Assessing the spatiotemporal dynamics of land-use/land-cover change and human-wildlife conflict around Hluhluwe-iMfolozi park, South Africa

1 Introduction

Assessing the spatiotemporal dynamics of land-use/land-cover (LULC) change and human-wildlife conflict (HWC) is vital for understanding how landscape changes shape the distribution of human-wildlife interactions. The accelerated destruction of natural habitats by human activities, particularly the expansion of settlements and agricultural lands (Gatew and Azadge, 2025), pastoralism and urbanization (Piano et al., 2020), and deforestation driven by tree cutting and firewood collection (Mekonen, 2020), increases the interactions between humans and wildlife (Blanco et al., 2020; Hallam and Harris, 2023). These interactions often lead to human-wildlife conflict globally (König et al., 2020), where human and wildlife habitats overlap (Abrahms et al., 2023). With the rising global human population, the incidence of HWC is intensifying (Bombieri et al., 2019; Bombieri et al., 2023).

In South Africa, reports of crop raiding by the African savanna elephants (Loxodonta africana) and baboons (Papio ursinus), as well as livestock predation by lions (Panthera leo) and spotted hyenas (Crocuta Crocuta), are well documented, with sporadic incidents of large-carnivore attacks on humans (Seoraj-Pillai and Pillay, 2016). Such incidents are part of a broader trend of increasing HWC in South Africa, with cases reported around protected areas (PAs). Similar patterns have been reported across South Asia and Africa, where land conversion and urbanization have reduced wildlife habitat for species such as African and Asian elephants (Elephas maximus) and leopards (Panthera pardus), potentially increasing interactions between wildlife and surrounding communities (Guarnieri et al., 2024; Neupane et al., 2017). Protected areas are increasingly emerging as prime locations for HWC occurrences, coinciding with wild animals’ movement during migration or foraging (Li et al., 2024). Wildlife corridors are important for preserving viable populations and ecological functionality in fragmented landscapes (Hefty and Korpowski, 2021; Serieys et al., 2021; Thatte et al., 2018). HWC intensifies because megafauna species require large ranges and considerable resources to sustain their populations and expand beyond PAs into human-dominated landscapes (Bharathy et al., 2022; Dhungana et al., 2019; Lubis et al., 2020; Siljander et al., 2020). Generally, the locations of known conflict zones are used to assess and predict conflict risks; however, less is understood about the spatial context of conflict and how the landscape may facilitate or hinder negative human-wildlife interactions (Bautista et al., 2021).

Therefore, recent advancements in Remote Sensing (RS) and Geographic Information Systems (GIS), particularly the availability of Sentinel-2 multispectral imagery, enable more accurate mapping of conflict hotspots, habitat fragmentation, and LULC dynamics. The breakthrough in acquiring and analyzing satellite data led to the development of many machine learning algorithms to harness geospatial data. Remote sensing utilizing Sentinel-2 imagery provides high-resolution multispectral data suitable for LULC classification (Jagannathan et al., 2025). Traditional machine learning methods often struggle to address numerous technical issues, such as spectral similarities, necessitating more robust, deep learning-based solutions (Yuan et al., 2021). Deep learning models are regarded as intelligent modelling approaches driving breakthroughs in land-use modelling (Gaafar et al., 2022).

While previous research (Billah et al., 2021; Gunawansa et al., 2024; Nad et al., 2022) has demonstrated the usefulness of geospatial techniques for examining LULC change and HWC, several limitations remain. Numerous studies have either focused on LULC change or conflict patterns separately, used static land-cover datasets or traditional machine-learning classifiers, or only partially evaluated how changes in landscape composition correspond to the spatial distribution of conflict over time. By combining deep learning-based multi-temporal LULC classification, post-classification change detection, kernel density estimation, and spatial overlay techniques, this study addresses these gaps by analysing the spatiotemporal dynamics of LULC change and human-wildlife conflict in and around Hluhluwe-iMfolozi Park (HiP). Specifically, the study seeks to: (1) classify and map land-use/land-cover from 2015 to 2025 using a deep-learning-based classification approach; (2) quantify and analyze LULC transitions between 2015 and 2025 using post-classification change detection; (3) assess and identify the spatial distribution of HWC incidents around HiP; (4) examine the spatiotemporal association of LULC change patterns and HWC hotspots using kernel density estimation and zonal analyses. By integrating spatiotemporal LULC analysis with human–wildlife conflict data, this study seeks to provide deeper insights into how landscape dynamics are spatially associated with conflict patterns, thereby informing more effective conservation and land management strategies in and around HiP.

2 Materials and methods

2.1 Study area

Hluhluwe-iMfolozi Park (HiP) is situated in the northern parts (28.07°S and 32.05°E) of KwaZulu-Natal, South Africa, encompassing approximately 960 km² (Figure 1). The park is fenced (Supplementary Figure S1) except for sections where the main rivers flow out of the park. It is bordered by settlements and agricultural fields, creating a complex interface between conservation areas and human activities. The park comprises two sections: Hluhluwe in the north and Mfolozi in the south, with the Corridor Road (R618) connecting the two.

The corridor is an open road that allows animals to cross between the two sections and serves as a passage between Mtubatuba and Hlabisa. HiP consists of three main rivers: the Black and White Mfolozi in the south, and the Hluhluwe River and its tributaries in the north. Rainfall is seasonal, with a mean annual rainfall ranging from 600 mm in the southern region to over 900 mm in the northern region of the park (Howison et al., 2017), which influences the vegetation patterns and water availability. Much of the area is dominated by woodland savannas mixed with shrub thickets (Mbatha and Xulu, 2018), forests in the northern part, and grasslands in the central parts. The park is primarily renowned for housing the “Big Five” (lion, leopard, elephant, buffalo, and rhino).

2.2 Workflow for methods

An overview of the methods developed in this study (Figure 2) begins with Sentinel-2 imagery acquisition and pre-processing (cloud masking and quality mosaicking), followed by feature extraction and training sample preparation, all performed in Google Earth Engine (GEE). Jupyter Notebook was used to implement machine learning and deep learning for LULC classification, accuracy assessment, post-classification change detection, and reclassification. Lastly, Kernel Density Estimation, Spatial overlay, Zonal Statistics, and map creation were performed using ArcGIS Pro.

2.3 Data acquisition and pre-processing

Sentinel-2 multispectral (Level 2A) atmospherically corrected surface reflectance (SR) imagery was acquired for the years 2020 and 2025, and Level-1C orthorectified top-of-atmosphere (TOA) reflectance for the year 2015 from Google Earth Engine (GEE). TOA was used for 2015 due to the limited availability of atmospherically corrected surface reflectance data for early Sentinel-2 acquisitions. The Level-1C data represent radiometrically calibrated top-of-atmosphere reflectance, whereas the Level-2A product includes standard atmospheric correction applied during the European Space Agency (ESA) pre-processing. All datasets were accessed in their pre-processed state via GEE. The years 2015, 2020, and 2025 were selected at approximately five-year intervals to capture meaningful medium-term LULC dynamics while ensuring the availability of cloud-free Sentinel-2 imagery and corresponding human–wildlife conflict records spatiotemporal analysis.

For each year, the images were filtered by acquisition date (30 June-31 December) using a 50% cloud-cover threshold. Cloud pixels were removed with the QA60 bitmask. The Normalized Difference Vegetation Index (NDVI) was computed from the near-infrared (B8) and red (B4) bands and added to each image in the collection to enhance composite quality and reduce scene boundaries. The pixels with the highest vegetation index values across the image stack are then selected using an NDVI-based quality mosaic approach. By prioritizing vegetated, cloud-free pixels, this method reduces cloud contamination and seam lines separating adjacent scenes. The final composite image was clipped to the study region, and images with smooth edges were subjected to a focal mean filter.

Spectral bands (Table 1) used in this study included 10-m-resolution Blue, Green, Red, and Near-infrared (NIR) bands, and 20-m-resolution shortwave infrared (SWIR) bands. Additionally, four spectral indices (Table 2) were generated to enhance discrimination between vegetation, built-up areas, and bare soil: Normalized Difference Vegetation Index (NDVI), Soil-Adjusted Vegetation Index (SAVI), Normalized Difference Built-up Index (NDBI), and Bare Soil Index (BSI). While the use of different processing levels has the potential to introduce minor differences in spectral reflectance that could influence multi-temporal change detection, these effects were minimised through consistent image pre-processing, cloud masking, selection of imagery from the same seasonal period, and the application of the same classification workflow and spectral indices across all study years.

TABLE 1. Spectral bands used for feature extraction in land use and land cover classification.

TABLE 2. Derived vegetation indices used in the classification of land use/land cover.

Training data for LULC classification were generated from a combination of visual interpretation of high-resolution imagery in GEE and knowledge of the study area. Representation polygons were manually digitized to capture the spectral variability within each LULC class (Table 3).

TABLE 3. Description of land use/land cover class types identified in the study area.

The classes considered in this study included natural forest, Woodland, Shrubland, grassland, cultivated land, barren land, built-up areas, and forest plantations (Table 3; Figure 3), based on their ecological and socio-economic relevance to HiP and surrounding communities. Such classes are essential for HWC analysis because they are associated with habitat suitability, prey availability, and the corridors used by wildlife during migration (Muleta Gurmessa et al., 2024). Multiple and spatially distributed polygons were collected per class type, with special care taken to avoid mixed pixels, especially along transitional zones where boundaries between classes were less distinct (e.g., woodland-shrubland or cultivated lands-grassland interfaces). The digitized polygons were subsequently converted into labelled reference datasets for classification, enabling robust separation of spectrally similar classes using reference data derived from field observations and visual interpretation. Additionally, the labelled samples were split into training (70%) and validation (30%) datasets for the development of the Random Forest, Support Vector Machine, Multi-layer Perceptron Artificial Neural Network, and the Two-dimensional Convolutional Neural Network models using stratified sampling with a fixed random state of 42, which was used to ensure reproducibility. A total of 52,380 reference samples were collected from digitized polygons representing the nine LULC classes (Table 4).

TABLE 4. Distribution of reference samples by LULC class and training–validation dataset split.

2.4 Land cover and land use classification

The study applied Sentinel-2 imagery to generate LULC maps at 5-year intervals, providing a detailed description of the landscape patterns surrounding the park. Furthermore, the study compares the performance of machine and deep learning algorithms, including Random Forest, Support Vector Machine, Multi-layer Perceptron Artificial Neural Network, and Two-dimensional Convolutional Neural Network. For land use and land cover classification and model validation, this study used Python 3 (Jupyter Notebook) with the Pandas, TensorFlow, Rasterio, Scikit-learn, and NumPy libraries, and mapped the outputs in ArcGIS Pro.

To enable a thorough performance comparison of LULC classification, four classifiers were selected to represent various machine learning paradigms. Given their shown efficacy in remote sensing applications, Random Forest (RF) and Support Vector Machine (SVM) were added as popular benchmark machine learning classifiers. The Convolutional Neural Network (CNN) was included to assess whether its ability to extract spatial features from image patches automatically improves classification performance compared with traditional machine learning and neural network approaches. At the same time, the Multi-Layer Perceptron Artificial Neural Network (MLP-ANN) was chosen to represent a traditional deep neural network architecture.

2.4.1 Random Forest

For this study, the RF model was trained on a training dataset comprising 300 decision trees (Ntree = 300) to increase classification stability and reduce variance through ensemble learning. With 11 input features (B2, B3, B4, B5, B8, B11, B12, NDVI, NDBI, BSI, and SAVI), three variables were taken into consideration at each node split (Mtry = 3). A subset of predictor variables equal to the square root of the total number of features was randomly selected at each node split, thereby reducing correlation among individual trees and increasing model diversity. The training process was run in parallel across all available processor cores to improve computational efficiency, and a fixed random seed was used to ensure reproducibility. After being trained on the training features and their accompanying class labels, the model was assessed on unseen test data during training. The model was used to identify satellite images block-wise to enhance computational speed and reduce memory constraints associated with large raster datasets. The spectral bands and derived indices for each block were recovered after the input image was divided into spatial subsets. These attributes were grouped into a two-dimensional array, where each row corresponded to a pixel and each column to a predictor variable. Before classification, invalid or missing pixel values were masked to ensure that only valid pixel data was used for prediction. The trained RF model was then applied to the valid feature set of each block to generate class labels. An internal majority vote across all decision trees’ forecasts was used to determine each pixel’s final class.

2.4.2 Support Vector Machine

The model was constructed within a processing pipeline that first applies feature standardization using a standard scaler to ensure that all input variables have a mean of zero and a variance of one, thereby promoting convergence and preventing any single feature from disproportionately influencing the model. The model’s hyperparameters were set to “C” = 10, which regulates the trade-off between maximizing the margin and minimizing classification error, and gamma = “scale,” which automatically modifies the kernel’s influence based on feature variance. The kernel coefficient gamma was computed directly from the input data, and probability estimation was turned off to reduce processing overhead. During training, the SVM approach determines the optimal separating hyperplanes in a converted high-dimensional feature space by maximizing the margin between classes. The SVM model was used to evaluate classification performance on unseen test data after training on the feature set and its associated class labels. Like the RF model, the model implementation processed massive geospatial datasets using a block-wise raster classification technique. By extracting spectral bands and calculating spectral indices, each image block was examined separately, yielding a feature matrix in which each pixel was represented as a multidimensional vector. To ensure reliable model input, data pre-processing involved removing invalid pixel values. The SVM classifier categorized pixels using the characteristics of valid pixels.

2.4.3 Multi-layer perceptron artificial neural network

The baseline classification model for this study was a simple MLP with a fully connected architecture. The input features consisted of estimated spectral indices and spectral bands. The dataset was categorized into nine land cover classes and then split into training and validation subsets using a predefined sampling procedure. The MLP model in TensorFlow, built with the Keras Sequential framework, used the Rectified Linear Unit (ReLU) activation function. The input layer was followed by two hidden layers with 128 and 64 neurons, respectively. To lessen overfitting, dropout regularization was applied at rates of 0.3 and 0.2 after each hidden layer. Nine neurons, one for each land cover type, made up the output layer, which generated class probabilities using a softmax activation function. The model was trained using the Adam optimizer (learning rate 0.001) and sparse categorical cross-entropy as the loss function, with classification accuracy serving as the performance indicator. Training was performed for up to 40 epochs with a batch size of 32. An early-stopping condition based on validation loss was used, with a patience of 5 epochs, to prevent overfitting and achieve the best possible generalization. The trained model was used to classify every pixel in the whole satellite image. To match the model’s input format, the raster was transformed into a (height, width, bands) format. After that, it was reorganized into a two-dimensional array, with each column representing a predictor variable (spectral bands and indices) and each row representing a single pixel. The raster data were homogenized before prediction using the same scaling parameters used during model training to ensure uniform feature distributions. The trained MLP model was used for batch processing (batch size = 5000) to compute class probabilities for each pixel, thereby improving computational performance and reducing memory usage.

2.4.4 Convolutional neural network

In this study, a 2D CNN was developed as the classification model for patch-based land-cover categorization using the TensorFlow Keras functional API. 32 × 32-pixel picture patches with 11 spectral and index-based features, including bands and derived indices, comprised the input data. These patches were processed as multi-channel images, with spectral bands and indices as feature channels, and convolutional techniques applied across the spatial dimensions (height and breadth). As a result, the model learned both spectral correlations and spatial patterns concurrently. A maximum of 20,000 patches were randomly selected to ensure computational efficiency. The CNN architecture consisted of three convolutional layers with 32, 64, and 128 filters each. After each convolutional layer, batch normalization and max pooling were applied to enhance feature extraction and reduce spatial dimensionality. The resulting feature maps were flattened, passed through a fully connected dense layer with 256 neurons, and then subjected to a dropout rate of 0.4 with the ReLU activation function to reduce overfitting. The output layer and a softmax activation function were used to classify pixels into 9 land-cover classes. The model was trained for 20 epochs with a batch size of 64 using categorical cross-entropy loss and the Adam optimizer (learning rate = 0.0001). To reduce memory usage, the trained model was applied to the raster dataset for full-scene classification using a block-wise processing approach. The input image was divided into smaller pieces (64 × 64 pixels) with overlapping borders to preserve spatial context during patch extraction. Patches for each pixel were generated, standardized using the previously fitted scaler, and then fed to the CNN in batches of 512 to obtain predictions. Batch processing and other memory management techniques were employed to ensure efficient computing.

2.5 Hyperparameter optimization

Initially, hyperparameter sets were based on widely used parameter settings documented in research on machine learning, deep learning, and remote sensing. The network size, dropout rate, learning rate, batch size, number of epochs, activation functions, loss functions, and optimization algorithms were among the hyperparameters for the CNN and MLP-ANN. To identify models that achieved strong validation performance while remaining computationally efficient, many configurations were empirically evaluated. The original parameter ranges and configurations were also informed by earlier remote sensing research (Ali and Johnson, 2022; Lam et al., 2023). The final configurations were applied uniformly across the 2015, 2020, and 2025 datasets and were chosen based on validation performance. Previous research on RF often suggests 500 trees as a reference value, which is frequently related to the square root of the number of input variables (Adugna et al., 2022; Belgiu and Drăguţ, 2016). As a result, various RF configurations were empirically examined; the final model used 300 decision trees, based on validation performance. For SVM, multiple parameter combinations were empirically evaluated after the original configuration adhered to the scikit-learn implementation’s default values. Based on validation performance, gamma = scale and the final SVM configuration were employed. The 2015, 2020, and 2025 datasets all used the same setups.

2.6 Model training and validation

RF, SVM, and MLP-ANN were trained on the complete extracted dataset (52,380 samples), which was randomly stratified and split into 36,666 training samples (70%) and 15,714 validation samples (30%), as shown in Table 4. In contrast, CNN was trained on a reduced dataset of 20,000 samples, comprising 14,000 training samples (70%) and 6,000 validation samples (30%), to improve computational efficiency (Supplementary Table S1). The training and validation samples were randomly selected from the same study area and were therefore not spatially independent. An accuracy assessment was conducted to evaluate the performance of the four models. The best-performing model was selected and used for further land-cover classification and change-detection analysis. For model evaluation, a confusion matrix was generated for each model using the test data. To further quantify the assessment, kappa coefficients (Kc) and overall accuracy (OA) provide a more robust assessment of the model’s performance. During this study, predicted class labels were generated by applying the trained models to the test set and selecting the class with the highest predicted probability for each sample. Additionally, a detailed classification report was generated, including class-wise predictions, recall (Producer accuracy), and F1 scores (User accuracy) to evaluate each class’s performance, as provided in Supplementary Tables S2–S4.

2.7 Post-classification change detection

Post-classification change detection was employed to determine land-cover changes between 2015-2020 and 2020-2025. This detection was conducted separately for areas within and outside the protected area, enabling comparison of land-cover dynamics between protected and non-protected landscapes. The pixels within the PA were analyzed to quantify internal LULC changes, whilst those outside the PA were assessed to characterize landscape transformations in the surrounding areas. This analysis also identified the dominant land-cover transitions using Intensity analysis to quantify category- and transition-level land-cover change rates across the study area. Selected key transitions were further disaggregated within and outside PA to evaluate their spatial association with potential wildlife movement. All the classified LULC rasters were reprojected to a common coordinate reference system (WGS 1984 to UTM Zone 36S) and resampled using nearest-neighbor interpolation to preserve categorical values and ensure spatial alignment for pixel-level correspondence before post-classification. The resulting transition map contained 81 possible class-to-class combinations (9 × 9).

Therefore, to improve interpretability and facilitate spatial analysis, these detailed transitions were reclassified into stable land-cover categories and grouped into change processes. Pixels that retained the same class between periods were classified as stable (e.g., stable grassland, stable cultivated land, stable woodlands). Transitions from natural vegetation classes (natural forests, grasslands, shrublands, and woodlands) to anthropogenic classes (cultivated land, forest plantations, barren land, and built-up areas) were grouped as vegetation loss. In contrast, transitions from anthropogenic to natural vegetation were grouped as vegetation gain. Changes involving waterbodies were grouped as waterbody changes. The remaining transitions were categorized as natural-to-natural (e.g., grassland to woodland) or anthropogenic-to-anthropogenic (e.g., barren land to cultivated land). The reclassification decreased the original 81 transition classes to a smaller set of ecologically and socioeconomically meaningful categories.

2.8 Human-wildlife conflict

2.8.1 HWC incident data

HWC incident records were obtained from Ezemvelo KZN Wildlife for 2019 to 2025 (Table 5). Before the 2019 incident, data could not be obtained because the records lacked spatial references. These incidents have been documented with location information, the wildlife involved, the area name, and the incident date (year, month, and day).

TABLE 5. Number of recorded human–wildlife conflict incidents and associated wildlife species.

2.8.2 Human-wildlife conflict and land use/land cover transition analysis

Human–wildlife conflict (HWC) point data (2019–2025) were analyzed to identify spatial patterns of conflict intensity. A kernel density estimation (KDE) was performed in ArcGIS Spatial Analyst using the Geodesic method. Output cell sizes were 0.001668° (2019–2020) and 0.002741° (2021–2025), automatically generated based on the conflict data. ArcGIS Pro automatically determined the search radius, producing a continuous raster of incident density (incidents/km²). For the LULC change from 2015 to 2020, only the HWC incidents from 2019 to 2020 were used. For the LULC change from 2020 to 2025, the entire HWC dataset was used. KDE analysis was conducted for each period separately to produce density surfaces that matched the LULC change period. The KDE raster was sorted into low (0.001-0.118), medium (0.119-0.236), and high (0.237–0.354) density classes to identify hotspots. KDE hotspots were overlaid on the land-use/land-cover change raster to examine their spatial association. Zonal Statistics as a Table to find the average KDE values for each LULC change class. This enabled measuring the intensity of conflict in relation to changes in land cover. This approach provided a comprehensive assessment of how landscape dynamics and protected area boundaries shape the spatiotemporal patterns of human-wildlife conflict.

3 Results

3.1 Land use and land cover classification

The comparison of the four classification algorithms showed variations in performance (Table 6), with the 2D-CNN classifier achieving the highest overall accuracy (OA) and kappa coefficient (Kc) across all years, with OA exceeding 99% and kappa values close to 1.0. The RF and MLP-ANN models also performed well, with OA values above 94% and Kc values above 0.94, while SVM showed slightly lower but still strong performance. The CNN demonstrated consistently high class-level performance across the three mapping years (Supplementary Tables S2–S4). Most LULC classes achieved precision, recall, and F1-scores of 1.00, with only minor reductions observed for cultivated land, built-up areas, and grasslands in 2025. The corresponding confusion matrices are provided in Supplementary Figures S2–S4.

TABLE 6. Accuracy assessment results for the land use/land cover classification.

Based on these results, the 2D-CNN was selected as the optimal classifier to produce the final LULC maps for the study area and period (Figure 4). The LULC maps for 2015, 2020, and 2025 showed changes in the spatial distribution of LULC classes over the 10-year study period. Built-up areas remained a dominant class across the study area throughout the period, although their spatial distribution changed between mapping years. Between 2015 and 2020, woodland expanded into several areas of the study landscape (Figures 4A,B). By 2025, shrubland had expanded across several areas, while cultivated land occupied a smaller spatial extent than in 2020 (Figure 4C).

3.2 Spatiotemporal change patterns of land use and land cover

Land use/land cover changes occurred between 2015-2020 and 2020–2025 across the study area, which were grouped into change processes (Figure 5). Anthropogenic-to-anthropogenic changes and vegetation gains were the dominant change processes during 2015–2020 (Figure 5A), whilst 2020-2025 (Figure 5B) was characterized by more natural-to-natural vegetation transition. For both 2015-2020 and 2020-2020, vegetation loss was spatially concentrated within stable built-up zones.

The analysis was conducted across the combined landscape comprising HiP and the surrounding non-protected areas. Therefore, the LULC areas and changes reported in Table 7 represent the entire extent of the analysis. Substantial gains and losses were observed between 2015-2020 and 2020-2025 (Table 7). Between 2015 and 2020, built-up areas increased by 394.99 km² (32.84%), representing the largest increase among the anthropogenic classes, while woodland and grassland increased by 157.23 km² (14.00%) and 133.83 km² (34.57%), respectively. In contrast, barren land declined by 215.62 km² (81.65%), representing the largest proportional reduction, followed by forest plantations, which decreased by 119.11 km² (39.53%). Between 2020 and 2025, the most pronounced change was a 334.35 km² (176.76%) expansion of shrubland, the largest increase in vegetation observed. Cultivated land declined by 257.13 km² (47.77%), while woodland decreased by 147.85 km² (11.55%). Built-up areas continued to expand, increasing by a further 85.89 km² (5.38%). Natural forests remained relatively stable during 2020–2025, changing by only −0.37 km² (−0.11%).

TABLE 7. Net change, annual rate of change, and area of land use/land cover (classes across the landscape.

LULC change between 2015–2020 and 2020–2025 shows distinct spatial patterns within and outside protected areas (Table 8). Within the protected area, LULC changes were generally smaller in magnitude than those observed outside the PA. Between 2015 and 2020, grasslands and woodlands recorded net gains, while waterbodies, barren land, and natural forests declined. During 2020–2025, natural forests increased by 33.00 km² and shrublands by 98.46 km², whereas woodlands decreased by 94.72 km². Outside the PA, changes were larger in magnitude, particularly between 2015 and 2020. Built-up areas increased by 417.24 km², while waterbodies decreased by 89.80 km². Between 2020 and 2025, built-up areas increased by a further 62.61 km², while cultivated land decreased by 257.12 km² and shrublands increased by 235.89 km². Grasslands increased by 82.44 km² during this period, whereas forest plantations declined by 148.25 km². These results indicate that the largest changes in anthropogenic land-cover classes occurred predominantly outside the PA, while changes within the PA were generally smaller in absolute area.

TABLE 8. Comparison of net land use/land cover change inside and outside protected areas.

Land-cover/land-use transitions occurred between 2015-2020 and 2020–2025 across the study area (Figure 6). The largest transition that occurred between 2015 and 2020 (Figure 6A) was from built-up to woodland (302.3 km²) and shrubland (103.8 km²). Woodland also gained area from cultivated land (231.1 km²), grassland (111.8 km²), and built-up land (172.5 km²).

In contrast, the 2020-2025 period (Figure 6B) was characterized by several large-magnitude transitions, with the largest occurring from woodland to shrubland (235.8 km²) and from woodland to grassland (217.2 km²), exceeding all others by a large margin. Other notable changes included cultivated land to woodland, natural forest to woodland, grassland to shrubland, and grassland to woodland.

3.3 Human-wildlife conflict and land cover/land use transitions

A total of 187 human–wildlife conflict (HWC) incidents were recorded during the study period, involving multiple wildlife species (Table 5). The spatial distribution of HWC incidents varied between protected and non-protected areas and between LULC categories that changed and those that did not (Table 9).

TABLE 9. Distribution of human–wildlife conflict (HWC) incidents inside and outside protected areas.

Human-wildlife conflict (HWC) incidents are primarily concentrated along the boundaries of the protected area (PA), with some occurring in nearby communities (Figure 7). The analysis of top land-cover transitions associated with human–wildlife conflict (HWC) reveals clear differences between the 2015–2020 and 2020–2025 periods (Figure 8). Between 2015 and 2020 (Figure 8A), HWC incidents were comparatively low and dispersed over a small number of transitions, with the highest frequency occurring in grasslands → grasslands, followed by barren land → woodlands and cultivated → cultivated transitions. Most other transitions documented only one incident, suggesting a less intense, more localized pattern of conflict. On the other hand, both the quantity and variety of HWC-related transitions significantly increased during the 2020–2025 period (Figure 8B). Woodland → woodland transitions predominate, followed by built-up → built-up and grasslands → woodlands, indicating that both expanding anthropogenic landscapes and stable natural habitats are important areas of conflict. The prevalence of transitions such as woodlands → grasslands and natural forests → shrublands further suggests that regions experiencing vegetation fragmentation and change are more often associated with higher HWC incident rates.

For both periods, the mean HWC density differed among LULC classes (Table 10). Grassland and woodland consistently displayed the highest densities among stable classes, rising from 1.215 to 1.328 incidents/km² and 1.015 to 1.155 incidents/km², respectively. Among LULC change classes, vegetation gain had the highest density in 2015–2020 (1.188 km²) but declined in 2020–2025 (1.059 km²). In comparison, Vegetation loss had a low density in 2015–2020 (1.069 km²) but increased in 2020–2025 (1.080 km²).

TABLE 10. Mean human–wildlife conflict (HWC) density per LULC class for 2015–2020 and 2020–2025.

Natural-to-natural transitions, however, remained relatively high, whereas Anthropogenic-to-Anthropogenic transitions decreased from 2020 to 2025 (Table 10). These findings are spatially supported by the KDE density map (Figure 9), which shows that HWC hotspots (red zones) are concentrated in grassland, woodland, and areas undergoing LULC change, demonstrating the strong spatial relationship between land cover dynamics and conflict intensity. The KDE results further indicate that areas of high HWC density remained spatially concentrated during both 2015–2020 (Figure 9A) and 2020–2025 (Figure 9B), although their spatial extent and configuration varied between periods. Thus, this indicates that conflict-prone areas were not restricted to a single period but remained concentrated in parts of the landscape over time.

4 Discussion

4.1 LULC classification performance

CNN achieved consistently high overall classification accuracy across the three study years. The consistently high class-specific performance of CNN across the three mapping years supports the overall classification results. However, the potential influence of spatial dependence between training and validation samples should be considered when interpreting the near-perfect accuracies. The higher performance of the CNN may be attributed to its ability to learn spatial patterns from image patches. At the same time, the inclusion of additional spectral indices may have further improved the separability of LULC classes. By contrast, RF consistently produced the second-highest accuracy, followed by MLP-ANN and SVM. Although these models achieved high classification accuracy, CNN demonstrated superior performance, suggesting that its ability to learn hierarchical spatial and spectral features provided an advantage for LULC classification. Together with the high-quality Sentinel-2 imagery, the inclusion of multiple spectral indices, and the use of balanced reference samples, these characteristics provide a relatively stable classification procedure across all study years. Thus, the results support the use of CNN for subsequent LULC mapping in this study and demonstrate that classifier performance depends on both the algorithm and the nature of the input features.

The observed performance is consistent with previous studies reporting advantages of CNNs for LULC and vegetation classification. Hasan et al. (2019), for example, reported higher accuracy for CNN than ANN and SVM, while Carranza-García et al. (2019) obtained overall accuracies ranging from 96.78% to 99.36% using CNN-based LULC classification. Similarly, Lam et al. (2023) reported higher accuracy for a patch-based CNN than MLP and RF. These findings support the potential advantage of CNNs where LULC classes exhibit complex spatial patterns and spectral similarities. However, the high accuracy observed in the present study should also be interpreted in light of the characteristics of the reference data and the validation strategy.

Despite the high classification accuracies, some uncertainty remains regarding model generalization. Previous studies have shown that random train–validation splits may overestimate classification performance when spatial autocorrelation exists between training and validation samples or when geographically independent validation is not employed (Bahn and McGill, 2013; Kattenborn et al., 2022; Karasiak et al., 2022; Roberts et al., 2017). Although the validation samples were independent in the sense that they were not used during model training, they were not necessarily spatially independent from the training data (Kattenborn et al., 2022). Therefore, while the CNN demonstrated strong classification performance, some uncertainty remains about its ability to generalize to geographically independent areas. This uncertainty is important when interpreting subsequent LULC transitions derived from the classified maps.

4.2 Landscape transformation around HiP

The observed LULC transitions indicate that the landscape surrounding Hluhluwe-iMfolozi Park (HiP) underwent various transformations during the study period. The 2015–2020 period was characterized largely by anthropogenic-to-anthropogenic transitions and vegetation gains, whereas 2020–2025 showed a greater contribution of natural-to-natural vegetation transitions. This shift suggests that landscape transformation was not uniform over time but instead reflected changes in the relative contributions of human land-use expansion and vegetation dynamics. The contrasting trajectories between protected and non-protected areas further indicate that land-use pressures and landscape management may operate differently across the PA boundary.

The expansion of built-up land and the transition of cultivated land to woodland indicate contrasting forms of landscape change. Built-up expansion is consistent with continued development pressure in the areas surrounding HiP, where conversion of natural and semi-natural vegetation can increase fragmentation of the remaining landscape. The transition may reflect a shift from active cultivation towards fallow or less intensively cultivated land, potentially allowing secondary grassland or woody vegetation to establish (Moyo and Ravhuhali, 2022). Abandoned or fallow-cultivated fields may initially be colonized by grasses and forbs and, over time, develop greater woody cover through secondary succession (Ruwanza, 2017). Changes in livelihood patterns and land-use choices may therefore be an important component in the observed conversion of cultivated land to more woody vegetation in communal landscapes such as those found in KwaZulu-Natal. However, the transition should be interpreted as a potential sign of land-use change and vegetation recovery rather than as proof of agricultural abandonment, because the current study did not directly quantify field abandonment or vegetation succession. The observed decreases in cultivated land, however, may reflect a shift from active cultivation to woody or semi-natural vegetation rather than a reduction in the potential for human influence.

The transition from woodland to built-up land is consistent with the expansion of human settlements and associated development pressure in areas surrounding HiP. Conversion of woodland to built-up land can fragment remaining semi-natural habitat and increase the interface between human-dominated and natural landscapes. Around HiP, this pattern may increase opportunities for wildlife to move through modified landscapes, particularly near the PA boundary. However, the present analysis identifies spatial associations rather than causal relationships between woodland conversion and HWC. The transition from woodland to grassland during 2020–2025 may reflect changes in vegetation structure associated with grazing, fire, or other environmental conditions (Graham et al., 2020). Fire and herbivory are important factors associated with the relative distribution of woody and herbaceous vegetation in savanna systems (Malongweni and Van Tol, 2022), while changes in grazing and fire regimes may promote shifts towards grass-dominated communities (Donaldson et al., 2017). Similarly, the loss of forest plantations could result from shifting land-use targets and management techniques in the surrounding area, as well as from changes in plantation management, harvesting, or subsequent land-cover conversion. Therefore, these transitions imply that the observed landscape changes may be caused by both ecological processes, such as disturbance and vegetation dynamics, and socio-economic processes, such as settlement growth, farming, grazing, and land-use management. However, these changes cannot be attributed to a single ecological or socioeconomic explanation because rainfall, plantation management, grazing intensity, and fire frequency were not measured precisely.

Because they reflect unexpected environmental changes within a brief period, some transitions need to be interpreted carefully. Specifically, between 2015 and 2020, there was a significant shift from built-up land to woodland concurrent with the transition from woodland to built-up land. Although development and vegetation loss are consistent with the conversion of woodland to built-up areas, actual land-cover recovery over the same period is less likely to account for the opposite shift. These reciprocal transitions could be caused by variations in spectral responses between image dates, mixed pixels, edge effects, or classification uncertainty. Therefore, without considering the uncertainty in the underlying classification maps, they should not be taken as direct proof of total land-cover conversion. This uncertainty is important because classification errors in either mapping year can propagate into the transition matrix, producing apparent changes where no actual land-cover conversion has occurred. The transition analysis should therefore be interpreted as representing mapped LULC changes subject to classification uncertainty rather than exact measurements of landscape transformation.

The expansion of shrubland during 2020–2025 is among the most important vegetation changes identified in the study. The transition from woodland to shrubland indicates a change in vegetation structure across parts of the landscape. It may reflect several interacting ecological processes, including woody encroachment, changes in grazing pressure, fire regimes, vegetation succession, or other forms of disturbance (Graham et al., 2020; Legge et al., 2019; Moyo and Ravhuhali, 2022). Reduced fire frequency and changes in herbivore pressure have been associated with shrub encroachment in savanna systems (Liao et al., 2018). Alternatively, the observed woodland-to-shrubland transitions may reflect vegetation succession and recovery following disturbance (Clark et al., 2020). Variations in wildlife visibility and mobility, fire behavior, predator-prey interactions, and fodder availability and accessibility may be linked to changes in vegetation structure, with implications for biodiversity and habitat use. Nevertheless, fire frequency, herbivore populations, grazing intensity, and vegetation succession were not directly measured in this study. Therefore, rather than conclusive proof of shrub encroachment, shrubland expansion cannot be attributed to a single process and may be a sign of shifting vegetation structure.

4.3 Landscape transitions and HWC patterns

The spatial distribution of HWC incidents indicates that conflict was concentrated within landscape interfaces, particularly around the boundary of HiP. The prevalence of disputes at PA borders underscores the significance of edge effects, high-risk interaction zones created by abrupt transitions between human-dominated landscapes and protected natural habitats. Several studies (Khosravi et al., 2024; Pitman et al., 2015; Suryawanshi et al., 2017) have revealed that conflict risk increases near the border of protected areas. Similarly, Tiwari et al. (2025) found that clusters of multiple conflict types occurred near forest edges and near wildlife habitats, suggesting a higher risk in areas adjacent to dense vegetation and the PA. These border regions function as conflict hotspots, with HWC occurring in nearby settlements, supporting the earlier interpretation that wildlife displacement extends beyond protected-area boundaries. The findings demonstrate that HWC extends beyond areas of direct habitat loss and is closely associated with the identified transitional landscapes.

The spatial distinction between the surrounding landscapes and the protected area provides further insight into the factors that may contribute to HWC. The relative stability of LULC within the PA suggests that conservation efforts are essential to preserving habitat integrity and controlling wildlife distribution. More predictable wildlife movement patterns can be supported by stable habitats within the PA, thereby reducing conflict in these areas. PAs are not totally impervious to environmental stresses, though, as observed by the decreases in important vegetation classes like woodlands and natural forests between 2015 and 2020. The greater magnitude of LULC change outside the HiP compared with within the protected area is consistent with different land-use pressures operating across the PA boundary. The conversion of natural vegetation to human-dominated land uses may be limited inside HiP by conservation management and restrictions on habitation and cultivation. The lack of comparable conservation rules outside the PA permits more freedom for land conversion, cultivation practices, and settlement growth. The disparate LULC trends seen inside and outside the protected region could be attributed to these variations. These modifications may enhance habitat suitability and may attract wildlife to migrate closer to the PA borders, leading them to venture outside in search of food and shelter and potentially resulting in confrontations with people (Bhat et al., 2024).

On the other hand, the greater extent and persistence of LULC changes observed outside the PA, particularly the expansion of built-up areas and decline in cultivated land, indicate a stronger degree of anthropogenic landscape transformation in the surrounding areas. Outside the PA, converting cultivated land to semi-natural vegetation is important because it creates transitional landscapes near human settlements. Such transitional areas may provide resources or movement pathways for wildlife while remaining adjacent to human activities, potentially increasing opportunities for human–wildlife encounters. This interpretation is consistent with Bharathy et al. (2022), who reported that high-risk areas were concentrated where agricultural and settlement areas occurred close to forested sites. Similarly, an analysis by Singh et al. (2025) indicated that the Buffer Zone showed an increase in dense forest, reflecting conservation efforts and natural regeneration; however, it also experienced substantial urban expansion and the depletion of water bodies, which exacerbated human-wildlife interactions in the Ranthambhore Tiger Reserve, India.

Spatial context for interpreting the distribution of HWC is provided by observed LULC transitions. The conversion of woodland and shrubland to developed land may result in the loss and fragmentation of semi-natural habitat. On the other hand, the development of transitional or regenerating zones may be indicated by changes from cultivated land to grassland or woods. These divergent shifts may be linked to a diverse environment where human-dominated and natural land uses coexist. These landscape arrangements may be pertinent to wildlife’s movement and utilisation of resources and may offer potential sites for interactions between humans and wildlife. However, rather than proving that particular LULC transitions directly produced HWC, these patterns show spatial associations. These habitat changes may alter wildlife movement and increase the likelihood that wildlife use areas adjacent to human-dominated landscapes. HWC incidents occurring near built-up land may indicate increased spatial overlap between wildlife habitat and human-dominated landscapes. The predominance of woodland-to-woodland transitions indicates that areas of relatively stable natural vegetation also remained important HWC locations. This suggests that conflict is not restricted to areas undergoing rapid land-cover conversion but may also occur within relatively stable habitats where wildlife activity and human presence occur in proximity.

In contrast, Bhat et al. (2024) found that agricultural/horticultural areas near the protected area perimeter appear to be the primary sites of human-wildlife interactions. Similarly, the persistence of built-up land indicates that established human-dominated areas remained an important component of the landscape. Where persistent built-up areas occur adjacent to natural or semi-natural vegetation, they may maintain long-term interfaces where human and wildlife activities overlap.

Therefore, rather than focusing only on specific land-cover classes, the association between LULC transitions and HWC is better understood in terms of changes in landscape layout. Wildlife may migrate between habitats and human-dominated areas through locations where woodland or other semi-natural vegetation occurs alongside cultivated or built-up land. In a similar vein, long-term human-wildlife interactions may occur in built-up regions near natural vegetation, even in the absence of a recent LULC transition. Therefore, the identified HWC hotspots may be places where habitat connectivity, landscape fragmentation, and proximity to human activity interact, likely influencing the geographic distribution of conflict. The concentration of HWC in both study periods further suggests that some conflict-prone areas may represent recurring human–wildlife interfaces rather than isolated or short-term hotspots. However, because hotspot persistence was assessed by comparing KDE patterns rather than through a formal spatial overlap analysis, this interpretation should be considered indicative rather than a quantitative measure of hotspot persistence.

The interpretation of HWC patterns should also account for differences among conflict species. The HWC database included various wild animals that differ substantially in habitat requirements, movement behavior, resource use, and mechanisms of conflict. Consequently, the same landscape transition may have different implications for different species. For example, the ecological significance of changes in vegetation structure may differ between herbivorous species and large carnivores, while crop-raiding species may be more closely associated with the distribution of cultivated land. Because the present study aggregated HWC incidents across species for KDE analysis, the identified hotspots represent areas of overall conflict concentration rather than species-specific responses to LULC change.

4.4 Ecological mechanisms underlying landscape- HWC patterns

The concentration of HWC around landscape interfaces is consistent with ecological mechanisms associated with habitat fragmentation and reduced connectivity around HiP. Studies suggest that conflicts are more likely in heterogeneous, fragmented landscapes characterized by mosaics of natural vegetation patches and croplands or livestock pastures (Miller, 2015; Tee et al., 2021; Davoli et al., 2022). The behavior of wildlife species, especially those with large space requirements, may change in fragmented landscapes due to reduced access to natural foods, reduced habitat connectivity, and increased anthropogenic pressures, which collectively increase interactions with humans (Acharya et al., 2017). These associations provide a basis for considering the ecological mechanisms that may underlie the observed conflict patterns. The availability of natural forage and wildlife habitat is greatly reduced by habitat degradation and conversion, pushing wild animals into human-modified landscapes where conflict is more likely (Naha et al., 2018; Sharma et al., 2017).

The prevalence of disputes at PA borders underscores the significance of edge effects, areas of potentially elevated interaction created by abrupt transitions between human-dominated landscapes and protected natural habitats. These border regions function as conflict hotspots, with HWC occurring in nearby settlements, indicating that wildlife–human interactions extend beyond protected-area boundaries. The findings demonstrate that HWC extends beyond areas of direct habitat loss and is closely associated with the transitional landscapes identified earlier in the discussion. Vegetation loss and disturbance may alter the availability and distribution of wildlife resources, potentially increasing the likelihood that animals use adjacent human-dominated areas. Agricultural fields can provide concentrated and accessible food resources, creating opportunities for crop-raiding and other forms of conflict (Bharathy et al., 2022). In the context of HiP, this mechanism may be particularly relevant where cultivated areas occur close to natural vegetation and wildlife movement routes.

Changes in water availability may represent another mechanism influencing wildlife movement and conflict. Previous studies have identified conflict hotspots near water sources and agricultural areas, highlighting the potential relevance of resource distribution to the spatial occurrence of wildlife–human interactions (Mahakata et al., 2025). In HiP, areas associated with changes in water availability may therefore represent locations where wildlife movement and resource use intersect with human activities. However, the present study did not directly measure wildlife dependence on these resources; therefore, this interpretation should be regarded as a possible mechanism rather than a demonstrated causal process.

A further method for deciphering the observed patterns is landscape connectedness. While residual semi-natural vegetation and fallow areas may offer alternative routes across human-dominated environments, fragmentation of natural vegetation can impede movement between habitat patches. These regions might promote ecological connection, but they might also make it more likely that wildlife movement will come into contact with populated areas, cultivated land, and other human endeavours. A source-sink paradigm can also be used to conceptually analyse the observed patterns, suggesting that reasonably stable habitats within HiP could serve as appropriate habitats. The nearby altered landscapes may also be utilised for resource acquisition or mobility. From a broader metapopulation perspective, maintaining connectivity among habitat patches is important for sustaining wildlife populations, whereas fragmentation may increase the frequency of animal crossings through human-dominated areas. These ecological concepts provide a useful framework for interpreting the observed spatial association between landscape configuration and HWC, although they were not directly tested in the present study.

4.5 Implications for conservation planning

The spatial concentration of HWC incidents around the boundaries of HiP and the greater magnitude of LULC transformation outside the protected area indicate that conservation planning should prioritise the protected-area–community interface, rather than applying mitigation measures uniformly across the landscape. Areas where built-up expansion, vegetation conversion, and HWC hotspots overlap represent priority zones for land-use monitoring and conflict mitigation. In these areas, local land-use planning should consider the proximity of settlements and cultivated land to remaining natural vegetation and wildlife movement areas when evaluating further development or land-use expansion. The observed expansion of built-up areas outside HiP also highlights the need for proactive monitoring of settlement growth adjacent to the protected area. Rather than responding to conflict after incidents occur, managers and relevant municipal planning authorities could use regularly updated satellite imagery to identify areas where settlement expansion is approaching existing vegetation patches or known HWC hotspots. Such areas could be prioritised for land-use planning, community engagement, and early conflict-prevention measures. This is particularly relevant where continued development may further reduce habitat connectivity or increase the spatial overlap between wildlife movement and human activities.

The concentration of HWC around landscape interfaces also supports targeting mitigation resources to identified persistent hotspot areas. For HiP, this could include prioritising ranger patrols, community-based reporting, early-warning systems, and appropriate deterrent measures in locations where HWC hotspots coincide with rapidly changing LULC. However, mitigation should be tailored to the species and type of conflict involved because elephants, carnivores, herbivores, and other species respond differently to landscape conditions. The spatial framework developed in this study can therefore support the identification of priority areas, while species-specific information should guide the choice of intervention. The expansion of shrubland between 2020 and 2025 also has implications for vegetation and habitat management. Areas undergoing persistent woodland-to-shrubland transitions should be monitored to determine whether the observed change represents natural succession, woody encroachment, changes in grazing pressure, fire-related dynamics, or other forms of disturbance. Such monitoring is important because changes in vegetation structure may have implications for forage availability, wildlife movement, fire regimes, and habitat suitability. Integrating periodic LULC monitoring with field-based vegetation assessments would allow managers to distinguish persistent ecological change from short-term or classification-related variation.

Maintaining landscape connectivity should also form part of conservation planning around HiP. Where natural or semi-natural vegetation occurs between the protected area and surrounding human-dominated landscapes, these areas may provide important movement pathways. Land-use planning should therefore consider retaining strategically located vegetation patches and corridors, particularly where they connect HiP to other suitable habitats. However, because wildlife movement and connectivity were not directly measured in this study, these areas should be treated as potential connectivity zones that require further ecological assessment rather than as confirmed wildlife corridors. Finally, the findings demonstrate the value of integrating remote sensing and HWC records into routine conservation monitoring. Rather than using LULC maps and HWC hotspot maps as separate products, managers could periodically update LULC classifications and conflict records to identify where landscape transformation and conflict hotspots persist, shift, or emerge. This would provide HiP and surrounding land managers with a spatially explicit basis for prioritising intervention areas and adapting management strategies as the human–wildlife interface changes.

5 Limitations of the study

Although the LULC classification achieved high overall accuracy across the three mapping years, classification errors may propagate into subsequent change-detection and spatial analyses. Misclassification in either of the years being compared can produce apparent transitions in the post-classification change matrix where no actual land-cover change has occurred. Similarly, classification uncertainty may affect the spatial overlay and interpretation of LULC transitions relative to HWC hotspots. Therefore, individual transition categories should not be interpreted as error-free representations of actual landscape dynamics. Although the four classifiers were evaluated using overall classification accuracy and related performance measures, their performance differences were not statistically tested. Consequently, the observed differences in accuracy should be interpreted as comparative performance rather than evidence of statistically significant differences between classifiers. Furthermore, although the validation samples were independent of the training samples, they were not explicitly spatially independent. Spatial autocorrelation between training and validation samples may therefore have resulted in optimistic estimates of model performance and may limit the generalisability of the classification to geographically independent areas.

An additional limitation concerns the use of different Sentinel-2 processing levels throughout the study period. The 2015 imagery was obtained as Level-1C top-of-atmosphere (TOA) reflectance, whereas the 2020 and 2025 imagery were obtained as Level-2A surface reflectance products. Although a consistent classification workflow and comparable image-selection criteria were applied, differences in atmospheric correction and reflectance characteristics between processing levels may introduce some inconsistencies in the spectral information used for classification. Consequently, some differences observed between mapping years may partly reflect differences in image processing rather than actual landscape change. The use of spectral indices and consistent classification procedures helped to reduce potential inconsistencies; however, this limitation should be considered when interpreting individual LULC transitions across the study period. Overall, the consistently high classification accuracies provide confidence in the broader LULC patterns identified. In contrast, classification uncertainty, validation design, classifier comparison, and differences in satellite processing levels should be considered when interpreting individual transitions and their spatial associations with HWC.

Another limitation of this research is the inconsistency and limited availability of HWC data. Reliable data were available only from 2019 to 2025, limiting the analysis of long-term relationships between land-use/land-cover changes and conflict occurrences. Additionally, the temporal coverage of the HWC records did not fully align with the LULC change intervals. Specifically, HWC records were only available for 2019–2020 to assess the 2015–2020 LULC transition. Consequently, the observed spatial relationships reflect the available conflict patterns in the latter part of the transition period rather than across the entire 2015–2020 interval. Consequently, the two analyses are not directly comparable. The findings should be interpreted as evidence of spatial association between LULC change and HWC within the available data period, rather than as a complete assessment of conflict dynamics throughout the full five-year interval. The apparent increase in conflict intensity may partly reflect longer observation periods.

The HWC database may be subject to reporting bias, as not all conflict incidents are reported or formally recorded. Variations in reporting effort across communities and administrative regions may also have resulted in the under-representation or over-representation of conflict hotspots. Despite these data limitations, consistent spatial patterns in HWC incidents suggest spatial associations between landscape structure and conflict hotspots, especially near protected areas. A further limitation is that wildlife movement was implied indirectly from spatial patterns of LULC change and HWC occurrence, rather than validated using direct movement data such as GPS telemetry. The study also focused primarily on environmental variables. It did not incorporate socio-economic factors, including livelihood dependence, crop types, or mitigation practices, which may provide additional context for interpreting HWC occurrence and reporting behavior. Finally, although significant spatial associations between LULC dynamics and HWC were identified, the study did not employ formal causal modelling; therefore, the findings should be interpreted as evidence of association rather than causation.

6 Recommendations for future research

According to the study’s findings, changes in land use and land cover, specifically the increase in built-up areas, the loss of natural habitats, and the fragmentation of ecological corridors, are spatially associated with changes in wildlife distribution patterns and an increased likelihood of human-wildlife interactions. Comparative, multi-regional studies should be the focus of future research to understand better how various species respond behaviorally to varying degrees of habitat disturbance. To inform land-use planning and conflict mitigation strategies, integrated and interdisciplinary approaches that combine ecological, spatial, and socio-economic data are needed, given the substantial socio-economic consequences of human–wildlife conflict, including threats to livelihoods, crop losses, and property damage. To capture fine-scale ecosystem dynamics and wildlife movement corridors, future research should leverage advances in high-resolution remote sensing, GPS-based wildlife tracking, and geospatial modelling. These methods would facilitate more focused, evidence-based conservation efforts and enhance the identification and forecasting of HWC hotspots.

Furthermore, accounting for livestock movement patterns, especially in the vicinity of protected areas, would provide important insights into how grazing dynamics interact with LULC changes and potentially contribute to conflict risk. Researchers should use higher temporal resolution and real-time conflict-monitoring data, along with other explanatory variables such as proximity to settlements, water sources, and road networks, to increase the robustness of subsequent analyses. This would reduce uncertainty about reporting bias and enhance understanding of the temporal and spatial determinants of conflict between humans and wildlife.

7 Conclusion

The findings of this study show notable spatiotemporal changes in land cover and land use surrounding Hluhluwe–iMfolozi Park during the study period, including shifts within plant classes, increases in built-up areas, and decreases in natural habitats. Because habitat fragmentation and land-use change may restrict wildlife mobility and increase the likelihood of human-wildlife contact, these landscape changes may be linked to patterns of human-wildlife conflict. The usefulness of sophisticated geospatial methods for ecological monitoring was highlighted by comparing machine learning and deep learning approaches, which showed that CNN-based models achieved higher accuracy in mapping complex land-cover types. By integrating high-resolution LULC data with spatially referenced HWC incidents, this study provides important insights into the spatial distribution of conflict hotspots. It supports the development of targeted mitigation strategies.

The findings highlight how landscape change is spatially associated with human-wildlife interactions, especially where natural ecosystems meet agricultural and urban areas. Even though protected areas remain essential for biodiversity conservation, growing pressure from nearby land-use changes underscores the need to consider landscape dynamics that extend beyond their borders. Overall, this study emphasizes the importance of combining ecological knowledge, remote sensing, and geospatial analysis to resolve conflicts between humans and wildlife. Planning at the landscape level to preserve habitat connectivity, control transitional land-use zones, and promote sustainable coexistence between human and wildlife populations will be necessary for effective mitigation.

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.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frsen.2026.1914197/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Ahmet Cilek, Çukurova University, Türkiye

Reviewed by: Isaac Kipkemoi, University of Embu, Kenya

Asima Khan, University of Leicester, United Kingdom