Detecting building changes (DBC) from dual-temporal remote sensing images is a vital tool for monitoring encroachments on state lands, detecting illegal constructions, and supporting sound planning for smart city development. This process also plays a significant role in enhancing the understanding of urban expansion and associated human activities. However, existing methods for DBC face several challenges, as they are highly susceptible to interference from spectral changes in building surface colors and shadows of high-rise structures caused by lighting variations and differences in imaging angles. These issues result in elevated error rates in identifying actual changes and reduced accuracy of the generated maps. Moreover, the reliance on localized spatial information and the limited capacity to represent extracted features often leads to incomplete building boundaries, particularly in densely built areas with significant overlap between adjacent structures, complicating the accurate delineation of building edges. To alleviate these problems, a novel Siamese method, referred to as Edge-CVT, was proposed. This method integrated convolutional neural networks with edge-guided vision transformers to accurately detect building changes while preserving the integrity of their boundaries in high-resolution satellite imagery. Specifically, a feature disparity boosting module (FDBM) was introduced as the core component of the Edge-CVT model. This module generated rich spatial and temporal features by combining local and global spatial information, thereby mitigating pseudo changes and reducing the impact of spectral interference. In addition, an edge-informed change module (EICM) was designed to direct the model’s focus toward the edges of changing buildings, enhancing geometric accuracy and maintaining the integrity of edge shapes. This module also enabled the effective separation of adjacent and overlapping building boundaries. We validated the effectiveness of Edge-CVT through extensive experiments conducted on four open-source DBC datasets, namely EGY-BCD, PRCV-BCD, LEVIRCD+, and BTRS-CD. The experimental results demonstrate that Edge-CVT outperforms state-of-the-art methods in both qualitative and quantitative evaluations, achieving F1-score of 90.12%, 88.85%, 94.26%, and 86.87% for the respective datasets. |
