Skip to content
Open access

Adaptive Boundary Consistency Learning for Open-Set Semi-Supervised Hyperspectral Image Classification

2026 · IEEE Transactions on Geoscience and Remote Sensing · Vol 64, pp. 5532617-5532617 · 3 citations · 56 references

Abstract

Most existing semi-supervised hyperspectral classification methods are built upon the closed-set assumption, which assumes that unlabeled samples originate exclusively from the known class space. However, this assumption is frequently invalid in real-world remote sensing scenarios. In addition, hyperspectral images (HSIs) exhibit complex spectral–spatial characteristics, facing challenges such as complex spatial boundaries, significant interclass spectral similarity, and intraclass variability. To address these issues, this article proposes an adaptive boundary consistency learning (ABCL) model for open-set semi-supervised HSI classification, which achieves low-cost and high-accuracy perception while also considering known-class classification and unknown-class rejection. First, to enhance the model’s ability to represent complex land-cover boundaries and spatial structural differences, an anisotropic position transformer (APT) is designed to introduce direction-sensitive spatial priors by decoupling horizontal and vertical positional biases. Second, to address interference from unknown classes in a large number of unlabeled samples, an adaptive entropy mask (AEM) is proposed to guide consistency learning, thereby improving the discriminability of the model’s features. Finally, to further optimize the feature space structure, a category recognition module (CRM) is introduced. By enhancing intraclass compactness and interclass separability, it provides a more stable representation basis for open-set discrimination. Experimental results on three public hyperspectral datasets, Indian Pines (IP), Salinas (SA), and LN02, demonstrate that with only ten labeled samples per known class, the overall accuracy (OA) of the proposed method reaches 79.56%, 88.20%, and 96.15%, respectively. Furthermore, the model exhibits stable advantages in metrics such as unknown, average accuracy (AA), and Kappa coefficient, outperforming various representative state-of-the-art methods. This study provides a low-cost, high-precision technical foundation with strong adaptability to unknown classes for target perception in HSIs.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.