Answers are grounded in this paper and cite their sources.
Existing open-set classification (OSC) methods for hyperspectral images (HSI) often treat unknown categories as a homogeneous cluster, failing to exploit spectral-spatial heterogeneity for fine-grained discrimination. To address this gap, a boundary-reinforced reciprocal point learning multi-classification network (RPLMN-BR) is proposed for robust HSI classification in open-set environments (OSE). First, a SeSaMLP feature extraction framework is designed to enhance discriminative spectral-spatial feature representation by integrating fully connected layers with spectral and spatial attention modules. Second, a weighted reciprocal point learning (RPL) module is fused with generative adversarial networks (GAN) to explicitly strengthen decision boundaries between known and unknown classes; a learnable dynamic threshold is introduced to adaptively construct a separable feature space, improving unknown instance detection. Finally, a correlation-aware multiclassification strategy is developed: global sample correlations (via Jaccard index) are used to build prototype relevance weights, and Pearson correlation coefficients between open-set instances and closed-set prototypes enable fine-grained discrimination of unknown categories. Extensive experiments on four public HSI datasets covering open-set domain adaptation (OSDA) and open-set multiclassification tasks demonstrate that RPLMN-BR achieves superior performance in both closed-set and open-set scenarios, outperforming state-of-the-art methods in harmonic open-set score, overall accuracy, average accuracy, and Kappa coefficient. The proposed framework provides a reliable solution for handling emerging unknown land cover types in remote sensing applications.
No takes yet. Share an insight, caveat, or question.
Xu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: