A multi-level damage assessment model based on change detection technology in remote sensing images

Automatic analysis technology of remote sensing imagery is crucial for effective building damage assessment following natural disasters. Change detection (CD) method, which compare pre- and post-disaster images, are frequently utilized to locate affected regions and identify the degree of disaster damage. However, when multiple disasters coincide, the characteristics of building damage become complex and variable. Some existing CD methods face challenges in determining the damage levels of buildings owing to the direct fusion of dual-temporal feature maps, which limits their ability to extract sufficiently detailed damage change information. To address the challenge, this research proposes a novel change detection model (named MDA-CD) for multi-level damage assessment. The architecture of MDA-CD is engineered with an Encoder-Bridge-Decoder configuration. In the Encoder stage, five different feature extraction modules are employed to enhance the representation of damaged buildings. Notably, the constructed global feature aggregation (GFA) module extracts the common features of various damaged buildings by calculating the spatial deformation similarity through long-range context modeling. Before passing the encoded feature maps to the Decoder, the Bridge stage is presented to capture key changes of interest between pre- and post-disaster feature maps by leveraging the concept of dividing semantic tokens from the transformer architecture. The constructed bitemporal image transformer compression (BITC) module first represents dual-temporal feature maps as key high-level semantic tokens, and then extracts refined damage features within the compact spatiotemporal space formed by these key tokens. In the Decoder stage, the subtle feature attention (SFA) module based on dual attention mechanisms is devised to aggregate damage features across different layers. In general, our model strengthens the representation capability of damage features, and maximizes the extraction of fine-grained change features. Compared with state-of-the-art methods, MDA-CD exhibits an enhanced discriminative ability for complex damage features, particularly for the categories of slight damage, ultimately improving the accuracy of damage assessment.