Abstract: Aiming at the problem of hyperspectral images (HSIs) few-shot classification, an unsupervised meta learning method is proposed in this paper. This method can utilize unlabeled samples for unsupervised meta learning, significantly reducing the dependence on a large number of labeled samples while effectively improving the accuracy of HSIs few-shot classification. Firstly, based on the idea of self-supervised learning, multiple different augmentation features of the same sample are generated through principal component analysis and data augmentation methods, to form a large number of different tasks and perform meta-training on the designed model. Then, a few labeled samples randomly selected from target HSIs are used to fine-tune the model, and all labeled samples are used for classification test to evaluate the few-shot classification performance of the model. Moreover, the voting strategy is introduced in the fine-tuning and classification phase, to further improve classification accuracy. The proposed method can perform sufficient unsupervised meta-training under the condition of zero labeled sample, breaking through the bottleneck and limitation of the supervised meta learning methods requiring a large number of source labeled samples. Experiments on three public datasets show that the proposed method can obtain better classification results than existing advanced methods in the few-shot scenario.

Key words: hyperspectral images classification, few-shot classification, unsupervised meta learning, deep learning, voting strategy