Orientational Clustering Learning for Open-Set Hyperspectral Image Classification
Hao Xu,Wenjing Chen,Cheng Tan,Hailong Ning,Hao Sun,Wei Xie
DOI: https://doi.org/10.1109/lgrs.2024.3432604
IF: 5.343
2024-08-28
IEEE Geoscience and Remote Sensing Letters
Abstract:Recently, some literature has begun to pay attention to the open-set problem in remote sensing application scenarios and studied various open-set hyperspectral image classification (OSHIC) methods. These OSHIC methods are usually based on deep neural networks, using the nondirectional Euclidean distance losses to constrain latent sample representations of known classes to be compact. Nonetheless, the potential effect of the spatial distribution of sample representations is ignored, resulting in degraded classification performance in OSHIC. In this letter, we propose an orientational clustering learning (OCL) method for OSHIC. First, in the feature space generated by the convolutional neural network, a class anchor strategy is employed to bring features of the same class closer while keeping features of different classes distant. Then, we utilize the orientational learning to further tighten the intraclass feature space. OCL directionally optimizes the spatial distribution of hyperspectral sample representations to improve the ability to identify known classes and distinguish unknown classes. Experiments show that the OCL achieves overall accuracies of 94.43%, 92.27%, and 76.94% on the Pavia University, Salinas, and Indian Pines datasets, respectively.
imaging science & photographic technology,remote sensing,engineering, electrical & electronic,geochemistry & geophysics