A SAR-based flood detection and situational awareness system for the Brahmaputra basin in Assam.
Optical satellite imagery (Sentinel-2, Landsat) can't see through cloud, and the Brahmaputra floods during monsoon season — exactly when the sky is cloud-covered. Sentinel-1 Synthetic Aperture Radar (SAR) is an active microwave sensor: it penetrates cloud and works at night, which makes it the only practical modality for monitoring floods as they happen rather than after the clouds clear.
- Ingest Sentinel-1 SAR scenes.
- Run a trained CNN segmentation model to produce pixel-level flood-water masks.
- Aggregate those masks up to administrative units (state/district).
- Serve the result as an interactive dashboard with a natural-language query layer on top, where every number the AI states is traceable back to an actual database query — no hallucinated statistics.
Early build — see the folder-level READMEs for what each part of the system does. This project is being built in phases; benchmark results and a full architecture write-up will land in benchmarks/ and docs/ as they're produced.
| Folder | What's in it |
|---|---|
data/ |
Dataset loaders, satellite data fetchers, the shared data contract |
models/ |
Segmentation model architectures and losses |
training/ |
Config-driven training loop, cluster job scripts |
benchmarks/ |
Evaluation harness and results |
inference/ |
Scene-scale tiled inference service |
api/ |
Node.js gateway and LLM agent layer |
frontend/ |
React + MapLibre dashboard |
ai-eval/ |
Evaluation harness for the AI layer specifically |
infra/ |
Docker, CI, and one-command local setup |
docs/ |
Architecture diagram and write-ups |