DADR-HCD: A Deep Domain Adaptation and Disentangled Representation Network for Unsupervised Heterogeneous Change Detection

Change detection, a critical and flourishing Earth observation technology, aims to identify changes through cross-temporal remote sensing images acquired over the same geographical area. With the widespread use in various change scenarios, it becomes essential to utilize heterogeneous images due to the high challenge of accessing the ideal homogeneous images. Nevertheless, domain shift, generated by different imaging factors (e.g., sensors, seasons, atmosphere, and illumination), makes it unable to compare the heterogeneous images directly. To address this problem, we propose a deep domain adaptation and disentangled representation network for unsupervised heterogeneous change detection (DADR-HCD), which bridges the domain gap from the perspective of causal mechanisms and compares the differences in the content feature space. In the training stage, the deep features of the input bitemporal images are further disentangled into the domain-invariant (content) features and domain-specific (style) features through an explicit image translation network. Furthermore, unlike comparing the differences at the image level or deep feature space, the change probability maps (CPMs) are directly calculated based on the content feature similarity in the prediction stage, which minimizes the style noise and avoids the asymmetry of image-level translation. Finally, the binary change maps (BCMs) are obtained using threshold segmentation and morphological postprocessing strategies. The comprehensive experimental results and detailed analysis on five typical datasets demonstrate the effectiveness and superiority of the proposed DADR-HCD network in the unsupervised heterogeneous change detection (Hete-CD) task.