Exploring Few-Labeled Classification Strategies for Hyperspectral Images Based on VIT and Semi-Supervised Learning

Most deep learning models require a large number of labeled samples to achieve good performance, which is often impractical in real-world scenarios. This paper explores the application of the Vision Transformer (ViT) models in few labeled samples hyperspectral image classification tasks. In response to the limited amount of hyperspectral image data, we propose an effective few labeled samples learning method aimed at improving classification accuracy and generalization capabilities. By introducing data augmentation, regularization techniques, and transfer learning strategies, we significantly enhance the performance of the ViT model in small-sample scenarios. The experimental results demonstrate that this method achieves superior classification performance compared to traditional methods on multiple hyperspectral datasets, providing a novel solution for the field of hyperspectral image analysis.