This is the authors' abstract. We don't add key points for this paper.
In hyperspectral image classification (HSIC), tasks can be categorized into closed-set and open-set classification based on whether the test data contains novel classes that are unseen during training. While closed-set classification has achieved saturated performance in recent years, real-world scenarios are inherently open. Existing open-set HSIC methods can be broadly classified into generative-based and discriminative-based methods. Generative models often exhibit instability in open-set scenarios, as they struggle to comprehensively model the complex and diverse characteristics of unknown classes. Discriminative models, on the other hand, typically rely on non-directional Euclidean distance constraints in the feature space, which may lead to loosely clustered features. Two key challenges in open-set HSIC are: (1) how to effectively leverage known classes to model the open space beyond seen categories, and (2) how to determine adaptive and learnable decision thresholds. Guided by these principles and the motivations, this paper proposes an orientational distribution learning network that integrates intra-class compactness and inter-class separability (ICS-ODL). Specifically, a dual-branch hierarchical partial convolutional network is adopted as the feature extractor, followed by supervised contrastive learning on the extracted features. Distance constraints are imposed between feature representations and their corresponding class centers, while intra-class and inter-class orientational constraints are introduced to refine the structure of the feature space. Finally, a class-wise unknown class rejection threshold is assigned to each category. Experiments conducted on open-set hyperspectral image datasets demonstrate that the proposed method achieves superior performance, validating its effectiveness in addressing the challenges of open-set HSIC.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: