Wind fields are crucial to the evolution and prediction of weather situations. Based on the medium wave channel data of GIIRS and the wind field data of ERA5, LightGBM is used to retrieve the three-dimensional atmospheric wind field in this paper. First, model feature variables are constructed. The two-step feature selection method is adopted for the optimal selection of GIIRS channels: (1) The blacklist of GIIRS channels is established; (2) Feature variables are selected by PFI method, and feature variables containing spatiotemporal information are constructed on the basis of forming optimal subsets of channels. Secondly, a three-dimensional wind field retrieval method based on LightGBM is constructed. Finally, LightGBM hyperparameter optimization and correlation retrieval experiments are carried out based on GIIRS encrypted data during Typhoon “Lekima”. The experimental results show that RMSE of the U and V components of the wind field in the test set is less than 1 m/s and 1.5 m/s respectively, compared with the ERA5 wind field data. The two-step feature selection method in this paper can realize the dynamic optimal selection of GIIRS channels.