Water-related disaster assessment using computer vision and satellite imagery

Abstract

The increasing frequency and severity of water-related disasters such as floods, tornadoes, hurricanes, and tsunamis in low- and middle-income countries exemplify the uneven effects of global climate change. The vulnerability of high-risk societies to natural disasters has continued to increase. Local damage assessments must be timely, exhaustive, and accurate to develop an effective and efficient adaptation strategy. We propose a novel deep-learning-based solution that uses pair of pre- and post-disaster satellite images to identify the water-related disaster-affected region. The model extracts features of pre- and post-disaster images and uses them to predict damage in the pair. We demonstrate that the model can successfully identify local destruction using less granular and less complex ground-truth data than previous segmentation models. When tested with various water-related disasters, our detection model reported an accuracy of 85.9% in spotting areas with damaged buildings. It also achieved a reliable performance in the case study on Hurricane Iota, showing its ability to be applied to real-disaster situations. Our deep learning-based damage assessment model can help direct resources to areas most vulnerable to climate disasters, reducing their impacts while promoting adaptive capacities for climate-resilient development in the most vulnerable regions.