FloodSight: Physics-Guided Machine Learning for Monthly Pluvial Flood Risk Forecasting in the Northeast US

Floods are the most expensive natural hazard in the United States. Pluvial flooding has been rising sharply, with extreme precipitation increasing 60% in the Northeastern US since 1950. Yet, monthly forecasts remain limited, with current dynamic methods focusing on short-term detection, while static tools like FEMA flood maps do not reflect climate variability. We present FloodSight, a physics-guided ML framework for monthly flood forecasting in the Northeast US, modeled on a novel geospatial dataset merging climate, terrain, and satellite flood observations at a 0.1° monthly resolution. Flood risk was modeled through a two-stage architecture: Stage 1 performed binary classification for flood occurrence, and Stage 2 regressed a novel flood magnitude index accounting for area, duration, and severity. An ensemble combining a physics-guided neural network with tree-based models captured 74% of flood events within the top 1% of predicted risk, achieving a 13x improvement in flood detection over FEMA zones at an equal spatial coverage. An additional Stage 3 incorporated socioeconomic vulnerability, prioritizing equitable investment in areas where both risk and limited resources exist. FloodSight can complement long-term maps with dynamic monthly flood prediction under climate change, allowing governments to allocate resources for flood mitigation in advance.