Mining footprint detection with multi-modal satellite data
Mining footprint delineation is a prerequisite for environmental compliance monitoring and land rehabilitation planning at national scale. Across Australia, one of the world's largest mining jurisdictions, mapping active and historical footprints previously required months of manual GIS analyst time per region, making continuous monitoring economically infeasible for regulatory bodies.
Single-modality optical imagery (Sentinel-2) is insufficient for robust footprint delineation. Cloud cover interrupts temporal consistency, spectral signatures of disturbed land overlap with natural bare soil and dry vegetation, and underground or sub-canopy operations are invisible to optical sensors. Integrating Synthetic Aperture Radar (Sentinel-1 SAR) and Digital Elevation Model (DEM) data introduces a multi-modal fusion problem with heterogeneous input dimensionality, temporal misalignment across sensors, and domain shift across ore types (iron ore vs. coal vs. bauxite) and vegetation biomes (arid vs. temperate).
Built a multi-modal deep learning pipeline fusing three independent input streams, optical (Sentinel-2, 13 bands), SAR (Sentinel-1 VV/VH backscatter), and DEM (elevation + slope derivatives). Stream-specific encoders extract modality-aligned features before a late-fusion aggregation head. Fine-tuned Prithvi, a NASA/IBM geospatial foundation model pre-trained on Sentinel-2 and Landsat, for mining-specific segmentation, representing the first published application of a geospatial FM to this task.
Built ingestion and preprocessing pipelines for 500GB+ of labeled multispectral imagery from Google Earth Engine. Coordinated peer-reviewed annotation with geographers across University of Queensland, UCL, and University of Nottingham, spanning multiple ore-type and biome classes, implementing inter-annotator agreement checks to ensure label quality at scale.
Addressed temporal change detection across multi-year image stacks, distinguishing active operations from rehabilitated or historically disturbed land, a distinction with direct legal and financial implications under Australian mining regulation.
Enables automated, scalable land-use accountability for environmental compliance and rehabilitation planning across continental-scale mining regions, replacing months of manual GIS analysis with a satellite-speed automated pipeline. Directly informs regulatory workflows for Australian state environmental agencies and international mining company ESG reporting.
- Monash University, Primary host (A/Prof Risqi U. Saputra, Prof Alex Lechner)
- University of Queensland, Remote sensing science and annotation
- UCL, Annotation and domain expertise
- University of Nottingham, Annotation and domain expertise