Deep Learning-Enabled Semantic Inference of Individual Building Damage Magnitude from Satellite Images

Abstract

Natural disasters are phenomena that can occur in any part of the world. They can cause massive amounts of destruction and leave entire cities in great need of assistance. The ability to quickly and accurately deliver aid to impacted areas is crucial toward not only saving time and money, but, most importantly, lives. We present a deep learning-based computer vision model to semantically infer the magnitude of damage to individual buildings after natural disasters using pre- and post-disaster satellite images. This model helps alleviate a major bottleneck in disaster management decision support by automating the analysis of the magnitude of damage to buildings post-disaster. In this paper, we will show our methods and results for how we were able to obtain a better performance than existing models, especially in moderate to significant magnitudes of damage, along with ablation studies to show our methods and results for the importance and impact of different training parameters in deep learning for satellite imagery. We were able to obtain an overall F1 score of 0.868 with our methods.