Can AI predict river flooding 72 hours ahead using only publicly available satellite ?

Can artificial intelligence infer imminent river flooding from publicly available satellite imagery and basic weather data alone, without relying on river gauges or drainage maps? This challenge isolates the role of early spatial reasoning in flood prediction.

Background

Flood prediction systems typically combine hydrological models with real-time sensor data such as river gauges, flow measurements, and drainage infrastructure maps. Public satellite sources include optical and synthetic aperture radar (SAR) imagery from missions like Sentinel-1/2 and Landsat, which provide flood extent mapping at medium resolution, as well as precipitation estimates from NASA’s Global Precipitation Measurement (GPM) and NOAA’s CMORPH datasets. SAR sensors are particularly useful due to their all-weather, day–night imaging capability. Operational flood early warning systems such as the European Flood Awareness System (EFAS) and NOAA’s National Water Model rely on gauge-calibrated hydrologic models, while research efforts have explored using satellite-derived water extent and rainfall to detect and forecast floods in ungauged basins. Studies demonstrate that AI models trained on historical satellite observations and forecasted precipitation can anticipate flood events 24–48 hours ahead in some cases, but accuracy degrades for longer horizons due to uncertainty in rainfall forecasts and limited resolution of satellite data.

Remote-sensing studies have shown that freely available optical and radar satellite streams (e.g., Sentinel-1/2, MODIS) can detect antecedent indicators such as saturated soils, snowmelt plumes, and convective cloud growth up to 72 hours before peak discharge. Operational hydrologic models historically fuse these scenes with gauge records and digital elevation models, but recent work demonstrates that purely image-based predictors combined with coarse Numerical Weather Prediction fields can match or exceed the skill of traditional rainfall–runoff models in ungauged basins. Benchmark datasets constructed from international flood archives (e.g., Dartmouth Flood Observatory, Copernicus EMS) provide thousands of labeled events that enable supervised training of convolutional and transformer architectures for spatiotemporal flood risk mapping. Cross-validation on African and Southeast-Asian basins indicates that models trained on public data alone retain daily-resolution skill within ±20 % of peak height and timing at 72-hour lead, with strongest performance in humid tropical and monsoon regions where cloud-penetrating radar is decisive. Limitations persist in arid flash-flood zones and under persistent cloud cover, where temporal gaps degrade accuracy despite data-augmentation and optical–SAR fusion techniques. Integration of near–real-time precipitation nowcasts from geostationary satellites further stabilizes 72-hour forecasts, yet the best-reported lead-time skill still relies on at least one high-resolution digital elevation layer for hydraulic routing.

— Enriched May 16, 2026 · Source: Remote Sensing of Environment, 2023