Comparative Analysis of Urban Flood Mapping Using Sentinel‐1 SAR Imagery and Machine Learning
TL;DR
Combining temporal change‐detection signals with landscape features using ML improved flood‐mapping performance, with average F1 scores of about 0.75 and a range of 0.58 in dense urban areas to 0.93 in peri‐urban regions.
Abstract
Flood mapping using remote sensing data has the potential to play a crucial role in monitoring and assessing flood events, particularly in urban environments, where complex infrastructure creates significant challenges. While Synthetic Aperture Radar (SAR) has promising potential for flood mapping, current methods struggle in the urban landscape due to the presence of heterogeneous surfaces and complex infrastructure, which can create double‐bounce effects and alter reflectance. This research advances urban flood mapping through a comprehensive comparative analysis of traditional change detection and Machine Learning (ML) methods, including the MultiSenseRF multi‐source feature‐fusion workflow. The study analyzes five geographically distinct urban areas—Houston, Texas, USA, Dhaka, Bangladesh, Iwaki, Japan, Lumberton, North Carolina, USA, and Beira, Mozambique—during flood events to evaluate method performance across diverse urban settings. Traditional change detection methods and ML techniques were evaluated alongside integrated approaches. Combining temporal change‐detection signals with landscape features using ML improved flood‐mapping performance, with average F1 scores of about 0.75 and a range of 0.58 in dense urban areas to 0.93 in peri‐urban regions. These findings provide a robust framework for improving urban flood mapping accuracy and have significant implications for urban flood management.