Seamless Augmentation of Satellite Imagery for Building Damage Segmentation

Abstract

To counter the increasing risk of natural disasters, a rapid and precise localization of affected buildings is critical for disaster relief. Given the diverse appearance of damage, caused by various building and disaster types, a manual damage assessment of satellite imagery is currently needed. Automating the comparison of imagery from before and after the onset of the disaster has the potential to significantly reduce precious time, enabling a mapping of arbitrarily large regions. To overcome the limited amount of labeled training data from past disasters, an augmentation procedure to seamlessly insert buildings is proposed. Therefore, a three-step process is followed: first, a selection of fusion and loss weighting schemes is studied to form a baseline model for the semantic segmentation of building localization and damage. In the second step, occlusion-based explainable AI methods are used to exemplify the importance of individual building regions. It is thereby shown that regions in close proximity to the outline of the building exhibit a major contribution for the classification of certain damage classes. Accordingly, buildings and their respective surroundings are augmented in the third step. Therefore, a combination of alpha-blending and Poisson Image Editing is used to study different compositions of inserted source buildings, selected from the xBD training dataset. Measured in terms of the xview2-metric, the proposed augmentation scheme enables an increase of 7.2% compared to the baseline.