Deep learning-based downscaling of global digital elevation models for enhanced urban flood modeling

Urban flood modeling depends heavily on the quality of Digital Elevation Models (DEMs). However, accurate, high-resolution DEMs are often expensive and not widely available, particularly in data-limited regions. Consequently, researchers frequently rely on Global Digital Elevation Models (GDEMs), which suffer from vertical biases and limited spatial resolution. This limitation is especially critical in urban settings, where detailed terrain features are essential for accurate flood prediction. In this study, we introduce a novel methodology that leverages Convolutional Neural Network (CNN) architecture (U-Net) and utilizes GDEMs and other publicly available datasets (e.g., Landsat-8, Sentinel-1, and Sentinel-2) to produce an enhanced DEM with a 5-meter spatial resolution. Using USGS high-resolution DEMs as a reference, our results demonstrate that our method is able to generate DEMs with significantly lower vertical biases (82.1% lower RMSE and 87.8% lower MAE) compared to GDEMs. Additionally, the model produces amore detailed representation of urban features that are essential for flood pattern analysis. By applying this improved DEM within a flood simulation model, we show that the Probability of Detection increases by 12% increase and the False Alarm Ratio decreases by 13% compared to GDEMs. These findings underscore the potential of using deep learning and multi-source data to improve DEM quality for more accurate urban flood modeling and management in data-limited regions.