Estimating the Footprint of Artisanal Mining in Africa

Abstract

Artisanal and small-scale mining (ASM) is a significant but often unregulated economic sector in sub-Saharan Africa. The lack of comprehensive monitoring of ASM activity hinders efforts to address its environmental and social impacts or promote its sustainable development. We address this gap by creating a large, high-resolution dataset of manually labeled ASM sites across five countries in the region. Using these data, we train machine learning models that integrate geographic features and satellite imagery to map ASM at scale. Our results demonstrate that these models can outperform on-the-ground surveyors in areas that have been previously monitored and can successfully detect ASM in previously unmonitored areas. We develop machine learning methods for domain adaptation in order to train models that can adapt to new spatial regions, a setting where standard machine learning approaches fail. Within these five countries, we predict that over 231,000 1 km$^2$ grid cells contain ASM activity, over forty times more than what is recorded by on-the-ground mapping efforts. Moreover, we estimate that 2% of the cells within protected areas reserved for conservation contain ASM. Using domain-adapted algorithms, we predict ASM across 15 additional countries in sub-Saharan Africa and, again, uncover the intrusion of ASM into protected areas and biodiversity hotspots.