Abstract：Compared to traditional methods， change detection methods based on deep learning have higher accuracy and stability. Most deep learning models belong to supervised methods and require a large number of standard samples with high cost of sample construction. On the other hand， the current unsupervised change detection methods have lower accuracy. Aiming at the problem， a self-supervised change detection method fused with remote sensing knowledge is proposed， and a self-supervised representation learning framework is designed combined with multi-view contrastive loss and vision transformer. In order to further improve the detection accuracy， NDVI and NDWI are weighted and fused with the regression error of supervised representation learning as remote sensing knowledge to further improve the change detection accuracy. The experiment conducted on the OSCD dataset of the University of Paris Thackeray in France shows that the F1 detection result of this method has an average improvement of over 2.76% compared to other unsupervised change detection algorithms， and has higher stability and robustness.
2024, Vol. 58 
