Adapting SAM for remote sensing change detection: feature alignment and discrepancy enhancement

Remote sensing change detection plays a vital role in resource management, environmental monitoring, infrastructure construction and agricultural production. However, change detection in remote sensing is often hindered by the scarcity of labeled data and pseudo-discrepancies caused by external factors such as lighting and seasonal variations between images captured at different times. In this work, we apply the segment anything model (SAM) model, known for its zero-shot learning capability, to the change detection task and propose the FADE-SAM model. This work includes a feature alignment module, which processes images to unify the visual styles of images taken at different times, thereby reducing interference from environmental factors. Additionally, we also introduce a discrepancy enhancement module to strengthen the representation of variable regions, effectively improving the accuracy and robustness of feature extraction. Our experiments on the LEVIR-CD dataset demonstrate the effectiveness of the proposed model, showing that our method outperforms existing approaches across relevant metrics.