Deep Learning‐Based Building Change Detection in Off‐Nadir Images via a Pixel‐Wise and Patch‐Wise Fusion Strategy

ABSTRACT Building change detection (BCD) is crucial for sustainable urban development and operational efficiency. High‐resolution remote sensing imagery offers numerous benefits for this task. However, it also presents challenges due to the intricate spectral and texture features and misalignment issues in bi‐temporal images. To overcome these obstacles, we have developed a new framework called PP‐BCD, which combines patch and pixel‐wise deep learning techniques to detect building changes in off‐nadir images. PP‐BCD utilizes PSPNet as the building detector to create pixel‐wise building masks and incorporates a custom pseudo‐siamese module (PSM) to filter unchanged building scenes, thus reducing the false change alerts resulting from image misalignment. This fusion strategy enables PP‐BCD to deliver pixel‐level and patch‐level BCD results simultaneously. We tested PP‐BCD in Zhuhai, China, using Pleiades satellite imagery with a spatial resolution of 0.5 m from 2019 to 2020. The experimental results indicate that PP‐BCD surpasses traditional post‐classification comparison methods while demonstrating competitive performance compared to mainstream deep learning‐based models. Importantly, our method relies solely on patch‐level change samples and pseudo‐sampling techniques to address the class imbalance effectively, thereby reducing sample collection costs in change detection tasks. The versatility and efficiency of PP‐BCD make it highly suitable for large‐scale BCD practical applications.