Abstract: In order to improve the accuracy of the microwave radiometer in detecting the atmospheric profile, two back propagation neural network (BPNN) models with different cloud information (C-BPNN model with cloud inlet and cloud outlet height added and Z-BPNN model with radar reflectivity factor added) were constructed using the ground-based microwave radiometer, millimeter wave cloud radar and radiosonde data provided by US ARM atmospheric observatory. The inversion results were compared with the BP neural network model without cloud information (BPNN₀). The results show that the retrieval errors of C-BPNN model and Z-BPNN model in any weather (with or without clouds) are smaller than those of BPNN₀ model; C-BPNN inversion results are more stable than the other two models. Through the analysis of the case with the best inversion results of the three models, it is found that the main error of C-BPNN and Z-BPNN models exists in the situation that there is no cloud at high altitude but the relative humidity has a jump, which indicates that the neural network model is sensitive to the initial weight and threshold value. Therefore, the BPNN model is optimized by genetic algorithm (GA). The inversion results after GA optimization show that BPNN₀ model and C-BPNN model have obvious optimization effects, however, the optimization effect of Z-BPNN model is not obvious.