Cloud pollution can easily decrease the accuracy of satellite infrared hyperspectral observation data, leading to the loss of a large amount of observation information. In this study, a method for retrieval of temperature profile on the cloud is proposed based on observation data of FY-4A/GIIRS with cloud conditions. The radiative transfer model is used to carry out simulation experiments of observation brightness temperature under conditions of clear sky and cloud, respectively. We statistically analyze the characteristics of simulated brightness temperature changes under different channels, determine the channel selection scheme according to the cloud top pressure, and realize the retrieval of the temperature profile on the cloud through the neural network algorithm. The ERA5 reanalysis data is used as the reference standard in the accuracy evaluation of the temperature profile retrieval. The experimental results show that the overall root mean square error is better than 1.5 K, and the retrieval temperature profile has a high accuracy, which effectively improves observation data usage rate of the FY-4A/GIIRS in the cloud under pollution.