Damage assessment with global differences and local attention

Abstract

Detecting the different changes caused by a natural disaster is critical for effectively directing humanitarian assistance and disaster response operations. However, it is challenging due to the large-scale disaster areas and complex ground environments. Existing assessment methods are usually labor-intensive and unsuitable for multiple disasters. In this paper, we propose a change transformer (CHTR) model for simultaneous building localization and multi-level change detection from dual-temporal satellite imagery. Based on the advantages that convolutional neural networks (CNNs) are good at learning detailed local features and the transformer can model long-range dependencies, we adopt a hybrid CNN-transformer architecture as the encoder. A natural disaster usually causes varying degrees of damage to buildings in a complex environment; thus, we propose a global difference module on the original features obtained by the CNN to capture the global change pattern and improve the overall awareness of the variations between dual-temporal images. Furthermore, a local gated attention module on the patches of features after the CNN is further developed to learn the local dependencies among the multi-level changes, which augments the discrimination of different changes. Extensive experiments on the largest building damage assessment dataset, xBD, demonstrate that the proposed CHTR model establishes new state-of-the-art results.