Open-set domain generalization for hyperspectral image classification via evidential uncertainty quantification

Official URL: https://dx.doi.org/10.1109/WHISPERS69515.2025.11501598

Open-set domain generalization for hyperspectral image classification tackles challenges from both unknown classes and cross-domain generalization without target adaptation. We propose a framework that extracts domain-agnostic features through attention-weighted frequency analysis, while capturing complementary spectral-spatial information via parallel pathways. It provides reliable open-set classification through uncertainty-aware pathway weighting. Experiments on Pavia University and Centre indicate performance levels comparable to state-of-the-art domain adaptation methods despite not requiring target domain access during training.