Graph-Weighted Contrastive Learning for Semi-Supervised Hyperspectral Image Classification

TL;DR

This work tackles semi-supervised hyperspectral image (HSI) classification with limited labels by moving away from costly pixel- or superpixel-based graph constructions. It proposes Graph-Weighted Contrastive Learning (GWCL), which builds a pixel-level graph and optimizes a joint loss $L = L_{gwcl} + \lambda L_{ce}$, where $L_{gwcl} = \frac{1}{|\mathcal{P}|} \sum_{(i,j)\in\mathcal{P}} s_{ij} \| z_i - z_j \|^2$ and $f(z_i,z_j|\theta) = \exp(- s_{ij} \| z_i - z_j \|^2)$ with $s_{ij}$ derived from a graph on $({\bm x}_i,{\bm x}_j)$. Graph construction uses ARM to reduce spectral dimensionality to $\beta$ and forms $\bm x_i=[\bm h_i'; m; n]$, with $s_{ij}=\exp(-\tfrac{1}{2}(\bm x_i-\bm x_j)^T \Sigma^{-1}(\bm x_i-\bm x_j))$ for $K$-NN neighbors. GWCL supports mini-batch training by precomputing the graph and updating on node subsets, while the backbone is a two-layer MLP with 180 hidden units. Empirical results on Indian Pines, Salinas, and University of Pavia show GWCL's superiority over strong baselines, validating effective pixel-level learning and the use of unlabeled data in semi-supervised HSI classification.

Abstract

Most existing graph-based semi-supervised hyperspectral image classification methods rely on superpixel partitioning techniques. However, they suffer from misclassification of certain pixels due to inaccuracies in superpixel boundaries, \ie, the initial inaccuracies in superpixel partitioning limit overall classification performance. In this paper, we propose a novel graph-weighted contrastive learning approach that avoids the use of superpixel partitioning and directly employs neural networks to learn hyperspectral image representation. Furthermore, while many approaches require all graph nodes to be available during training, our approach supports mini-batch training by processing only a subset of nodes at a time, reducing computational complexity and improving generalization to unseen nodes. Experimental results on three widely-used datasets demonstrate the effectiveness of the proposed approach compared to baselines relying on superpixel partitioning.

Figures (10)

• Figure 1: Different approaches to applying graphs in semi-supervised HSI classification include: (a) Pixel-level graph has high computational and storage requirement; (b) Superpixel-level graph often suffers from pixel misclassification due to inaccuracies in superpixel partitioning; (c) Our proposed GWCL method eliminates the need for superpixel partitioning and supports mini-batch training, offering a more efficient and accurate solution.
• Figure 2: Indian Pines in pseudocolor and the ground-truth map with class information.
• Figure 3: Salinas in pseudocolor and ground-truth map with class information.
• Figure 4: University of Pavia in pseudocolor and ground-truth map with class information.
• Figure 5: The ground-truth and classified maps of different methods on the Indian Pines dataset. (a) the ground-truth map; (b) IFRF ($92.65\%$); (c) ARM ($94.18\%$); (d) SGL ($94.31\%$); (e) MSSGU ($95.93\%$); (f) DMSGer ($95.25\%$); (g) ConGCN ($96.44\%$); (h) GWCL ($98.81\%$) .