<p>Within the technical framework of formulating change detection as semantic segmentation, we propose a conditional discriminative adversarial change detection method (CDACD), which utilizes the conditional adversarial module (CAM) to improve the performance of the detector. From the perspective of the objective function, CDACD adopts the cross-entropy loss for semantic segmentation with an additional adversarial regularization term, which serves as a valuable supplement to the pure cross-entropy loss. From the perspective of the optimization process, the CAM is inspired by conditional generative adversarial networks (cGANs): within the CAM, we enable the change detector (a discriminative network) and another conditional discriminator to compete against each other and improve jointly. The CAM is conditioned on image differences. Through iterative adversarial optimization, it drives the change detector to generate masks consistent with real changes. Unlike existing methods limited by predefined change categories, CDACD focuses on learning the intrinsic spatio-temporal discrepancies between bi-temporal images via the CAM. Furthermore, the CAM demonstrates plug-and-play compatibility across diverse change detection methods, consistently improving sensitivity to subtle changes. Extensive experiments on four benchmarks confirm CDACD’s robustness and accuracy over state-of-the-art methods.</p>