Reproducible research artifact — not a live monitoring product
Detecting river-corridor disturbance signatures consistent with reported sand extraction
SandWatch studies whether openly available Sentinel-2 imagery can identify physical disturbance signatures — exposed sediment, altered water extent, channel change — in river corridors that independent press, NGO, and academic sources have reported as sites of sand or aggregate extraction. It evaluates this under a spatially honest leave-one-river-out holdout across 7 real river reaches.
What this project is not
This is not a system for detecting illegal mining. No verified, adjudicated ground truth for illegal extraction exists or is used here. Labels come from documented reporting at reach/district granularity (Level C — see the labeling handbook), and a disturbance signature can equally arise from floods, construction, agriculture, or natural channel migration. Every figure and number on this site traces to a file under results/ in the source repository — see the Methods and Data pages for full provenance.
5 case-study reaches
Real river reaches with documented press/NGO/academic reports of extraction activity — e.g. Yamuna (Shamli, India), Krishna (Vijayawada, India), Mekong (Chau Doc, Vietnam).
2 comparison reaches
Reaches with no located extraction report (a formally banned river set in Kerala, and a U.S. federally protected riverway) — weak negatives, not verified absences.
Headline result status
| Model | Accuracy | Balanced accuracy | F1 |
|---|---|---|---|
| gradient_boosting | 0.625 | 0.400 | 0.769 |
| logistic_regression | 0.462 | 0.296 | 0.632 |
| majority | 0.786 | 0.500 | 0.880 |
| random_forest | 0.634 | 0.405 | 0.776 |
| spectral_threshold_bsi | 0.535 | 0.530 | 0.646 |
n=7 reaches. Per-fold precision/recall are not separately meaningful with a single held-out reach per fold — see Models for the full, honest breakdown and Methods for why.