Using deep learning to combine satellite observations, topographic information and rainfall spatial data for large-scale flood predictions

Climate change is increasing the frequency and extent of flooding events, and a forecast-based warning system can help mitigate risk. Deep learning offers a promising alternative to physics-based models for fast and detailed flood forecasts on large scales, provided that enough data is available. We combine spatial data on rainfall, topography, geology, land use, and modelled groundwater from open sources to train a U-Net neural network against satellite observations of surface water. The model predicts correctly large pluvial and fluvial flooding events that were not included during training. Better performance is expected with treating separately the various types of flooding. The main potential lies in the scalability of the proposed method, the easy interpretability of the results and the possibility to integrate various types of data (e.g. rainfall forecasts).