Monitoring Ecosystem Dynamics Using Machine Learning: Random Forest-Based LULC Analysis in Dinder Biosphere Reserve, Sudan ^†

Abstract

Dinder Biosphere Reserve (DBR), a UNESCO-recognized biodiversity hotspot in Sudan, faces escalating land-use pressure. We analyzed land cover changes from 2019 to 2024 using Sentinel-2 imagery processed in Google Earth Engine. A Random Forest classifier identified five land cover classes: water, built-up areas, vegetation, bare land, and crops. The transition matrix revealed significant changes over this period. About 1501 km² of vegetation and 1648 km² of cropland were converted to bare land. Built-up areas lost 95 km² to bare land. Bare land remained largely unchanged (4749 km²), while water bodies were the most stable (13,473 km² unchanged). Only minor transitions involved water (27.6 km² to vegetation, 15.2 km² to bare land). Notably, 411 km² of cropland and 1773 km² of bare land transitioned to vegetation, indicating some regrowth. These land cover changes reflect a dynamic interplay between degradation and recovery processes; however, the results should be interpreted with caution due to potential classification inaccuracies, seasonal variation in imagery, and absence of field validation. Continued satellite monitoring is essential to guide adaptive land management and safeguard ecosystem function in DBR.

1. Introduction

Analyzing and modeling changes in land use and land cover (LULC), alongside their related ecosystem service values, has become an essential global objective for promoting sustainable land management policies [1]. LULC changes significantly alter the structure and functionality of terrestrial ecosystems on global scale [2]. These transformations often involve shifts in how landscapes are utilized and function, taking forms such as urbanization, deforestation, agricultural expansion, and infrastructure development [3]. Such changes, driven by factors like population growth, economic activities, and agricultural intensification, profoundly affect landscapes, contribute to climate change, threaten biodiversity, and influence human health and well-being [4]. In resource-dependent regions like Sudan [5], understanding these dynamics is crucial for developing effective conservation strategies and sustainable land-use practices [6].

Protected areas are integral to both national and international conservation strategies. They play a critical role in maintaining natural ecosystem functions and conserving biodiversity [7]. In semi-arid and savanna regions such as the Dinder Biosphere Reserve, the increasing frequency of droughts and reduced rainfall significantly impact vegetation cover. Vegetation cover is crucial for sustaining wildlife habitats, mitigating erosion and desertification, and regulating both local and global climates through carbon balance and greenhouse gas moderation [8,9,10]. These ecosystem services highlight the importance of vegetation cover for sustainable management.

LULC change is significantly influencing natural resource management, environmental modeling, and agricultural planning. However, detecting and modeling these changes remain a complex and data-intensive task within the remote sensing field due to the need for processing extensive historical and current datasets, real-time scenario interactions, and spatial environmental data analysis [11]. Advances in remote sensing technologies, combined with the increasing availability of satellite systems over recent decades, have exponentially expanded access to spatial data [12,13]. This growth provides unparalleled opportunities for detailed LULC analysis, despite managing these extensive datasets remains a significant challenge.

Modern Earth observation imagery and advanced classification algorithms have facilitated progress in LULC research [14]. Traditional classification methods like maximum likelihood and spectral angle analysis, while they are foundational, often fall short in achieving the high accuracy demanded today [15]. Since the launch of Landsat-1 in 1972, machine learning algorithms have played an increasing and central role in LULC classification, including the use of neural networks, support vector machines, and decision trees [16]. Neural networks, in particular, have proven to be robust and highly effective for classifying large-scale multispectral datasets [17].

Among ensemble learning techniques, Random Forest (RF) stands out as a robust method, effectively combining weak classifiers to form a strong decision-making framework that excels in LULC classification [16,18]. RF has consistently demonstrated high accuracy in producing reliable land cover maps [19]. Several recent studies have successfully applied RF for LULC classification across diverse landscapes. For instance, in 2021, Reyes-Palomeque et al. utilized RF to classify land cover in tropical semi-deciduous and semi-evergreen forests, achieving an overall accuracy of 91% and 88.37%, respectively [20]. Similarly, Chowdhury in 2024 employed RF for urban–rural land use mapping, highlighting its ability to handle complex datasets with high-dimensional features [21]. Additionally, Lawer in 2024 demonstrated the effectiveness of RF in semi-arid savanna ecosystems, emphasizing its robustness in handling mixed vegetation classes and noisy data [22]. The advent of cloud-based platforms such as Google Earth Engine (GEE) has further revolutionized LULC analysis, allowing researchers to efficiently process large datasets and generate high-resolution spatial maps [23]. These technologies facilitate spatio-temporal analyses that inform conservation planning and policy development [24].

This study employs Random Forest-based LULC analysis to monitor ecosystem dynamics in Dinder Biosphere Reserve. By leveraging advanced remote sensing and machine learning tools, the research contributes to understanding the drivers of ecosystem changes while offering actionable insights for sustainable management and restoration in one of Sudan’s most ecologically significant conservational areas.

2. Methodology: Study Area and Analysis Framework

The Dinder Biosphere Reserve in eastern Sudan, located between 12°26′ and 12°42′ North latitude and 34°48′ and 35°02′ East longitude, covers an area of approximately 10,291 km². It is divided into three management zones: the Core Zone, which is strictly protected; the Buffer Zone, where resource use is regulated; and the Transition Zone, which is subject to high human pressure (Figure 1). The reserve experiences a tropical climate with an average annual rainfall of 775 °C and temperatures ranging from 18 to 30 °C [25]. The dominant tree species in the area include Acacia seyal, Balanites aegyptiaca, and Terminalia brownii. Land-use change continues to threaten the ecological balance of the reserve [26].

This study assessed land cover change from 2019 to 2024. The acquisition and processing of Sentinel-2 imagery were performed in Google Earth Engine (GEE). Median composites for each year, 2019 and 2024, were created using dry and wet season images (<20% cloud cover), selecting 10 spectral bands. Land cover was classified into five categories (Water, Bare land, Vegetation, Built-up, Crops) using a Random Forest (RF) classifier (500 trees, 70/30 train–validation split). Accuracy was assessed with confusion matrices to evaluate the quality of the classification results [27]. Change detection was conducted via (1) image differencing and (2) unique class encoding ((2019 × 100) + 2024) for pixel-level transition analysis. All outputs were visualized in QGIS.

3. Results and Discussion

The LULC classification of Sentinel-2 images from 2019 to 2024 identified five major land cover categories within the Dinder Biosphere Reserve (DBR): water, built-up areas, vegetation, bare land, and crops (Figure 2). The total land area analyzed was approximately 25,323.35 km².

Table 1. LULC transition matrix 2019–2024 in DBR: percentage (% of total area).

Reference Class	Water	Built-Up	Vegetation	Bare Land	Crops	Total
Water	53.20	0.00	0.11	0.06	0.00	53.37
Built-up	0.00	0.15	0.04	0.38	0.02	0.59
Vegetation	0.01	0.09	5.00	5.93	0.35	11.39
Bare land	0.01	0.35	7.00	18.75	0.16	26.27
Crops	0.01	0.18	1.62	6.51	0.05	8.38
Total	53.24	0.78	13.78	31.62	0.58	100.00

summarizes the dynamics of LULC category over the five-year period. The land cover transition matrix revealed significant patterns of change within the study area. Water exhibited the highest stability, with approximately 13,472.84 km² remaining unchanged, and only minor conversions to vegetation (27.65 km²) and bare land (15.19 km²), suggesting minimal fluctuation in water bodies during the study period. Bare land showed relative stability (4749.29 km²), yet simultaneously contributed a substantial area (1772.62 km²) to vegetation, indicating potential regeneration or afforestation efforts.

Vegetation, while moderately stable (1267.40 km²), underwent the most dynamic transitions, losing 1500.69 km² to bare land, which likely reflects deforestation or degradation. Additionally, significant portions of croplands transitioned to bare land (1648.01 km²), pointing to land abandonment or reduced agricultural activity. Built-up areas experienced notable loss, with 95.29 km² reverting to bare land, possibly due to destruction or de-urbanization. Smaller transitions such as crops to vegetation (410.94 km²) and water to bare land (15.19 km²) further underscore the ecological dynamism of the landscape. Overall, the matrix highlights a complex interplay of degradation and recovery processes, influenced by both natural and anthropogenic factors.

The land cover transitions between 2019 and 2024 reflect both degradation and recovery processes within the study area. Notable vegetation and cropland losses to bare land suggest land degradation or abandonment, while gains in vegetation from bare land may indicate regeneration or improved environmental conditions. These changes may also suggest reduced human pressure or increased rainfall, as observed in Sudan by Ahmed et al. [28]. However, the magnitude of change should be interpreted with caution, as classification errors, seasonal imagery differences, and lack of field validation may hinder obtaining optimal accuracy of results [19]. These land cover transitions underscore the value of continuous remote sensing-based monitoring to inform sustainable land management and adaptive planning in protected ecosystems like DBR.

The classification model achieved an Overall Accuracy of 0.77, indicating that 77% of the total classified pixels were correctly assigned to their respective classes when compared to the reference data. While this level of accuracy demonstrates reasonable performance, further improvements may be explored to address potential sources of misclassification, particularly in areas with overlapping spectral signatures or heterogeneous land cover patterns.