A graph contrastive learning network for change detection with heterogeneous remote sensing images

Land cover change detection (LCCD) with heterogeneous remote sensing images (Hete-RSIs) is an attractive topic in the community of remote sensing applications. Intuitively, Hete-RSIs are acquired with different remote sensors, and they cannot be compared directly for LCCD because of the different imaging modalities. In this paper, a graph contrastive learning network (GCLN) is proposed for LCCD with bitemporal Hete-RSIs. First, with the motivation of smoothing the noise and utilizing contextual information, the k-nearest neighbor algorithm is used to improve the spectral homogeneity of the pixels within a superpixel. Then, a pairwise graph is constructed on the basis of each superpixel from spectral similarity and dissimilarity perspectives, and a graph feature learning network is designed to learn the near-far dependencies of graph features for change detection. Finally, the similarity and dissimilarity loss functions are coupled as a contrastive loss function to expand the difference between similar and dissimilar features. Comparisons with seven advanced methods on five pairs of Hete-RSIs demonstrate the feasibility and superiority of the proposed GCLN for LCCD with Hete-RSIs. For example, the improvements on the five datasets are 3.63$\%$, 8.47$\%$, 4.17$\%$, 8.23$\%$, and 4.98$\%$ in terms of overall accuracy. The code of the proposed approach can be available at: https://github.com/ImgSciGroup/2024-GCLN.