Global Surface Coal Mine Segmentation with Multi-Spectral Sentinel-2 Imagery: Overcoming Spectral Dilution using Geospatial Foundation Models
Surface coal mines are a diffuse, undermonitored source of methane, whose reduction could deliver measurable near-term climate benefit. A lack of precise spatial boundaries and reliance on global emission inventories (based on point coordinates with kilometer-scale spa- tial uncertainty), limit their continuous spaceborne attribution and monitoring. To bridge this gap, we introduce the Global Surface Coal Mine (GSCM) Polygons dataset, obtained by spatially match- ing the Global Coal Mine Tracker to the Global Mining Footprint and linking candidate polygons to cloud-free Sentinel-2 imagery. We benchmark semantic-segmentation models from convolutional, vision-transformer, and geospatial-foundation-model families. On spatially isolated test tiles containing mine pixels, a pre-trained Prithvi-EO-2.0 model with six visible and infrared bands obtains the best mine recall (0.6975) and mIoU (0.6971). In contrast, adding channels to an ImageNet-pretrained SegFormer causes mine IoU to fall to 0.0094 at 15 bands. We term this spectral dilution. While the positive-only evaluation does not establish global mine discovery or operational readiness, the results support a practical baseline for delineating known surface mines.
Authors
- Piyush Yadav (ORCID: https://orcid.org/0000-0002-4872-0205)
- Chengxing Zhai (ORCID: https://orcid.org/0000-0002-0291-4522)
- Hui Su
Institutions
- Hong Kong University of Science and Technology (HK)
- University of Hong Kong (HK)
Publication Details
- Journal
- EarthArXiv (California Digital Library)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.31223/x5jj99
- Primary Topic
- Remote-Sensing Image Classification
- Type
- preprint