Disaster Assessment from Satellite Imagery by Analysing Topographical Features Using Deep Learning

This paper explores the application of deep learning techniques in the task of assessing disaster impact from satellite imagery. Identifying the regions impacted by a disaster is critical for effective mobilization of relief efforts. Satellite images, with their vast coverage of ground surface, are a valuable resource that can be leveraged for this purpose. However, the task of analysing a satellite image to detect regions impacted by disaster is challenging. In recent years, the increasing availability of satellite images of a place presents an opportunity to utilize deep learning on these images to provide a preliminary insight on the impact of a disaster after its occurrence. We particularly focus on water related disasters like floods and hurricanes. To identify the impacted regions, we employ Convolutional Neural Networks to semantically segment topographical features like roads on pre and post-disaster satellite images and find the regions with maximal change. However, this approach is less applicable on rural landscapes due to the sparse distribution of topographical features like roads. To address such cases of imagery from rural areas, we propose a bitemporal image classification approach to compare pre and post disaster scenes and directly identify if the regions are impacted or not. On testing against a ground truth satellite image from DigitalGlobe with labeled regions depicting the impacts of Hurricane Harvey, the flooded road extraction approach achieved an accuracy of 0.845 with a F1-Score of 0.675. Similarly, the bitemporal image classification approach registered an accuracy of 0.94 when tested against a rural landscape impacted by South Asian Monsoon Flooding of 2017.