The study area is no ordinary landscape. Hluhluwe-iMfolozi Park spans roughly 960 square kilometers of woodland savanna, shrub thicket, forest, and grassland, and is home to the Big Five: lion, leopard, elephant, buffalo, and rhino. The park is fenced except where its main rivers flow out, but it is bordered by settlements and agricultural fields, creating a complex interface between conservation and human activity. Reports of crop raiding by African savanna elephants and baboons, livestock predation by lions and spotted hyenas, and sporadic large-carnivore attacks on humans are well documented in South Africa, and such incidents are increasingly reported around protected areas as wildlife move during migration or foraging.

To capture the shifting landscape, the team acquired Sentinel-2 multispectral imagery for each study year, filtering acquisitions between late June and December with a 50 percent cloud-cover threshold and removing cloud pixels using the QA60 bitmask. They computed the Normalized Difference Vegetation Index from the near-infrared and red bands and applied an NDVI-based quality mosaic, prioritizing vegetated, cloud-free pixels to reduce cloud contamination and seam lines. Alongside the raw spectral bands — blue, green, red, near-infrared, and shortwave infrared — they generated four spectral indices: NDVI for vegetation greenness, SAVI adjusted for soil brightness, NDBI for built-up and impervious surfaces, and BSI for bare soil. These features were designed to sharpen the separation between vegetation, settlements, and exposed ground.

Training data came from manual digitization of high-resolution imagery combined with field knowledge, producing 52,380 reference samples across nine land-cover classes: waterbodies, cultivated land, forest plantations, barren land, built-up areas, natural forests, grasslands, shrublands, and woodlands. The samples were split 70/30 into training and validation sets using stratified sampling with a fixed random seed for reproducibility. Four classifiers were then pitted against one another: Random Forest with 300 decision trees, a Support Vector Machine with standardized features and a margin parameter of 10, a Multi-layer Perceptron with two hidden layers of 128 and 64 neurons and dropout regularization, and a two-dimensional Convolutional Neural Network processing 32-by-32-pixel patches as multi-channel images.

The CNN won decisively. It achieved an average overall accuracy of 99.92 percent and a kappa coefficient of 0.9991, outperforming Random Forest (above 95 percent), the Multi-layer Perceptron (above 94 percent), and the Support Vector Machine (above 93 percent). The convolutional architecture’s advantage lies in its ability to learn both spectral correlations and spatial patterns concurrently, extracting hierarchical features from image patches through three convolutional layers with 32, 64, and 128 filters, batch normalization, and max pooling. Most land-cover classes achieved precision, recall, and F1-scores of 1.00, with only minor reductions for cultivated land, built-up areas, and grasslands in 2025. The authors note, however, that training and validation samples were drawn from the same study area and were not spatially independent, so the near-perfect accuracies should be interpreted with some caution regarding generalization to geographically independent regions.

The classified maps exposed a decade of dramatic change. Built-up areas expanded by nearly 395 square kilometers between 2015 and 2020 — a 32.84 percent increase — and continued growing thereafter. Cultivated land declined steeply, falling by 257 square kilometers between 2020 and 2025, a 47.77 percent reduction. Shrubland exploded by 334 square kilometers in the same period, the largest vegetation gain observed, while woodland increased between 2015 and 2020 before contracting. Barren land collapsed by more than 81 percent in the first interval, and forest plantations shrank by nearly 40 percent. The largest transitions between 2020 and 2025 were woodland converting to shrubland and woodland converting to grassland, suggesting shifting vegetation structure possibly linked to grazing pressure, fire regimes, or succession — though the study did not directly measure these drivers.

Crucially, the magnitude of change differed sharply across the protected-area boundary. Outside the park, built-up areas gained 417 square kilometers in the first interval and a further 62 square kilometers in the second, while cultivated land fell by 257 square kilometers and shrubland surged by 236 square kilometers. Inside the protected area, changes were generally smaller in absolute area, reflecting the stabilizing effect of conservation management and restrictions on habitation and cultivation. The researchers grouped the 81 possible class-to-class transitions into ecologically meaningful categories — stable, vegetation loss, vegetation gain, natural-to-natural, and anthropogenic-to-anthropogenic — and found that anthropogenic activities and vegetation gains dominated earlier transitions, while natural vegetation transitions drove later changes.

Then came the overlay with conflict. Ezemvelo KZN Wildlife provided 187 human-wildlife conflict incident records from 2019 to 2025, each with location, species, area name, and date. Elephants accounted for 79 incidents and lions for 71, with buffalo, wild dogs, baboons, leopards, Nile crocodiles, white rhinos, and spotted hyenas making up the remainder. Kernel density estimation revealed that conflict events were increasingly linked to grassland, woodland, and transitional landscapes, and were concentrated along protected-area boundaries. Between 2015 and 2020, incidents were comparatively low and dispersed across few transitions; by 2020 to 2025, both the quantity and variety of conflict-related transitions had increased markedly, with woodland-to-woodland transitions predominating, followed by built-up-to-built-up and grassland-to-woodland conversions.

The mean conflict density differed among land-cover classes in both periods. Grassland and woodland consistently displayed the highest densities among stable classes, rising from 1.215 to 1.328 incidents per square kilometer and from 1.015 to 1.155 incidents per square kilometer respectively. Among change classes, vegetation gain had the highest density in the first interval but declined in the second, while vegetation loss increased slightly. Natural-to-natural transitions remained relatively high throughout, suggesting that conflict is not restricted to areas of rapid conversion but also occurs within relatively stable habitats where wildlife activity and human presence overlap in proximity.

The spatial pattern points to edge effects as a key mechanism. Abrupt transitions between human-dominated landscapes and protected natural habitats create high-risk interaction zones, and the concentration of incidents near settlements supports the interpretation that wildlife displacement extends beyond protected-area boundaries. The authors frame their findings in terms of habitat fragmentation and reduced connectivity: conversion of natural vegetation to human-dominated land uses may push animals into modified landscapes where conflict is more likely, while agricultural fields can provide concentrated food resources that attract crop-raiding species. Changes in water availability may represent another mechanism influencing wildlife movement, though the study did not directly measure wildlife dependence on these resources.

For conservation planning, the implications are targeted rather than uniform. The researchers argue that mitigation should prioritize the protected-area-community interface, focusing ranger patrols, community-based reporting, early-warning systems, and deterrent measures where conflict hotspots coincide with rapidly changing land cover. They also recommend proactive monitoring of settlement growth using regularly updated satellite imagery, so that areas where development approaches existing vegetation patches or known hotspots can be identified before incidents occur. Maintaining landscape connectivity — retaining strategically located vegetation patches and corridors between the park and surrounding habitats — should form part of the planning framework, though the authors caution that wildlife movement was inferred indirectly from spatial patterns rather than validated with GPS telemetry.

The study is candid about its limitations. Classification errors may propagate into change-detection analyses, producing apparent transitions where no actual conversion occurred. The 2015 imagery used top-of-atmosphere reflectance while later years used atmospherically corrected surface reflectance, introducing potential inconsistencies in spectral information. Conflict records were only reliably available from 2019 onward, and reporting bias may under- or over-represent certain hotspots. The study also aggregated incidents across species, so identified hotspots represent overall conflict concentration rather than species-specific responses. Nevertheless, the consistent spatial patterns — conflict clustering at boundaries, hotspots overlapping active change zones, and increasing frequency over time — provide a robust evidence base for the spatial association between landscape dynamics and human-wildlife interactions.

What emerges is a portrait of a landscape in flux, where the collision of conservation and development is not an abstract threat but a measurable, mappable phenomenon. As built-up areas expand and vegetation structure shifts, the interfaces where wildlife and people meet are multiplying. The study demonstrates that sophisticated geospatial technologies — deep learning classification, kernel density estimation, and spatial overlay — can pinpoint conflict hotspots with enough precision to guide targeted mitigation. For Hluhluwe-iMfolozi Park and the communities that surround it, the message is clear: the human-wildlife interface is changing, and managing that change will require looking beyond the park’s fences to the landscape as a whole.

Subject of Research: Human-wildlife conflict around Hluhluwe-iMfolozi Park is spatially linked to land-cover change, with hotspots concentrated along protected-area boundaries.

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

Article References: Njomi, O., Gwala, O. Z., Altoom, M. B., Ramoelo, A., Chandran, R., Avtar, R., & Gebreslasie, M. (2026). Assessing the spatiotemporal dynamics of land-use/land-cover change and human-wildlife conflict around Hluhluwe-iMfolozi park, South Africa. Frontiers in Remote Sensing, 7, Article 1914197. https://doi.org/10.3389/frsen.2026.1914197

Image Credits: AI Generated

DOI: 10.3389/frsen.2026.1914197

Keywords: human-wildlife conflict, deep learning, Sentinel-2, land-use change, protected areas, South Africa, conservation, Assessing, spatiotemporal, dynamics, land-use, land-cover