Minority-Aware Semi-Supervised Learning for Class-Imbalanced Hyperspectral Image Classification with Limited Labels

Abstract

Hyperspectral image (HSI) classification under limited labeled data is particularly challenging due to severe class imbalance, which often causes learning algorithms to favor majority classes and results in poor minority-class recognition. To address these challenges, this paper proposes a Minority Preserving Graph-Based Semi-Supervised Learning (MA-SSL) framework with Superpixel-Level Markov Random Fields (MRFs) for hyperspectral image classification. The proposed approach first segments an HSI into homogeneous superpixels and constructs a superpixel graph to capture spectral-spatial relationships. A Laplacian graph-based semi-supervised learning model is then employed to propagate label information from a small set of labeled samples, while a class-weighted optimization strategy emphasizes minority classes and mitigates imbalance during label propagation. To further enhance spatial consistency and preserve minority-class regions, a minority aware superpixel-level MRF refinement is introduced, preventing minority labels from being overwhelmed by neighboring majority classes. The proposed framework is evaluated on the Indian Pines, Pavia University, and Salinas benchmark datasets under a limited-label setting in which only 10% of samples per class are labeled. Experimental results demonstrate that the proposed method achieves overall accuracies of 90.42%, 91.49%, and 93.15% and minority-class recalls of 71.60%, 98.50%, and 90.47% on the respective datasets. Comparative analysis further shows that the proposed framework consistently outperforms SVM+SMOTE and conventional graph-based semi-supervised learning methods in both overall classification performance and minority-class recognition. These results confirm that the integration of minority-aware graph learning and superpixel-level MRF refinement provides an effective and robust solution for class imbalanced HSI classification under scarce-label conditions, making it well suited for practical remote sensing applications.