Cross-Country Transferability of Deep Learning Models for Artisanal and Small-Scale Mining Mapping in Ghana and Côte d’Ivoire

Artisanal and small-scale mining (ASM) is expanding across West Africa, driving deforestation, water contamination, and land degradation that require scalable monitoring approaches. Deep learning (DL) models applied to Sentinel imagery have shown acceptable performance for ASM mapping within individual countries, but their ability to transfer across national boundaries remains poorly understood. This study evaluates four DL segmentation models for ASM mapping in Ghana and Côte d’Ivoire, two major gold-producing countries with contrasting mining landscapes. The models were trained under three geographic scenarios of Ghana only, Côte d’Ivoire only, and both countries combined, using the SmallMinesDS benchmark dataset for Ghana and a newly compiled dataset for Côte d’Ivoire, both combining Sentinel-1 SAR and Sentinel-2 optical imagery. Models performed best when trained and tested on the same country, with U-Net++ (EfficientNet-B3) reaching IoU = 0.748 (F1 = 0.856) on Ghana. On the harder Côte d’Ivoire domain, models trained on Ghana alone transferred poorly (IoU = 0.488, F1 = 0.656), but training on both countries together gave the best result (IoU = 0.568, F1 = 0.724). Overall, the geography of the training data, rather than the model architecture or the use of SAR data, was the main factor driving performance, though some differences may also reflect the datasets being collected at different times.