Cross-Domain Few-Shot Learning with Label Propagation for Hyperspectral Image Classification

In practical applications, the acquisition cost of hyperspectral images is high, and acquiring substantial labeled data is difficult, which results in a restricted number of labeled samples using for training. Few-shot learning (FSL) serves as an efficient method to tackle the classification challenges of hyperspectral images (HSI) when labeled samples are scarce. It acquires transferable knowledge from abundant labeled data, that can be transferred to new domains. However, the distributional differences between domains with abundant labeled data and those with limited labeled data make the generalizable knowledge learned from the former less applicable to the classification tasks of the latter. To address this issue, this paper proposes a cross-domain few-shot hyperspectral image classification method based on label propagation. Firstly, the method employs a feature extraction module to capture complex information within samples, then applies a graph structure to the label propagation algorithm. The graph structure effectively captures local relationships between samples and avoids over-reliance on distant samples, thereby reducing the risk of overfitting. Furthermore, it incorporates a Tri-training structure to learn from different perspectives and complement each other, ultimately achieving classification of cross-domain few-shot data. Experimental results on hyperspectral datasets with Chikusei as the source domain, and Pavia University and Salinas as the target domains demonstrate that our method is effective and has good performance.