HaorFloodAlert: ML Ensemble for 72-Hour Flood Prediction in Bangladesh Haor Wetlands
There's a new machine learning tool called HaorFloodAlert that predicts flooding in the Sunamganj Haor of Bangladesh for the next 72 hours. This area is about 8,000 km² and is particularly at risk from flash floods, which can harm the boro rice crops grown there. Traditional river flood models don’t work well in this flat region. Researchers found that including temperature data improved the model's accuracy by 6.9 percentage points since floods often happen in warmer months. The tool uses data from the Sentinel-1 satellite, giving about 36 hours of warning. Validation showed an impressive 84-91 percent match with real events, and the system has a three-tier alert setup. The findings were published on arXiv under ID 2605.20167.
Key facts
- HaorFloodAlert forecasts 72-hour flood probability for Sunamganj Haor (8,000 km²).
- Temperature inflated accuracy by 6.9 pp as a seasonal cheat code; data was deseasonalized.
- Upstream Barak River Sentinel-1 SAR proxy from Silchar, Assam gives 36 hours lead time.
- Otsu-thresholded SAR change detection validates at 84-91% spatial match.
- Operational ensemble: RF 0.5625 + XGBoost 0.4375.
- Achieves 89.6% LOOCV accuracy, 87.5% recall, 0.943 AUC-ROC on 77 Sentinel-1 events.
- Three-tier alert pipeline implemented.
- Published on arXiv (ID 2605.20167).
Entities
Institutions
- arXiv
Locations
- Bangladesh
- Sunamganj Haor
- Silchar
- Assam
- Barak River