Residual Attention-Enhanced Bitemporal Change Detection with Pseudo-Change Suppression

With the increasing resolution of remote sensing images, surface change detection has become more detailed yet more challenging due to pseudo-changes caused by seasonal variations and imaging conditions. Traditional change detection methods, including pixel-level and object-level approaches, often assume highly similar image data, limiting their effectiveness in complex environments. While deep learning techniques have enhanced detection accuracy, convolutional neural network (CNN)-based models still suffer from vanishing gradients and high computational costs, hindering large-scale applicability. To address these issues, we propose a bitemporal change detection method that integrates residual attention mechanisms with differential operations to suppress pseudo-changes and enhance feature extraction. The model employs channel attention to assign adaptive weights to spectral features and utilizes residual connections to strengthen temporal feature representation. Furthermore, differential operations generate fine-grained difference maps, effectively distinguishing real changes from pseudo-changes. Experimental results demonstrate that our method improves detection robustness and precision, offering a more reliable solution for high-resolution remote sensing change detection.