Feature Mutual Reconstruction for Semi-Supervised Few-Shot Remote Sensing Image Scene Classification

Few-shot remote sensing image scene classification (RSISC) faces the challenge of too few labeled samples. In recent years, semi-supervised few-shot learning (SSFSL) have expanded the labeling information by assigning pseudo-labels to unlabelled samples. However, obtaining high-confidence pseudo-labels with limited labeling samples also remains challenging. For this problem, this work proposes a feature mutual reconstruction method for semi-supervised few-shot RSISC. Specifically, the proposed feature mutual reconstruction module uses the similarity between unlabeled and labeled data to reconstruct labeled data, and the reconstructed labeled data classify the unlabeled data by joint representation error to obtain high-confidence pseudo-labels. In addition, the proposed prototype contrast correction module adjusts the spatial position of the prototypes by constructing a contrast learning model through pseudo-labels. Experimental results on three public datasets demonstrate that our method can effectively distinguish unlabelled data and achieve excellent classification results.