Road-focused flood depth and velocity prediction using deep learning for transportation resilience

Flooded roadways threaten mobility and public safety, underscoring the need for high-resolution, road-specific flood prediction tools. This study developed a deep learning framework to predict flood depth and velocity on transportation networks using geomorphic and hydrologic features. A UNet-based Convolutional (UNet Conv) model and a UNet Convolutional Long Short-Term Memory (UNet ConvLSTM) model were implemented for flood depth and velocity, respectively. By incorporating roadway-focused weighted loss functions, the models prioritized accuracy on road pixels and achieved strong performance (Depth: R 2 ≈ 0.89, RMSE ≈ 0.07 m; Velocity: R 2 ≈ 0.93, RMSE ≈ 0.08 m/s). The framework generalized well to unseen storm conditions (R 2 > 0.75). Unlike conventional hydrologic and hydraulic models, this method eliminates the need for time- and resource-intensive modeling setups, thereby facilitating rapid flood prediction using readily available data. These results demonstrate the potential of road-focused deep learning to support flood forecasting, transportation resilience planning, and emergency response.