Mining asset detection (MAD)
Description #
Built during my civil service at EAWAG, MAD uses the YOLO object-detection model to locate mining activity around the globe from Sentinel-2 satellite imagery.
Mining sites are mapped inconsistently and often not at all, which makes it hard to study their environmental footprint at scale. Sentinel-2 covers the entire land surface every few days and is free, so the question was whether a detector could find the sites directly in the imagery.
Data pipeline. The bulk of the work was getting from raw satellite archives to something trainable: downloading terabytes of Sentinel-2 granules, filtering by cloud cover, selecting and normalising the relevant spectral bands, stitching and re-tiling scenes into fixed-size chips, and aligning them with georeferenced labels of known mining sites.
Modelling. Trained and evaluated YOLO variants on the resulting dataset, comparing band combinations and tiling strategies, and running inference over large unlabelled regions to surface candidate sites.