Abstract: To accurately discriminate the maturity of tobacco leaves in field with low labeling cost, a semi-supervised image classification model based on adaptive threshold pseudo-labeling and consistency regularization was developed. By introducing pseudo-label comprehensive quality, evaluation indexes of the sequential confidence and sequential stability of the samples were evaluated, and a dynamic self-adaptive pseudo-labeling threshold was designed, which was self-adjustable to adapt to different learning states of the model, thereby to introduce more low-confidence samples and provide richer effective sample information to the model. In addition, the model is constrained to maintain prediction consistency of unlabeled samples via consistency regularization of the unlabeled samples enhanced by different data. The developed model was compared with traditional semi-supervised models, and the results showed that the developed model could effectively and accurately classify the maturity of tobacco leaves in field while reducing the manual labeling costs. Under the condition that only 30% of the initialized labeled samples were used, the recognition accuracies of the developed model on the three datasets of lower, middle, and upper leaves reached 91.1%, 90.3%, and 87.9%, respectively, which were better than the performance of the traditional semi-supervised models. This technology supports the accurate and rapid maturity discrimination of tobacco leaves in field.