Using High-Resolution Satellite Imagery and Deep Learning to Map Artisanal Mining Spatial Extent in the Democratic Republic of the Congo

Highlights: What are the main findings? We developed a deep learning framework to map Artisanal and Small-scale Mining (ASM) in the DRC, supported by a pseudo-ground-truth dataset derived from ASM field observations through a multi-stage processing pipeline. The Late Fusion model combining Planet-NICFI optical and Sentinel-1 SAR data achieved the best performance (F1 = 0.73; overall accuracy = 88.4%) for ASM detection. What is the implication of the main finding? The integration of optical and SAR data enhances ASM detection under conditions of limited ground truth and persistent cloud cover. The resulting high-resolution maps can provide an operational tool for regional monitoring, policy support, and sustainable resource management. Artisanal and Small-scale Mining (ASM) significantly impacts the Democratic Republic of Congo’s (DRC) socio-economic landscape and environmental integrity, yet its dynamic and informal nature makes monitoring challenging. This study addresses this challenge by implementing a novel deep learning approach to map ASM sites across the DRC using satellite imagery. We tackled key obstacles including ground truth data scarcity, insufficient spatial resolution of conventional satellite sensors, and persistent cloud cover in the region. We developed a methodology to generate a pseudo-ground truth dataset by converting point-based ASM locations to segmented areas through a multi-stage process involving clustering, auxiliary dataset masking, and manual refinement. Four model configurations were evaluated: Planet-NICFI standalone, Sentinel-1 standalone, Early Fusion, and Late Fusion approaches. The Late Fusion model, which integrated high-resolution Planet-NICFI optical imagery (4.77 m resolution) with Sentinel-1 SAR data, achieved the highest performance with an average precision of 71%, recall of 75%, and F1-score of 73% for ASM detection. This superior performance demonstrated how SAR data’s textural features complemented optical data’s spectral information, particularly improving discrimination between ASM sites and water bodies—a common source of misclassification in optical-only approaches. We deployed the optimized model to map ASM extent in the Mwenga territory, achieving an overall accuracy of 88.4% when validated against high-resolution reference imagery. Despite these achievements, challenges persist in distinguishing ASM sites from built-up areas, suggesting avenues for future research through multi-class approaches. This study advances the domain of ASM mapping by offering methodologies that enhance remote sensing capabilities in ASM-impacted regions, providing valuable tools for monitoring, regulation, and environmental management.