resolution was retained throughout the classification process, and wetland change was quantified separately within each temporal phase by converting classified pixels to area in ArcGIS Pro 3.6.0, thereby avoiding uncertainties associated with cross-sensor resampling. However, as shown in Table 5 and Table 6, the S2 data enabled the improved delineation of fragmented wetlands, narrow riparian corridors, small hydrologically connected depressions, and land-cover transitions when compared with L5 imagery. L5 imagery tends to overlap land-cover classes, leading to underestimation of smaller wetland patches and boundary inaccuracies, while S2 imagery reduces mixed-pixel effects. Consequently, many small or narrow wetlands embedded within agricultural and riparian landscapes may have remained undetected or poorly classified in Landsat-based analyses (Figure 6).

The integration of Sentinel-1 SAR backscatter, spectral indices, and GLCM texture measures enhanced the delineation of wetlands by providing complementary information beyond optical imagery alone. SAR variables improved the identification of saturated soils, flooded vegetation, and hydrologically connected wetland environments, particularly in areas where spectral confusion existed among wetlands, agriculture, and riparian vegetation. Likewise, GLCM texture metrics (contrast, entropy, and IDM) improved the characterization of vegetation structure, surface roughness, wetland edge conditions, and heterogeneous riparian gradients that are not fully represented by spectral information alone. RF VI analysis indicated that terrain variables (DEM and slope) were the most influential predictors across all RF classifications, while spectral indices including MNDWI, NDMI, LSWI, AWEI, and MSAVI consistently showed strong contributions to Sentinel-based wetland mapping by improving the identification of inundated and moisture-rich environments. In contrast, SAR backscatter and SAR-derived texture variables exhibited comparatively lower variable importance. One plausible explanation is that the dominance of row-crop agriculture (approximately 75% of the watershed) masked wetland signals and reduced the relative contribution of SAR variables in distinguishing wetlands from surrounding agricultural land. Nevertheless, SAR information remains valuable as a complementary predictor, particularly under cloud-contaminated conditions and for characterizing hydrological variability, and this hypothesis warrants further investigation in other intensively managed agricultural watersheds.

Between 2000–2025, the Sangamon River watershed experienced significant hydrologic alteration and ecological vulnerability with a 54.85% wetland decline along with losses of 46.51% forest cover and 36.80% water bodies from the total watershed area relative to their respective 2000 extents. Table 8 and Table 9 shows that wetlands remain comparatively stable during 2000–2008, with a significant shift of 50.25% between 2008 and 2017, while the forest cover has also proportionally declined. During the 25-year period, urban/developed land expanded by 72.92%, while agriculture/grassland/barren land continued to increase, collectively intensifying watershed modification. This transition is directly associated with hydrologic and habitat connectivity, loss of biodiversity, sediment transport, low nutrient retention, riparian instability, increased erosion, and declining overall ecosystem resilience in the watershed.

The observed wetland decline is consistent with the long-term landscape transformation reported across central Illinois. Although this study did not explicitly quantify the individual drivers of wetland conversion, the spatial patterns are consistent with previous studies identifying agricultural intensification, particularly the expansion of corn and soybean production, together with extensive agricultural drainage practices (subsurface tile drainage), as major contributors to hydrologic alteration and wetland degradation in Illinois [12,13]. Tile drainage systems rapidly remove excess soil moisture, shorten hydroperiods, and convert seasonally inundated wetlands into productive agricultural land. Moreover, because approximately 75% of the watershed is dominated by row-crop agriculture, many small and fragmented wetlands are increasingly isolated and subjected to hydrological modification [64,65,66]. Continued urban expansion further compounds these impacts through increased impervious surfaces and altered drainage networks. Collectively, these processes reduce wetland extent and connectivity while degrading critical ecosystem services, including groundwater recharge, flood attenuation, water-quality, carbon sequestration, and habitat provision for aquatic and migratory species, underscoring the dominant influence of anthropogenic land-use change on long-term wetland dynamics within the watershed.

On the computational analysis side, this study classified and compared a set of 18 ML and DL models. Each model was independently calibrated based on its algorithm-specific parameters, input predictor variables, sampling strategy, and spatial classification settings, to achieve an optimized output. For ML-based models, these calibrations include numbers of trees adjustment (500–550), learning rates, bag fractions (0.7), kernel functions, and regularization parameters. For DL-based models, we used standard architecture-specific optimization including normalization, dropout regularization, batch normalization, learning-rate scheduling, class weighting, and patch dimensions. Each model was rerun with multiple adjustments until an improved classification accuracy is achieved. All DL models were trained for 100 epochs using GPU acceleration in Google Colab, allowing stable convergence and improved optimization. Hyperparameters were iteratively refined until satisfactory classification performance was achieved.

Computational complexity varied considerably among the evaluated classifiers. Object-based RF and GTB models required comparatively lower computational resources and shorter training times while achieving competitive classification performance, particularly for Landsat imagery. In contrast, the semantic segmentation models (U-Net, Attention U-Net, and especially the transformer-based SegFormer) required substantially greater GPU memory, longer training times, and larger storage capacity because of patch-based processing, hierarchical feature extraction, and multi-stage optimization. Within GEE, additional computational challenges included memory limitations and timeouts associated with SNIC segmentation, reduceConnectedComponents() operations, and object-level feature extraction from large multi-source predictor stacks integrating S2, S1 SAR, terrain variables, spectral indices, and GLCM metrics. These challenges were mitigated by dividing the workflow into separate training and classification tasks and importing intermediate training assets from GEE assets. Exporting large TFRecord datasets and DL inputs also required substantial cloud storage and careful verification of patch generation to ensure data consistency. Although DL models incurred substantially higher computational costs than conventional ML approaches, SegFormer consistently achieved the highest classification performance, whereas object-based RF and GTB provided an effective balance between computational efficiency and classification accuracy, making it an attractive alternative for operational watershed-scale wetland monitoring where computational resources may be limited.

As established by previous studies, object-based models consistently produced improved classification robustness and reduced uncertainty through segmentation-based spatial aggregation. The object-based GTB classifier produced the strongest overall performance among the ML models, followed closely by RF and SVM, demonstrating the effectiveness of segmentation-based feature aggregation for complex agricultural landscapes. Among the DL models, SegFormer consistently achieved the highest classification performance for both Landsat- and Sentinel-based datasets, outperforming DNN, U-Net, Attention U-Net, and all ML classifiers by effectively capturing both local spatial features and long-range contextual information. The improved Sentinel-based classifications further highlight the value of integrating multispectral imagery, SAR, terrain variables, spectral indices, and texture information within a unified GeoAI framework for accurate wetland and LULC mapping. However, the improved performance of transformer-based semantic segmentation was accompanied by substantially higher computational requirements. Training U-Net, Attention U-Net, and SegFormer for 100 epochs required GPU acceleration, patch-based processing, larger storage capacity, and considerably longer training times than conventional ML models. In contrast, object-based GTB and RF provided competitive classification performance with substantially lower computational cost and shorter processing times, particularly for Landsat imagery. These findings suggest that classifier selection should consider both mapping objectives and available computational resources. For operational watershed-scale wetland monitoring where computational efficiency is a priority, object-based RF and GTB offer an effective balance between accuracy and processing efficiency, whereas SegFormer is more suitable for applications requiring maximum classification accuracy and improved delineation of fragmented wetlands and heterogeneous land-cover boundaries.

These findings demonstrate that integrating multi-source Earth observation data with advanced GeoAI-based ML and DL frameworks substantially improves the delineation of fragmented wetlands and heterogeneous land-cover transitions by exploiting both local spatial detail and broader contextual information. While the proposed framework demonstrates strong potential for watershed-scale wetland mapping in intensively managed agricultural landscapes, its transferability to other geographic regions should be evaluated through future multi-region validation. Continued advances in cloud-native geospatial computing, GPU accessibility, distributed processing, and scalable AI frameworks are expected to further improve the operational feasibility, reproducibility, and scalability of GeoAI-based wetland monitoring, restoration planning, and climate-resilient watershed management.

5. Conclusions

The wetland change analysis revealed significant 54.85% wetland and 46.51% forest losses in addition to continued expansion of agricultural and urban land uses during the 25-year study period, making this watershed more ecologically vulnerable for water quality, flood regulation, biodiversity conservation, and climate resilience. From a planning, conservation, and environmental management perspective, this study provides one of the first comprehensive evaluations of wetland dynamics in the Sangamon River watershed using an integrated and comparative GeoAI framework.

This study demonstrates that integrating multi-source Earth observation datasets, including Landsat 5 TM, S2 optical imagery, SAR, terrain variables, spectral indices, and texture measures within an integrated and comparative GeoAI framework substantially improves wetland delineation and LULC classification in complex agricultural watersheds. The comparative evaluation of ML (RF, GTB, and SVM) and DL (DNN, U-Net, Attention U-Net, and SegFormer) models showed that classification performance was strongly influenced by both spatial representation and learning architecture. Object-based ML methods consistently outperformed their pixel-based counterparts, while the transformer-based SegFormer achieved the highest overall classification performance for both Landsat- and Sentinel-based datasets by effectively exploiting local spatial detail together with broader contextual information. Object-based RF and GTB provided an effective balance between classification accuracy and computational efficiency, making it well suited for operational watershed-scale wetland monitoring where computational resources may be limited. These findings highlight that classifier selection should consider both mapping objectives and available computational infrastructure. Future studies should validate its applicability across diverse geographic regions, wetland types, and environmental conditions to support scalable, evidence-based watershed management.