Machine Learning Based Flood Risk Modeling Using Features from Satellite, Socioeconomic, Geographic, and Building Data

Abstract

Due to the effects of climate change coupled with increased urbanization, many cities will be experiencing more frequent and intense flooding in the future. As a result, it would be very beneficial for urban planners to have a low-cost and efficient modeling tool that can determine the flood risk at a granular level such as the census tract. Boston is one such coastal urban city that will experience an increase in flooding. Since each census tract in Boston is unique and varies in population and land use, urban planners and policy makers must know which areas in Boston are the most vulnerable to provide them with resources. This research proposes a machine learning based model that evaluates the flood risk of census tracts in Boston. The overall flood risk of a census tract is determined by aggregating relevant features such as land cover data from aerial satellite imagery via semantic segmentation methods, elevation, slope, and flow accumulation. In addition to these flood hazard features, we also integrate flood vulnerability features from socioeconomic data and building information for each census tract.