Comparative machine learning classification of a heterogeneous Ramsar wetland using Sentinel-1 and Sentinel-2 imagery with quantity and allocation disagreement diagnostics

Strategic and active management of protected and heterogeneous wetland landscapes requires timely, accurate, and spatially explicit land-use and land-cover (LULC) information. This study presents a comprehensive landscape-level characterisation of LULC across South Africa’s first proclaimed World Heritage and Ramsar site, the iSimangaliso Wetland Park (IWP), and its surrounding areas. Multi-season Sentinel-2 imagery acquired in 2025 (autumn: March‐May; spring: September‐November) was integrated with Sentinel-1 backscatter, topographic variables, and ancillary datasets to enhance class separability in this complex environment. A total of 6,588 training samples and 2,358 validation samples were used to classify eight LULC classes, including open water, wetland, planted forest, natural forest, cropland, grassland, barren, and built-up areas, using four machine learning algorithms: Random Forest (RF), Support Vector Machine (SVM), Classification and Regression Trees (CART), and K-Nearest Neighbours (KNN). Classification performance was evaluated using confusion‐matrix metrics (overall, producer’s, and user’s accuracy), complemented by quantity disagreement (QD) and allocation disagreement (AD) to distinguish between proportional and spatial allocation errors. RF achieved the highest overall accuracy (0.91), followed by SVM (0.88), CART (0.87), and KNN (0.84). Class‐based analysis showed that spectrally distinct classes, such as open water, achieved very high accuracy (<0.90), whereas heterogeneous classes, including wetlands and grassland, exhibited lower accuracy due to spectral similarity. Misclassification errors were spatially clustered, primarily along wetland‐upland boundaries. Uncertainty metrics further indicated low classification confidence in areas with mixed vegetation and hydrologically dynamic conditions. The methodological novelty of this study lies in integrating full spatial coverage classification with disagreement-based accuracy analysis and spatially explicit uncertainty scrutiny, providing a robust evaluation and baseline for long-term LULC monitoring and conservation planning in complex wetland systems.

The iSimangaliso Wetland Park (IWP) in KwaZulu-Natal, South Africa is a renowned UNESCO World Heritage Site, boasting distinct ecological diversity that encompasses complex wetland landscapes, exceptional biodiversity, and the scenic beauty of coastal dunes, forests, lakes, and estuaries (Kuseni et al., 2025). These features make the park an attraction valued by tourists, visitors, and socio-economic investors alike. It represents one of the most biologically diverse and ecologically important protected areas in South Africa and was declared the country’s first UNESCO World Heritage Site in 1999 (South African History Online, 2018). Ecologically, iSimangaliso offers valuable ecosystem services, including flood control, water purification, groundwater recharge, shoreline stabilisation, and carbon sequestration, thereby facilitating climate adaptation for surrounding communities. It supports a great number of plant and animal species, including endangered and endemic taxa, and is internationally important for migratory water-birds, nesting marine turtles, and estuarine-dependent fish species (UNESCO World Heritage Centre), underscoring its role in both regional and global conservation networks.

Economically, the park is a huge cornerstone of nature-based tourism in the province of KwaZulu-Natal, generating employment (World Bank, 2009), inspiring small and medium enterprises, lifting local livelihoods in a historically disadvantaged rural area, while also promoting South Africa’s international conservation and tourism profile (Department of Forestry, Fisheries and the Environment, 2023). In a social and cultural context, iSimangaliso combines conservation with community development and land restitution, encouraging collective stewardship models that strengthen the link between resource management, cultural heritage, and sustainable development (IWP, 2022). Collectively, these characteristics make the park a nationally distinguished landscape for biodiversity conservation, climate adaptation, and participatory socio-economic development. In addition, the area stands out as a globally significant natural asset, whose protection yields long-term gains far beyond its boundaries (UNESCO World Heritage Centre).

The region is sculpted by the combined effects of geological, hydrological, climatic, and biological processes that sustain one of southern Africa’s most biodiverse landscapes. It is believed that Quaternary sea-level fluctuations shaped its parallel-dune systems and sandy plains, which are responsible for soil development, river networks, and landscape heterogeneity (Botha and Porat, 2007). The heterogeneity is further maintained by fire and occasional flooding that disturb regimes that support the savanna–grassland–forest mosaics by regulating forest encroachment and redistributing nutrients (Bond et al., 2005). Hydrological processes characterized by shallow aquifers, seasonal rainfall, and surface–groundwater interactions replenish extensive wetlands, lakes, and peat systems, controlling nutrient cycling, productivity, and biodiversity (Taylor et al., 2006). The convergence of tropical and subtropical biomes in this region is the basis for distinct ecotonal settings that boost species turnover and facilitate high levels of plant endemism, particularly in ecosystems developed on infertile sandy soils (Wyk and Smith, 2001). Climate variability and anthropogenic pressures such as land-use change and water abstraction, impact ecosystem dynamics, altering natural feedbacks and compromise long-term ecological stability (Grundling et al., 2013).

Unfortunately, many studies have reported that the park has undergone considerable landscape changes and significant wetland losses over the past few years (Adeeyo et al., 2022). As anthropogenic pressures, land-cover change, and climate variability amplify their influence on the IWP (Orimoloye et al., 2020), efficient land-use and land-cover (LULC) monitoring becomes increasingly important, especially for understanding the characteristic balance between socio-economic and environmental elements and for equitable resource allocation. For these reasons, the park attracts researchers from a variety of disciplines, including land-use and land-cover classification (Zeng et al., 2024).

Reliable LULC maps are essential for understanding spatial patterns, monitoring landscape dynamics, and assessing the impacts of anthropogenic and climatic pressures on ecosystems (Bassa et al., 2016; Nedd et al., 2021; Carter et al., 2024). These datasets support evidence-based land-use planning, identify degradation hotspots, inform conservation interventions, and underpin modelling of land–atmosphere interactions such as hydrological processes and climate variability, particularly within ecologically sensitive wetland systems (Nedd et al., 2021; Bairwa et al., 2025). However, accurately mapping LULC in wetland environments remains challenging because of high spatial heterogeneity, seasonal variability, and spectral confusion among land-cover classes. This challenge is particularly evident in the iSimangaliso Wetland Park (IWP), where wetlands, open water, coastal forests, grasslands, and anthropogenic features occur in close proximity. Recent assessments further indicate that the lack of comprehensive and up-to-date spatial datasets continues to constrain effective land-use planning and management within the IWP (Department of Agriculture, Land Reform and Rural Development, DALRRD, 2024). Traditional environmental classification approaches, which often rely on Boolean thresholds, may further oversimplify and distort transitional landscapes, thereby limiting the accurate representation of ecologically dynamic environments.

Recent advances in remote sensing and Machine Learning (ML) have substantially improved the generation of accurate and detailed LULC maps in complex landscapes. The integration of multi-source satellite imagery, including Sentinel-2, Landsat, and high-resolution datasets such as WorldView-2, with ancillary datasets such as digital elevation models, Height Above Nearest Drainage (HAND), and global settlement layers has enabled more effective large-scale and temporally consistent LULC assessments within cloud-based platforms such as Google Earth Engine (GEE) (Macarringue et al., 2022). In addition, spectral indices including NDVI, OSAVI, NDMI, NDWI, AWEI_nsh, NDBI, and BSI have improved thematic discrimination by capturing vegetation vigour, moisture gradients, and built-up characteristics (Zhang et al., 2022; Nagaraj and Kumar, 2024; Bogale et al., 2025). Recent studies have further shown that multi-source data fusion significantly enhances classification performance in wetlands and flood-prone environments (Sarker and Jahan, 2026; Islam et al., 2024; Kauser et al., 2026).

Among ML approaches, Random Forests (RF), Support Vector Machines (SVM), Classification and Regression Trees (CART), and K-Nearest Neighbours (KNN) are widely used for LULC classification because they can handle complex, high-dimensional remote sensing data. RF, an ensemble supervised classifier that combines decision-tree and aggregation approaches (Breiman, 2001), is recognised for its high classification accuracy, robustness to noise, and capacity to model complex nonlinear land-cover transitions across heterogeneous landscapes (Jin et al., 2018; Shafizadeh-Moghadam et al., 2021; Amini et al., 2022; Adugna et al., 2022). The algorithm generates multiple decision trees from randomly selected subsets of training data and variables, with final classifications determined through majority voting (Predić et al., 2018). Similarly, SVM has demonstrated strong performance in separating spectrally similar classes and handling limited training samples (Valkenborg et al., 2023; Pavithra, 2025), while CART has proven effective for object identification and automated land-cover classification within the GEE environment (Shelestov et al., 2017; Budreski et al., 2007; Pantaleoni et al., 2009). In some cases, CART has even outperformed RF and SVM in classification accuracy (Pan et al., 2025). These classifiers have consistently demonstrated reliable performance in heterogeneous and multisource remote sensing applications (Mutale et al., 2024). Furthermore, integrating multisource datasets such as optical and SAR imagery has been shown to significantly improve classification accuracy using traditional ML methods such as RF and SVM, often achieving performance comparable to more computationally intensive approaches (Liu et al., 2019).

Although advanced techniques such as Extreme Gradient Boosting and deep learning have gained increasing attention in remote sensing, they generally require larger training datasets, extensive hyperparameter tuning, and higher computational resources, which may limit their practicality within cloud-based environments such as GEE. Deep learning approaches may also present challenges related to model complexity, interpretability, and high data requirements in remote sensing applications (Ball et al., 2017). Consequently, RF, SVM, CART, and KNN provide an effective balance between classification accuracy, interpretability, computational efficiency, and scalability for large-scale LULC mapping in complex environments such as the IWP.

However, despite these advances, some critical gaps remain in LULC research on wetland environments. First, the literature lacks a comparative evaluation of multiple ML classification algorithms across heterogeneous wetland environments. Large-scale reviews and meta-analyses revealed that existing studies predominantly rely on single classifiers or are restricted to pairwise comparisons, limiting understanding of algorithm reliability, flexibility, and contextual sensitivity in diverse wetland environments (Zerrouk et al., 2025; Effah et al., 2025; Yuan et al., 2025). While individual studies have applied specific classifiers in the IWP, such as RF (Slagter et al., 2020), RF and oblique RF (Bassa et al., 2016), and Maximum Likelihood Classification (Orimoloye et al., 2019), a few have systematically compared multiple ML approaches under consistent input conditions. This narrows the chances of identifying optimal models for complex wetland landscapes, where similar spectral signatures and seasonal dynamics can significantly impact classifier performanceSecond, disagreement metrics, such as Quantity Disagreement (QD) and Allocation Disagreement (AD), are rarely applied in wetland classification studies. Most existing research continues to rely heavily on traditional confusion-matrix-derived metrics, which do not adequately distinguish between errors arising from incorrect class proportions and those arising from spatial misallocation of pixels. These metrics, including overall accuracy, producer’s accuracy, and user’s accuracy, provide valuable summaries of classification performance but fail to clearly quantify the spatial distribution of errors. As a result, they may hide important misallocations between classes that share similar spectral signatures, an issue frequently documented in wetland systems, where vegetation and water bodies repeatedly overlap in spectral signatures (Slagter et al., 2020). Pontius and Millones (2011) presented these metrics as a robust alternative to other error-matrix-based measures by breaking down total disagreement into quantity and spatial components. Subsequently, Pickard et al. (2017) employed these metrics to standardise multiple land-use models, revealing trade-offs between errors in class proportions and spatial allocation. In the same way, Guizani et al. (2025) combined QD and AD with overall accuracy and F1-score to establish a more diagnostic evaluation of classification performance in urban environments.

Although previous studies have highlighted the challenges of LULC classification in the IWP, particularly spectral confusion, mixed pixels, and reduced accuracy in densely vegetated areas (Slagter et al., 2020; Bassa et al., 2016; Pal and Mather, 2005; Deventer et al., 2017), there remains limited evidence of integrated approaches that combine satellite imagery data, ancillary variables, and advanced validation frameworks at a full landscape scale. Moreover, these studies have focused on specific sub-regions rather than providing a comprehensive spatial characterisation of the entire park and its surroundings. Therefore, this study aims to classify and quantify the spatial distribution of LULC types across the IWP and its surrounding landscape, establishing a strong baseline for long-term ecological monitoring and conservation planning. The study integrates Sentinel-2 spectral bands and indices with Sentinel-1 polarisation data and ancillary variables (HAND, slope, elevation, global human settlement data, and distance to roads) within a multi-season framework (autumn: March–May 2025; spring: September–November 2025). Multiple ML algorithms, RF, SVM, CART, and KNN, are evaluated comparatively to assess their performance in a highly heterogeneous wetland system. Importantly, the study incorporates quantity disagreement (QD) and allocation disagreement (AD) as supplementary validation metrics, thereby advancing beyond traditional accuracy assessments and providing a more comprehensive evaluation of classification performance. This integrated methodological framework addresses key gaps in wetland LULC research and enhances mapping reliability in one of Africa’s most ecologically complex protected landscapes.

Study area

The IWP spans approximately 320,000 hectares and extends along 220 km of coastline from north-eastern KwaZulu-Natal, South Africa, to the Mozambique border. It was declared a Wetland of International Importance under the Ramsar Convention in 1986 and later became South Africa’s first world heritage site in 1999 (South African History Online, 2018). The park encompasses several key Ramsar sites, including Lake Sibaya, the Kosi Bay Lake System, the St Lucia Lake System, and the Turtle Beaches and Coral Reefs (DALRRD, 2024). These areas hold significant ecological, economic, cultural, and recreational value. The unique characteristics of the park have fostered the development of various socio-economic and cultural activities, some of which have threatened the ecological value of this vital wetland system.

IWP is characterised by a subtropical climate with warm, humid summers and mild, drier winters, with mean annual rainfall ranging from approximately 800 to 1,200 mm, mainly experienced between October and March. Mean annual temperatures range between 17 °C and 27 °C, with coastal effects controlling extremes (Whitfield and Baliwe, 2013). Hydroecologically, the system is highly dynamic, comprising interconnected lakes, estuaries, wetlands, and coastal aquifers that are heavily dependent on and influenced by intermittent freshwater inflows, groundwater seepage, and marine exchange (UNESCO World Heritage Centre, 2024; iSimangaliso Wetland Park Authority, 2022).

Anthropogenic activities, combined with climate change, have been a central factor in the ongoing shifts in land use and land cover (Orimoloye et al., 2019). These land conversions have resulted in the expansion of planted forests, tourism infrastructure, and human settlements throughout the region. Such developments pose direct threats to biodiversity, compromise the integrity of wetlands, and disrupt the hydrological functioning of aquatic ecosystems (Vivier and Cyrus, 2009; Smithers et al., 2017; Dlamini et al., 2021). This situation highlights the urgent need for accurate land-cover data to inform effective planning and conservation strategies that preserve the integrity of this globally significant world heritage site. Figure 1 illustrates the study area and its setting.

Data collection and pre-processing

Multi-season Sentinel-2 imagery obtained between 1 March and 30 May 2025 (41 scenes) and between 1 September and 2 November 2025 (24 scenes) was used to capture wet- and dry-season conditions within the study area. Sentinel-2 surface reflectance imagery was acquired from the Copernicus Sentinel-2 archive accessible in Google Earth Engine and used to derive spectral predictors, including the Blue, Green, Red, Near-Infrared (NIR), Shortwave Infrared 1 (SWIR1), and Shortwave Infrared 2 (SWIR2) bands, Only images with cloud cover below 5% were included to ensure quality. The bands of interest (B2–B12) were processed for scaling and further renamed for analysis. Median compositing was performed for each season to reduce speckle noise. Subsequently, the median early- and late-season images were stacked to create a double-season composite... Vegetation, moisture, water, built-up, and bare-soil indices were subsequently calculated, including the Normalized Difference Vegetation Index (NDVI) (Mbatha and Xulu, 2018), Optimized Soil-Adjusted Vegetation Index (OSAVI) (Rondeaux et al., 1996), Normalized Difference Moisture Index (NDMI) (Taloor et al., 2021), Normalized Difference Water Index (NDWI) (Jovanovic et al., 2014), Automated Water Extraction Index (AWEI nsh) (Jiang et al., 2014), Normalized Difference Built-up Index (NDBI) (Mudede et al., 2020), and Bare Soil Index (BSI) (Salas and Kumaran, 2023). The equations used to derive these indices were adopted from their respective source publications.

To improve the discrimination of wetland and non-wetland land-cover classes, Sentinel-1 Synthetic Aperture Radar (SAR) Ground Range Detected imagery obtained from the Copernicus Sentinel-1 archive was incorporated through the VH polarization band and the normalized polarization difference index (VV − VH)/(VV + VH), following Yue et al. (2025).

A detailed pre-processing workflow was applied to Sentinel-1 data, starting with the conversion from decibels to natural units, followed by the removal of thermal noise to eliminate sensor-based artefacts. Radiometric terrain corrections were applied, using incidence angle adjustments to mitigate topographic effects. A 30-m median filter was used to remove speckle noise, followed by conversion back to decibels (dB) and clipping to the region of interest. Seasonal compositing was performed to create early and late SAR mosaics. This study did not prepare a classification of separate monthly images. Instead, median composites for autumn (March–May) and spring (September–November) were produced to provide a stable basis for a land cover map for 2025. Seasonal compositing masks residual cloudiness, improving classification representations by reducing intra-seasonal noise and underscoring persistent spectral signatures, mainly for vegetation and built-up land covers (Qiu et al., 2023). The Vertical-transmit/Vertical-receive/Vertical transmit, Horizontal receive (VV/VH) index was computed to facilitate distinction among vegetation, built-up areas and wetlands.

Apart from indices, some auxiliary datasets were also integrated as extra predictor variables to enhance classification accuracy. Elevation and slope were derived from the 30 m Shuttle Radar Topography Mission (SRTM) digital elevation model (USGS/SRTMGL1_003) (Farr and Kobrick, 2001). Hydrological conditions were represented using the Global 30 m Height Above Nearest Drainage (HAND) dataset (users/gena/GlobalHAND/30m/hand-1000) (Donchyts et al., 2016). Anthropogenic influences were characterised using the Global Human Settlement Layer Built-up Surface product (JRC/GHSL/P2023A/GHS_BUILT_S) (Pesaresi and Politis, 2023) and a distance-to-road variable derived from the OpenStreetMap road network obtained through the Humanitarian OpenStreetMap Team (HOTOSM). All datasets were processed and analysed in the Google Earth Engine cloud computing environment.

The last steps prior to classification involved resampling the continuous and categorical layers using bilinear interpolation to a 10 m spatial resolution corresponding to Sentinel-2. To ensure consistency. Additionally, all bands were normalised (0–1) in preparation for machine learning algorithms.

Feature importance using SHAP

To improve model interpretability and reduce potential bias arising from multicollinearity among predictor variables, feature importance was further evaluated using SHAP (SHapley Additive exPlanations) (Lundberg and Lee, 2017). SHAP goes beyond traditional importance measures by quantifying the marginal contribution of each feature, while accounting for interactions among variables. For each classifier, SHAP analysis was done to interpret global feature importance and local prediction-level contributions. By enabling consistent cross-model comparison, this approach provides improved insight into the role of spectral bands, indices, and ancillary variables on land cover classification. Feature contributions were also evaluated across temporal subsets to capture seasonal variability in predictor importance.

Classification in GEE

The complete analysis workflow was performed in GEE (available at https://code.earthengine.google.com/) using the JavaScript API. Sentinel-1, Sentinel-2, and auxiliary datasets were processed entirely within GEE. Classification was implemented using built-in ee. Classifier models. No software outside the GEE was used, ensuring a fully cloud-based and reproducible workflow. A fixed random seed and stratified sampling were applied to ensure reproducibility of training and validation splits. The LULC classes were used as strata, and random samples were assigned equally across all classes to ensure balanced representation. A total of 2,400 samples were generated, with 300 per class, for the following categories: Natural forest, Water, Planted forest, Barren land, Cropland, Wetland, Grassland, and Built-up areas. Equal sample allocation was preferred over proportional allocation to ensure sufficient representation of minority classes and facilitate fair comparison across the classifiers.

The dataset was subdivided into training stratified samples (TSS) and validation stratified samples (VSS) to support both model development and accuracy assessment. Reference data were derived from a visual interpretation of high-resolution imagery such as Google Earth. Equal sample allocation across classes was deliberately implemented to reduce class imbalance, enhance classifier learning, and ensure robust comparison among machine learning algorithms. This sampling design is supported by established best practices in LULC (Olofsson et al., 2014; Congalton and Green, 2019).

Table 1 shows the distribution of training and validation samples. SVM and KNN classifiers had 215–238 samples per class in the training dataset and 62–85 samples set aside for validation. In contrast, the RF training subset contained 219–242 samples per class, whereas 58–81 samples per class were reserved for validation. CART had a range of 216 to 237 training data samples and 63 to 84 validation samples (Table 1). These classifiers harnessed 20 predictor variables, encompassing spectral signatures (Blue, Green, Red, NIR, SWIR1, SWIR2), a complement of spectral indices (NDVI, NDWI, NDMI, NDBI, BSI, OSAVI, AWEI_nsh), topographic derivatives (elevation, slope, HAND), settlement information (GHS Built-up), radar-derived properties (VH, VV/VH), and proximity to roads, to distinguish land-cover classes.

Classifier attributes descriptions

Random Forest

In our study, the RF model was tuned using the number of trees (ntrees), minimum leaf size (minLeaf), and bagging fraction (bagFraction). For the number of trees the study searched (50, 100, 150, and 200), we set the minimum leaf size (1, 5, and 10) and the bagging fraction (0.5 and 0.7). We found the best combination of parameters to be ntrees = 100, minLeaf = 1, and bagFraction = 0.7. We additionally set maximum nodes to be 500, and 7 seeds for reproducibility. Overall, the model tuning achieved an overall accuracy of 91% on multiple attempts.

Support Vector Machines

Similar to RF, we tuned SVM using cost values (cost) (1, 5, 10 and 20) as well as gamma values (gamma) (0.1, 0.5 and 1). Additionally, a Radial Basis Function (RBF) kernel was used with cost = 20 and gamma = 1.

K-nearest neighbours

In our study, the KNN model was built using a classifier. smile KNN with k = 5, which means that each pixel was assigned a class by relating it to the most common class among its five nearest neighbours in the high-dimensional feature space. The model classification system compared each pixel in the composite image to the labelled training samples using the Euclidean distance within the 20-dimensional feature space. The model selected the five nearest training samples (neighbours) and assigned the pixel to the class most commonly represented among them.

Classification Regression Trees

We used Classifier.smileCart, with a maximum tree size of 100 nodes to avoid overfitting and a minimum leaf population of 10 samples for stability. The classifier used a set of 20 predictors, including spectral bands, radar-derived metrics, spectral indices, and ancillary data variables, to separate the land-cover categories. During the training phase, the algorithm repeatedly split the training dataset by identifying predictor variables and threshold values that most effectively clarified the class at each node. This process continued until one of two conditions was satisfied: either the maximum of 100 nodes had been reached or the minimum of 10 leaves had been attained. Ultimately, the resulting decision tree applied a structured set of rules to classify each pixel into one of the LULC types, utilising spectral properties, derived indices, and environmental characteristics.

Hyperparameter tuning

A grid search was applied for hyperparameter optimisation, evaluating multiple combinations of model parameters, and the final configuration was selected considering the highest validation accuracy. Table 2 shows the classifiers and corresponding hyperparameter.

Validation

A hold-out validation subset derived from the overall sample pool,was used for validation to prevent biased assessment of the classification. The dataset comprised reference points representing all land-cover classes. The points were superimposed on the classified image, and the assigned classes were compared with the corresponding reference classes. Several accuracy metrics were computed from the confusion matrix to quantify classification performance. These were the overall accuracy (OA), producer’s accuracy (PA), and the user’s accuracy (UA). Overall accuracy measures the proportion of correctly classified observations. where: n i i = correctly classified observations for class i

N = total number of observations

A producer’s accuracy measures how well a certain land-cover class (or category) has been classified from the reference data perspective. It is related to omission error.

For class i where. n i i = correctly classified observations for class i

n i + = total reference samples for class i i.e., total observations in row i

User’s accuracy measures the reliability of the classified map from the user’s perspective. It is related to the commission error.

For class i where. n i i = correctly classified observations for class i

n + i = total reference samples into class i i.e., total observations in column i (Stehman et al., 2009).

Beyond these traditional measures of performance, this study applied the guidelines introduced by Pontius and Millones (2011) to calculate QD and AD. QD refers to the extent to which the classified map has falsely predicted the overall area of land in each land-cover class, either as undervalued or overvalued.

Where N is the total number of samples, n i j number of samples in row i (reference) and column j (map) r i = ∑ j = 1 k n i j i.e., the reference totals and c i = ∑ j = 1 k n j i i.e., map totals, and K is the number of classes in the confusion matrix.

AD highlights spatial misallocations of classes, assuming accurate class total areas

The two metrics are more explicit in distinguishing misclassifications caused by differences in class proportions and those arising from misallocations of classes (Pontius and Santacruz, 2014; Pontius and Millones, 2011). According to Olofsson et al. (2014), they are very useful, particularly in land-use and land-cover classifications, where misclassification errors are area-specific and important.

Figure 2 shows the complete workflow that was followed.

Performance of classifiers and validation

Table 3 shows the combined assessment of Producer’s Accuracy (PA), User’s Accuracy (UA), and F1-score. These metrics provide a class-specific evaluation of omission and commission errors.

RF reveals strong class-level patterns in classification performance. Natural Forest, Water, and Built-up consistently achieved high accuracies across all metrics. For example, Natural Forest produced PA = 0.982; UA = 0.950; and F1 = 0.966, suggesting strong class distinction and small omission and commission errors. RF and SVM achieved higher and more balanced PA, UA and F1 Score values across most LULC classes, whereas CART and KNN showed greater variability and generally lower accuracies. Based on the three metrics for the RF model, the natural forest and built-up classes consistently achieved high classification accuracies, whereas cropland and grassland showed comparatively lower performance. A similar trend was observed across the other classifiers employed in this study. To provide a simple yet high-level summary of how well a classification map aligns with the reference data, we report the overall accuracy, quantity-disagreement, and allocation-disagreement values as shown in Table 4.

Table 4 that RF achieved the highest overall accuracy, with an OA value of 0.911, followed by SVM (0.884), CART (0.866), and KNN (0.844). This indicates that RF produced the greatest proportion of correctly classified validation samples relative to the reference data. Regarding quantity disagreement, RF recorded the lowest QD value (0.017), indicating a minimal discrepancy between the predicted and reference class proportions. Higher QD values were observed for SVM (0.031), CART (0.040) and KNN (0.029), suggesting greater differences in the overall quantities of land-use/land-cover classes predicted by these models. Allocation disagreement values showed a similar trend. RF again produced the lowest AD value (0.071), reflecting fewer spatial misallocations of class labels after accounting for quantity differences. SVM and CART recorded moderate allocation disagreements of 0.084 and 0.093, respectively, whereas KNN exhibited the highest allocation disagreement (0.125), indicating greater spatial misplacement of classes. Overall, the results in Table 3 indicate that RF outperformed the other classifiers in terms of overall accuracy while also minimising both quantity and allocation disagreement.

Spatial pattern of LULC features in the IWP

Figure 3 shows the classification outcomes of the four classifiers. The maps depicts a relatively similar pattern of land use and land cover in the IWP. It is evident from all classifiers that the park is characterised by classes of natural origin, with natural forests, grassland, wetland, water, and barren land collectively being the most representative units of the overall system. Natural forests and grassland are the most prominent and widespread features throughout much of the park. The coastal section of the system is clearly occupied by three main water bodies: Lake St Lucia in the south, Lake Sibaya in the centre, and Kosi Bay in the far north towards Mozambique. An assemblage of wetlands is widespread from the south to the interior, mostly around Lake St Lucia. The IWP is also characterised by a distinct land-use footprint, with planted forest blocks occurring in pockets from the south through the central east to the northern section of the IWP. All classifiers identified common types of landscape units in a relatively consistent manner, with the RF algorithm producing the most spatially coherent classification and displaying the most-even class structure with minimal salt-and-pepper noise. SVM also produced strong, visually reliable results, whereas CART and KNN showed less consistency, particularly across the diverse vegetation types in the northern areas. The inset maps illustrate these variations, showing that CART and KNN produced more dispersed patches, especially along the wetland–grassland and forest–cropland margins.

Quantitative analysis of LULC features in the iSimangaliso Wetland Park

Figure 4 shows the predicted area coverage of LULC classes outcome for each algorithm, The predicted area coverage of LULC classes varied slightly among the four machine learning classifiers (RF, SVM, KN, and CART), but the overall spatial dominance patterns were largely consistent. Natural forest occupied the largest proportion of the IWP landscape across all classifications, covering 2,711 km 2 (32%) with RF, 2,645 km 2 (31%) with CART, 2,499 km 2 (29%) with KNN, and 2,810 km 2 (33%) with SVM. This essentially implies a common agreement among classifiers in detecting forested areas.

Cropland occupied an area estimated to range from 751.8 km 2 (9%) under CART to 1,007.7 km 2 (12%) under SVM. RF and KNN generated estimates of 881.5 km 2 (10%) and 963.5 km 2 (11%), respectively. These variations reflect the classifier’s ability to separate cropland from spectrally similar classes, including grassland and planted forest. Planted forest coverage also revealed clear variations, ranging from 843.1 km 2 (10%) using SVM to 1,252.3 km 2 (15%) using KNN. CART and RF predicted 1,079.7 km 2 (13%) and 885.4 km 2 (10%), respectively. The higher proportion of planted forest area identified by KNN suggests overgeneralization of this class relative to other vegetation types.

Water bodies, wetlands, barren land and built-up areas accounted for a relatively smaller proportion of the study area. The parts occupied by water ranged narrowly between 497.9 km 2 (6%) and 583.7 km 2 (7%) across all classifiers, suggesting a strong distinction in water features in the input data. Likewise, wetlands covered a limited area, with estimates ranging from 342.7 km 2 (4%) under CART to 442.8 km 2 (5%) under KNN. SVM and RF produced similar values of nearly 5% of the total area. The small variations among classifiers may be attributed to spectral confusion between wetlands and adjacent vegetated areas. Barren land was denoted as one of the least extensive classes in coverage, with areas of 339.3 km 2 (4%) using CART and 475.9 km 2 (6%) using SVM. RF and KNN formed intermediate predictions of 419.1 km 2 (5%) and 392.7 km 2 (5%), respectively. Built-up areas showed the greatest variation among classifiers, ranging from 222.3 km 2 (3%) under KNN to 586.1 km 2 (7%) under CART. RF and SVM estimated areas of 422.8 km 2 (5%) and 341.3 km 2 (4%), respectively. This variability suggests that classifiers perform differently in detecting heterogeneous and spatially fragmented urban features.

Assessment of per-class accuracy and model agreement

Figure 5 illustrates a confusion matrix based on validation points per model per class while Figure 6 shows pairwise inter-model agreement matrices. Both figures help to identify class-dependent errors and algorithm-specific uncertainty. These class-level diagnostics are increasingly gaining momentum and recognition as essential for strong land-cover assessment in heterogeneous landscapes (Nicolau et al., 2023).

Random Forest (RF) shows the most stable and internally reliable performance across all classes. Validation samples are closely concentrated along the correct class axes (Figure 5), suggesting strong class discrimination and minimal misclassification. This is evident particularly for Water, Wetland, and Grassland. Similar patterns have been widely reported for ensemble classifiers in complex and fragmented environments. Natural forests are classified with high reliability, although occasional misclassifications occur in planted forests. Support Vector Machine (SVM) performs comparably but shows greater diffusion across heterogeneous classes, notably Cropland and Wetland, indicating reduced robustness within-class variability. CART and KNN show considerably higher misallocation, especially in transitional classes, indicating their low capacity to model non-linear and spectrally mixed class boundaries, consistent with previous comparative evaluations (Huang et al., 2025; Li et al., 2023).

Misclassification patterns are not random but rather strongly directional. Wetland and Cropland classes are repeatedly redistributed into Grassland and among themselves, resulting in higher omission errors due to spectral and ecological overlap. On the other hand. Grassland shows higher commission resulting from the inclusion of misclassified Wetland and Cropland pixels. Built-up areas show a greater overlap with Grassland and Barren in weaker classifiers, suggesting spectral mixing in thinly developed landscapes. Water maintained low omission and commission errors across all models, suggesting its strong, spectrally distinct attribute. A trend commonly observed in the literature on optical remote sensing classifications (Nicolau et al., 2023).

Inter-model agreement matrices (Figure 6) confirm that classification uncertainty is also based on the different classifiers. High agreement between RF and other classifiers is observed for Water (87%–96%) and Grassland (up to 84%), indicating stable, reproducible classification of spectrally distinct classes. Agreement drops abruptly for heterogeneous classes, where RF–CART agreement for Cropland and RF–KNN agreement for Built-up both decline to 45%, indicating substantial variability in class representation throughout all algorithms. Similarly, a planted forest is often absorbed into a natural forest in ensemble classifiers. Moderate agreement for Wetland (71%–82%) also shows sensitivity to environmental gradients and mixed spectral responses (Huang et al., 2025).

Comparative feature importance interpretation

Figure 7 shows Pearson correlation matrix for feature selection illustrating relationships among multispectral bands, spectral indices, radar backscatter, and environmental variables.

There is a strong redundancy among predictor variables, with high positive correlations across visible bands (Blue–Green: 0.96; Green–Red: 0.95) and across SWIR bands (SWIR1–SWIR2: 0.95). Perfect correlations exist among vegetation indices such as NDVI and OSAVI (r = 1.00). On the contrary, strong negative relationships are recorded between NDMI and NDBI (r = −1.00) and between NDVI and NDWI (r = −0.78).

Figure 8 shows feature importance trends across the four classifiers. RF emphasises Green, Elevation, and GHS Built-Up, whereas SVM places greater weight on SWIR1, NDVI, and NDBI. CART places greater weight on SWIR2 and NIR, whereas KNN displays a stronger influence from Elevation and SAR variables such as VH, consistent with findings from multi-feature classification approaches integrating spectral and ancillary data. These patterns correspond with the feature importance results (Figure 5), where models emphasising spectral indices and SWIR bands (SVM) show improved separation of vegetation and moisture classes, while models influenced by topographic and SAR variables (KNN) produce more heterogeneous outputs. Areas of class mixing are most evident in transitional zones between cropland, grassland, and wetlands, particularly in the northern inset regions, where fine-scale heterogeneity is pronounced. 3.4.

Spatial patterns of misclassification and uncertainty

Figure 9 presents agreement–disagreement maps comparing RF with SVM, CART, and KNN classifications. Areas shown in colour represent locations where the classifier pairs assigned the same land-cover class, whereas white areas indicate disagreement between classifiers.

Agreement is strongest within regular and spectrally stable land-cover types, particularly natural forests, wetlands, and water bodies (panels a2, b2, c2). These areas appear as large, continuous regions with relatively limited fragmentation, suggesting that the classifiers constantly identified these classes without struggle.

On the other hand, disagreement is concentrated mainly within heterogeneous and transitional landscapes. Notable disagreement zones occur along forest–grassland transitions, cultivated–wetland boundaries, and urban fringe areas (panels a1, b1, c1). These regions are characterised by fragmented land-cover patterns and mixed spectral signatures, which increase class distortions and classification uncertainty.

The spatial extent of disagreement varies among classifiers. The RF–SVM comparison shows the smallest and most localised disagreement patterns (panel a1), indicating a strong similarity between the two classifiers. Greater disagreement is observed in the RF–CART comparison, particularly within cultivated and built-up areas (panel b1). The RF–KNN comparison shows the strongest spatially dispersed disagreement around heterogeneous regions (panel c1), suggesting relatively lower consistency with RF classifications.

The enlarged sections (a3–a4, b3–b4, c3–c4) further reflect that disagreement consistently occurs within fragmented and transitional environments, regardless of classifier pair. This suggests that landscape heterogeneity is the primary driver of classification uncertainty, while differences between classifiers play a secondary role.

RF emerged as the best-performing classifier in this study, achieving the highest overall accuracy (OA = 0.911) and the lowest quantity disagreement (QD = 0.017) and allocation disagreement (AD = 0.071) among all evaluated algorithms (Table 4). These results show that RF was more effective compared to SVM (OA = 0.884; AD = 0.084), CART (OA = 0.866; AD = 0.093), and KNN (OA = 0.844; AD = 0.125) at both predicting class proportions and correctly assigning land-cover classes. The superior performance of RF can be attributed to its ensemble structure, where bootstrap combination reduces overfitting and random feature selection reduces inter-tree correlation, enabling the model to capture the spectral heterogeneity, mixed pixels, and nonlinear class boundaries characteristic of wetland landscapes (Breiman, 2001; Jocea et al., 2025). The lower allocation disagreement achieved by RF further highlights its ability to produce more spatially consistent classifications, making it suitable for classifying ecologically heterogenous wetland environments.

The notably lower allocation disagreement produced by RF is particularly informative, as it indicates improved spatial placement of land-cover classes rather than merely higher overall accuracy. In wetland environments, where gradual shifts between vegetation and open water are common, accurate spatial allocation is important for ecological interpretation and management. The superior spatial consistency achieved by RF, therefore, underscores its suitability for mapping complex wetland landscapes. Comparable patterns have been widely reported in wetland mapping studies across floodplains, coastal systems, and savanna–wetland interfaces, where RF consistently outperforms single-tree and distance-based classifiers under conditions of high environmental heterogeneity (Adam et al., 2010; Mahdianpari et al., 2019; Amani et al., 2019). Collectively, these studies indicate that RF’s strength lies not only in classification accuracy but in its ability to capture the spatial and ecological complexity inherent to wetland systems.

SVM achieved a comparatively high overall accuracy, demonstrating its effectiveness in high-dimensional spectral feature spaces and under limited training sample conditions. However, the higher allocation disagreement recorded in this study demonstrates that classification success in heterogeneous wetland landscapes depends not only on global class separability but also guided by algorithm’s capacity to accommodate local spatial variability. SVM constructs optimal hyperplanes that maximise the global margin between classes, but this method assumes relatively well-defined class boundaries. In wetland systems, where vegetation types, moisture gradients, and land–water interfaces often transition gradually rather than in isolation, this assumption is violated, resulting in distortions in ecotonal boundaries.

This behaviour was noted in previous wetland classification studies, which have shown that spectral gradients driven by seasonal hydrological variability and mixed vegetation groupings interfere with class separability for margin-based classifiers such as SVM (Mahdianpari et al., 2019; Cherian and Kanjirappuzha, 2024). Thus, while SVM remains effective for spectrally distinct land-cover types, it struggles in transitional environments, highlighting an inherent limitation when applied to complex wetland mosaics.

The weaker performance of CART and KNN suggests that more important methodological constraints are at play in representing continuous ecological gradients. CART, as a single decision tree model, is highly sensitive to training data variability and prone to overfitting, limiting its ability to capture classes across heterogeneous landscapes. The high allocation disagreement observed in this study by CART suggests an inability to capture the gradual spectral transitions characteristic of wetland ecotones, where class boundaries are fundamentally unclear rather than hierarchical. KNN had the lowest overall accuracy and the highest allocation disagreement, underscoring its weakness in the presence of spectral mixing and class overlap. KNN relies on local similarity assumptions; its performance diminishes in landscapes such as iSimangaliso Wetland Park, where neighbouring pixels often denote different land cover types due to fine-scale environmental heterogeneity.

These limitations are well documented in other LULC studies, particularly in ecosystems characterised by strong spatial mixing and land-use complexity (Munyati, 2000; Bhungeni et al., 2024; Michael et al., 2026; Nguyen et al., 2026). The concentration of misclassification across all algorithms within transitional boundaries, such as wetland boundaries, reflects a combination of algorithmic failure and the underlying ecological situations of gradual transitions and seasonal inundation dynamics. From a classification perspective, these conditions worsen allocation disagreement; whereas from an ecological perspective, they highlight zones of high functional connectivity and environmental variation. Similar convergence between classification uncertainty and ecological significance has been reported in global wetland mapping efforts, where areas of high misclassification frequently coincide with ecologically sensitive transition zones (Junk et al., 2013; Huang et al., 2025).

The ecological and management implications of these findings extend beyond differences in classification accuracy and illustrate how algorithmic attributes can form interpretations of landscape dynamics. The high spatial consistency achieved by Random Forest (RF) enhances the reliability with which important ecological processes, such as wetland shrinkage or extension, forest invasion, and grassland degradation, can be detected and monitored over time. In wetland systems, landscape change usually occurs incrementally along diffuse boundaries. This calls for an accurate spatial representation that distinguishes genuine ecological transitions from classification artefacts. Consequently, RF has proved reliable in reducing spatial fragmentation, improving confidence in assessments of ecosystem health, and supporting evidence-based conservation planning objectives within protected areas.

In contrast, the more fragmented classification outcomes from CART and KNN also have critical implications for landscape interpretation. By overstating spatial discontinuities, these algorithms may miscalculate habitat connectivity and distort landscape structure. The results may lead to wrong valuations of ecological integrity. Such misrepresentations are misleading and problematic in conservation contexts, where management choices increasingly rely on spatial metrics derived from classified maps to prioritise intervention areas (Nagendra et al., 2015).

The regular distribution of classification uncertainty across ecotonal zones further underscores their role as focal points of both ecological complexity and environmental sensitivity. These transitional boundaries are characterised by gradual spectral gradients, seasonal inundation dynamics, and structurally diverse vegetation communities, making them particularly responsive to hydrological and climate variability, as well as to anthropogenic pressures such as agricultural expansion. In the context of iSimangaliso Wetland Park, where ecosystem functioning is largely dependent on hydrological connectivity and landscape connectivity, these zones serve as early indicators of change rather than simple sources of classification error. Integrating high-accuracy classification outputs, such as those produced by RF, with targeted field validation in ecotonal areas consequently offers a practical pathway to improve monitoring frameworks and support adaptive management strategies.

In broader terms, this study supports findings from recent literature that classifier selection should be guided not only by overall accuracy but also by the algorithm’s capability to capture spatial disagreement, ultimately representing ecological heterogeneity (Pontius et al., 2025) Increasingly, research findings emphasise that pixel-level accuracy metrics alone can disguise spatially structured errors in complex landscapes (Maxwell et al., 2023; Pande et al., 2024). The convergence of classification uncertainty with ecologically dynamic zones supports the idea that mapping error can yield ecological insight rather than being a limitation in transitional settings such as wetlands. The use of cloud-based platforms such as Google Earth Engine improves the applicability of this method by enabling scalable, transparent, and reproducible land-use and land-cover mapping, particularly in data-limited regions (Tamiminia et al., 2023). Future research should build on these findings through explicitly multi-temporal analyses capable of capturing seasonal dynamics and long-term ecological change, which are especially critical in climatically sensitive wetland systems (Mutale et al., 2024). Building on these findings, future research should adopt explicitly multi-temporal approaches to improve capture of seasonal dynamics and long-term ecological change, which are serious issues in hydrologically variable sensitive wetland systems.

This study demonstrates the methodological potential of integrating publicly available Sentinel-2 and Sentinel-1 imagery with GEE and several machine-learning algorithms for effective land-use and land-cover classification in ecologically complex wetland systems. A notable contribution of this work lies in the integrated evaluation of thematic accuracy and spatial reliability, achieved by combining overall accuracy with quantity and allocation disagreement metrics. This framework goes beyond conventional accuracy assessment by clearly recognising spatial misallocation, which is particularly relevant in heterogeneous and ecologically sensitive landscapes such as the iSimangaliso Wetland Park. Among the evaluated classifiers, Random Forest consistently dominated others, achieving superior performance, not only the highest overall accuracy but also minimising both quantity and allocation disagreement. This indicates a stronger capacity to accurately predict class extents and to spatially allocate land cover classes in regions characterised by mixed pixels, gradual transitions, and strong ecotonal gradients. In contrast, Support Vector Machine showed robust yet lower performance, while Classification and Regression Trees and K-Nearest Neighbours displayed higher spatial disagreement, particularly for classes with common spectral signatures such as wetlands, grasslands, and transitional vegetation. These results highlight that classification performance is strongly influenced by metrics other than accuracy alone. Despite these contributions, the limitations identified should be acknowledged. The analysis adopted in this study, while effective at reducing short-term noise, may overlook finer intra-seasonal dynamics important to certain wetland processes. The absence of independently documented or official reference area estimates limited the evaluation of which classifier most accurately represented true LULC class extents. Future studies should incorporate reliable reference-area datasets to improve the assessment of classifier-specific area-estimation accuracy and sensitivity. Furthermore, a lack of field validation exercises may contribute to uncertainty estimates for ecologically critical classes. Future research should refine this framework through multitemporal and time-series analyses to better capture seasonal variability, disturbance events, and long-term ecological change in wetland systems. Integrating field data and climate variables, and exploring hybrid or deep learning techniques, offer promising paths to improve both classification accuracy and ecological interpretability. In the wider sense, this study highlights the need for classifier selection approaches that prioritise spatial agreement and ecological realism, supporting more informed management and conservation decision-making in threatened wetland landscapes.