An early-stage research and engineering project to explore whether an AI model, built by fine-tuning an existing open-source Earth-observation foundation model, can help detect buried subsurface caves and voids in karst / limestone terrain using open-source satellite imagery.
Optical satellites cannot see underground. What we can realistically detect from space are the surface and near-surface signals that correlate with karst voids: ground subsidence/deformation (InSAR), surface karst morphology (sinkholes, dolines, uvalas), thermal anomalies, topographic depressions, and vegetation/moisture anomalies. This project treats "cave detection" as detecting these observable proxies and flagging areas of elevated probability — not as literally imaging a buried chamber. Confirmation always requires airborne LiDAR, geophysics, or ground survey.
- Learn from known caves. Use registries of known karst caves as positive training labels, paired with matched negative (non-karst / no-cave) controls.
- Fine-tune an open-source foundation model. Start from a pretrained geospatial model (e.g. Prithvi-EO, Clay, SatMAE, Satlas) rather than training from scratch.
- Multi-modal input. Fuse Sentinel-1 SAR/InSAR + Sentinel-2 multispectral + DEM + thermal where useful.
- Infer on new regions. Apply the trained model to unsurveyed karst areas to produce ranked candidate locations for follow-up.
docs/
methodology.md Distilled, actionable plan (start here)
research/
01-prior-art-competitive-survey.md Is anyone already doing this? (verdict: no)
02-data-sufficiency.md Enough open data to train + validate? (verdict: yes)
03-methodology-deep-research.md Verified deep-research report (feasibility, models, pipeline)
| Question | Verdict |
|---|---|
| Already solved by an existing tool? | No — open problem; buried voids are physics-limited from satellite. |
| Enough open data to train? | Yes — proxy labels (sinkholes/dolines) are abundant, far above fine-tuning needs. |
| Enough open data to validate / run? | Yes retrospectively; imagery is free and global. Confirming novel finds needs fieldwork. |
Recommended stack: fine-tune Prithvi-EO-2.0 (or Clay) with LoRA
via TerraTorch, framed as segmentation of surface proxies, validated with
spatial block cross-validation. See docs/methodology.md.
🔬 Research phase complete. Problem space, data sources, candidate models, and a realistic pipeline are documented. Next step (engineering: data-download + label assembly + fine-tuning pipeline) begins only on explicit go-ahead.
Note on sensitivity: precise cave locations can be legally protected and ecologically/culturally sensitive (e.g. bat habitats, archaeology). This project handles location data responsibly and does not aim to publish precise coordinates of vulnerable sites.
