Structural-Spectral Graph Convolution with Evidential Edge Learning for Hyperspectral Image Clustering

TL;DR

This paper tackles unsupervised clustering of large-scale hyperspectral images by operating on superpixel graphs and addressing two core challenges: insufficient spectral feature extraction by conventional GNNs and the presence of incorrect edges in the superpixel topology. It introduces a structural-spectral graph convolutional operator (SSGCO) to co-extract spatial and high-order spectral features, and an evidence-guided adaptive edge learning (EGAEL) module to refine the superpixel graph with learned and empirical edge weights, all within a BYOL-style contrastive clustering framework that combines neighborhood alignment and prototype contrast. The approach achieves consistent improvements across four benchmark datasets, with substantial gains in ACC and other clustering metrics, and is supported by extensive ablation and hyperparameter analyses. The proposed method offers scalable, effective clustering for large HSIs and provides practical guidance for edge refinement and spectral-spatial representation learning, though it acknowledges limitations in capturing intra-superpixel 2D structure and points to future work along those lines.

Abstract

Hyperspectral image (HSI) clustering groups pixels into clusters without labeled data, which is an important yet challenging task. For large-scale HSIs, most methods rely on superpixel segmentation and perform superpixel-level clustering based on graph neural networks (GNNs). However, existing GNNs cannot fully exploit the spectral information of the input HSI, and the inaccurate superpixel topological graph may lead to the confusion of different class semantics during information aggregation. To address these challenges, we first propose a structural-spectral graph convolutional operator (SSGCO) tailored for graph-structured HSI superpixels to improve their representation quality through the co-extraction of spatial and spectral features. Second, we propose an evidence-guided adaptive edge learning (EGAEL) module that adaptively predicts and refines edge weights in the superpixel topological graph. We integrate the proposed method into a contrastive learning framework to achieve clustering, where representation learning and clustering are simultaneously conducted. Experiments demonstrate that the proposed method improves clustering accuracy by 2.61%, 6.06%, 4.96% and 3.15% over the best compared methods on four HSI datasets. Our code is available at https://github.com/jhqi/SSGCO-EGAEL.

Figures (7)

• Figure 1: Motivation of the proposed method: (a) clustering accuracy of different superpixel feature representation networks on Pavia University dataset, (b) the impact of deleting incorrect edges on Pavia University dataset.
• Figure 2: The diagram of the proposed method: (a) the contrastive clustering framework consists of an online network $f(\cdot)$ and a target network $f'(\cdot)$, which take superpixels and augmented views as inputs. Neighborhood alignment (NA) and prototype contrast (PC) are adopted as training objectives; (b) Structural-spectral graph convolutional operator (SSGCO) sequentially performs 1D convolution (1D-Conv) and graph convolution in each layer to extract both spatial and spectral features of superpixels; (c) In evidence guided adaptive edge learning (EGAEL), each edge is represented by concatenating the soft assignments of its two endpoint nodes, and input into an MLP to get the predicted edge weight $w^{pre}$. The empirical edge weight $w^{emp}$ computed with node clustering confidence and similarity is adopted as guidance. The adjacency matrix is momentum updated based on $w^{pre}$.
• Figure 3: Clustering maps on Indian Pines: (a) ground truth, (b) K-means, (c) SSC, (d) NCSC, (e) SGLSC, (f) DGAE, (g) EGRC, (h) S$^3$-ULDA, (i) S$^2$DL, (j) SDST, (k) SPGCC, (l) SAPC, (m) Ours.
• Figure 4: The influence of different numbers of superpixels ($M$) and different numbers of network layers ($L$) on: (a) Indian Pines, (b) Pavia University, (c) Botswana and (d) Trento.
• Figure 5: The influence of different weights of $\mathcal{L}_{PC}$ ($\alpha$) and different weights of $\mathcal{L}_{E}$ ($\beta$) on: (a) Indian Pines, (b) Pavia University, (c) Botswana and (d) Trento.