A comparative analysis of deep learning methods for building change detection in high-resolution satellite imagery

The rapid advancement of remote sensing technology has facilitated the extensive use of high-resolution satellite imagery for building change detection (BCD), supporting applications such as urban planning, disaster assessment, and land monitoring. While deep learning (DL) methods have shown strong potential in BCD tasks, comparative studies on their performance remain limited. This study conducts a systematic evaluation of representative DL-based BCD approaches on high-resolution imagery, using standard metrics to assess their accuracy and generalization. The goal is to identify effective models, highlight current limitations, and provide insights for future improvements.We compare direct-prediction and taskdecoupling architectures, assess multimodal (optical and SAR) fusion, and highlight scenarios where each method excels or falters. The results clarify performance gaps, confirm the benefits of task decoupling and data fusion, and outline directions for more robust BCD algorithms.