Abstract
Urban flooding is intensifying under climate change and urbanization, demanding efficient deep learning-based prediction. However, such models are commonly trained to minimize data-fitting loss alone, with limited incorporation of physical constraints on surface water flow. As a result, they may learn spurious statistical relationships and provide little insight into the factors governing predicted inundation, limiting their practical value for flood risk management. This study develops a physics-guided and interpretable deep learning framework for compound rainfall-tidal flood prediction in a representative coastal island setting, where flood dynamics are strongly shaped by interactions between rainfall forcing, terrain controls, and tidal boundary conditions. The framework integrates a U-Net encoder-decoder for spatial feature extraction, along with Bidirectional LSTM branches and multi-head self-attention, to encode rainfall and tidal time series. Physics-guided loss functions that enforce gradient consistency and spatial smoothness are introduced via staged weight scheduling to improve physical plausibility while preserving predictive accuracy. Model interpretability is further achieved using Integrated Gradients to quantify feature contributions, with robustness confirmed by K-fold stability analysis and independent ablation experiments. Results show that drainage junctions emerge as the dominant predictor under the present architecture–task setting, while terrain-related variables jointly account for the majority of attribution, indicating that the model captures key hydraulic and topographic controls on inundation. The resulting importance hierarchy, in which drainage junctions and geomorphological features jointly exceed the contribution of elevation, is preserved in a paired no-physics baseline trained under an otherwise identical configuration, with a cross-model Spearman rank correlation of ρ = 1.00 under the mean-depth target. This robustness identifies the hierarchy as a property of the architecture-task pairing rather than of the physics-guided loss terms; the latter act instead as an output-level hydraulic regulariser, reducing gradient-consistency violations by 16.7% at a marginal 0.51% reduction in R². These findings demonstrate that integrating physical constraints with gradient-based attribution analysis can yield more credible and interpretable deep learning predictions for compound urban flooding.