面向遥感语义变化检测的时空变化协同感知方法

Semantic Change Detection (SCD) aims to accurately identify complex land cover changes and simultaneously determine their corresponding semantic categories from bitemporal remote sensing images. Existing methods typically employ separate change localization and semantic recognition branches to accomplish SCD tasks. However, these frameworks often fail to fully integrate and leverage semantic information to guide the change detection process, which reduces their robustness in complex scenarios, seasonal and illumination variations, and multiscale small-target changes. To address this limitation, this study proposes a Semantic-guided Spatio-Temporal Collaborative Perception Network (SemSTNet) that incorporates semantic guidance into the precise identification of change regions while considering spatiotemporal interactions of bitemporal features. The proposed SemSTNet consists of three key components. First, a symmetric spatiotemporal demodulation differential module (SSDM) is constructed. It employs bidirectional feature modulation to achieve cross-temporal feature mutual calibration and explicitly model intrinsic spatiotemporal difference patterns using 3D central difference convolution, thereby extracting discriminative difference features from bitemporal images. Second, a Semantic Gated Fusion Decoder (SGFD), which introduces a semantic similarity dynamic weighting mechanism, is designed. It generates semantic priors through cosine similarity maps and enhances the saliency representation of genuine change regions via adaptive gated fusion. Third, a Contrastive Change Loss (CCL) is proposed to construct dual supervision signals. It constrains the semantic consistency of unchanged regions in the feature space and enlarges the interclass distance between changed and unchanged regions through margin-based contrastive learning. Extensive experiments are conducted on two SCD datasets: the SECOND dataset and the JL1 dataset. On the SECOND dataset, SemSTNet achieves the best performance with an mIoU of 73.62%, a separated kappa (SeK) of 24.19%, and an F1 score of 64.27%, surpassing the second-best method CdSC by 0.30%, 0.67%, and 0.57%, respectively, while maintaining lower parameter count (27.85 M vs. 38.6 M) and reduced memory consumption (424.17 MB vs. 482.9 MB). On the JL1 dataset, SemSTNet again achieves the highest scores across all metrics, with an mIoU of 86.51%, a SeK of 59.02%, and an F1 of 88.01%, demonstrating strong robustness against complex seasonal spectral variations. Ablation studies on the SECOND dataset confirm the effectiveness of each proposed component, with the full model improving mIoU from 70.60% to 73.62% and SeK from 19.09% to 24.19% over the baseline. The proposed SemSTNet effectively breaks the conventional separation between change detection and semantic recognition by establishing a collaborative optimization mechanism among SSDM, SGFD, and CCL. Experimental results on multiple public datasets validate the superiority of the proposed method in change localization accuracy and semantic recognition consistency, particularly in scenarios with frequent land cover type transitions and significant scale variations. Future work will focus on exploring lightweight collaborative architectures and incorporating multimodal remote sensing data to further enhance practical applicability.