A machine-learning system built on public data is delivering 72-hour flash-flood warnings to farmers in Bangladesh's haor wetlands — and it worked in a live test this spring.
The Sunamganj Haor is a bowl-shaped basin covering 8,000 km² with fewer than twelve working gauges, leaving rice farmers almost blind to the floods that arrive each spring just before the boro harvest. HaorFloodAlert addresses this with a free-data stack: Sentinel-1 radar imagery (which penetrates storm clouds), rainfall records and forecasts, soil moisture readings, and a modeled signal from the upstream Barak River in India that adds roughly 36 hours of lead time. A critical preprocessing step removes a seasonal bias baked into earlier models — raw temperature correlates with floods only because it follows the calendar, not the water. A monthly climatological anomaly correction cuts that temperature-label correlation from r=0.570 to r=−0.031. Tested against 77 events and 12.3 years of official gauge records by leave-one-out cross-validation, the Random Forest and XGBoost ensemble reaches 90.9% accuracy, an F1-score of 89.2%, and an AUC of 0.939. In a live ten-day run during May–June 2026, the system raised a high-risk alert roughly three days before the river approached its danger level. Warnings are sent in Bengali via SMS, e-mail, and WhatsApp.
The seasonal-bias fix is the underappreciated contribution here. Most flood models trained on temperature data are quietly learning that floods happen in spring — not that conditions are dangerous right now. Removing that shortcut makes the model actually predictive rather than calendrical. The Barak River upstream signal is equally clever: it buys lead time that local sensors alone cannot provide.
Flood early-warning systems are not new, but most depend on dense gauge networks and government infrastructure that rural Bangladesh simply does not have. The public, seeded pipeline means any researcher or NGO can reproduce every number — which raises the obvious question of why it took this long for someone to wire this together.