Performance of a Machine Learning-Based Flood Prediction Model Using DEMs of Different Spatial Resolutions

Urban flooding is one of the most common natural disasters across the world. The impacts of flooding in cities have been exacerbated due to modifications in the watershed from rapid urbanization and climate change. The frequency and intensity of urban flooding are becoming more severe over time; therefore, accurate forecasting models and tools have a significant role in planning, design, management, and operations to prevent and mitigate flooding. This study investigates how the spatial resolution of Digital Elevation Models (DEMs) affects the accuracy of flood depth prediction using a deep learning Convolutional Neural Network (CNN) method. To do this, first several synthetic hydrographs are produced as training input variables for the CNN model. Next, the inundation depth simulated by the LISFLOOD-FP model is used as a target variable while considering DEMs with various spatial resolutions that are derived from a 1-meter LIDAR-based Digital Terrain Model (DTM). The performance of the trained CNN models is compared with the outputs of LISFLOOD-FP for case study flood events. Results indicated that the CNN-based flood depth prediction model trained by lower resolution DEMs outperforms the hydrodynamic model and can be used in real-time predictions and uncertainty quantification for faster response.