Detecting semantic changes from VHR remote sensing images by integrating semantic correlations and change priors

Investigating semantic change detection (SCD) is critical for tracking the evolution of urban and rural conversion. However, existing methods applied to SCD face two pivotal challenges: (1) modeling of semantic correlations among bi-temporal features is insufficient, leading to substantial pseudo-changes induced by seasonal and temporal differences; and (2) change information lacks effective enhancement, causing incomplete and missed detection. To overcome these challenges, we propose a change prior-guided network, namely CPGNet, which employs a multi-branch paradigm that integrates supplemental changed information to accurately identify diverse types of land cover changes in very high-resolution (VHR) remote sensing images. Initially, a Siamese network is introduced to simultaneously extract robust semantic representations from bi-temporal images. Next, a semantic correlation module is designed to capture and process the interconnections present in bi-temporal features, thereby alleviating pseudo-changes resulting from season variance and time disparities. We then incorporate an auxiliary CD branch to capture changed prior information. Finally, a multi-view change feature fusion module based on Bayesian theory is developed to enhance salient changed features to improve the under-detection of changed areas, particularly for changed objects with small sizes and complex shapes. We applied the proposed method to the semantic change detection dataset (SECOND), JiLin-1 (JL-1) dataset, and Xiamen (XM) dataset, and compared the results to ten typical SCD methods. Our method yielded superior accuracy, outperforming existing methods by 5.97%–40.54%, 9.28%–65.98%, and 0.99%–13.77% in SeK value on the JL-1, XM, and SECOND datasets, respectively. Moreover, the detected results exhibited well boundary consistency and internal integrity. Our data and code are accessible at https://github.com/long123524/CPGNet . • CPGNet is developed to detect semantic changes from VHR remote sensing images. • An auxiliary change detection branch is used to provide changed prior information. • An SCM for modeling semantic correlations between bi-temporal features. • The proposed MCFF module improves the under-detection of changed areas. • CPGNet performs better than ten existing methods across varied changed scenarios.