ChangeAD: Enhanced Remote Sensing Change Detection via Bi-Temporal Alignment and Differential Feature Integration

Remote sensing change detection (RSCD) aims to identify semantic changes in bi-temporal images, with broad applications in disaster assessment, urban expansion analysis, and other fields. With advancements in computer vision, deep learning-based methods have significantly improved the performance of RSCD tasks. However, existing approaches primarily focus on feature representation, overlooking the interaction between bi-temporal features, which leads to suboptimal performance when dealing with imaging condition disparities (such as seasonal changes, lighting variations, and image distortions). To address these limitations, we introduce ChangeAD, a novel framework that leverages Bi-temporal Alignment (BTA) to align bi-temporal features, mitigating false changes caused by imaging condition differences. We further integrate Differential Feature Integration (DFI) and Cross-level Feature Integration (CFI) modules to extract and integrate a differential feature pyramid. This approach utilizes the integrated differential feature maps, rich in semantic and edge contour information, to address the identification challenges posed by complex details and texture features in high-resolution images. Experimental evaluations on the LEVIR-CD and S2-Looking datasets demonstrate that ChangeAD achieves F1 scores of 78.6% and 83.6%, respectively, showcasing its effectiveness in handling the complexities of high-resolution remote sensing Images.