Enhancing Flash Flood Prediction Accuracy with Bi-LSTM and Satellite Rainfall Estimates

Flash floods pose a serious threat in hilly catchments, where steep slopes and intense rainfall often trigger rapid water level rises that endanger lives and infrastructure. In this study, a deep learning-driven flash flood early warning system is proposed and evaluated for the BWDB (Bangladesh Water Development Board) station at Muslimpur on the Jhalukhali River, Sunamganj, Bangladesh. A supervised bidirectional Long Short-Term Memory (Bi-LSTM) neural network was used to forecast river water levels 25 time steps ahead, or with a 5-day lead (each day is divided into five time steps: 6:00, 9:00, 12:00, 15:00, and 18:00). The model incorporates a wide range of input features, including precipitation data using half-hourly IMERG satellite rainfall estimates for five geographical locations, temporal indicators (hour, day of year, month, and monsoon flags), lagged values, rolling statistics, and cumulative precipitation. Its architecture leverages two input streams: one processing 25 past time steps for each feature using Bi-LSTM layers, and the other projecting 25 future time steps, with outputs concatenated to predict the final output. For increased sensitivity to outlier events, a peak-weighted loss function was specifically employed. The model demonstrated robust predictive capability, with coefficients of determination (R²) between 0.88 and 0.93, mean absolute errors between 32.5 and 41 cm, and root mean square errors between 36.5 and 54.5 cm for flood events in April 2017, May 2019, and May 2020. Visualization of predicted and measured water levels confirmed that the system effectively replicated both gradual changes and abrupt flood maxima, accurately simulating the rapid increases in water level characteristic of flash floods. These results highlight the ability of deep learning models utilizing satellite data to provide more reliable and enhanced preemptive alerts than traditional early warning systems.