An Object-Level Active Semantic-Driven Posterior Probability Change Detection Method for Remote Sensing Images with Limited Annotations

Change detection in high-resolution remote sensing imagery is susceptible to the adverse effects of within-class spectral heterogeneity, shadows, and complex spatial structures. Meanwhile, change detection methods based on land-cover posterior probabilities generally rely on reliable semantic training information, making complete manual interpretation costly. To address these issues, this study develops a change detection framework that integrates SEEDS-guided SAM object construction and active learning-based interpretation with FCM-CSBN-CVAPS posterior probability inference. First, representative points of SEEDS superpixels are used as prompts for SAM to generate candidate objects and construct a mutually exclusive object partition. Second, based on the spectral statistical features of the superpixels and an XGBoost classifier, active learning is employed to select highly uncertain samples for limited manual annotation, thereby obtaining object-level semantic information. Finally, FCM is used to characterize pixel-level spectral uncertainty, while CSBN incorporates spatial context to estimate posterior probabilities of land-cover classes. Bi-temporal change measurement and changed-area extraction are then performed in the posterior probability space. Experimental results show that the proposed framework achieves stable change detection performance by manually annotating only approximately 10% of the superpixels in the target scenes. On DS1, DS2, and DS3, the Kappa coefficients reach 83.87%, 86.05%, and 90.21%, respectively, while the corresponding F1 scores are 87.62%, 91.78%, and 91.31%, achieving the highest Kappa and F1 scores among the evaluated methods under the specified experimental settings. These results indicate an effective balance between false-alarm suppression and missed-detection control for change detection using limited target-scene semantic annotations. In adaptive evaluation experiments, the mean Kappa and F1 scores reach 84.58% and 87.84%, respectively, further confirming the effectiveness and stability of the proposed framework across diverse remote sensing scenarios.