Uncertainty-Aware Prototype Learning for Open-Set Hyperspectral Image Classification

Hyperspectral image classification, aiming to assign an unique label for each pixel, has garnered much more attention in recent years. However, current most classification approaches assume that all classes between training and test sets are consistent, which limits their applications for identifying the unknowns. In this article, an uncertainty-aware prototype learning model (UAPLM) is proposed for open-set hyperspectral image classification. First, a spectral-spatial convolutional network is utilized to extract the high-order semantic information from input data. Then, feature consistency and sample uncertainty are estimated to obtain the probability of an unknown class. Finally, the known class probabilities obtained by the fully connected layer and the unknown class probabilities are integrated to yield the final classification result. Experiments on several datasets demonstrate that the proposed UAPLM can better recognize the known and unknown samples compared to other open-set classification methods.