SARE-CD: Structure–Appearance Routed Experts for Building Change Detection in Remote Sensing Imagery

In building change detection from high-resolution remote sensing imagery, genuine building changes typically appear as local alterations in roof contours and spatial layouts, whereas differences in solar elevation, atmospheric conditions, and sensor calibration introduce cross-temporal appearance drift in unchanged regions. When the two types of difference are jointly encoded in a single feature space, appearance drift tends to dominate the network response. We therefore propose SARE-CD, a dense, pixel-wise, soft-gated heterogeneous mixture-of-experts (MoE) framework that combines scene-level bias correction with functionally distinct expert modeling. At the feature front end, SARE-CD generates affine parameters from the channel statistics of the bitemporal change representation and applies them in residual form to suppress systematic radiometric drift. The Structure, Appearance, and Adaptive Experts then model geometric boundary changes, local appearance perturbations, and composite differences, respectively, and a pixel-wise router assigns expert weights at each location, with routing-entropy and weak load-balancing regularization used to stabilize the division of labor among the experts. On LEVIR-CD+, S2Looking, and SYSU-CD, SARE-CD achieves F1 scores of 91.58%, 65.78%, and 84.46%, all of which are the highest among the compared methods. In addition, controlled perturbation experiments show that F1 retention remains high under hue shifts, soft shadows, and small misregistration, whereas extreme additive brightness offsets lead to a notable F1 decline on S2Looking.