Remote sensing image scene classification has been extensively studied, and a significant proportion of solutions for this task are based on deep learning techniques nowadays. Although powerful, deep learning models usually require large amounts of labeled data for training, and labels for remote sensing images are usually hard to obtain and expensive. To alleviate this problem, semi-supervised learning has been proposed to train classifiers using both labeled and unlabeled samples. In this paper, we propose an improved method for labeled sample products based on consistency regularization. This method performs sample selection through a dynamic threshold mechanism based on a class-related complexity index. In the field of remote sensing image scene classification, this method can be used in the framework of semi-supervised learning to perform better sample set expansion according to the semantic complexity of different classes.