Gaussian Mixture Model Based Pseudo-label Learning for Semisupervised Hyperspectral Image Classification
Abstract
Semisupervised learning can alleviate the heavy reliance on labeled samples for hyperspectral image (HSI) classification while suffering from the critical problem of the accurate identification of reliable pseudo-labeled data. The problem is mainly caused by the fact that the model may predict an incorrect label with high confidence and vice versa, especially in the case of limited labeled samples. To solve this problem, a new Gaussian mixture model based pseudo-label learning method (GMM-PLL) is proposed for HSI classification (HSIC). First, we innovatively model the reliability of classification results using a Gaussian mixture model modeling (GMM-M) module. Benefiting from this, reliable pseudo-labeled samples can be adaptively selected based on the probabilistic model, rather than a rigid reliance on prediction confidence. Then, considering that a limited number of labeled samples cannot sufficiently train the classification model, we construct a class-balanced contrastive learning (CBCL) mechanism to fully exploit unlabeled information for training. As a consequence, pseudo-label sample generation and classification model training are performed alternately, which produces more reliable pseudo-labels and better models. Finally, satisfactory classification performance can be achieved based on the well-trained model. Experimental results demonstrate the effectiveness and superiority of the proposed method. The code is available at https://github.com/zzm-hnu/GMM-PLL