Natural Disaster Building Damage Assessment Using a Two-Encoder U-Net

Abstract

When a natural disaster occurs, damaged regions rely on timely damage assessments to receive relief. Currently, this is a slow and laborious process, during which emergency response groups conduct on-the-ground evaluations to form fiscal estimates. This project attempts to expedite relief efforts by applying novel computer vision algorithms to satellite images to quickly and accurately estimate physical and fiscal damage caused by natural disasters. This paper investigates a modified U-Net architecture to jointly localize buildings, classify damage, and establish change detection. In particular, a second encoder is added to the U-Net architecture to simultaneously process pre- and post-event imagery, with both encoders sharing weights. In this way, the decoder is trained to both locate buildings and classify damage estimates by formulating it as a single semantic segmentation problem – producing damage estimates in one pass without needing to re-visit pixels (i.e. detection + classification tasks). Finally, a downstream task is added that provides a pixel-based financial model capable of outputting financial costs according to the United States National Grid (USNG) coordinate system through an interactive web application.