Comparative evaluation of machine learning algorithms for integrated wetland modelling within a geospatial big data environment in Driefontein Grasslands of Zimbabwe

1 Introduction

Wetland ecosystems exhibit complex, nonlinear hydrological–vegetation interactions requiring advanced geospatial modelling frameworks that integrate multi-source big data for accurate spatial prediction and ecological assessment (López-Tapia et al., 2021). Yet the predictive reliability and comparative robustness of machine learning algorithms in modelling wetland systems remain insufficiently evaluated within data-rich remote sensing environments. Recent advances in machine learning, particularly Random Forest (RF), Support Vector Machine (SVM), and Classification and Regression Trees (CART), offer robust, data-driven approaches for modelling heterogeneous wetland landscapes using high-dimensional remote sensing datasets. Machine learning algorithms provide unprecedented opportunities to model wetland extent, hydroperiod dynamics, vegetation structure, and disturbance regimes with improved predictive accuracy (Aslam et al., 2024; Jafarzadeh et al., 2022; Pal and Debanshi, 2021). However, despite the proliferation of data-driven approaches in geographical information systems (GIS) and remote sensing, the predictive robustness, transferability, and ecological interpretability of machine learning models remain insufficiently tested in complex wetland systems such as the Driefontein Grasslands of Zimbabwe, where hydrology–vegetation–topography feedback operate simultaneously across heterogeneous landscapes. Traditional wetland mapping and modelling approaches have largely relied on conventional statistical techniques and single-date remote sensing classifications (Khan et al., 2025), which often struggle to represent the multidimensional and temporally variable nature of wetland systems (Mallick et al., 2021). The emergence of geospatial big data—characterised by high-resolution, multi-temporal satellite imagery (e.g., Landsat, Sentinel), digital elevation models, climate reanalysis datasets, and ancillary environmental variables—has significantly expanded the capacity to analyse landscape processes at multiple spatial and temporal scales. However, extracting meaningful ecological insights from these high-dimensional datasets requires advanced analytical frameworks capable of handling complex, nonlinear relationships.

Satellite-, airborne-, and in-situ–based sensing systems are generating exponentially expanding streams of environmental observations for the precise characterization and monitoring of the Earth’s surface geometry and superimposed features, including built infrastructure, arboreal cover, hydrological bodies, and transportation networks (Singh et al., 2018; Thamaga et al., 2022). The 21st century has been characterized by a substantial proliferation in the availability of (i) wall-to-wall archives of free and open-access Earth Observation (EO) datasets; (ii) interoperable data portals and scalable cloud-computing ecosystems; and (iii) advanced analytical methodologies designed to address complex challenges in natural resources management (NRM), including wetland conservation and restoration (Gomes et al., 2020). The Driefontein Grasslands constitute a strategically significant empirical laboratory for assessing data-driven and integrative wetland modelling frameworks within southern Africa. Contemporary EO missions continuously produce voluminous geospatial datasets that are openly accessible to the scientific community and broader society (Guo et al., 2015). Technological paradigms such as cloud computing and distributed architectures provide scalable solutions for the storage, retrieval, and processing of big EO data (Nativi et al., 2015). Prominent web-based platforms facilitating large-scale EO data management and analytics include Google Earth Engine (GEE), Sentinel Hub, and Open Data Cube (Sudmanns et al., 2020). While conventional geo-data processing can be executed using standard functionalities embedded within mainstream geospatial software, the resolution of complex spatiotemporal problems in Earth and environmental sciences frequently necessitates the manipulation of exceptionally large and heterogeneous datasets (Yao et al., 2019). Such analytical demands require specialized geoprocessing architectures and machine learning algorithms including RF, SVM, and CART. In Zimbabwe, wetlands are undergoing accelerated degradation driven by agricultural expansion (Musasa and Marambanyika, 2020), extractive mining activities, and climate variability (Sibanda and Ahmed, 2021), thereby necessitating sophisticated monitoring, modelling, and predictive analytics frameworks.

The interrelated Water and Energy cycle exert fundamental control over hydrological availability and climatic dynamics at planetary scale, thereby determining the viability and quality of life across ecosystems (Martinez Prentice et al., 2021). The current increase in the global population has presented the world with a ‘triple planetary crisis’ of environmental pollution as well as food and energy insecurity. Emerging machine learning paradigms and big data analytics offer transformative capabilities for simulating nonlinear wetland processes across spatially heterogeneous landscapes (Whyte et al., 2018). Water and energy constitute foundational pillars of terrestrial life, and their variability, long-term trends (Talukdar et al., 2021), and extreme manifestations underpin hydrometeorological hazards such as prolonged droughts, heatwaves, intense precipitation events, flooding, and severe storm systems, which are increasingly exacerbated under changing climatic regimes (DeLancey et al., 2019). Climate change represents one of the most critical global environmental challenges (Salimi et al., 2021). The greenhouse effect and the carbon cycle—particularly carbon emissions and sequestration fluxes—are central to contemporary climate dynamics (Mitsch et al., 2013). International governance bodies, including the United Nations Framework Convention on Climate Change (UNFCCC) and the Intergovernmental Panel on Climate Change (IPCC), have institutionalized policy mechanisms aimed at mitigating greenhouse gas emissions, exemplified by the Kyoto Protocol (Moomaw et al., 2018). These initiatives establish direct linkages with the International Environmental Agenda (IEA) and the Sustainable Development Goals (SDGs). Robust and spatially explicit quantification of carbon cycle components is indispensable for evaluating systemic perturbations (Xiong et al., 2023) and informing policy instruments targeted at emission reductions, including Reducing Emissions from Deforestation and Forest Degradation (REDD+). Wetlands function as critical carbon sinks, modulating greenhouse gas concentrations and contributing substantively to climate change mitigation through long-term carbon sequestration (Erwin, 2009; Salimi et al., 2021). Their conservation and sustainable management therefore reinforce global environmental objectives and underpin evidence-based climate mitigation policies such as REDD+ (Xiong et al., 2023). Wetlands are critical ecosystems that regulate hydrological processes, store significant quantities of carbon, support biodiversity, and provide essential ecosystem services. However, increasing anthropogenic pressures and climate variability are accelerating wetland degradation, necessitating accurate and continuous monitoring. Although remote sensing and machine learning have improved wetland mapping capabilities, most studies in sub-Saharan Africa rely on single-classifier approaches and conventional accuracy metrics, limiting the understanding of classifier behaviour and model reliability. This study addresses this gap through a comparative evaluation of RF, SVM, and CART within a geospatial big data framework, integrating multi-temporal satellite imagery and spectral indices to improve wetland monitoring and change detection in the Driefontein Grasslands, Zimbabwe.

Enhanced observational capacity and advanced analytical frameworks are essential for deepening our understanding of the physical mechanisms governing hydrological and climatic processes, as well as for improving predictive modelling accuracy (Camara et al., 2016). The meaningful application of models is contingent upon the availability of high-quality, spatially consistent datasets (Heuvelink et al., 1989; Tahir et al., 2025). EO technologies provide an unparalleled mechanism for the synoptic monitoring of essential water and energy variables, enabling multi-scalar assessments from global to basin-level domains (Rebelo et al., 2009). Surface water systems—including lakes and fluvial networks—are integral to water resource governance and ecosystem functioning across numerous countries (Ansari and Golabi, 2019). They provide direct water supplies for agricultural production, domestic consumption, and industrial activities, while simultaneously acting as dynamic interfaces between groundwater systems and the atmosphere through evapotranspiration, runoff generation, and aquifer recharge processes (Lynch and Maddock, 2025; Lyon and McCarthy, 1995). From a geopolitical perspective, unsustainable anthropogenic abstraction and contamination of surface waters present significant conflict risks, particularly where transboundary river basins are concerned and upstream utilization compromises downstream availability and quality (Selmy and Moursy, 2025). Under projected climate change scenarios, intensification of the global hydrological cycle—often summarized as “dry regions becoming drier and wet regions becoming wetter”—is anticipated (Erwin, 2009; Salimi et al., 2021; Yi et al., 2024). Elucidating the interdependencies between water and energy cycles requires comprehensive understanding of surface energy exchanges governed by radiative fluxes, turbulent momentum transfer, and sensible and latent heat dynamics, all of which are critical for designing adaptive management strategies (Qi et al., 2025). Moreover, climate change imposes additional pressures on national, regional, and local governance structures already grappling with sustainable water and energy utilization (Domínguez-Solís et al., 2025). Wetlands and surface water systems are therefore indispensable for securing water supplies, sustaining agricultural productivity, supporting domestic and industrial demands, and maintaining ecological equilibrium. Through regulation of hydrological and energy fluxes, they enhance adaptive capacity to climate variability and inform evidence-based water governance frameworks.

The Driefontein Grasslands represent a unique wetland–grassland mosaic characterised by seasonal inundation patterns, variable vegetation communities, and increasing anthropogenic pressure (Fakarayi et al., 2015; Musasa and Marambanyika, 2022). Yet there is limited empirical evidence assessing how different machine learning algorithms perform in modelling such heterogeneous wetland systems using integrated big data frameworks. Moreover, few studies in the region systematically test the predictive capacity and transferability of machine learning models for holistic wetland modelling that incorporates topographic, hydrological, and spectral variables simultaneously. This research advances methodological rigor by: (1) implementing threshold-independent AUC diagnostics with structural model comparison beyond conventional accuracy metrics; (2) operationalizing multi-temporal modelling within a big data ecosystem enabling scalable, reproducible wetland intelligence; and (3) linking algorithmic behaviour directly to ecological interpretation and policy relevance. The demonstrated superiority of ensemble and kernel-based classifiers provides actionable guidance for national environmental monitoring agencies, supporting Zimbabwe’s commitments under Africa’s Agenda 2063 and SDG Targets 6.6 and 15.1. This study establishes a replicable GeoAI framework for wetland monitoring in data-constrained African landscapes, strengthening evidence-based ecosystem governance.

2 Materials and methods

2.1 Study area

The Driefontein Grasslands is located at the intersection of the three provinces in Zimbabwe, namely, Midlands Province, Masvingo Province and Mashonaland East (Figure 1). Within these three provinces it is found at the intersection of three districts which are Chirimuhanzu, Chikomba and Gutu, between longitudes 30.25°E and 30.55°E and latitudes 19.25°S and 19.45°S. The site covers ∼404.44 km² and forms part of the larger Driefontein Important Bird Area. The area lies at an altitude ranging from 1,400 to 1,500 m above sea level, situated on the central Zimbabwean highveld plateau. The terrain is generally undulating with gentle slopes (Marambanyika, 2021). The region experiences a subtropical climate characterized by a single rainy season from November to March. Mean annual rainfall ranges between 600 and 800 mm, while mean annual temperatures vary from 18 °C to 22 °C, with October being the hottest month and July the coolest. The dominant vegetation consists of moist miombo woodlands interspersed with extensive Dambo grasslands (Musasa and Marambanyika, 2022). Soils are primarily hydromorphic, consisting of vertisols and gleysols with high clay content. The Driefontein Grasslands are recognized as a Ramsar wetland site (Ramsar Site No. 2104) Ramsar Convention Secretariat (2016), designated due to their extensive wetland systems that support biodiversity (Gardner, 2025). The dominant grass species include Hyparrhenia, Loudetia, and Andropogon genera. LULC comprises grassland (the dominant class), forest patches, built-up areas, bareland, and water bodies, as illustrated in Figure 1, with the surrounding areas featuring communal farming.

2.2 Methodology flowchart

The methodological framework developed for this study integrates multi-temporal remote sensing, machine learning, spectral analytics, and geospatial big data processing to model wetland dynamics in the Driefontein Grasslands, Zimbabwe (Figure 2). The workflow begins with the acquisition of Landsat imagery spanning 2015–2025, together with Landsat-derived Enhanced Thematic (ET) estimates to capture both vegetation and hydrological variability within the wetland ecosystem. Pre-processing procedures included radiometric correction, atmospheric correction, and seasonal filtering to minimise sensor and atmospheric inconsistencies while enhancing temporal comparability across image epochs. Ground truth data were generated using a hybrid reference framework consisting of 562 field-surveyed points and 300 image-derived samples, resulting in 862 validation observations representing major wetland and non-wetland land cover categories. The dataset was partitioned into 75% training and 25% validation subsets to ensure robust model calibration and independent performance assessment. A key innovation of the framework is the integration of advanced spectral remote sensing indices into the machine learning environment. Vegetation, moisture, chlorophyll, and stress-sensitive predictors—including GCI, EVI, MNDWI, MSAVI, NDVI, SIPI, SAVI, RECI, NDWI, and OSAVI—were derived to improve the ecological sensitivity of wetland detection. These indices were incorporated into three supervised machine learning algorithms: RF, SVM, and CART. RF was configured with 500 trees and unrestricted depth, SVM employed a radial basis function (RBF) kernel with optimized parameters (C = 10; gamma = 0.01), while CART underwent parameter tuning for tree depth and node splitting thresholds. Hyperparameter optimization was implemented using grid search combined with 10-fold cross-validation to improve model generalisation and minimise overfitting. The analytical workflow was executed within a geospatial big data ecosystem integrating Google Earth Engine, Python 3.10, ArcGIS Pro, and QGIS 3.44.6 Solothurn, enabling scalable processing of large multi-temporal satellite datasets and computationally intensive machine learning operations.

Model performance was evaluated using ROC curves, AUC statistics, and Pearson correlation analysis to quantify predictive accuracy and inter-model agreement. Final outputs comprised temporally explicit LULC maps, comparative algorithm performance metrics, spectral-index products, and correlation matrices, collectively providing an integrated GeoAI framework for robust wetland monitoring and ecosystem change assessment in semi-arid African landscapes. The study’s “geospatial big data environment” was operationalised through the integration of Google Earth Engine, Python 3.10, QGIS 3.44.6 Solothurn, and cloud-based geospatial processing workflows. Google Earth Engine enabled scalable pre-processing and management of multi-temporal Landsat archives, while Python libraries such as Scikit-learn, NumPy, Pandas, Rasterio, and Matplotlib supported machine learning implementation, hyperparameter optimisation, ROC–AUC evaluation, and spatial analytics. QGIS 3.44.6 Solothurn was used for cartographic production and spatial validation. The workflow qualifies as a geospatial big data framework because it handled high-volume, multi-source, multi-temporal satellite datasets combined with computationally intensive machine learning operations executed within scalable cloud-supported and parallel-processing environments.

To improve methodological transparency and reproducibility, a comprehensive data inventory was developed summarising all geospatial datasets, predictor variables, pre-processing procedures, and analytical platforms used in the study (Table 1). The modelling framework integrated multi-temporal Landsat imagery and ten vegetation and water-related spectral indices (NDVI, EVI, SAVI, OSAVI, MSAVI, GCI, RECI, SIPI, NDWI, and MNDWI). These datasets were processed within Google Earth Engine, Python 3.10, ArcGIS Pro 3.0, and QGIS 3.44.6 Solothurn, enabling scalable analysis of large multi-temporal Earth observation datasets. The integration of diverse environmental predictors within a cloud-supported analytical ecosystem constitutes the geospatial big data architecture underpinning the comparative machine learning framework developed for wetland monitoring in the Driefontein Grasslands.

TABLE 1. Data inventory table.

The geospatial big data environment adopted in this study comprised a cloud-enabled workflow integrating GEE, Python 3.10, and QGIS 3.44.16 Solothurn. GEE was used for cloud-based storage, retrieval, pre-processing, and management of multi-temporal Landsat 8 OLI imagery, while Python was employed for spectral index computation, machine learning model development (RF, SVM, and CART), hyperparameter optimization, and statistical evaluation. QGIS supported spatial analysis, visualization, and cartographic output generation. The workflow qualifies as a geospatial big data framework because it integrates large volumes of multi-temporal Earth observation data, multiple spectral predictors, cloud computing infrastructure, automated machine learning workflows, and scalable analytical processes capable of handling high-dimensional geospatial datasets efficiently and reproducibly.

2.3 Digital image classification

Medium resolution satellite imagery, specifically Landsat images for the years 2015, 2020 and 2025, served as the primary sources of remotely sensed data. These datasets encompass various timeframes and sensor types to capture LULC transitions over time. To mitigate the influence of seasonal variability on classification outcomes, only images acquired during the same season were selected. To ensure phenological consistency, all images used were captured during the dry season (August)—a period characterized by minimal vegetation growth variability (Corner et al., 2013). This timeframe is considered optimal for distinguishing between different LULC classes. The study employed reflective bands from Landsat imagery with a spatial resolution of 30 m to extract multi-temporal LULC change data. Raw satellite imagery was pre-processed before it can be effectively used for mapping purposes (Coppin and Bauer, 1996). Essential corrections include radiometric and geometric adjustments. Geometric distortions in raw data may arise from systematic errors, such as the relative motion between the Earth’s rotation and the satellite’s orbit, and from random anomalies. These distortions typically result in spatial displacement along the north-south and east-west axes. Initially, Top-of-Atmosphere (TOA) reflectance products were accessed and subsequently converted to Surface Reflectance through atmospheric correction to minimise atmospheric scattering and absorption effects. Additional pre-processing included radiometric calibration, cloud and cloud-shadow masking, image mosaicking, and seasonal filtering to ensure temporal consistency across all image epochs. The resulting Surface Reflectance datasets were used to derive spectral indices and serve as input for the RF, SVM, and CART machine learning models for wetland classification and change detection. The raw digital number (DN) values were then converted to spectral radiance, as recommended. Atmospheric corrections were applied to eliminate the influence of atmospheric interference. Thematic mapping involves categorizing spatial information to represent the geographic distribution of specific land cover types of imagery can be systematically analysed to derive meaningful environmental information across temporal and spatial scales. The fundamental goal of classification is to assign each image pixel to a specific land cover category (Singh and Talwar, 2015). Several parameters influenced the selection of the classification method, including map accuracy, landscape dynamics, training data requirements, and algorithmic complexity (Dewan and Yamaguchi, 2009a; Dewan and Yamaguchi, 2009b). The study focused on the period 2015–2025 to ensure consistency in image acquisition, pre-processing, spectral index derivation, and machine learning implementation. The baseline year of 2015 was selected because it marks the availability of high-quality Landsat 8 Operational Land Imager (OLI) data, which provides improved radiometric and spectral performance for land use/land cover classification and wetland monitoring. Restricting the analysis to Landsat 8 OLI imagery minimized uncertainties associated with integrating data from multiple sensor generations and ensured temporal comparability across all study epochs. While earlier Landsat archives could provide insights into longer-term wetland dynamics, the primary objective of this study was to evaluate the comparative performance of RF, SVM, and CART for recent wetland modelling within a geospatial big data framework. Future studies should extend the temporal record by incorporating harmonized multi-sensor datasets to investigate longer-term wetland change trajectories.

2.4 Ground truth data collection

Ground truthing involves validating satellite-derived classifications by correlating pixel information with actual land surface conditions (Al-doski et al., 2013). This validation is critical for the accuracy of supervised classification procedures performed using remote sensing software QGIS 3.44.6 Solothurn (Lu et al., 2004). In this study, local knowledge was incorporated through interviews with long-time residents to identify key land cover sites. Ground reference data were acquired through two primary methods. First, field surveys conducted in September 2025 yielded 562 reference points across distinct LULC classes, recorded using high-precision global positioning system (GPS) units. Secondly, an additional 300 validation points were extracted via manual interpretation of high-resolution imagery from Google Earth and aerial photographs. This resulted in a total of 862 ground truth points. In accordance with Al-doski et al. (2013) and Lu et al. (2004), each land cover class was represented by at least 12 reference points, each corresponding to a minimum of 50 pixels. Seventy-five percent of the reference dataset was used for image classification, with the remaining portion reserved for accuracy validation. The 2025 classification was validated using 562 field-surveyed reference points collected during field campaigns conducted in September 2025, complemented by 300 reference samples derived from high-resolution imagery. For the 2015 and 2020 classifications, validation was undertaken using time-specific reference samples interpreted from historical Google Earth imagery and other temporally corresponding image archives. This approach ensured that each classification year was evaluated using reference data representative of the respective period, thereby minimizing temporal bias and improving the reliability of the accuracy assessment. The complete distribution of training and validation samples for each land-cover class is presented in Table 2.

TABLE 2. Training points collected in GEE for classification accuracy.

2.5 Change detection and identification

Examining LULC changes is essential for implementing effective land management interventions. In remote sensing, change detection is employed to analyse variations in the condition of objects over time using multi-temporal spatial datasets. The fundamental concept of utilizing satellite imagery for change detection lies in the principle that alterations in LULC are expressed as changes in spectral brightness. There are three primary methods of change detection in remote sensing: image subtraction, image ratio approach, and change detection after categorisation (Collins and Woodcock, 1996). Image subtraction detects changes in pixels based on their grey values. As the name suggests, the image ratio method is used to determine the ratios of pixels in each band of an image. Furthermore, it is the most apparent method after independent image processing and classification for different periods; hence, post-classification change detection was selected for this work. To calculate the area change, these maps were converted from raster to polygon due to the high accuracy of polygons in determining areas.

2.6 Random forest (RF)

Within the framework of the present study, RF was implemented as a core ensemble learning approach for integrated wetland modelling in the Driefontein Grasslands under a geospatial big data environment. Originally developed by Breiman, RF operates by constructing multiple classification trees from bootstrapped subsets of the training data and randomly selected predictor variables (Breiman, 2001). The final class allocation is determined through majority voting across the ensemble, thereby reducing variance and improving predictive stability. RF is particularly suitable for wetland detection due to its non-parametric and non-linear nature, which enables it to model complex spectral–environmental interactions characteristic of wetland–grassland mosaics. Furthermore, the algorithm provides variable importance measures, allowing for quantitative assessment of the contribution of spectral indices, topographic derivatives, hydrological proxies, and ancillary geospatial predictors to wetland discrimination. This capability is critical in understanding the drivers of wetland dynamics within semi-arid African landscapes. To ensure optimal predictive performance, hyperparameter tuning was performed using grid search coupled with 10-fold cross-validation on the training dataset. The search domain included the number of trees (n_estimators: 100, 200, 500), maximum tree depth (max_depth: None, 10, 20, 30), minimum samples required for node splitting (min_samples_split: 2, 4, 6), and minimum samples per terminal leaf (min_samples_leaf: 1, 2, 4). The optimal configuration—500 trees, unrestricted depth, a minimum of two samples per split, and one sample per leaf—was selected based on the highest mean ROC-AUC value. This optimized RF model demonstrated strong discriminatory capacity and temporal stability, supporting its suitability for large-scale, data-intensive wetland monitoring applications in Zimbabwe and comparable African environments.

2.7 Support vector machines (SVM)

In this study, SVM was implemented as a supervised learning algorithm for integrated wetland modelling within the Driefontein Grasslands geospatial big data framework. SVM is fundamentally a margin-based classifier that identifies the optimal separating hyperplane between classes by maximizing the distance between support vectors belonging to different categories. In the context of wetland detection, the algorithm discriminates between wetland and non-wetland land cover types by exploiting multidimensional feature space representations derived from spectral, topographic, and hydrological predictors. SVM is particularly effective in high-dimensional remote sensing applications because it can model complex, non-linear relationships using kernel functions. In this research, a radial basis function (RBF) kernel was adopted to accommodate the spectral heterogeneity and transitional boundaries typical of wetland–grassland interfaces. Hyperparameter optimisation was conducted using grid search with 10-fold cross-validation to enhance generalisation performance. The search space included the regularisation parameter (C: 0.1, 1, 10, 100), which controls the trade-off between margin maximisation and classification error, and the kernel coefficient (gamma: 0.001, 0.01, 0.1, 1), which determines the influence of individual training samples. The optimal configuration (C = 10, gamma = 0.01) achieved the highest cross-validated ROC–AUC score, indicating strong discriminative power. The integration of SVM within a geospatial big data environment represents a significant methodological advancement for wetland monitoring in Zimbabwe. By processing large volumes of multi-source satellite data and terrain attributes, SVM facilitates scalable, repeatable, and dynamically updateable wetland classification. Moreover, its capacity to model nonlinear environmental controls enhances predictive reliability in semi-arid African landscapes, thereby strengthening decision-support systems for sustainable wetland management and ecosystem resilience planning.

2.8 Classification and regression trees (CART)

In this study, CART was implemented as a baseline tree-based learning algorithm within the comparative framework for integrated wetland modelling in the Driefontein Grasslands under a geospatial big data environment. CART is a supervised, non-parametric decision tree approach that recursively partitions the feature space into homogeneous subsets based on threshold rules derived from predictor variables. At each node, the algorithm selects the variable and split point that maximally reduces impurity—commonly measured using the Gini index for classification tasks—thereby constructing an interpretable hierarchical decision structure. In the context of wetland detection, CART classifies pixels into wetland and non-wetland categories using multidimensional geospatial predictors, including spectral bands, vegetation and moisture indices, topographic derivatives (elevation and slope), and proximity to hydrological features. Its rule-based structure enables transparent interpretation of environmental thresholds governing wetland occurrence. To enhance model performance and reduce overfitting, hyperparameter optimisation was conducted using grid search with 10-fold cross-validation. The tuning process evaluated maximum tree depth (e.g., None, 10, 20, 30), minimum samples required for node splitting (2, 4, 6), and minimum samples per leaf (1, 2, 4). The optimal configuration was selected based on the highest cross-validated ROC–AUC score, balancing model complexity with generalisation capacity. Although CART offers computational efficiency and high interpretability—advantages in operational geospatial workflows—it is inherently sensitive to training data variability and may exhibit reduced stability in spectrally heterogeneous wetland–grassland mosaics. Within this study’s comparative design, CART serves as a benchmark against ensemble (RF) and margin-based (SVM) approaches, thereby strengthening the methodological robustness of integrated wetland modelling in semi-arid African environments.

2.9 Model evaluation and accuracy assessment

Model performance in this study was evaluated using the ROC curve and the AUC metric (Fawcett, 2006) within a Python 3.10 computational environment. The ROC–AUC framework was selected because it provides a threshold-independent measure of classifier discrimination capability, making it particularly suitable for wetland detection in spectrally heterogeneous and class-imbalanced environments such as the Driefontein Grasslands. All models (RF, SVM, and CART) were trained using the optimised hyperparameters described earlier, and evaluation was performed on an independent validation dataset withheld during model training. Probabilistic class outputs were generated using the predict_proba() function (for RF and CART) and probability calibration for SVM (via probability = True). These predicted probabilities were then compared against ground-truth labels to compute ROC curves using the roc_curve() function from the sklearn. metrics module. The AUC values were subsequently derived using the auc() and roc_auc_score() functions. The ROC curve plots the True Positive Rate (Sensitivity) against the False Positive Rate (1 – Specificity) across varying classification thresholds. The AUC statistic quantifies the model’s ability to distinguish wetland from non-wetland classes, with values approaching 1.0 indicating excellent discrimination and values near 0.5 suggesting random classification performance. All computations were implemented using widely adopted scientific libraries, including NumPy, Pandas, Scikit-learn, and Matplotlib, ensuring methodological transparency and reproducibility within a scalable geospatial big data workflow. Cross-validation (10-fold) was integrated during model optimisation to prevent overfitting and to ensure robust generalisation across spatially distributed training samples. The adoption of ROC–AUC as the primary performance metric represents a methodological advancement over conventional overall accuracy and Kappa statistics commonly used in African LULC studies (Jafarzadeh et al., 2022; Kipterer et al., 2025). By employing a threshold-independent diagnostic framework, this research provides a more rigorous and policy-relevant evaluation of wetland classification reliability, thereby strengthening confidence in model-derived spatial intelligence for ecosystem monitoring and sustainable land management planning. Because the study involved multi-class land use/land cover classification (forest, grassland, water, built-up, and bare land), ROC–AUC evaluation was implemented using a one-vs-rest (OvR) approach. Under this framework, each class was individually evaluated against all remaining classes, generating class-specific ROC curves and AUC values. The reported AUC statistics for RF, SVM, and CART represent macro-averaged AUC scores, calculated as the arithmetic mean of the individual class AUC values. This approach provides a threshold-independent and balanced assessment of classifier performance across all land-cover categories, regardless of class prevalence, thereby enabling robust comparison among the machine learning algorithms.

To complement the ROC–AUC analysis, conventional classification accuracy metrics were also evaluated. Overall classification accuracy ranged from 84.35% to 87.78%, while Kappa coefficients varied between 0.82 and 0.86, indicating strong agreement between classified outputs and reference data (Table 3). The consistently high Kappa values demonstrate that the machine learning models achieved substantial to almost perfect classification agreement beyond chance. RF produced the highest overall accuracies and Kappa statistics across the study periods, followed closely by SVM, while CART exhibited comparatively lower performance. These findings are consistent with the ROC–AUC results and further support the suitability of ensemble and kernel-based classifiers for wetland mapping in heterogeneous landscapes.

TABLE 3. Classification accuracy assessment of RF, SVM, and CART models.

2.10 Spectral remote sensing indices

To improve wetland discrimination and strengthen the predictive capability of the RF, SVM, and CART models, this study integrated a comprehensive suite of remotely sensed spectral indices derived from multi-temporal Landsat imagery for 2015, 2020, and 2025. The indices included the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Soil Adjusted Vegetation Index (SAVI), Optimized Soil Adjusted Vegetation Index (OSAVI), Modified Soil Adjusted Vegetation Index (MSAVI), Green Chlorophyll Index (GCI), Red Edge Chlorophyll Index (RECI), Structure Insensitive Pigment Index (SIPI), Normalized Difference Water Index (NDWI), and Modified Normalized Difference Water Index (MNDWI). The incorporation of these indices was directly linked to the prior machine learning analysis because each index enhanced the separability of wetland and non-wetland classes by capturing specific biophysical and hydrological characteristics of the landscape. Vegetation-sensitive indices such as NDVI, EVI, SAVI, MSAVI, OSAVI, GCI, and RECI improved detection of wetland vegetation vigour, chlorophyll concentration, biomass distribution, and moisture-related plant stress. Water-sensitive indices including NDWI and MNDWI enhanced identification of inundated surfaces and hydrologically active zones, while SIPI contributed to vegetation stress detection associated with wetland degradation. Together, these predictors strengthened the discriminatory performance observed in the ROC–AUC evaluation, particularly for RF and SVM, which effectively exploited nonlinear relationships among variables.

2.11 Limitations of the study

Several limitations should be considered when interpreting the findings of this study. First, the analysis relied primarily on optical Landsat 8 OLI imagery, which may be affected by atmospheric conditions, seasonal variability, and spectral confusion among wetland classes despite the application of pre-processing corrections. Second, the study focused on the Driefontein Grasslands only; therefore, the transferability of the results to other wetland systems with different ecological and hydrological characteristics requires further validation. Third, although overall model performance was assessed using ROC–AUC and inter-model correlation analysis, detailed class-specific accuracy assessments were not undertaken. Finally, while the workflow was implemented within a geospatial big data environment, differences in computational infrastructure and data availability may influence its application in other regions. Future studies should integrate Synthetic Aperture Radar (SAR) data, higher-resolution imagery, detailed class-level validation, and multi-site assessments to improve model robustness and generalizability across African wetland ecosystems.

3 Results

3.1 Comparison of RF, SVM and CART

The comparative LULC outputs derived from RF, SVM, and CART reveal substantial inter-model variability in class allocation across 2015, 2020, and 2025 in the Driefontein Grasslands wetlands (Figure 3; Table 4). RF and SVM exhibit broadly consistent trajectories, particularly in 2020 where forest extent peaks (104.05 km² in both models), suggesting improved discrimination of woody vegetation during that epoch. However, by 2025, RF and SVM indicate a marked expansion of bare land (16.59 km² and 2033 km², respectively) coupled with contraction of forest and grassland, implying progressive wetland degradation and surface exposure. In contrast, CART produces markedly different spatial distributions, with substantial forest expansion in 2025 (74.39 km²) and fluctuating water extents, indicating sensitivity to spectral heterogeneity and potential overfitting to localized patterns. Built-up areas decline sharply between 2015 and 2020 in RF and SVM but rebound moderately by 2025, reflecting dynamic anthropogenic pressures. Overall, ensemble-based RF and margin-based SVM demonstrate greater temporal coherence, whereas CART exhibits instability across epochs. The results suggest accelerated conversion of vegetated wetland surfaces to bare substrates by 2025, signalling hydrological stress and anthropogenic disturbance within the Driefontein wetland system.

TABLE 4. Areal extent of LULC classes in the Driefontein wetlands for the years 2015, 2020, and 2025 (km²), as classified by RF, SVM, and CART algorithms.

3.2 Model evaluation and accuracy assessment

The AUC metrics derived from the constructed ROC curves demonstrate consistently superior performance of RF across all epochs (2015–2025), with values ranging from 0.880 to 0.891 (Table 5; Figures 4a–c), indicating excellent discriminatory capability in wetland LULC classification. SVM follows closely, improving progressively from 0.850 (2015) to 0.873 (2025), reflecting enhanced generalization capacity under increasing data volume and spectral heterogeneity. In contrast, CART records comparatively lower AUC values (0.749–0.789), suggesting moderate predictive reliability and higher susceptibility to classification noise. Temporal trends show marginal performance gains for all models by 2025, consistent with improved training sample representation and algorithm tuning within the geospatial big data framework. Pairwise Pearson correlation (Figure 5) further reveals moderate positive association between RF and SVM (r = 0.49), indicating partial convergence in classification behaviour, while RF–CART (r = −0.45) and SVM–CART (r = −0.19) correlations are negative, highlighting structural divergence in CART outputs. Collectively, the results confirm that ensemble-based RF and margin-optimised SVM offer more stable and robust predictive capacity for integrated wetland modelling than single-tree CART (Figure 6), particularly in spectrally dynamic and fragmented wetland–grassland mosaics.

TABLE 5. AUC summary table (derived from constructed ROC curves).

Class-wise accuracy assessment further confirmed the superiority of RF and SVM over CART (Tables 6, 7). For the years 2015, 2020 and 2025 classification, RF achieved the highest User’s Accuracy (87.0%–90.0%), Producer’s Accuracy (86.5%–90.5%), and F1-scores (87.0%–90.2%) across all land-cover classes. SVM produced slightly lower but comparable accuracies, with F1-scores ranging from 85.2% to 89.0%. CART consistently recorded the lowest performance, particularly for built-up and grassland classes where F1-scores fell below 84%. Water and forest classes exhibited the highest classification accuracies across all models, whereas built-up and bare land classes showed relatively greater commission and omission errors due to spectral similarities with surrounding land-cover types. These findings are consistent with the ROC–AUC results and further demonstrate the enhanced capability of ensemble and kernel-based classifiers for wetland mapping in heterogeneous landscapes.

TABLE 6. User’s accuracy (UA), Producer’s accuracy (PA), and F1-Score (%) for 2015, 2020 and 2025 classification results.

TABLE 7. Commission error (CE) and omission error (OE) (%) for 2015, 2020 and 2025 classification results.

Pearson correlation analysis was used to quantify the degree of agreement between the spatial classification outputs generated by the RF, SVM, and CART algorithms (Figure 5). Correlation values range from −1 to +1, where positive values indicate similar spatial classification patterns, values close to zero indicate weak agreement, and negative values indicate contrasting classification structures. The moderate positive correlation between RF and SVM (r = 0.49) suggests that both algorithms identified broadly similar wetland and land cover patterns, whereas the negative correlations between RF and CART (r = −0.45) and between SVM and CART (r = −0.19) indicate substantial differences in how CART represented landscape features. These results highlight the greater consistency between RF and SVM and reinforce the superior reliability of RF for wetland modelling in the Driefontein Grasslands.

3.3 Spectral remote sensing indices

The multi-temporal spectral index analysis reveals substantial spatial and temporal variability in vegetation condition, hydrological dynamics, and wetland ecological integrity across the Driefontein Grasslands between 2015, 2020 and 2025 (Figures 7a, b). Vegetation-sensitive indices, including NDVI, EVI, SAVI, OSAVI, and MSAVI, consistently show stronger reflectance signatures within the central and southwestern wetland complexes, indicating areas of relatively high biomass productivity and persistent vegetation cover. However, progressive spatial fragmentation and localized declines in greenness intensity are evident by 2025, suggesting increasing ecological stress and vegetation degradation within portions of the wetland system. Water-sensitive indices (NDWI and MNDWI) indicate declining hydrological connectivity in several peripheral zones, while core wetland areas maintain relatively stable moisture signatures throughout the study period. The persistence of higher MNDWI values within central depressions suggests that some hydrologically active wetland nuclei remain functional despite broader environmental pressures. Similarly, chlorophyll-related indices such as GCI and RECI demonstrate spatial heterogeneity in vegetation physiological condition, with declining spectral intensity in some eastern sectors indicating possible nutrient stress, vegetation senescence, or anthropogenic disturbance. SIPI patterns reveal increasing vegetation stress across portions of the landscape, particularly toward 2025, which corresponds with the expansion of bare land and reduced grassland extent observed in the RF and SVM classifications. These spectral transformations align closely with the machine learning outputs and ROC–AUC performance trends, confirming that the integrated predictor-variable framework enhanced model discrimination capability.

Overall, the indices collectively demonstrate that wetland condition in the Driefontein Grasslands is undergoing spatially uneven ecological transformation characterized by localized vegetation decline, hydrological alteration, and increasing surface exposure. The integration of multiple vegetation and moisture indices therefore provided critical environmental intelligence for improving wetland detection accuracy and strengthening the ecological interpretability of machine learning-based wetland modelling.

4 Discussion

4.1 Comparison of RF, SVM and CART

The results demonstrate that machine learning algorithm selection substantially influences wetland land use/land cover (LULC) characterization in the Driefontein Grasslands. Random Forest (RF) and Support Vector Machine (SVM) produced more coherent and temporally stable classifications than Classification and Regression Trees (CART), reflecting their superior ability to model nonlinear relationships within heterogeneous wetland environments. Wetlands are characterized by complex interactions among vegetation, water, soil moisture, and seasonal inundation, resulting in highly variable spectral signatures. RF and SVM are well suited to such conditions because they effectively exploit multidimensional feature spaces and reduce classification uncertainty associated with mixed pixels (Kipterer et al., 2025).

The temporal patterns observed across the classified maps indicate important ecological changes within the wetland system. RF and SVM both identified an increase in built-up areas and bare land between 2015 and 2025, accompanied by fluctuations in grassland and forest cover. The expansion of built-up surfaces reflects increasing human settlement and agricultural activities around the wetland margins, while the growth of bare land may indicate vegetation degradation, soil exposure, and localized hydrological stress. Similar trends have been reported in Southern African wetlands where agricultural encroachment, recurrent droughts, and land-use intensification have altered wetland vegetation composition and ecosystem functioning (Matlhodi et al., 2021). The persistence of water bodies across all periods suggests that core wetland areas remain intact, although surrounding vegetated zones appear increasingly vulnerable to disturbance. The substantially larger Bare Land extent predicted by SVM in 2025 compared with RF reflects differences in classifier decision boundaries rather than actual landscape extent. SVM employs a margin-based optimization approach that can be highly sensitive to spectral overlap among bare land, sparsely vegetated grassland, and degraded wetland surfaces, particularly under semi-arid conditions. Consequently, some transitional pixels exhibiting mixed spectral responses were assigned to the Bare Land class by SVM, whereas RF distributed these pixels among grassland and other vegetation classes based on ensemble voting across multiple decision trees. The higher AUC values obtained by RF indicate greater robustness in handling heterogeneous wetland environments and complex non-linear relationships among predictor variables. Therefore, the discrepancy should be interpreted as an expression of algorithmic sensitivity to class separability rather than evidence of a five-fold increase in actual bare land extent. This finding further highlights the importance of comparing multiple classifiers when assessing environmentally sensitive and spectrally complex wetland systems.

A particularly important finding is the divergence of CART from both RF and SVM. Pearson correlation analysis revealed a moderate positive relationship between RF and SVM, indicating reasonable agreement in spatial representation. In contrast, CART exhibited negative correlations, implying fundamentally different classification structures. These negative relationships suggest that areas identified as a particular class by CART were frequently assigned to different classes by RF and SVM. This divergence is likely attributable to the hierarchical nature of single decision trees, which partition feature space through a sequence of binary splits. In spectrally heterogeneous wetland environments, small variations in predictor values can lead to markedly different decision pathways, increasing sensitivity to noise and reducing generalization capacity. Consequently, CART may overemphasize local spectral variability while failing to capture broader ecological gradients across the landscape. This behaviour is consistent with previous remote sensing studies demonstrating that single-tree classifiers are less robust than ensemble approaches when mapping complex ecosystems (Jafarzadeh et al., 2022). The negative Pearson correlations observed between CART and the other classifiers indicate that CART identified substantially different spatial patterns of wetland distribution and land-cover transitions. From an ecological perspective, this divergence suggests that CART was less effective in representing the complex vegetation–water gradients that characterize the Driefontein wetland system. Wetlands are inherently heterogeneous environments containing mixtures of open water, saturated soils, emergent vegetation, grasslands, and degraded surfaces. Because CART relies on a sequence of binary decision rules, it may oversimplify these ecological gradients and assign spectrally similar transitional zones to different classes than RF and SVM. Consequently, areas experiencing seasonal inundation, partial vegetation cover, or wetland degradation may be represented inconsistently. The negative correlations therefore reflect not only methodological differences but also contrasting ecological interpretations of landscape structure. In contrast, RF and SVM better captured the continuous and non-linear interactions among vegetation condition, moisture availability, and land-cover characteristics, resulting in more ecologically realistic wetland representations. These findings suggest that CART may be less suitable for monitoring dynamic wetland ecosystems where gradual ecological transitions rather than discrete boundaries dominate landscape patterns. Unlike RF, which aggregates hundreds of trees to minimize variance, CART relies on a single decision structure and is therefore more vulnerable to overfitting. The negative correlations observed in this study provide quantitative evidence of this limitation and represent an important methodological insight. Rather than merely showing lower accuracy, CART generated a substantially different interpretation of wetland structure and change. This finding highlights the importance of evaluating not only classification performance but also inter-model agreement when selecting algorithms for wetland monitoring.

Overall, the results suggest that RF and SVM provide more reliable representations of wetland dynamics in semi-arid African landscapes. Their consistency across time periods and stronger agreement with one another indicate greater suitability for long-term monitoring, ecosystem assessment, and environmental decision-making in data-rich geospatial environments.

4.2 Model evaluation and accuracy assessment

The ROC–AUC results provide strong evidence that RF and SVM are more effective than CART for wetland mapping in the Driefontein Grasslands. RF consistently achieved the highest AUC values, followed by SVM, while CART recorded lower values. According to established remote sensing performance thresholds, AUC values approaching 0.90 indicate excellent discrimination between classes, whereas values below 0.80 represent comparatively weaker predictive performance (Talukdar et al., 2020). The consistently high performance of RF demonstrates its ability to accurately distinguish spectrally similar wetland classes despite temporal variability and environmental heterogeneity. The superior performance of RF can be attributed to its ensemble architecture, which combines predictions from multiple decision trees and reduces classification variance (Halimu et al., 2019). This characteristic is particularly advantageous in wetlands where vegetation, water, and exposed soil often occur within the same pixel neighbourhood. Similarly, the strong performance of SVM reflects its capacity to identify optimal class boundaries within high-dimensional feature spaces through kernel transformation. The gradual increase in SVM AUC values over time suggests that the algorithm benefited from improved feature representation and increasing separability among land-cover classes.

The comparatively weaker performance of CART is reflected not only in lower AUC values but also in its structural disagreement with the other classifiers. The negative Pearson correlations indicate that CART produced spatial patterns that differed substantially from those generated by RF and SVM. This observation is significant because it demonstrates that classifier choice can influence ecological interpretation. For example, divergent class allocations can alter estimates of wetland extent, vegetation condition, and degradation trajectories. Consequently, relying solely on traditional accuracy measures may obscure important differences in model behaviour. Previous wetland-mapping studies in Southern Africa have similarly reported superior performance of RF and SVM compared with traditional decision-tree approaches, particularly in environments characterized by seasonal flooding, vegetation heterogeneity, and complex land-cover transitions (Matlhodi et al., 2021). The results of the present study therefore reinforce broader regional evidence while extending it through the incorporation of inter-model correlation analysis. By combining ROC–AUC diagnostics with Pearson correlation metrics, the study evaluates both predictive performance and structural agreement among classifiers. This provides a more comprehensive assessment than conventional approaches based solely on overall accuracy or Kappa statistics.

The findings have practical implications for operational wetland monitoring. Accurate and consistent classification is essential for detecting environmental change, identifying degradation hotspots, and supporting ecosystem management interventions. The superior stability of RF across all temporal periods indicates that it is the most appropriate classifier for monitoring dynamic wetland systems within geospatial big data environments. Furthermore, the observed divergence of CART highlights the need for caution when applying single-tree methods in spectrally complex landscapes. Collectively, these results demonstrate that model evaluation should extend beyond accuracy metrics to include analyses of classifier agreement and structural consistency. Such an approach improves confidence in ecological interpretations and provides a stronger scientific basis for wetland conservation planning. The framework presented here offers a robust methodology for future wetland studies in Zimbabwe and across Southern Africa, where reliable geospatial intelligence is increasingly required to support sustainable ecosystem management and long-term environmental monitoring.

4.3 Spectral remote sensing indices

The integration of multi-temporal spectral indices significantly strengthened the predictive and ecological interpretability of wetland modelling within the Driefontein Grasslands. Wetlands are inherently dynamic ecosystems characterized by fluctuating hydrological regimes, heterogeneous vegetation assemblages, and strong seasonal variability. Consequently, reliance on conventional spectral bands alone often limits accurate discrimination of wetland classes in semi-arid African environments. By incorporating vegetation, chlorophyll, moisture, and soil-adjusted indices into the RF, SVM, and CART workflows, this study advanced the capacity of machine learning algorithms to capture complex ecohydrological processes governing wetland dynamics (Mukomberanwa et al., 2024). Vegetation-sensitive indices such as NDVI, EVI, SAVI, MSAVI, and OSAVI proved particularly valuable in distinguishing wetland vegetation from surrounding grassland and degraded surfaces. The persistence of high vegetation index values within central wetland zones indicates sustained biomass productivity and moisture availability, while declining values in peripheral regions suggest increasing ecological stress. These findings are consistent with studies demonstrating that vegetation indices provide strong proxies for wetland health, primary productivity, and ecosystem resilience in water-limited landscapes. The observed decline in vegetation vigour toward 2025 further supports the broader LULC evidence of wetland degradation and bare land expansion detected by RF and SVM. Water-sensitive indices (NDWI and MNDWI) contributed substantially to hydrological discrimination by enhancing the identification of inundated surfaces and saturated soils. MNDWI demonstrated strong sensitivity to persistent wetland nuclei, supporting previous research showing its effectiveness in separating water bodies from built-up and vegetated surfaces. The persistence of hydrologically active cores despite surrounding degradation suggests partial ecological resilience within the Driefontein system, although the contraction of peripheral moisture zones indicates increasing hydrological fragmentation. This has important implications for biodiversity conservation and ecosystem service sustainability.

A major novelty of this study lies in the integrated use of a broad spectral-index ensemble within a geospatial big data machine learning framework tailored to African wetland ecosystems. Most previous wetland studies in Southern Africa have relied on a limited number of indices, typically NDVI and NDWI, or have focused solely on classification accuracy. In contrast, this research incorporated chlorophyll-sensitive (GCI, RECI), vegetation stress (SIPI), moisture-sensitive (NDWI, MNDWI), and soil-adjusted vegetation indices simultaneously, thereby enabling multidimensional characterization of wetland ecological condition. This represents a methodological advancement in GeoAI-driven ecosystem monitoring. Another important contribution is the linkage between spectral dynamics and algorithmic behaviour. The superior ROC–AUC performance of RF and SVM demonstrates that ensemble and kernel-based algorithms effectively exploit nonlinear relationships among spectral predictors, whereas CART exhibited comparatively weaker generalisation capacity. The study therefore provides new evidence that integrating diversified spectral indicators within advanced machine learning architectures improves both classification reliability and ecological interpretation in heterogeneous wetland–grassland mosaics.

From a broader policy perspective, the findings contribute directly to sustainable wetland governance in Zimbabwe and across Africa. Accurate monitoring of wetland degradation supports evidence-based ecosystem management aligned with African Union Agenda 2063 aspirations for environmentally sustainable development. Furthermore, the research contributes to SDG Target 6.6 through improved monitoring of water-related ecosystems and SDG Target 15.1 through enhanced conservation intelligence for inland freshwater habitats. Importantly, this study demonstrates how geospatial big data ecosystems combining Google Earth Engine, Python-based machine learning, and multi-temporal satellite archives can operationalize scalable wetland intelligence systems for data-constrained regions. The methodological framework therefore offers transferable solutions for continental-scale wetland monitoring, climate adaptation planning, and ecosystem restoration initiatives in Africa and beyond.

5 Conclusion

This study comparatively evaluated Random Forest (RF), Support Vector Machine (SVM), and Classification and Regression Trees (CART) for integrated wetland modelling within a geospatial big data environment in the Driefontein Grasslands, Zimbabwe. The results demonstrated that RF consistently achieved the highest predictive performance and temporal stability, followed by SVM, while CART exhibited lower predictive accuracy and significant structural divergence from the other classifiers. The combined use of ROC–AUC diagnostics and Pearson inter-model correlation analysis provided a more comprehensive assessment of classifier behaviour than conventional accuracy metrics alone. A major contribution of this study was the integration of multiple spectral vegetation and water indices, including NDVI, EVI, SAVI, OSAVI, MSAVI, GCI, RECI, SIPI, NDWI, and MNDWI, within a machine learning framework. These indices enhanced the discrimination of wetland vegetation, moisture conditions, chlorophyll dynamics, and vegetation stress, thereby improving classification performance and ecological interpretation. The results revealed notable changes in wetland land-cover composition between 2015 and 2025, including increasing built-up areas and localized expansion of bare land, suggesting growing environmental pressures and landscape transformation within the wetland system. Several limitations should be acknowledged. The study relied primarily on optical Landsat imagery, focused on a single wetland system, and did not undertake detailed class-specific accuracy assessment. Additionally, differences in computing environments may influence workflow transferability. Therefore, the observed land-cover changes should be interpreted as indicators of ecological change rather than direct measurements of wetland degradation processes. Future research should integrate Synthetic Aperture Radar (SAR), higher-resolution imagery, detailed class-level accuracy assessments, and multi-site validation across African wetlands. Such advances would strengthen operational wetland monitoring and support evidence-based conservation, restoration planning, and sustainable ecosystem management under Agenda 2063 and SDG Targets 6.6 and 15.1.

6 Recommendations

Based on the findings of this study, several recommendations are proposed to enhance wetland monitoring and management in Zimbabwe and similar African environments.

• Adopt Random Forest for operational wetland monitoring: Given its superior predictive performance and temporal stability, RF should be considered the preferred machine learning algorithm for routine wetland mapping and change detection in heterogeneous wetland ecosystems.
• Integrate multi-source Earth observation data: Future studies should combine optical imagery with Synthetic Aperture Radar (SAR) data to improve the detection of inundated areas and reduce limitations associated with cloud cover and seasonal variability.
• Establish continuous wetland monitoring systems: Environmental agencies should implement regular geospatial monitoring programmes using machine learning and remote sensing technologies to identify early signs of wetland degradation and support timely intervention measures.
• Expand monitoring to other wetlands: The modelling framework developed in this study should be tested and validated across multiple wetland systems in Zimbabwe and sub-Saharan Africa to assess its transferability and regional applicability.
• Strengthen policy integration: Wetland monitoring outputs should be incorporated into national land-use planning, ecosystem restoration programmes, climate adaptation strategies, and reporting frameworks supporting Agenda 2063 and SDG Targets 6.6 and 15.1.
• Improve validation procedures: Future research should undertake detailed class-specific accuracy assessments and incorporate larger field-based datasets to further enhance classification reliability and ecological interpretation.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Xiaojun Li, Southwest Jiaotong University, China

Reviewed by: Harineck Mayamiko Tholo, Malawi University of Science and Technology, Malawi

John Kapoi, Dedan Kimathi University of Technology, Kenya