Detecting demolished buildings after a natural hazard using high resolution RGB satellite imagery and modified U-Net convolutional neural networks

Abstract

Collapsed buildings are usually linked with the highest number of human casualties reported after a natural disaster; therefore, quickly finding collapsed buildings can expedite rescue operations and save human lives. Recently, many researchers and agencies have tried to integrate satellite imagery into rapid response. The U.S. Defense Innovation Unit Experimental (DIUx) and National Geospatial Intelligence Agency (NGA) have recently released a ready-to-use dataset known as xView that contains thousands of labeled VHR RGB satellite imagery scenes with 30-cm spatial and 8-bit radiometric resolutions, respectively. Two of the labeled classes represent demolished buildings with 1067 instances and intact buildings with more than 300,000 instances, and both classes are associated with building footprints. In this study, we are using the xView imagery, with building labels (demolished and intact) to create a deep learning framework for classifying buildings as demolished or intact after a natural hazard event. We have used a modified U-Net style fully convolutional neural network (CNN). The results show that the proposed framework has 78% and 95% sensitivity in detecting the demolished and intact buildings, respectively, within the xView dataset. We have also tested the transferability and performance of the trained network on an independent dataset from the 19 September 2017 M 7.1 Pueblo earthquake in central Mexico using Google Earth imagery. To this end, we tested the network on 97 buildings including 10 demolished ones by feeding imagery and building footprints into the trained algorithm. The sensitivity for intact and demolished buildings was 89% and 60%, respectively.