Disaster Damage Assessment from Satellite Imagery

By Navjot Kaur

Much of life-saving humanitarian aid and disaster response work is guided by post-disaster damage assessments that are obtained through analysis of satellite images of an affected area. Most recent work in this field uses either a Siamese network or a difference module in the spatial domain to assess damage, and applies attention mechanisms on the temporal features independently instead of their difference. In this project, we propose a novel transformer-based network for damage assessment. This proposed network achieves state-of-the-art performance on a large-scale disaster damage dataset for building localization and damage classification.

Project report | GitHub repoProject poster