The data variational assimilation method is based on the assumption that the observation error is untiased, so the bias correction is one of the important links in the quality control of satellite data. In this paper, the research based on ensemble learning on the bias correction of the brightness temperature of the midwave infrared channel of FY-4A/GIIRS is carried out. Random Forest, XGBoost, Decision Tree and Extra Tree are used as the base models for the ensemble learning. After optimizing the hyperparameters of the base model, the generalized error minimization method is used to integrate the base model regression results. Based on the encrypted clear-sky field-of-view data during Typhoon Lekima, the correction effects of the ensemble learning, the base model, and the offline method on the brightness temperature bias of the GIIRS channel are compared. The experimental results show that all the correction methods used in this paper achieve good results Among all the methods, the ensemble learning has the best correction effect. Among the air mass predictor, geographical (longitude and latitude) information contributes a lot to the base model. The methods in this paper can be extended to the bias or error correction of other information.