This study introduces an integrated framework for short-term hydrological forecasting and spatial flood risk mapping by coupling Long Short-Term Memory (LSTM) neural networks with the Height Above the Nearest Drainage (HAND) topographic model. Focused on the Tamanduateí River basin in São Paulo, Brazil—a critical fastresponse catchment—the methodology addresses the urgent need for high-precision, low-latency alerts in densely urbanized areas. By utilizing precipitation and stage height time series, the LSTM model predicts river levels up to 60 minutes ahead, using 10-minute intervals and a 6-hour antecedent window. Results indicate exceptional predictive performance, yielding coefficients of determination (???? 2 ) of 0.9946, 0.9757, and 0.8956 for 10, 30, and 60-minute lead times, respectively. Complementarily, the HAND model generates stratified susceptibility maps to delineate inundation extents across various river stages. These outputs provide interpretable, actionable visualizations that enhance preventive decision-making and emergency response planning in vulnerable urban environments.