Flood Uncertainty Estimation Using Deep Ensembles

We propose a probabilistic deep learning approach for the prediction of maximum water depth hazard maps at high spatial resolutions, which assigns well-calibrated uncertainty estimates to every predicted depth. Efficient, accurate, and trustworthy methods urban flood management have become increasingly important due higher rainfall intensity caused by climate change, expansion cities, changes in land use. While physically based models can provide reliable forecasts location catchment, their computational burden is hindering application large areas resolution. been used address this issue, disadvantage that they are often perceived as “black-box” overconfident about predictions, therefore decreasing reliability. Our model learns underlying phenomena priori from simulated hydrodynamic data, obviating need manual parameter setting new event test time. The only inputs needed time forecast parameters terrain such digital elevation predict with complete events. validate accuracy generalisation capabilities our through experiments on dataset consisting catchments within Switzerland Portugal 18 patterns. method produces 1 m resolution achieves mean absolute errors low 21 cm extreme cases above m. Most importantly, we demonstrate able an estimate map, thus increasing model’s trustworthiness during flooding