Multi-temporal land use/land cover dynamics (2005–2025) and machine learning algorithm comparison in and around kruger national park

1 Introduction

Land use and land cover (LULC) change is a critical component of global environmental change, influencing ecosystem structure, biodiversity, and biogeochemical cycles (Nandi et al., 2026). Over recent decades, rapid anthropogenic activities such as agricultural expansion, urbanization, and infrastructure development have significantly altered natural landscapes, particularly in ecologically sensitive regions (Afuye et al., 2024). These changes are widely recognized as major drivers of biodiversity loss, ecosystem degradation, and disruptions to ecosystem services (Santos et al., 2021; Umuhoza et al., 2025) mentions that, in sub-Saharan Africa, savanna ecosystems are highly dynamic and sensitive to both climatic variability and anthropogenic pressures, including land-use change, grazing intensity, and alterations in fire regimes, which together shape vegetation structure and ecosystem functioning (Umuhoza et al., 2025). Kruger National Park, one of Africa’s largest and most renowned conservation areas, represents a complex mosaic of grassland, woodland, and shrubland ecosystems that support diverse wildlife populations (Umuhoza et al., 2025). Despite its protected status, the park and its surrounding buffer zones are increasingly influenced by external pressures, including land transformation, population growth, and resource exploitation (Kuempel et al., 2021; Jones et al., 2018; Smit et al., 2024). Consequently, understanding the spatial and temporal patterns of land-use and land-cover (LULC) change within and around the park is essential for informing biodiversity conservation strategies, improving ecosystem management, and supporting sustainable land-use planning in the region (Nandi et al., 2026).

Given the scale and complexity of these landscape dynamics, robust and consistent monitoring approaches are required to capture both spatial and temporal changes in land cover (Nxumalo et al., 2026). Remote sensing has emerged as a powerful tool for monitoring LULC dynamics over large spatial and temporal scales (Mungai et al., 2025). Satellite platforms such as Landsat 5, Landsat 8, and Sentinel-2 provide continuous, freely available imagery suitable for long-term environmental monitoring (Hosseini et al., 2026). These datasets enable consistent analysis of landscape changes from the early 2000s to the present, facilitating the detection of trends in vegetation dynamics, land degradation, and human-induced transformations (Nxumalo et al., 2026). Accurate classification of LULC from satellite imagery is fundamental to such analyses. Traditional parametric classification methods have largely been replaced by machine learning approaches, which offer improved performance in handling complex, nonlinear relationships in high-dimensional data (Bayable et al., 2025). Among these, algorithms such as Random Forest, Support Vector Machine, K-Nearest Neighbors, and Artificial Neural Network have been widely applied in LULC classification studies. Random Forest is known for its robustness to noise and overfitting, while Support Vector Machine performs well in high-dimensional feature spaces (Mutale et al., 2024). K-Nearest Neighbors offers simplicity and effectiveness in certain contexts, whereas Artificial Neural Networks can capture complex nonlinear patterns (Venkateswarlu and Gangula, 2024).

Despite the increasing use of machine learning in LULC mapping, there remains a need for systematic comparisons of multiple algorithms applied to long-term, multi-temporal datasets, particularly in ecologically important regions such as Kruger National Park (Bogale et al., 2025). A lot previous studies focus on single-date classifications or limited comparisons of algorithms, which limits the generalizability of their findings (Maxwell et al., 2018; Belgiu and Drăguţ, 2016; Yin et al., 2014; Talukdar et al., 2020; Venkateswarlu and Gangula, 2024). Furthermore, fewer studies integrate long-term time-series analysis with detailed change detection, particularly in recent periods of accelerated environmental change (Mishra et al., 2020; Tassi and Vizzari, 2020; Karasiak et al., 2022).

This study directly addresses these gaps by conducting a comprehensive multi-temporal analysis of LULC changes from 2005 to 2025 in and around Kruger National Park, with a particular focus on change detection between 2005 and 2015, 2015–2025, and 2005–2025. Four machine learning algorithms: Random Forest, Support Vector Machine, Gradient Tree Boosting, and K-Nearest Neighbors are evaluated and compared using consistent training and validation datasets. The necessity for this study stems from the ongoing pressures on this vital conservation area and the limitations of existing research in providing a comparative and holistic assessment of advanced classification techniques over long periods of time. The underlying focus of this research is on the global importance of biodiversity hotspots, which face dynamic anthropogenic and climate challenges. The study offers robust insights into policy formulation, conservation planning, and resource management in similar protected areas globally. The objectives of this study are to: (1) generate accurate LULC maps for multiple time periods (2005, 2015, and 2025), (2) assess and compare the performance of different machine learning algorithms (Random Forest, Support Vector Machine, Gradient Tree Boosting and K-Nearest Neighbors) using consistent training and validation datasets, and (3) quantify spatial patterns of land cover change, particularly in relation to human pressures and ecosystem dynamics. The anticipated findings of this research are expected to contribute to improved methodological understanding of LULC classification techniques, while also providing valuable insights into landscape changes in one of Africa’s most important conservation regions, such as Kruger National Park. Such information is critical for informing land management strategies, supporting biodiversity conservation, and guiding policy decisions in the face of ongoing environmental change.

2 Materials and methodology

2.1 Study area

The study was conducted in and around Kruger National Park, a major protected savanna ecosystem located in north-eastern South Africa (Carruthers, 1995). The study area extends approximately 40 km around Kruger National Park across portions of South Africa, Zimbabwe, and Mozambique within the broader Great Limpopo Transfrontier Conservation Area (GLTFCA) as shown in (Figure 1). The study area covers approximately 59,677.403 km² and represents one of the largest transboundary conservation landscapes in southern Africa, characterized by extensive savanna ecosystems, rich biodiversity, and interconnected protected areas managed collaboratively across international boundaries (Mpofu et al., 2026; Nyathi et al., 2025). Kruger National Park is situated within the semi-arid Lowveld savanna region and is characterized by diverse landscapes, including open grasslands, woodland savannas, shrublands, riverine vegetation, and riparian ecosystems (Venter et al., 2003).

The climate of the study area region is predominantly semi-arid to subtropical, with hot, wet summers and mild, dry winters. Mean annual rainfall ranges from approximately 450–750 mm, while summer temperatures frequently exceed 30 °C. Vegetation is dominated by mopane woodland, mixed bushveld, acacia savanna, shrubveld, and seasonal grasslands (Tobin and Kulmatiski, 2018; Hejda et al., 2022).

The study area supports exceptional biodiversity and forms part of one of Africa’s most significant wildlife conservation regions. The ecosystem hosts a wide variety of fauna, including elephants, lions, leopards, rhinoceroses, buffaloes, giraffes, zebras, and numerous antelope species (Chase et al., 2016; Ferreira and Funston, 2010). The region is internationally recognized for conservation and ecotourism. It forms a critical ecological corridor that facilitates wildlife movement among South Africa, Mozambique, and Zimbabwe within the Great Limpopo Transfrontier Conservation Area (GLTFCA) (Naidoo et al., 2024; Mpofu et al., 2026). Despite its protected status, the region has experienced increasing environmental pressure due to agricultural expansion, settlement growth, infrastructure development, deforestation, and climate variability (Laurance et al., 2014; Di Marco et al., 2018). Moreover, these pressures have contributed to substantial land-use and land-cover (LULC) changes in recent decades, particularly along the park boundaries and on communal lands surrounding the conservation area. Consequently, the area provides an ideal environment for evaluating machine learning algorithms for LULC classification and change detection across heterogeneous landscapes and transboundary ecosystems (Taubert et al., 2018; Mpofu et al., 2026).

2.2 Land use/land cover dataset

A two-pronged methodology, combining remote sensing analysis and field-based verification, was employed to estimate land use/land cover (LULC) dynamics over time in this study. Remote sensing technologies are widely recognized as effective tools for monitoring landscape change and environmental dynamics due to their ability to provide consistent, multi-temporal, and large-scale spatial data (Unger Holtz, 2007; Smith, 2014; Wulder et al., 2019). These approaches enable the detection and analysis of spatial and temporal variations in land cover, supporting accurate assessment of environmental change and land-use dynamics.

2.3 Methodological workflow

2.4 Acquisition of satellite data

LULC data were derived from multi-temporal Landsat satellite imagery processed in Google Earth Engine. This cloud-based geospatial analysis platform enables large-scale, long-term environmental monitoring through efficient time-series analysis of satellite data (Wulder et al., 2019; Gorelick et al., 2017). Landsat imagery was selected due to its long-term continuity, moderate spatial resolution (30 m), and proven suitability for land cover classification.

Multi-sensor datasets were utilized to ensure temporal coverage, including Landsat 5 Thematic Mapper for 2005; Landsat 8 OLI/TIRS for 2015, and Landsat 9 OLI-2/TIRS-2 for 2025. These time steps, selected at approximately 10-year intervals, allow for consistent assessment of long-term LULC dynamics.

All datasets were accessed and processed in Google Earth Engine and selected to represent full-year composites, ensuring adequate temporal coverage of seasonal variability and reducing the influence of short-term environmental fluctuations, such as cloud contamination and phenological differences. This approach improves the robustness and comparability of LULC classification results across the study period.

2.5 Pre-processing

Pre-processing of satellite imagery is a fundamental step in land use and land cover (LULC) classification, as it ensures radiometric consistency, reduces atmospheric distortions, and improves comparability across multi-temporal datasets (Song et al., 2001; Zhu, 2017). In this study, all pre-processing operations were conducted in Google Earth Engine using Landsat Surface Reflectance products. Atmospheric effects were addressed using Level-2 Surface Reflectance data, which are already corrected with established atmospheric correction algorithms. Specifically, the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) was applied to Landsat 5 Thematic Mapper data. At the same time, the Landsat Surface Reflectance Code (LaSRC) was used for Landsat 8 OLI/TIRS and Landsat 9 OLI-2/TIRS-2 datasets. These physics-based radiative transfer models correct for atmospheric scattering and absorption, ensuring that the derived reflectance values accurately represent surface conditions rather than atmospheric interference. The use of surface reflectance products is widely recommended for multi-temporal land use and land cover (LULC) analysis because it ensures radiometric consistency across long time-series datasets derived from different Landsat sensors. This consistency is essential for reliable detection of land-cover changes over time and for improving comparability between images acquired across different years (Zhu, 2017; Flood, 2014).

2.6 Feature extraction

Feature extraction plays a critical role in improving the performance of land use and land cover (LULC) classification by enhancing the separability of different surface types. In remote sensing-based analyses, the selection of appropriate features directly influences classification accuracy and model robustness. In this study, both original spectral bands from Landsat 5 Thematic Mapper, Landsat 8 OLI/TIRS, and Landsat 9 OLI-2/TIRS-2 imagery, as well as derived spectral indices, were utilized to capture the biophysical characteristics of the landscape (Lu & Weng, 2007; Jensen, 2015). The Normalized Difference Vegetation Index (NDVI) was used to quantify vegetation density and health (Xue and Su, 2017). In contrast, the Normalized Difference Water Index (NDWI) was applied to enhance the detection of surface water bodies and moisture content (Feyisa et al., 2020). The Normalized Difference Built-up Index (NDBI) was employed to improve the identification of built-up and urban areas (Zha et al., 2003; Sinha et al., 2016). The Soil-Adjusted Vegetation Index (SAVI) was used to minimize soil brightness effects in areas with sparse vegetation cover (Huete, 2012; Ren et al., 2018), while the Enhanced Vegetation Index (EVI) was applied to improve vegetation characterization by reducing atmospheric influences and minimizing soil background effects (Ren et al., 2018; Huete, 2012; Villamuelas et al., 2016) (Table 1).

TABLE 1. Vegetation indices (VIs) were used in this work and computed on Google Earth Engine (GEE).

2.7 Land use/land cover classes

Eight land use and land cover (LULC) classes were defined to represent the dominant landscape characteristics within and around Kruger National Park, as shown in Table 2. These classes were selected based on regional ecological conditions, previous LULC studies in savanna environments, and their relevance to environmental monitoring and land management. Such classification schemes are commonly structured to capture landscape heterogeneity and transitional vegetation patterns typical of savanna ecosystems, where land cover boundaries are gradual and strongly influenced by climatic variability and human activities (Babu and Nair, 2025; MohanRajan et al., 2020; Macarringue et al., 2022).

TABLE 2. The LULC classes considered in this study description.

2.8 Training and validation data

In this study, training samples were generated using a combination of field-based observations (GPS coordinates collected during field surveys within the study area), high-resolution imagery, and expert visual interpretation in Google Earth Engine. These multiple data sources ensured accurate labeling and improved the representativeness of each LULC class.

The dataset was divided into 70% for training and 30% for validation. This split provides a balance between model learning and independent performance evaluation. Importantly, the same training and validation datasets were used across all classification algorithms (Random Forest, Support Vector Machine, Gradient Boosting, and K-Nearest Neighbors) to ensure a fair and unbiased comparison of model performance. The integration of field observations, high-resolution imagery, and expert knowledge enhances the reliability of training datasets and reduces classification uncertainty. Previous studies have demonstrated that such hybrid approaches provide robust training data for machine learning-based LULC classification (Maxwell et al., 2018; Zhu, 2017).

Training samples were initially digitized as polygons representing homogeneous land-cover areas using high-resolution imagery in Google Earth Engine. Polygon-based sampling allows for more accurate delineation of land cover classes by capturing spatial context and minimizing labeling errors.

Subsequently, pixel-based samples were extracted from these polygons using a stratified sampling approach, converting them into point features for classification. This approach ensures that each training sample corresponds to a single pixel, thereby improving classification accuracy and maintaining consistency with pixel-based machine learning algorithms.

The combined use of polygon digitization and point-based sampling has been widely recommended in remote sensing studies to enhance training data quality and reduce classification uncertainty (Maxwell et al., 2018; Belgiu and Drăguţ, 2016).

Training samples were digitized as polygons and subsequently converted into pixel-based samples using the sampleRegions function in Google Earth Engine. A random sampling approach was applied to split the dataset into 70% training and 30% validation subsets, ensuring unbiased model evaluation. A total of 1,163 reference samples were collected for this study and distributed across the eight LULC classes, namely, water bodies, agricultural land, built-up areas, woodland/shrubland, grassland, bare land, natural forests, and plantation forests. To minimize sampling bias, samples were selected from homogeneous land-cover areas and were spatially distributed throughout the study area to capture the variability of each class. Following the 70/30 random split, 70% of the samples were used for model training and the remaining 30% were reserved for independent validation. The distribution of samples among the LULC classes is presented in Table 3.

TABLE 3. Distribution of training and validation samples.

2.9 Classification algorithms

Four machine learning algorithms were implemented for land-use and land cover (LULC) classification on Google Earth Engine. These algorithms were selected to represent different learning paradigms, including ensemble methods, margin-based classifiers, and instance-based approaches, allowing for a comprehensive comparison of classification performance.

To ensure a consistent and unbiased comparison among RF, SVM, GTB, and KNN classifiers, identical input variables were used for all models. Six spectrally equivalent Landsat bands (Blue, Green, Red, NIR, SWIR1, and SWIR2) together with selected spectral indices (NDVI, SAVI, EVI, NDBI, and NDWI) were employed as predictor variables for all classification experiments.

2.9.1 Random forest

Random Forest is an ensemble classification algorithm that constructs a large number of decision trees using bootstrap samples of the training data. Each tree is built using a random subset of input features, and the final classification is determined by majority voting across all trees.

Random Forest is particularly effective in remote sensing applications due to its ability to handle high-dimensional data, non-linear relationships, and noisy inputs. It is also less prone to overfitting than single decision trees and provides stable, accurate results across diverse landscapes (Belgiu and Drăguţ, 2016; Breiman, 2001).

2.9.2 Support vector machine

Support Vector Machine (SVM) is a supervised machine learning algorithm widely used for classification tasks that identifies optimal decision boundaries (hyperplanes) by maximizing the margin between different classes in the feature space. The theoretical foundation of SVM was introduced by Vapnik (Vapnik, 1998) and has since become a standard method in pattern recognition and remote sensing applications (Cortes and Vapnik, 1995; Vapnik, 1998; Zeng et al., 2019). To address non-linear relationships in the data, the Radial Basis Function (RBF) kernel was employed, which transforms the input data into a higher-dimensional space, thereby improving class separability. This kernel is widely used for land use and land cover (LULC) classification due to its strong performance in complex, heterogeneous landscapes (Maxwell et al., 2018, Feizi et al., 2022). SVM is particularly suitable for high-dimensional datasets and performs well even with limited training samples, making it a robust choice for remote sensing-based classification problems.

2.9.3 Gradient tree boosting

Gradient Tree Boosting is an advanced ensemble learning technique that builds decision trees sequentially, with each new tree trained to correct the errors of the previous model by minimizing a specified loss function via gradient descent. This approach is part of the broader family of boosting algorithms, which are designed to improve predictive performance by combining weak learners into a strong predictive model (Friedman, 2001; Chen and Guestrin, 2016). Gradient boosting methods are highly effective in capturing complex, non-linear relationships in data and often achieve higher predictive accuracy than traditional classification algorithms in remote sensing applications (Maxwell et al., 2018; Belgiu and Drăguţ, 2016). However, they are more sensitive to parameter tuning, overfitting, and noisy data than Random Forest models, which are generally more robust and stable across diverse datasets.

2.9.4 K-nearest neighbors

K-Nearest Neighbors (KNN) is a non-parametric, instance-based machine learning algorithm that classifies each pixel based on the majority class of its k nearest neighbors in feature space. The method assigns class labels by measuring similarity between samples using distance metrics, most commonly Euclidean distance. Unlike parametric models, KNN does not build an explicit predictive model during training; instead, it stores all training samples and performs computation at the classification stage (Progga et al., 2023; Hastie, 2009).

In remote sensing applications, KNN is valued for its simplicity and interpretability, particularly in land use and land cover (LULC) classification tasks involving well-separated feature spaces. However, its performance can be affected by high computational cost, sensitivity to irrelevant features, and reduced efficiency when applied to large datasets (Maxwell et al., 2018). Despite these limitations, KNN remains a useful benchmark classifier in comparative studies of machine learning algorithms for geospatial analysis.

2.10 Accuracy assessment

The performance of the land use and land cover (LULC) classification models was evaluated using an independent validation dataset derived from the 30% hold-out samples. Classification accuracy was assessed using a combination of standard statistical metrics, including the confusion matrix, Overall Accuracy (OA), Kappa coefficient, Producer’s Accuracy, and User’s Accuracy as presented in Table 3. These metrics provide complementary information for evaluating both overall model performance and class-level classification reliability, particularly in multi-class classification problems common in remote sensing (Foody, 2020; Olofsson et al., 2014).

The confusion matrix is a widely used evaluation tool in remote sensing that compares predicted class labels with reference (ground truth) data. It provides detailed information on classification errors, including omission and commission errors for each land cover class, thereby enabling a more comprehensive assessment of classification performance. Recent studies emphasize its continued importance as a standard framework for accuracy assessment in LULC mapping and machine-learning-based classification (Maxwell et al., 2018; Foody, 2020), alongside its integration with advanced performance metrics, such as F1-score, to improve interpretability.

Overall Accuracy (OA) is the proportion of correctly classified pixels among the total number of validation samples and provides a general measure of model performance. It is calculated as the ratio of correctly classified observations to the total number of observations, as shown in Table 3. Although OA is widely used in remote sensing studies, it does not account for agreement that occurs by chance and therefore should be interpreted alongside complementary accuracy metrics (Foody, 2020; Olofsson et al., 2014).

The Kappa coefficient measures the level of agreement between predicted and reference classifications while correcting for chance agreement. It has been widely applied in remote sensing to evaluate classification performance, particularly in multi-class problems (Foody, 2020). However, recent studies have highlighted the limitations of the Kappa statistic and recommend interpreting it alongside other metrics, such as Precision, Recall, and F1-score, for a more comprehensive evaluation of classification.

2.11 Change detection

Change detection is a fundamental technique in remote sensing used to quantify and analyze temporal changes in land use and land cover (LULC) by comparing multi-date satellite images. It enables the identification of spatial patterns of land transformation, such as vegetation loss, urban expansion, and agricultural development over time (Singh, 1989; Lu et al., 2004). In this study, post-classification comparison was adopted as the primary change detection approach due to its simplicity and effectiveness in minimizing errors arising from atmospheric and sensor differences between image dates, as reported in (Mas, 1999; Coppin et al., 2004).

The post-classification comparison method involves independently classifying satellite images from different years and then comparing the classified maps pixel-by-pixel to generate change information. This approach allows for the production of transition matrices that quantify the conversion from one land cover class to another over time. It is widely recognized as one of the most reliable techniques for LULC change analysis because it provides “from-to” change information, which is essential for understanding land transformation processes (Hussain et al., 2013; Lu et al., 2004). In this study, classified LULC maps for 2005, 2015, and 2025 were used to generate change detection outputs for three-time intervals: 2005–2015, 2015–2025, and 2005–2025. A transition matrix was developed for each period to quantify land-cover conversions among classes, as shown in Tables 4–6. The matrices were derived using raster algebra in ArcGIS Pro by combining two classified raster layers to produce unique transition codes representing “from-to” class changes.

TABLE 4. LULC area changes (km² and %) between 2015 and 2025 derived from RF, SVM, GTB, and KNN classifiers.

TABLE 5. LULC area changes (km² and %) between 2005 and 2025 derived from RF, SVM, GTB, and KNN classifiers.

Bare land showed strong reductions across all methods, suggesting conversion to other land uses. Natural forests generally increased in RF, GTB, and KNN, whereas SVM, showed a decline. Woodland/shrubland exhibited mixed trends, highlighting classification uncertainty across algorithms. Plantation forests showed a consistent decrease in most classifiers, indicating ongoing land cover transformation over the study period.

TABLE 6. Transition matrix for land use/land cover changes between 2005 and 2015 (km²).

W−body: Water body, AG−Land: Agriculture Land, B−Up: Built−Up, W−Land/Sh−Land: Woodland/shrubland, G−Land: Grassland, B−Land: Bare land, N−Forests: Natural forests, P−Forests: Plantation forests.

The Random Forest (RF) classified maps (Figure 2) were selected for the final change detection analysis due to their highest overall classification accuracy and Kappa coefficient among the tested machine learning algorithms (RF, SVM, GTB, and KNN). This ensured that the change detection results were based on the most reliable classification outputs. Furthermore, the resulting transition matrices were used to compute land cover gains, losses, and persistence across the study periods. These outputs provide detailed insights into the dynamics of LULC changes and support the interpretation of spatial patterns of environmental change within the study area.

3 Results

3.1 Classification performance of machine learning algorithms

3.1.1 Overall classification accuracy and kappa coefficient results

Four different classifiers were applied to three different satellite images to compare their performance. The performance of these four machine learning classifiers: Random Forest, Gradient Tree Boosting, Support Vector Machine, and K-Nearest Neighbors, was evaluated using independent validation data derived from a 70/30 training–validation split. However, Overall accuracy and the Kappa coefficient are the most commonly used methods to evaluate the accuracy of any classifier for classification. Table 7 and Figure 3 present the Overall Accuracy (OA) and Kappa coefficient results for the four machine learning classifiers, namely, RF, GTB, SVM, and KNN, for the years 2005, 2015, and 2025.

TABLE 7. Overall classification accuracy and kappa coefficient results for the study period.

For 2015, RF, again produced the best classification results, maintaining a high OA, of 89% and Kappa coefficient of 0.87. GTB, and KNN, also demonstrated strong performance, recording OA values of 86% and 85%, with corresponding Kappa coefficients of 0.84 and 0.83, respectively. SVM, remained the least accurate classifier, with an OA of 72% and Kappa coefficient of 0.66.

In 2005, the Random Forest (RF) classifier achieved the highest classification performance with an OA of 92% and a Kappa coefficient of 0.87, indicating excellent agreement between the classified and reference data. Gradient Tree Boosting (GTB) followed with an OA of 88% and a Kappa value of 0.82, while K-Nearest Neighbors (KNN) achieved an OA of 86% and a Kappa coefficient of 0.79. In contrast, Support Vector Machine (SVM) produced the lowest classification accuracy, with an OA of 73% and a Kappa coefficient of 0.55, indicating comparatively weaker classification performance.

Similarly, in 2025, RF maintained stable, consistent performance, with an OA of 89% and a Kappa coefficient of 0.87. GTB closely followed with an OA of 87% and a Kappa coefficient of 0.85, while KNN achieved an OA of 84% and a Kappa coefficient of 0.81. SVM again recorded the lowest performance among all classifiers, with an OA of 72% and a Kappa coefficient of 0.667.

3.1.2 Confusion matrix, Producer’s accuracy, and User’s accuracy

Figure 4 presents the User’s Accuracy (UA) and Producer’s Accuracy (PA) values derived from the confusion matrices for the RF, SVM, GTB, and KNN classifiers for 2005, 2015, and 2025. The accuracy assessment demonstrates varying classification performance among the four machine learning algorithms across different LULC classes.

In 2005, the Random Forest (RF) classifier achieved consistently high UA and PA values across most land cover classes, indicating strong classification reliability and reduced omission and commission errors. Water bodies and plantation forests recorded the highest accuracies, with UA values of 0.95 and 0.96, and PA values of 0.92 and 0.97, respectively. Agricultural land, built-up areas, and natural forests also showed strong performance under RF and Gradient Tree Boosting (GTB), with most accuracies exceeding 0.80. In contrast, Support Vector Machine (SVM) produced comparatively lower accuracies for several classes, particularly grassland (PA = 0.19), wetland/shrubland (UA = 0.45), and natural forests (PA = 0.28), indicating substantial classification confusion. K-Nearest Neighbors (KNN) demonstrated moderate classification performance, generally producing accuracies between those of RF/GTB and SVM. For 2015, classification performance improved for most classes across all classifiers. Water bodies achieved near-perfect accuracies for all algorithms, with both UA and PA values reaching 0.99 in several cases. RF and GTB again demonstrated superior and stable performance, particularly for agricultural land, built-up areas, and plantation forests, where most UA and PA values exceeded 0.85. SVM showed improvement compared to 2005, especially for grassland, where both UA and PA reached 1.00; however, lower accuracies were still observed for bare land (PA = 0.46) and natural forests (PA = 0.59). KNN maintained relatively good performance, particularly for agricultural land and built-up classes, though with slightly lower accuracies than RF and GTB.

Similarly, in 2025, RF and GTB continued to provide the most reliable classification results. Water bodies recorded perfect or near-perfect accuracy across all classifiers, with UA and PA values reaching 1.00 for RF, GTB, and KNN. Agricultural land and plantation forests also achieved very high accuracies, particularly under RF and GTB, with UA values up to 0.99 and PA values exceeding 0.90. SVM continued to exhibit weaker performance for several classes, especially wetland/shrubland and bare land, despite some improvement in grassland classification. Overall, RF consistently produced the highest and most balanced UA and PA values across the study period, indicating its robustness and suitability for LULC classification in heterogeneous landscapes.

3.2 Spatial distribution of land use/land cover

The classified land use and land cover (LULC) maps for 2005, 2015, and 2025 revealed substantial spatial and temporal dynamics across the study area. Overall, the results indicate a progressive expansion of anthropogenic land uses over time, particularly characterized by the growth of built-up areas around settlement zones and the expansion of agricultural land.

In contrast, natural vegetation classes, including woodland/shrubland and grassland, exhibited varying degrees of reduction and spatial fragmentation in several parts of the study area, reflecting increasing anthropogenic pressure on natural ecosystems. However, relatively stable vegetation patterns were observed in protected conservation areas, particularly in Kruger National Park, suggesting effective ecological preservation and management in these zones.

3.2.1 Spatial distribution of land use/land cover in 2005

The spatial distribution of LULC classes for 2005, derived from multiple classifiers, revealed a landscape predominantly dominated by natural vegetation, particularly woodland/shrubland, followed by grassland and bare land classes. Anthropogenic land uses such as agriculture and built-up areas were relatively less extensive but already evident in proximity to settlement zones.

Table 7 and Figure 5 summarize the area distribution of LULC classes derived from Random Forest (RF), Support Vector Machine (SVM), Gradient Tree Boosting (GTB), and K-Nearest Neighbor (KNN) classifiers, the 2005 LULC patterns results indicate that woodland/shrubland was the dominant land cover class across all classifiers, with RF estimating the highest extent (31,129.60 km²) and SVM showing a relatively lower estimate (26,162.70 km²). This dominance reflects the presence of extensive natural vegetation cover in the study area during 2005.

Grassland and bare land also accounted for substantial portions of the landscape, indicating a heterogeneous natural environment with varying vegetation density and exposed surfaces. Bare land estimates varied considerably among classifiers, with SVM producing the highest value (12,129.80 km²), suggesting potential overestimation due to spectral confusion between sparsely vegetated and exposed soil areas.

Agricultural land and built-up areas were relatively limited in extent compared to natural classes; however, RF results (2,139.21 km² and 746.40 km², respectively) indicate early stages of human expansion, particularly around settlement zones.

KNN consistently produced higher estimates for anthropogenic classes, suggesting a tendency toward over-classification of developed land. Water bodies and plantation forests occupied the smallest proportion of the landscape across all classifiers, although some variation exists due to spectral mixing in heterogeneous riparian zones.

3.2.2 Spatial distribution of land use/land cover in 2015

The 2015 classified LULC maps indicate notable spatial and compositional changes compared to 2005, reflecting increasing anthropogenic influence and progressive transformation of natural landscapes. Overall, the results show an expansion of built-up and agricultural land classes, accompanied by a relative reduction and fragmentation of natural vegetation in several parts of the study area.

Table 8 and Figure 6 present the spatial distribution of LULC classes for 2015 derived from the Random Forest (RF), Support Vector Machine (SVM), Gradient Tree Boosting (GTB), and K-Nearest Neighbor (KNN) classifiers.

TABLE 8. Land use/land cover areas (km²) derived from RF, SVM, GTB, and KNN classifiers for 2005.

The 2015 results reveal a more pronounced shift toward anthropogenic land uses compared to 2005. Built-up areas increased substantially across all classifiers, with KNN estimating the highest extent (2,879.91 km²) and RF indicating a moderate but consistent expansion (2,117.92 km²). This growth reflects ongoing urban expansion and settlement development, particularly around peri-urban zones.

Similarly, agricultural land showed considerable spatial expansion, especially under KNN (3,530.16 km²), suggesting intensification of farming activities and land conversion from natural vegetation. RF and GTB produced estimates that were relatively comparable (2,211.36 km² and 2,136.80 km², respectively), indicating a more stable and reliable classification of agricultural areas. Woodland/shrubland remained the dominant land cover class across all classifiers; however, a noticeable reduction in RF estimates from 2005 suggests a gradual conversion of natural vegetation into agricultural and built-up land. Interestingly, SVM produced substantially higher woodland/shrubland estimates (39,121.10 km²), indicating potential overclassification due to spectral confusion between vegetated and mixed land-cover types.

Grassland and bare land classes also exhibited variability across classifiers, with KNN yielding higher grassland estimates (11,743.60 km²), whereas SVM produced lower estimates, likely due to differences in class separability and model sensitivity to spectral heterogeneity. Water bodies and plantation forests remained the smallest land cover classes, showing relatively minor spatial changes. However, slight variations among classifiers suggest localized spectral mixing and classification uncertainty in riparian and managed plantation areas.

3.2.3 Spatial distribution of land use/land cover in 2025

The 2025 classified LULC maps indicate continued spatial transformation of the landscape, characterized by further expansion and redistribution of both anthropogenic and natural land cover classes. Compared to previous years, the results suggest ongoing pressure on natural ecosystems, particularly through the expansion of agricultural land and modifications in vegetation cover structure.

Table 9 and Figure 7 present the spatial distribution of LULC classes for 2025 derived from the Random Forest (RF), Support Vector Machine (SVM), Gradient Tree Boosting (GTB), and K-Nearest Neighbor (KNN) classifiers.

TABLE 9. Land use/land cover areas (km²) derived from RF, SVM, GTB, and KNN classifiers for 2025.

The 2025 results indicate continued landscape transformation, with clear evidence of expansion in both agricultural and built-up land classes across most classifiers. Agricultural land shows a notable increase in extent, particularly in KNN (3,612.99 km²) and RF (2,403.34 km²), suggesting intensified land conversion and agricultural expansion driven by increasing human demand for land resources. Built-up areas also exhibit continued growth, with KNN estimating the highest extent (1,520.57 km²), followed by RF (1,150.60 km²). This trend reflects ongoing urbanization and settlement expansion, particularly in peri-urban and accessible zones. However, SVM significantly underestimates the extent of built-up areas (287.44 km²), indicating potential spectral confusion between impervious surfaces and other land cover types.

Woodland/shrubland remains the dominant land cover class across all classifiers; however, RF and GTB results indicate a slight decline compared to 2015, suggesting gradual vegetation conversion and fragmentation. In contrast, SVM continues to overestimate woodland/shrubland (39,472.90 km²), reflecting its tendency to merge spectrally similar vegetated classes. Grassland shows an overall increase in RF and GTB outputs, indicating possible transition zones between forested and agricultural systems. Bare land, however, shows a general decrease in RF estimates compared to 2015, which may reflect land rehabilitation, vegetation regrowth, or conversion to other land-use categories. Water bodies and plantation forests remain relatively stable, with minor fluctuations across classifiers, although KNN slightly overestimates plantation areas compared to RF and SVM, likely due to classification uncertainty in managed vegetation zones.

3.3 Comparative analysis of land use/land cover changes

3.3.1 Comparative analysis of land use/land cover changes (2005–2015) using RF, SVM, GTB, and KNN classifiers

Table 9 and Figure 2 present the changes in land use/land cover classes between 2005 and 2015 derived from RF, SVM, GTB, and KNN classifiers. The results indicate substantial variations in class-specific area changes among the classifiers, although several common trends were observed.

Built-up areas increased considerably across all classifiers during the study period. The RF classifier estimated an increase of 1,371.5 km² (184%), while the KNN classifier recorded the highest increase of 1,740.8 km² (153%). This trend indicates rapid urban expansion and settlement growth within the study area. Natural forests showed notable increases in RF, GTB, and KNN classifications, with RF and GTB estimating increases exceeding 44%. However, SVM indicated only a marginal increase (3%), suggesting differences in classifier sensitivity for forest discrimination. However, Woodland/shrubland exhibited contrasting trends among classifiers. RF, GTB, and KNN estimated declines ranging from 5% to 10%, whereas SVM indicated a substantial increase of approximately 50%, reflecting classification inconsistencies among algorithms.

3.3.2 Comparative analysis of land use/land cover changes (2015–2025) using RF, SVM, GTB, and KNN classifiers

Table 10 presents the LULC changes between 2015 and 2025 derived from the RF, SVM, GTB, and KNN classifiers. Built-up areas declined across all classifiers, while grassland showed substantial increases, particularly in SVM and GTB. Woodland/shrubland increased in RF, SVM, and KNN classifications, whereas GTB indicated a slight decline. Bare land exhibited contrasting trends, with a major increase observed in SVM but decreases in RF, GTB, and KNN. Natural and plantation forests generally declined across most classifiers, indicating continued vegetation pressure during the study period.

TABLE 10. LULC area changes (km² and %) between 2005 and 2015 derived from RF, SVM, GTB, and KNN classifiers.

3.3.3 Comparative analysis of land use/land cover changes (2005–2015) using RF, SVM, GTB, and KNN classifiers

Table 10 summarizes the long-term LULC changes between 2005 and 2025 derived from the RF, SVM, GTB, and KNN classifiers. Water bodies consistently declined across all classifiers, indicating persistent hydrological reduction. Built-up areas showed an overall increasing trend in RF, GTB, and KNN, reflecting urban expansion, while SVM indicated a decrease. Grassland increased substantially across all classifiers, particularly under GTB and KNN.

3.4 Change detection analysis using RF classifier

The transition matrix derived from the ArcGIS Pro RF-classified LULC maps revealed substantial land cover transformations between 2005 and 2015 across the study area.

3.4.1 Change detection analysis (2005–2015) using RF classifier

The RF-based transition matrix revealed substantial land use and land cover transformations between 2005 and 2015 across the study area (Figure 8; Table 12). Woodland/shrubland remained the dominant and most persistent land cover class, with approximately 17,829.15 km² unchanged during the study period. Considerable persistence was also observed in grassland (3,829.45 km²), bare land (2,161.24 km²), natural forest (1,902.55 km²), and plantation forest (823.28 km²).

Despite this stability, significant land cover conversions occurred among vegetation-related classes. Woodland/shrubland was extensively transformed into grassland (5,442.51 km²), natural forest (2,346.86 km²), and bare land (1,942.00 km²), indicating vegetation modification and degradation processes. Grassland also showed substantial conversion into woodland/shrubland (4,815.16 km²), reflecting vegetation succession and spatial redistribution of vegetative cover.

Natural forests experienced notable conversion into woodland/shrubland (2,905.69 km²) and agricultural land (540.14 km²), suggesting forest fragmentation and agricultural encroachment. Agricultural land displayed dynamic interactions with surrounding vegetation classes, particularly transitions into woodland/shrubland (714.61 km²) and natural forest (398.62 km²). Built-up areas also exhibited considerable interaction with woodland/shrubland and bare land, indicating increasing anthropogenic activities and settlement expansion within the study area.

Overall, the 2005–2015 transition analysis highlights substantial vegetation transformation, agricultural redistribution, and urban expansion, with dominant transitions occurring among woodland/shrubland, grassland, agricultural land, and natural forest classes.

3.4.2 Change detection analysis (2005–2015) using RF classifier

The transition matrix analysis for the 2015–2025 period revealed continued land cover dynamics across the study area (Table 11), with woodland/shrubland remaining the dominant land cover class. Approximately 17,511.38 km² of woodland/shrubland remained unchanged, indicating the relative persistence of natural vegetation during the study period.

TABLE 11. Transition matrix for land use/land cover changes between 2015 and 2025 (km²).

W−body: Water body, AG−Land: Agriculture Land, B−Up: Built−Up, W−Land/Sh−Land: Woodland/shrubland, G−Land: Grassland, B−Land: Bare land, N−Forests: Natural forests, P−Forests: Plantation forests.

TABLE 12. Transition matrix for land use/land cover changes between 2005 and 2025 (km²).

W−body: Water body, AG−Land: Agriculture Land, B−Up: Built−Up, W−Land/Sh−Land: Woodland/shrubland, G−Land: Grassland, B−Land: Bare land, N−Forests: Natural forests, P−Forests: Plantation forests.

TABLE 13. Accuracy assessment equations for LULC used in this study.

Significant transitions were observed among vegetation-related classes. Woodland/shrubland experienced major conversion to grassland (4,944.30 km²), bare land (3,580.27 km²), and natural forest (3,101.51 km²), reflecting ongoing vegetation modification and landscape transformation. Similarly, grassland showed substantial conversion into woodland/shrubland (4,721.83 km²), suggesting vegetation succession and spatial redistribution of vegetative cover.

Agricultural land remained relatively stable, with 473.08 km² retained as agriculture, although notable transitions into woodland/shrubland (594.29 km²) and natural forest (590.87 km²) were also recorded. Built-up areas demonstrated continued expansion and interaction with surrounding land cover classes, particularly woodland/shrubland and bare land, indicating increasing anthropogenic influence and settlement growth within the study area. Natural forests experienced considerable conversion into woodland/shrubland (1,752.33 km²), while plantation forests largely remained stable, with 901.02 km² persisting during the study period. Water bodies also exhibited moderate transitions with surrounding land cover classes, likely associated with hydrological variability and seasonal environmental changes.

Overall, the 2015–2025 transition analysis highlights continued vegetation dynamics, urban expansion, and land degradation processes across the study area, with dominant transitions occurring among woodland/shrubland, grassland, agricultural land, and bare land classes.

3.4.3 Overall change detection analysis (2005–2025) using the RF classifier

The overall transition matrix for the 2005–2025 period revealed substantial long-term land use and land cover transformations across the study area (Table 14). Woodland/shrubland remained the dominant and most persistent land cover class, with approximately 18,293.78 km² remaining unchanged over the 20-year period. Significant persistence was also observed in grassland (2,839.34 km²), bare land (1,855.69 km²), natural forest (1,458.27 km²), and plantation forest (790.26 km²).

TABLE 14. Land use/land cover areas (km²) derived from RF, SVM, GTB, and KNN classifiers for 2015.

Despite this persistence, major land cover conversions occurred among vegetation-related classes. Woodland/shrubland experienced considerable transformation into grassland (5,846.45 km²), natural forest (2,725.62 km²), and bare land (2,610.17 km²), indicating ongoing vegetation modification and landscape restructuring. Similarly, grassland showed substantial conversion into woodland/shrubland (4,812.11 km²), suggesting dynamic vegetation succession processes within the study area.

Agricultural land exhibited notable interactions with surrounding vegetation classes, particularly transitions into woodland/shrubland (642.77 km²) and natural forest (632.12 km²), which may reflect vegetation recovery and shifting land management practices. Built-up areas demonstrated continued expansion and spatial interaction with woodland/shrubland and bare land classes, highlighting increasing anthropogenic activities and settlement development across the region. Natural forests experienced significant conversion into woodland/shrubland (1,843.22 km²), while plantation forests largely remained stable, with 790.26 km² persisting during the study period. Water bodies also showed moderate fluctuations and transitions with adjacent land cover classes, likely associated with hydrological and environmental variability.

Overall, the long-term RF-based transition analysis indicates that the study area underwent substantial vegetation transformation, land degradation, agricultural redistribution, and urban expansion between 2005 and 2025. The dominant transitions involved exchanges among woodland/shrubland, grassland, agricultural land, and natural forest, reflecting the dynamic nature of land-cover processes within the transboundary landscape.

4 Discussion

4.1 Overview of the main findings

This study evaluated long-term Land Use/Land Cover (LULC) dynamics in and around Kruger National Park between 2005 and 2025 using four machine learning classifiers and Random Forest (RF)-based transition analysis. Overall, the findings demonstrate that RF provided the most accurate and consistent classification performance among the evaluated algorithms (Table 13; Figures 3, 9), while Gradient Tree Boosting (GTB) and K-Nearest Neighbors (KNN) also produced reliable results. In contrast, Support Vector Machine (SVM) exhibited comparatively lower classification performance, particularly for spectrally similar vegetation classes. The superior performance of RF is consistent with its well-documented ability to model complex, nonlinear relationships, handle high-dimensional datasets, and reduce overfitting through ensemble learning, making it particularly suitable for heterogeneous landscapes characterized by high spectral variability.

The spatial distribution and change detection analyses (Figures 2, 4–7; Tables 4–10, 14) revealed substantial landscape transformation over the 20-year study period. Woodland/shrubland remained the dominant land-cover class throughout the study period, reflecting the ecological importance of the savanna ecosystem within the Greater Limpopo Transfrontier Conservation Area. Nevertheless, this vegetation experienced progressive fragmentation and conversion into agricultural land, grassland, and bare land, indicating increasing anthropogenic pressure on natural ecosystems. Concurrently, agricultural land and built-up areas expanded, particularly in areas surrounding settlements and communal lands, while water bodies exhibited an overall declining trend. These patterns collectively indicate that the landscape is undergoing continuous ecological and socio-economic transformation driven by interactions between natural processes and human activities.

The transition matrix analysis further demonstrated that, despite considerable land-cover persistence within the protected core of Kruger National Park, substantial exchanges occurred among woodland/shrubland, grassland, agricultural land, and natural forest classes. These transitions highlight the dynamic nature of savanna ecosystems, where vegetation succession, land degradation, agricultural expansion, and environmental variability operate simultaneously. The relative stability observed within the protected areas, compared with the more pronounced changes along park boundaries and adjacent communal landscapes, further emphasizes the important role of protected-area management in maintaining ecosystem integrity despite increasing external pressures.

Overall, the findings demonstrate that integrating machine learning techniques with multi-temporal Landsat imagery provides a robust framework for monitoring long-term LULC dynamics across complex transboundary conservation landscapes. Beyond identifying patterns of landscape transformation, the study contributes to understanding the interactions between environmental processes and anthropogenic activities, providing valuable information to support biodiversity conservation, sustainable land management, and evidence-based planning within Kruger National Park and similar protected savanna ecosystems. The following sections discuss these findings in greater detail by interpreting the performance of the machine learning algorithms, examining the environmental and socio-economic drivers of the observed changes, comparing the results with previous national and international studies, and outlining their implications for conservation and future research.

4.2 Performance of machine learning algorithms for LULC classification

The comparative assessment of the four machine learning classifiers demonstrated clear differences in their ability to classify the heterogeneous land use and land cover (LULC) classes within the Kruger National Park landscape. Overall, the Random Forest (RF) classifier consistently achieved the highest Overall Accuracy and Kappa coefficient throughout the study period (Table 13; Figures 3, 8), outperforming Gradient Tree Boosting (GTB), K-Nearest Neighbors (KNN), and Support Vector Machine (SVM). These findings confirm that ensemble-based machine learning approaches provide a robust framework for LULC mapping in heterogeneous savanna environments. Similar results have been reported by Breiman (2001), Rodríguez-Galiano et al. (2012), and Belgiu and Drăguţ (2016), who demonstrated that RF effectively models complex nonlinear relationships while minimizing overfitting.

The superior performance of RF is closely related to the ecological characteristics of the study area. Kruger National Park and the surrounding Greater Limpopo Transfrontier Conservation Area consist of complex mosaics of woodland, shrubland, grassland, agricultural land, bare land, riparian vegetation, plantation forests, and expanding settlements. Such heterogeneous landscapes generate considerable spectral variability due to differences in vegetation structure, soil background, seasonal phenology, and land management. RF effectively accommodates these complex spectral patterns without requiring assumptions about data distribution, making it particularly suitable for heterogeneous ecosystems.

In contrast, SVM consistently produced lower classification accuracies, particularly for woodland/shrubland, bare land, and natural forest classes (Figure 3). This reduced performance is likely associated with spectral similarity among vegetation classes and the fragmented nature of the savanna landscape, where mixed vegetation communities and transitional ecotones increase classification uncertainty. The confusion matrix further confirmed that RF produced more balanced Producer’s and User’s Accuracy values across nearly all land-cover classes, indicating lower omission and commission errors than the other classifiers.

Despite the strong performance of RF, the results should be interpreted considering the limitations of medium-resolution Landsat imagery. The 30 m spatial resolution inevitably results in mixed pixels, particularly along woodland–grassland transitions, riparian zones, and fragmented peri-urban landscapes where multiple land-cover types occur within a single pixel. These limitations may contribute to classification uncertainty regardless of the algorithm used. Nevertheless, Landsat remains one of the most valuable datasets for long-term environmental monitoring because of its consistent historical archive and global coverage, enabling robust multi-decadal analyses.

Overall, the findings demonstrate that RF is the most suitable classifier for long-term LULC monitoring in heterogeneous savanna ecosystems such as Kruger National Park. Its ability to accommodate spectral heterogeneity and nonlinear relationships resulted in more reliable classification outputs than SVM, GTB, and KNN, supporting its application for regional environmental monitoring and change detection. The interpretation of these results should also consider the intrinsic limitations of applying advanced pixel-based machine learning classifiers, such as RF and GTB, to medium-resolution Landsat imagery within a highly fragmented and heterogeneous savanna ecosystem. Although RF consistently produced the highest classification accuracy, the 30 m spatial resolution limits the ability of pixel-based classifiers to resolve fine-scale landscape features and mixed pixels occurring along ecotone boundaries, particularly woodland–grassland transitions, riparian corridors, and fragmented peri-urban environments. These spectral mixtures may have contributed to some of the classification uncertainty and the observed anomalies in built-up and vegetation classes during the 2015–2025 period. Nevertheless, Landsat remains one of the most appropriate datasets for long-term regional LULC monitoring because of its consistent historical archive, while future studies may benefit from integrating higher-resolution optical and radar imagery to improve the discrimination of complex land-cover classes.

4.3 Interpretation of long-term land use/land cover dynamics

The spatial distribution maps (Figures 4–6), comparative LULC analyses (Tables 7–9, 14), and RF-based transition matrices (Tables 4–6; Figure 2) demonstrate that the landscape surrounding Kruger National Park underwent substantial transformation between 2005 and 2025. Although woodland/shrubland remained the dominant land-cover class, the transition analysis revealed progressive fragmentation and conversion into agricultural land, grassland, and bare land, indicating increasing anthropogenic pressure combined with natural ecological processes. Similar long-term vegetation changes have been reported across southern African savanna ecosystems, where land-cover dynamics are driven by interactions among climate variability, land-use intensification, and population growth.

The persistence of woodland/shrubland within the protected core of Kruger National Park, contrasted with greater landscape modification around park boundaries, highlights the effectiveness of protected-area management while emphasizing the vulnerability of surrounding communal lands. Agricultural land and built-up areas expanded throughout the study period, reflecting increasing demand for food production, settlement development, and infrastructure. Although these changes support socio-economic development, they may also reduce habitat connectivity and place additional pressure on biodiversity if not carefully managed.

The observed decline in water bodies is environmentally significant because water availability regulates vegetation productivity, wildlife distribution, and ecosystem resilience within semi-arid savanna ecosystems. This trend may reflect the combined influence of rainfall variability, increasing temperatures, changing hydrological conditions, and growing human demand for water resources. Continued monitoring of surface water is therefore essential for understanding long-term environmental change within the Greater Limpopo Transfrontier Conservation Area.

An important finding of this study is the substantial bidirectional transition between woodland/shrubland and grassland identified in the RF transition matrices (Tables 4–6). These exchanges suggest that the savanna ecosystem remains highly dynamic, with vegetation succession, regeneration, disturbance, and seasonal variability occurring simultaneously. Consequently, some land-cover transitions likely represent natural ecological processes, whereas others reflect anthropogenic landscape modification.

Overall, the long-term LULC dynamics indicate that the landscape surrounding Kruger National Park is undergoing continuous ecological transformation driven by the interaction of environmental processes and human activities. While the protected core remains comparatively stable, increasing land conversion and vegetation fragmentation in adjacent areas may progressively affect habitat connectivity, ecosystem resilience, and biodiversity conservation. These findings highlight the importance of continuous LULC monitoring using robust machine learning approaches to support adaptive conservation management and sustainable land-use planning.

4.4 Influence of local environmental and socio-economic drivers on land use/land cover dynamics

The observed land use and land cover (LULC) changes within and around Kruger National Park are closely associated with the interaction of environmental processes and socio-economic activities operating across the Greater Limpopo Transfrontier Conservation Area (GLTFCA). As one of the largest transboundary conservation landscapes in southern Africa, the study area is characterized by a semi-arid savanna ecosystem where vegetation dynamics are strongly influenced by rainfall variability, temperature fluctuations, fire regimes, herbivory, and increasing human activities (Venter et al., 2003; Tobin and Kulmatiski, 2018). Consequently, the spatial patterns identified in this study (Figures 2, 4–7; Tables 4–10, 14) should be interpreted within the broader ecological and socio-economic context of this dynamic landscape.

One of the principal environmental drivers influencing LULC dynamics is climate variability. The study area experiences marked interannual variability in rainfall, with annual precipitation ranging from approximately 450–750 mm and summer temperatures frequently exceeding 30 °C. Such climatic conditions strongly regulate vegetation productivity, surface water availability, and ecosystem resilience. Periods of below-average rainfall or prolonged drought can reduce vegetation cover, increase the extent of bare land, and contribute to the contraction of surface water bodies, while wetter periods may promote vegetation regeneration and woodland recovery. The bidirectional transitions observed between woodland/shrubland and grassland in the RF transition matrices (Tables 4–6) are therefore likely to reflect not only anthropogenic land conversion but also the natural ecological dynamics of savanna ecosystems responding to climatic variability and vegetation succession.

Fire also represents an important ecological process shaping vegetation structure throughout Kruger National Park. Natural and prescribed fires play a fundamental role in maintaining the balance between woody vegetation and grassland by influencing tree recruitment, grass biomass, nutrient cycling, and species composition. In combination with herbivory by large mammals such as elephants, buffalo, and other grazers, fire contributes to the continuous restructuring of vegetation communities across the landscape. Consequently, some of the transitions observed among woodland/shrubland, grassland, and natural forest classes may reflect long-term ecological processes rather than permanent land degradation alone. Recognizing these natural disturbance regimes is therefore essential when interpreting LULC changes within protected savanna ecosystems.

In addition to environmental factors, increasing human activities have become major drivers of landscape transformation, particularly in areas surrounding the protected boundaries of Kruger National Park. The expansion of agricultural land and built-up areas observed in this study reflects growing population pressure, increasing demand for food production, settlement development, and infrastructure expansion within surrounding communal lands. These findings are consistent with previous studies demonstrating that agricultural expansion remains one of the leading causes of vegetation loss and habitat fragmentation across southern Africa (Laurance et al., 2014; Di Marco et al., 2018). The conversion of woodland/shrubland into agricultural land identified in the transition matrices (Tables 4–6) suggests that land-use intensification continues to alter ecosystem structure beyond the boundaries of formally protected areas.

The transboundary nature of the GLTFCA further contributes to the complexity of land-cover dynamics. Because the study area spans portions of South Africa, Mozambique, and Zimbabwe, differences in land-use policies, governance systems, socio-economic conditions, and resource management practices may influence the spatial distribution and intensity of land-cover change. While Kruger National Park itself benefits from relatively strong conservation management, adjacent communal and cross-border landscapes often experience varying degrees of agricultural expansion, fuelwood harvesting, grazing pressure, and settlement growth. These differences help explain why greater landscape transformation was observed around park boundaries than within the protected core, emphasizing the importance of coordinated transboundary management strategies for maintaining ecological connectivity and biodiversity conservation.

Tourism also represents an important socio-economic factor within the region. Kruger National Park is one of Africa’s premier ecotourism destinations, generating substantial economic benefits while supporting conservation initiatives. However, tourism development may also stimulate infrastructure expansion, including roads, accommodation facilities, and associated services, particularly near park entrances and surrounding settlements. Although tourism provides strong incentives for biodiversity conservation, balancing economic development with environmental protection remains essential to ensure the long-term sustainability of the ecosystem.

Overall, the findings indicate that the observed LULC dynamics cannot be attributed to a single driving force. Instead, they result from the combined influence of climatic variability, ecological processes, agricultural expansion, settlement growth, infrastructure development, and transboundary socio-economic interactions operating across the Greater Limpopo Transfrontier Conservation Area. Understanding these interacting drivers is fundamental for developing effective conservation strategies, improving land-use planning, and strengthening adaptive ecosystem management in one of Africa’s most ecologically significant protected landscapes.

4.5 Comparison with national and international studies

The findings of this study are consistent with previous research demonstrating that machine learning algorithms are effective tools for mapping long-term land use and land cover (LULC) dynamics in heterogeneous landscapes. Similar to studies conducted in South African savanna ecosystems, the observed fragmentation of woodland/shrubland and expansion of agricultural and built-up areas indicate increasing anthropogenic pressure associated with population growth, agricultural intensification, and infrastructure development (Venter et al., 2003; Laurance et al., 2014; Di Marco et al., 2018). Recent studies within the Greater Limpopo Transfrontier Conservation Area have also reported greater landscape transformation around protected-area boundaries, while the protected core remains relatively stable due to effective conservation management (Mpofu et al., 2026; Nyathi et al., 2025).

Comparable trends have been reported in other African conservation landscapes, including Zimbabwe, Mozambique, Kenya, and Tanzania, where agricultural expansion, settlement growth, and climate variability have contributed to vegetation fragmentation and habitat modification. These similarities suggest that the LULC dynamics observed in this study reflect broader environmental and socio-economic processes affecting savanna ecosystems across sub-Saharan Africa.

From a methodological perspective, the superior performance of the Random Forest classifier agrees with previous studies showing that ensemble learning algorithms consistently outperform conventional classifiers in complex and heterogeneous environments (Breiman, 2001; Rodríguez-Galiano et al., 2012; Belgiu and Drăguţ, 2016). The high classification accuracy achieved by RF demonstrates its suitability for long-term environmental monitoring using multi-temporal Landsat imagery.

The present study contributes to the existing literature by integrating the comparative evaluation of four machine learning classifiers with RF-based transition matrix analysis over a 20-year period within a large transboundary conservation landscape. This integrated approach provides a more comprehensive understanding of both classification performance and long-term landscape dynamics, offering valuable information for biodiversity conservation, land-use planning, and environmental management in protected areas.

4.6 Implications for conservation, land management and policy

The findings of this study have important implications for biodiversity conservation, sustainable land management, and environmental policy within Kruger National Park and the Greater Limpopo Transfrontier Conservation Area (GLTFCA). Although the protected core of Kruger National Park remained relatively stable, the continued expansion of agricultural land and built-up areas, together with the fragmentation of woodland/shrubland (Figures 2, 4–7; Tables 4–6), indicates increasing anthropogenic pressure on surrounding landscapes. These changes may reduce habitat connectivity, alter ecosystem processes, and threaten biodiversity if not effectively managed.

The superior performance of the Random Forest classifier demonstrates the value of integrating machine learning with multi-temporal satellite imagery for long-term environmental monitoring. Accurate and continuous LULC monitoring can support the early detection of vegetation degradation, agricultural encroachment, urban expansion, and hydrological changes, enabling timely conservation and land management interventions.

The RF-based transition matrices further provide valuable information on land-cover persistence and conversion pathways, which can assist conservation agencies in identifying priority areas for ecological restoration and evaluating the effectiveness of existing management strategies. In transboundary conservation landscapes such as the GLTFCA, these findings can strengthen collaborative environmental planning among South Africa, Mozambique, and Zimbabwe.

Overall, this study demonstrates that combining machine learning with long-term Earth observation data provides a practical framework for supporting sustainable land-use planning, biodiversity conservation, and evidence-based environmental decision-making in protected and transboundary ecosystems.

4.7 Study limitations and future research

Although this study demonstrated the effectiveness of machine learning for long-term LULC classification and change detection, several limitations should be acknowledged. First, the analysis relied on Landsat imagery with a 30 m spatial resolution. While Landsat provides an invaluable long-term archive for environmental monitoring, its medium spatial resolution may result in mixed pixels, particularly within heterogeneous savanna landscapes where vegetation, agricultural land, and built-up areas frequently occur within the same pixel. This limitation may contribute to classification uncertainty, especially along ecotones and fragmented peri-urban environments.

Second, although training samples were developed using field observations, high-resolution imagery, and expert visual interpretation within Google Earth Engine, uncertainty associated with historical reference data cannot be completely eliminated. Nevertheless, the use of independent validation data and multiple accuracy assessment metrics, including Overall Accuracy, Producer’s Accuracy, User’s Accuracy, and the Kappa coefficient, provides confidence in the robustness of the classification results.

Future research should investigate the integration of higher-resolution remote sensing datasets, such as Sentinel-1, Sentinel-2, PlanetScope, UAV imagery, and LiDAR, together with advanced deep learning approaches, including Convolutional Neural Networks (CNNs) and Vision Transformers, to improve the discrimination of complex land-cover classes. Incorporating climate, socio-economic, and ecological datasets would also provide a more comprehensive understanding of the drivers of land-use change and strengthen conservation planning in transboundary protected areas.

Despite these limitations, this study demonstrates that Random Forest combined with multi-temporal Landsat imagery provides a reliable and effective framework for monitoring long-term LULC dynamics in heterogeneous savanna ecosystems. One important limitation of this study is the use of medium-resolution Landsat imagery (30 m), which may introduce mixed-pixel effects and classification uncertainty in highly heterogeneous savanna landscapes. Although these limitations are inherent to long-term Landsat-based analyses, the dataset remains one of the most reliable sources for multi-decadal environmental monitoring because of its continuous historical archive. Future research should integrate higher-resolution optical, SAR, LiDAR, or UAV datasets with advanced deep learning approaches to improve the discrimination of fine-scale landscape features and reduce classification uncertainty.

5 Conclusion

This study investigated the long-term Land Use/Land Cover (LULC) dynamics in and around Kruger National Park between 2005 and 2025 using multi-classifier machine learning algorithms and Random Forest (RF)-based change detection. The study was undertaken to improve understanding of landscape transformation in this ecologically important region and to identify a reliable machine learning approach for long-term LULC monitoring.

The results demonstrated that the Random Forest (RF) classifier consistently outperformed Gradient Tree Boosting (GTB), K-Nearest Neighbors (KNN), and Support Vector Machine (SVM), achieving the highest classification accuracy and Kappa coefficient across all study years. The LULC analysis revealed substantial landscape changes during the study period, characterized by the fragmentation of woodland/shrubland, expansion of agricultural and built-up areas, changes in grassland distribution, and a decline in water bodies. These patterns indicate increasing anthropogenic pressure, land use intensification, and environmental change within and around Kruger National Park.

These findings support the study’s initial objective of evaluating machine learning techniques for monitoring LULC dynamics and demonstrate that ensemble-based algorithms such as Random Forest provide a robust and reliable approach for mapping complex and heterogeneous savanna ecosystems. The observed land cover changes highlight the need for continuous monitoring to better understand the interactions between human activities and ecosystem processes in protected and surrounding landscapes.

Overall, this study provides valuable scientific evidence for sustainable land management, biodiversity conservation, and environmental planning in and around Kruger National Park. The findings can support conservation authorities, land-use planners, and policymakers in developing strategies that balance socio-economic development with ecosystem protection. Future research should integrate higher spatial resolution satellite imagery, additional environmental variables, and advanced deep learning approaches to further improve classification accuracy and the monitoring of landscape dynamics in heterogeneous environments.

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: Xinxin Wang, Fudan University, China

Reviewed by: Trio Santoso, Lampung University, Indonesia

Rakesh Ranjan Thakur, KIIT University, India