Integrating Net Rainfall Calculation in Deep Learning-Based Surrogate Modeling Frameworks for 2D Flood Prediction

Abstract

[Abstract]: This study proposes a novel deep learning (DL)-based surrogate model that incorporates the calculation of net rainfall using the SCS-CN method, providing a flexible framework for evaluating the influence of rainfall events under different antecedent moisture conditions (AMC). The proposed framework involves establishing a ground truth model (Iber-SWMM) and defining the necessary terrain features and rainfall patterns for training the surrogate. A benchmark surrogate model using only gross rainfall, replicating methodologies from previous studies, is also developed for comparison. The trained models are then applied to predict water depth maps using test rainfall patterns under different scenarios, both with and without net rainfall. The results demonstrate that the proposed surrogate model reduces the computational times of Iber-SWMM by 2 to 4 orders of magnitude while outperforming the benchmark surrogate in all the measures. It presents satisfactory accuracy in water depth prediction, with 80% to 95% of predictions within a -0.2 to 0.2 m error range and hit ratios between 0.87 to 0.91 in terms of flooded pixels in the more extreme events. These outcomes are comparable to those achieved by a physics-based model on one of the test events. The study also suggests future lines for refinement.