A flood forecast for the calls that come next

Flood112 uses XGBoost decision trees to predict whether flood-related 112 calls reach a threshold in the next hour across Catalonia’s 947 municipalities. Inputs include rain, radar, warnings, population and recent calls.[1]

The August 2026 paper’s rainfall filter retained 15,081 calls across 714 days in 2018–2025. At least three stations exceeding 3 mm in 30 minutes qualified a day. Results describe retrospectively selected rainy days.[1]

In high-density municipalities (≥120 inhabitants/km²), a positive hour meant at least three calls. Mean Critical Success Index was 28.77%, 24.35 percentage points above the FF-EWS radar-warning baseline. CSI divides hits by hits, misses and false alarms. The model caught 50.90% of positive hours; 40.78% of its alarms were correct.[1]

For high-density hours with calls in the preceding three hours, CSI was 41.84%; without them, 11.10%. Results averaged ten tests advancing through time. Low-density municipalities (<20 inhabitants/km²) scored 0.83% CSI, with a one-call threshold.[1]

A 2024 Catalan study had already used machine learning to predict daily ranges of weather-related emergency incidents. Its inputs included wind, rain, waves and snow as well as local characteristics. The incident record also made a practical distinction: several phone calls could describe a single event requiring an emergency response.[2]

WMO’s guidance stresses joint design with the people who act on forecasts. Here that would mean agreeing what a predicted busy hour should trigger: retained staffing, closer monitoring of particular municipalities or preparation for further reports. The operational value would come from that decision while the phones are already ringing.[3]

Report an error or suggest a correction