Multi-level Graph Subspace Contrastive Learning for Hyperspectral Image Clustering

TL;DR

This work tackles hyperspectral image clustering by addressing the lack of global-local interaction modeling in prior methods. It introduces Multi-level Graph Subspace Contrastive Learning (MLGSC), which builds dual feature views (spectral-spatial and texture), applies attention pooling to obtain a robust global graph representation, and leverages node-level and graph-level contrastive losses to fuse local and global information across views. A self-expression-based affinity learning stage further refines the clustering, culminating in spectral clustering on the learned affinity matrix $W = \tfrac{1}{2} (|C| + |C|^T)$. Empirically, MLGSC achieves state-of-the-art OA/NMI/Kappa on Indian Pines, Pavia University, Houston-2013, and Xu Zhou, validating its effectiveness and robustness for unsupervised HSI clustering with practical impact for large-scale remote sensing analytics.

Abstract

Hyperspectral image (HSI) clustering is a challenging task due to its high complexity. Despite subspace clustering shows impressive performance for HSI, traditional methods tend to ignore the global-local interaction in HSI data. In this study, we proposed a multi-level graph subspace contrastive learning (MLGSC) for HSI clustering. The model is divided into the following main parts. Graph convolution subspace construction: utilizing spectral and texture feautures to construct two graph convolution views. Local-global graph representation: local graph representations were obtained by step-by-step convolutions and a more representative global graph representation was obtained using an attention-based pooling strategy. Multi-level graph subspace contrastive learning: multi-level contrastive learning was conducted to obtain local-global joint graph representations, to improve the consistency of the positive samples between views, and to obtain more robust graph embeddings. Specifically, graph-level contrastive learning is used to better learn global representations of HSI data. Node-level intra-view and inter-view contrastive learning is designed to learn joint representations of local regions of HSI. The proposed model is evaluated on four popular HSI datasets: Indian Pines, Pavia University, Houston, and Xu Zhou. The overall accuracies are 97.75%, 99.96%, 92.28%, and 95.73%, which significantly outperforms the current state-of-the-art clustering methods.

Figures (9)

• Figure 1: Motivation of the MLGSC. First, the spectral-spatial feature views ($A^{1}$) and texture feature views ($A^{2}$) are constructed based on the input HSI data, and the data augmentation is performed on the dual views to obtain the node representations of the dual views using GCN and the global representations of the views through attention pooling. Then, based on the node representations and global representation of the two feature views, multi-level graph contrastive Learning is performed to obtain the joint local-global graph representation. Finally, spectral clustering is applied to the affinity matrix.
• Figure 2: Multi-level Contrastive Learning Framework.
• Figure 3: Clustering visualization on Indian Pines dataset: (a) original image, (b) ground truth, (c) k-means, (d) SSC, (e) 12-SSC, (f) EGCSC, (g) HGCSC, (h) GR-RSCNet, (i) our proposed MLGSC.
• Figure 4: Clustering visualization of Pavia University dataset: (a) original image, (b) ground truth, (c) k-means, (d) SSC, (e) 12-SSC, (f) EGCSC, (g) HGCSC, (h) GR-RSCNet, (i) our proposed MLGSC.
• Figure 5: Clustering visualization of Houston2013 dataset: (a) original image, (b) ground truth, (c) k-means, (d) SSC, (e) 12-SSC, (f) EGCSC, (g) HGCSC, (h) GR-RSCNet, (i) our proposed MLGSC.