EfficientSAM-Assisted Network With Feature Interaction and Priori Guidance for Remote Sensing Change Detection

Change detection is one of the important methods for earth observation, and with the rapid development of deep learning, more effective methods are provided for the change detection task. However, many methods do not pay attention to the correlation between Bi-temporal remote sensing images and cannot effectively extract change information, which leads to the problems of poor recognition of pseudo-changes as well as the appearance of internal voids with inadequate expression of change features. In this letter, we propose an EfficientSAM-assisted feature interaction and a priori guidance network (EfficientSAM-FG) to address these issues. We enhance feature extraction and generalization capabilities by fine-tuning EfficientSAM. We introduce a multi-scale feature interaction module (MFI) to learn spatial information correlations between diachronic features. We design a change guided module (CGM) in the feature fusion stage to guide multiscale feature fusion with prior knowledge. Finally, we employ a space attention module (SAT) to obtain a location attention map, enhancing change region features and suppressing background region features. In order to verify the effectiveness of our method, we compare it with other methods on two datasets, LEVIR-CD and WHU-CD, and the experimental results prove that EfficientSAM-FG can achieve better results, with F1 of 92.15% and IOU of 85.41% on LEVIR-CD and 94.50%, 89.42% on WHU-CD.