Graph Contrastive Learning With High-Order Relations for Hyperspectral Image Clustering

Hyperspectral image (HSI) clustering has shown remarkable potential in Earth observation, as it can utilize the rich spectral information for classification without depending on labels. Deep clustering methods based on contrastive learning have drawn significant attention from researchers due to their high performance. However, existing contrastive learning-based methods are unable to capture high-order positive samples and frequently misclassify negative samples. To address this issue, we propose a novel Graph Contrastive Learning method with High-order Relations (GCLHR) for HSI clustering. Specifically, we construct an affinity matrix and employ multiple random walks to capture the distant positive samples of each sample while avoiding capturing false negative samples. Meanwhile, the similarity measurement also filters out the false negative samples to a certain extent. Before formal training, we use a 3D autoencoder based pre-training to capture deep spectral and spatial features to provide the input of the graph convolutional network. Moreover, GCLHR utilizes clustering semantic information to eliminate noisy edges for view augmentation and designs a triplet self-supervised learning module to enhance the consistency between views. Experiments carried out on multiple HSI datasets demonstrate that our proposed GCLHR method attains excellent and reliable clustering performance.