Optimising U-Net models for large-scale rainfall-runoff modelling

Background High-resolution flood simulations are critical for risk management, yet conventional numerical solvers remain computationally prohibitive for real-time applications.Methods We propose a deep learning-based super-resolution framework that couples the TRITON hydraulic simulator with a U-Net architecture. The model is trained on paired constant-rainfall simulations using coarse depth maps and fine topography. Sensitivity analyses are conducted on topographic feature selection and domain decomposition strategies to optimize performance.Results Despite being trained on constant-rainfall scenarios, the U-Net model successfully generalizes to realistic, time-varying rainfall events while preserving high-resolution details.