Natural Disaster Analytics using High Resolution Satellite Images

Throughout history, natural calamities have taken a toll on both property and life. Forest fires, hurricanes, floods, earthquakes, and tornadoes are all responsible for substantial damages, and often render access to the affected areas difficult. One of the challenges, particularly in remote areas, is accurately assessing the severity of the disaster. In this paper, we propose a solution to the challenge of determining the severity of property damage through the analysis building change within satellite imagery. We determine the severity of the loss by training deep neural networks to count buildings which were destroyed in multitemporal satellite imagery of the affected area. We demonstrate, through experimental results, that a model composed of a Single Shot Detector (SSD) and a Feature Pyramid Network (FPN) trained with Focal Loss is able to detect buildings through satellite imagery with improved performance compared to conventional object detection models. For evaluation, we use the xView dataset, which consists of high resolution satellite images containing building labels. We evaluate three different object detection models, namely: (i) SSD, (ii) Faster R-CNN (Regional-based Convolutional Neural Network), and (iii) SSD+FPN with Focal Loss. Our findings show that SSD+FPN with focal loss model achieves mAP improvements of 15% and 20% increase in detecting buildings in comparison to Faster R-CNN and standard SSD models, respectively. The improved mAP metric is a reflection of our more accurate detection and localization of buildings from remotely sensed imagery.