Application of machine learning-based surrogate models for urban flood depth modeling in Ho Chi Minh City, Vietnam

Abstract

Rapid flood prediction in coastal urban areas is an important but challenging task. However, multi-driver floods in coastal areas and their non-linearity in physical processes are hard to represent in physics-based numerical models (PBNMs). In this study, we investigated the performance of surrogate machine learning (ML) models and their flood prediction capability. Initially, we utilize the MIKE+ coupled 1D–2D model to simulate coastal urban flooding in one of the severely flood-affected areas of Ho Chi Minh City (HCMC), Vietnam. Then, nine ML models, including AdaBoost (AB), Decision Tree (DT), Gaussian Process (GP), k-Nearest Neighbors (KNN), Linear Discriminant Analysis (LDA), Naive Bayes (NB), Neural Network (NN), Random Forest (RF), and Support Vector Machine (SVM) are employed to surrogate the PBNM flood prediction performance and engaged to predict flood depths of the study area domain. 806 simulation scenarios of MIKE+ modeling having a spatial grid of 1107 ×1513, grid size = 2 m, extracting 270,000 inundation points to generate input data for nine ML models are used to simulate surface flood depths for the study area. Results show three ML models, GP, RF, and NN, outperform the remaining models, with R2 value of 0.997, 0.996, and 0.995, respectively. Thus, applying ML models can significantly reduce the simulation time by a PBNM, improve accuracy, and potentially be adopted for real-time forecasting and emergency management.