Real-Time Flood Risk Prediction Using Temporal Transformer-based Deep Learning and Multisource Remote Sensing Data

The rapid and unexpected nature of flash floods leaves communities with minimal time to respond. Accurate forecasting of such events is crucial, as it allows for timely warnings, early evacuations, and preparations, significantly reducing potential losses. This study applies a transformer-based Deep Learning (DL) with Remote Sensing (RS) validation to predict floods with at different lead times. The Guadalupe River basin in Texas was selected as a case study due to the 2025 catastrophic flash flood, one of the deadliest floods in U.S. history. For forecasting purposes, discharge (Q) data from a downstream USGS station were chosen as the target variables. Precipitation, air temperature and moisture and vegetation indices were extracted for upstream station from RS sources. These RS products were validated against corresponding ground-based observations to ensure quality as dynamic inputs for flood prediction. Input climatic variables were structured based on different times windows (t-1, t-3, t-5, t-7, t-15, and t-30), representing conditions at 1, 3, 5, 7, 15 and 30 days before the current downstream state (t). different input sequences were utilized in a Multivariate Predictive Transformer Network, a DL model designed to make predictions from a temporal multivariate sequential dataset. The flood events were identified by exceedance from the 90 th , 95 th and 99th percentile threshold for Q within the 2003 to 2025 period. The predictive performance of the model was assessed regarding both the accuracy in predicting the flooding events and timing of impact. Furthermore, the model's capability to forecast floods with annual exceedance probabilities of 2, 5, 10, 25, 50, and 100 was thoroughly evaluated. Prediction results show that developed model can accurately detect the flood occurrence up to 5 days notice (Accuracy: 0.96). While the accuracy drops in next time lags, the metric shows comparable results with similar studies. The outcomes of this study, which integrates advanced transformer-based predictive modeling with remote sensing data, can substantially improve early forecasting and management strategies for floods as a detrimental natural hazard phenomenon.