Spectral-spatial open-set recognition with adaptive contrastive semantic reconstruction for hyperspectral image classification

Official URL: https://dx.doi.org/10.1109/SIU71813.2026.11636693

Open-set recognition in hyperspectral imagery requires classifying known land-cover classes while rejecting samples from unseen categories. Existing reconstruction-based approaches often discard cross-band spectral correlations, provide insufficient per-class separation, and rely on fixed thresholds that limit adaptability. This paper proposes a unified framework that addresses all three limitations. A lightweight pointwise convolution restores spectral correlations after grouped embedding. A hinge-based contrastive loss separates bottleneck representations across class autoencoders, reducing inter-class confusion. The global mean reconstruction error provides a complementary detection signal, and fusion weights are calibrated using the Fisher discriminability ratio. An adaptive threshold anchors the decision boundary to the training score distribution, eliminating manual tuning. Experiments on the Houston benchmark demonstrate improved overall accuracy, kappa coefficient, and unknown-class detection compared with state-of-the-art methods.