对比学习结合DenseNet的高光谱图像开放集分类

Key Points

• To develop a method for hyperspectral image classification that improves recognition of unknown classes while maintaining accuracy for known classes.
• Proposed a contrastive learning approach combined with DenseNet for feature extraction.
• Employed a spectral feature extraction module to obtain original spectral features.
• Used a transitional module to compress spectral channels for clearer class boundary distribution.
• Mapped output features to a spatial feature extraction module for extracting spatial dimensions.
• Implemented ResNet to capture local spatial structure features and enhance spatial perception.
• Achieved higher land cover classification accuracy for unknown classes: 68.81% on Houston2013, 69.24% on Pavia University, and 59.26% on WHU-Hi-LongKou datasets.
• Overall classification accuracies were 89.49%, 95.06%, and 95.03%, respectively.
• Demonstrated improved model performance in distinguishing ambiguous boundary features.