U-FLOOD – Topographic deep learning for predicting urban pluvial flood water depth

This study investigates how deep-learning can be configured to optimise the prediction of 2D maximum water depth maps in urban pluvial flood events. A neural network model is trained exploit patterns hyetographs as well topographical data, with specific aim enabling fast predictions depths for observed rain events and spatial locations that have not been included training dataset. architecture widely used image segmentation (U-NET) adapted this purpose. Key novelties are a systematic investigation which inputs should provided deep learning model, hyper-parametrization optimizes predictive performance, evaluation performance were considered training. We find input dataset only 5 variables describe local terrain shape imperviousness optimal generate depth. Neural architectures between 97,000 260,000,000 parameters tested, 28,000,000 found optimal. U-FLOOD demonstrated yield similar existing screening approaches, even though assessment performed natural unknown network, generated within seconds. Improvements likely obtained by ensuring balanced representation temporal rainfall dataset, further improved datasets, linking dynamic sewer system models.