Bipartite Spatial Transformer Difference Network for Unsupervised Change Detection in Misaligned Images
Ling Yan Hu, Qichao Liu, Jia Liu, Zhihui Wei, Liang Xiao · IEEE Transactions on Geoscience and Remote Sensing · 2025
Unsupervised change detection algorithms typically identify changed areas by comparing pixels or ground objects in co-registered bi-temporal images. However, due to small viewpoint variations in misaligned images and unknown changed regions between bi-temporal images, image registration becomes extremely challenging, and misaligned pixels may lead to a mass of false alarms. To address this issue, we develop a bidirectional spatial transformer network (BidSTN) capable of identifying the spatial shifts of corresponding key points in pre- and post-event images. It allows us to compute the coordinate offset relationships between misaligned images with small viewpoint variations and achieve spatial interpolation and registration using a thin-plate spline function. Building on the BidSTN, the bipartite spatial transformer difference network (BSTDN) is constructed by integrating encoders and decoders with feature difference minimization and cross-reconstruction constraints. It concurrently learns the distribution-consistent features of bi-temporal images in the feature space and their coordinate offset relationships, enabling the spatial registration and change detection of heterogeneous images within a single learning process. Furthermore, we incorporate a series of iterative refining modules (RMs) based on the aligned images and deep features. These modules enhance the accuracy of change detection by learning real changes within two distinct feature spaces and continuously refining the change map. Finally, by cascading the BSTDN and RMs, an unsupervised cross-modal registration-detection joint framework is constructed. Extensive experiments on both real and simulated datasets validate the framework’s superiority in suppressing misalignment artifacts, with F1 scores improving by an average of 15.83% (real dataset) and 11.12% (simulated dataset) compared to the best baseline methods. Code is available at https://github.com/lh-rs/BSTDN.