对比学习结合 DenseNet 的高光谱图像开放集分类

Hyperspectral image classification typically assumes that the training and test data share identical categories and that no unknown classes appear in the test set. However，this assumption is rarely satisfied in practical applications. In addition，the subtle inter-class differences inherent in hyperspectral data often lead to overlapping feature distributions and consequent decision boundary ambiguity. To address these issues，an open-set classification method for hyperspectral images is proposed that integrates contrastive learning with DenseNet. First，a spectral feature extraction module is employed to obtain the original spectral features，and multi-level feature interaction is realized through DenseNet. A transition module is further applied to compress spectral channels，thereby yielding clearer class boundary distributions. Second，the extracted spectral features are mapped to a spatial feature extraction module to obtain spatial-domain representations，where ResNet is adopted to capture local spatial structural information and enhance spatial perception. Subsequently，contrastive learning is introduced to reinforce intra-class compactness and inter-class separability，and is combined with a hard-sample mining mechanism to optimize ambiguous boundary features and improve the model’s discriminative capability for boundary-region samples. Experiments conducted on the Houston 2013，Pavia University，and WHU Hi-LongKou datasets demonstrate that the proposed method achieves superior ground-cover classification performance on unknown categories，with accuracies of 68. 81%，69. 24%，and 59. 26%，respectively. Meanwhile，overall accuracies of 89. 49%，95. 06%，and 95. 03% are obtained，indicating that the recognition of unknown categories is effectively enhanced while maintaining high classification accuracy for known categories.