Local Damages from Hurricanes: Application of Machine Learning with Satellite Data, Katie Kolodner, UG '24 (3963520)
Though hurricanes are some of the most destructive natural disasters, with the frequency of intense and damaging hurricanes increasing due to climate change, current estimates of damages from hurricanes are a source of uncertainty. To better estimate home damage from hurricanes, this study constructs a novel dataset of building damages from hurricane wind and water using satellite-based observations of regions impacted by hurricanes. With imagery provided by Maxar’s Open Data Program, damages inflicted on houses were evaluated by classifying buildings before and after a hurricane and analyzing the differences. The xView2 Challenge dataset of satellite and aerial imagery, labels, and targets was used to train a convolutional neural network (CNN) to predict damages incurred by houses after a hurricane. The processed results reveal successful building localization and damage prediction relative to the targets in the training set and to out-of-sample prediction. They further reveal the importance of imagery color normalization and training on various terrain and damage types. The model achieved an overall score of 0.699, taking into account a combined dice loss coefficient of 0.847 and F1 score of 0.636. Continued work involves connecting building damage scores to monetary damage and developing wide-scale testing through model refining and experimentation on additional hurricanes captured by Maxar and other satellite image sources. This data-intensive interdisciplinary research effort will make significant advancements to improve our understanding of costs and damages of hurricanes, benefitting researchers and allowing policymakers to provide more effective pre-disaster risk management programs and post-disaster recovery assistance.