The Problem

Build a model that uses satellite data to predict flood, cyclone, or landslide risk in a chosen high-risk region, producing an actionable spatial risk map rather than a single classification score. Use a public satellite dataset and a defined historical time window — not a live feed.

Background

Floods, cyclones, and landslides cause massive damage in high-risk regions, and early-warning systems that rely on sparse ground sensors often miss or lag the event. Satellite imagery offers wide, repeated coverage but is underused for genuinely predictive (not just after-the-fact) risk mapping because turning raw satellite bands into an actionable spatial risk score is non-trivial.

Core Requirements

1. Pick exactly one hazard type (flood OR cyclone OR landslide) and one specific region — do not attempt all three generically.
2. Use a publicly available satellite dataset (e.g. Sentinel Hub/Copernicus, NASA MODIS/Landsat, ISRO Bhuvan open data) and disclose which one, its resolution, and its revisit time.
3. Go from raw/preprocessed satellite bands relevant to the hazard (e.g. NDWI for flood extent, cloud-top temperature for cyclone intensity, slope + precipitation for landslide susceptibility) to a spatial risk score across the region.
4. Validate the model against at least one real historical disaster event present in the dataset's time window: does the model's risk score actually rise before or during that known event?
5. Produce a clear visual output — a risk heatmap overlaid on a map of the region — and be explicit about the lag between a satellite pass and your prediction.

Bonus / Stretch

• Fuse in a second public data source (rainfall gauge data, elevation/DEM data, or historical disaster records) for a multi-modal risk score.
• Add a simple threshold-based alert system (e.g. flag when predicted risk crosses a set level).