Deep learning with upstream-aware terrain representations and rainfall adapter for generalizable pluvial flood prediction

Abstract

Deep learning-based surrogate models can substantially accelerate pluvial flood simulations. However, most existing approaches struggle to generalize across heterogeneous terrains or modified topographies, limiting their applicability for real-world flood risk assessment unless retrained for each specific terrain. In this work, we investigate how deep learning flood surrogates can be designed to achieve robust spatial generalization across previously unseen terrain configurations without retraining. We train an attention-enhanced U-Net on diverse terrain and introduce upstream-aware input maps that propagate terrain properties into downstream regions, allowing the model to capture long-range spatial dependencies and supporting the learning of generalizable spatial features. In addition, we add architectural design choices to improve the model’s capacity to generalize beyond the training domain. We further propose a two-stage training strategy with an adapter-like rainfall encoder that enables efficient adaptation to previously unseen rainfall intensities while preserving spatial generalizability. The resulting model predicts maximum water depths and is evaluated on unseen catchments in Germany, Austria and Denmark, as well as on artificially modified terrains, such as terrains with added dikes. Our results demonstrate that combining a large training dataset with upstream-aware terrain representations and a generalization-oriented architectural design enables accurate flood predictions across unseen terrains, reaching an RMSE of 0.05m and CSI0.3 of 0.83 over all test domains.