Evaluation of mesh resolution effects on real-time flood inundation prediction using dimensionality reduction and deep learning

Abstract

This paper investigates a real-time flood inundation prediction method that combines deep learning and dimensionality reduction, with a focus on evaluating the effects of mesh resolution on prediction accuracy and computational efficiency. Singular Value Decomposition (SVD) is employed for dimensionality reduction, and a Deep Neural Network (DNN) is adopted as the prediction model. The previously developed method is applied to three mesh resolutions, namely 25 m, 15 m, and 5 m, and its prediction accuracy a nd computational efficiency are compared. The results indicate a tendency to achieve comparable prediction accuracy across the three mesh resolutions under the present analysis conditions, while the reduction in computation time becomes more pronounced as the mesh resolution increases. In particular, for the 5 m mesh, the method achieved an approximately 13,000-fold speedup compared with the conventional flood inundation simulation, with a prediction time of approximately 5.1 seconds.