Mine Site and Tailings Monitoring Using Sentinel-2 and Machine Learning

Since the Brumadinho tailings-dam collapse (Brazil, 2019-01-25) operators and regulators want independent maps of mine footprints and tailings facilities. This project maps mining land on free Sentinel-2 Level-2A images (30 tiles of 10.24 km in 11 regions on a standard laptop CPU, 16 GB RAM) with classical index thresholds and machine learning, tested on regions never seen in training, against the Maus et al. (2022) polygons, the only reference. Pooled F1 with regions held out: RF + context 0.627, U-Net 0.622, Random Forest 0.555; classical baselines Tuned thresholds 0.591, Spec rule 0.528, trivial all-mining 0.460. Against the tuned thresholds the U-Net gains +0.031 (95 % CI -0.044 to +0.111): the interval includes 0, so it is not shown to beat them; against the spec rule it gains +0.094 (95 % CI +0.013 to +0.176), interval excludes 0. Leakage demo, labelled as such and not as a result: the Random Forest scores 0.763 on random pixels, 0.656 on west / east halves and 0.645 on site-grouped folds, against 0.555 with regions held out; the U-Net 0.766 on random chips against 0.622 with regions held out. RF marks 95.8 % of the Calama desert tile as mining (almost all of it). Brumadinho: 2.32 km2 connected to the dam (below spec 2.5-3.5 km2); control pair 0.23 km2 connected, 3.54 km2 in the window against 2.50. Jagersfontein: first date after the failure (2022-09-16) NOT flagged, z -1.00; 1 flag(s) before it; median pond 0.342 to 0.143 km2. One event is one case. Research and teaching prototype: not a dam-safety, compliance or legal boundary system.

Category: Mining Engineering

Mining Engineering Project MN06 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.