Disaster risk modeling in India is fragmented. Cyclone prediction, earthquake early warning, landslide susceptibility, and wildfire risk are each maintained by separate agencies using incompatible data formats and methodologies. A district collector facing a compound event — say, a heavy-rain landslide trigger during active wildfire season — has no single source of truth.
The RESQ Design Philosophy
RESQ (Real-time Emergency & Situational Intelligence) was designed around three constraints:
- Open data only — Every dataset must be publicly accessible: NDMA hazard catalogues, MODIS satellite fire detections, Open-Meteo historical climate records, IMD rainfall grids.
- Reproducible pipeline — The entire ETL and modelling flow is committed to GitHub. Any researcher can re-run it on updated data.
- Actionable outputs — Predictions are linked to NGO directories for the six highest-risk states, so a prediction immediately maps to response resources.
The Medallion ETL Architecture
We structured the data pipeline as a three-layer medallion store:
- Bronze — Raw ingest from each source, stored with provenance metadata.
- Silver — Cleaned, normalised, spatially-joined records with consistent district-level identifiers.
- Gold — Feature-engineered training sets with 11-year rolling window climate statistics, terrain features from SRTM, and historical hazard frequency counts.
Dual-Output XGBoost
The core model is a dual-output XGBoost: one head classifies hazard type (4 classes: cyclone, earthquake, landslide, wildfire), and a second head estimates severity (3 ordinal levels: minor, moderate, severe). We trained on the gold-layer dataset spanning 2012–2023, holding out 2024 for temporal validation.
The key insight from evaluation: hazard type prediction is significantly more reliable than severity prediction, particularly for earthquake severity where the relationship between climate features and magnitude is weak. We reflect this uncertainty explicitly in the UI rather than hiding it behind a single confidence score.
What's Next
We are adding a compound hazard detection layer — identifying district-level periods where multiple hazard probabilities exceed threshold simultaneously — and integrating Sentinel-2 satellite imagery for near-real-time landslide susceptibility updates. Collaborators with satellite imagery expertise welcome.