Data-driven rapid flood prediction mapping with catchment generalizability

• A data-driven model is proposed for urban-scale rapid pluvial flood prediction. The extrapolates to new catchments that were not included in the training data. Two options are compared process catchment areas of different sizes and shapes. Different parameters, including input kernel sizes, investigated. Data-driven machine learning models have recently received increasing interest resolve computational speed challenge faced by various physically-based simulations. few studies explored application these develop new, fast, applications fluvial prediction, extent mapping, susceptibility assessment. However, most focused on development specific areas, drainage networks or gauge stations. Hence, their results cannot be directly reused other contexts unless extra data available further trained. This study explores generalizability potential convolutional neural (CNNs) as prediction models. proposes a CNN-based can with topography once investigates two options, patch- resizing-based showed predicts accurately “unseen” significantly less time when obtained also suggest patch-based option more effective than terms accuracy. In addition, all experiments shown flow velocity accurate water depth, suggesting accumulation sensitive global elevation information velocity.