Prototype-Based Pseudo-Label Refinement for Semi-Supervised Hyperspectral Image Classification

Pseudo-label (PL) learning-based methods usually regard class confidence above a certain threshold for unlabeled samples as PLs, which may result in PLs still containing wrong labels. In this letter, we propose a prototype-based PL refinement (PPLR) for semi-supervised hyperspectral image (HSI) classification. The proposed PPLR filters wrong labels from PLs using class prototypes, which can improve the discrimination of the network. First, PPLR uses multihead attentions (MHAs) to extract the spectral-spatial features, and designs an adaptive threshold that can be dynamically adjusted to generate high-confidence PLs. Then, PPLR constructs class prototypes for different categories using labeled sample features and unlabeled sample features with refined PLs to improve the quality of PLs by filtering wrong labels. Finally, PPLR further assigns reliable weights (RWs) to these PLs in calculating their supervised loss, and introduces a center loss (CL) to improve the discrimination of features. When ten labeled samples per category are utilized for training, PPLR achieves the overall accuracies of 82.11%, 86.70%, and 92.50% on the Indian Pines (IP), Houston2013, and Salinas datasets, respectively.