Using High-Resolution Satellite Imagery and Deep Learning to Map Artisanal Mining Spatial Extent in the DRC

By Francesco Pasanisi

Abstract
Artisanal and Small-scale Mining (ASM) significantly impacts the Democratic Republic of Congo’s (DRC) socio-economic and environmental integrity. This work presents an investigation into the application of remote sensing and deep learning to segment ASM areas. A critical part of this project was to address the lack of reference data for this study area. The creation of such dataset was crucial for the training and validation of deep learning models. Utilizing the U-Net architecture, the effectiveness of integrating high-resolution optical data and SAR data was evaluated. The Late Fusion approach adopted enhanced ASM area segmentation, achieving an average precision of 71%, recall of 75% and F1-score of 73%. These metrics highlight the model’s accuracy and the value of data fusion in improving segmentation results, revealing that SAR data textural features and optical data spectral information provide complementary and useful information that is critical for accurate ASM segmentation. Furthermore, the developed model was deployed on a targeted area of the DRC territory
to show its utility in a real-world scenario. The map created with the developed model scored an overall accuracy of 88.4%. However, this work also identified challenges such as distinguishing ASM sites from built-up areas. These findings advance the domain of ASM detection, offering methodologies that can enhance the monitoring and management capabilities of remote sensing within ASM-impacted regions.