The influence of atmospheric background data on liquid water path retrieval results is discussed based on the airborne GVR and BP neural network algorithm. It provides a basis for reasonably selecting training samples to obtain more accurate liquid water observation data and is beneficial to understand the detection scope of the retrieval algorithm. Multiple historical sounding data are selected and classified by historical data time series length, season, and region. Different training sample sets are established to train BP neural networks to obtain the corresponding retrieval equations. The sample test set is selected to calculate the retrieval accuracy of each type of retrieval equations. The influence of atmospheric background data difference on liquid water path retrieval results is analyzed by retrieval accuracy comparison. The results show that the spatial and temporal differences in the atmospheric background of the training samples influence the retrieval results. The effect of atmospheric background differences on the retrieval error can be reduced by increasing the length of historical-sounding data. However, it does not work when the time series length reaches a certain extent. Seasonal classification can reduce the impact of atmospheric background differences on retrieval error, but data classification reduces the sample size in practice. For the historical data of a certain time series length, classification according to the season cannot effectively improve the retrieval accuracy of the liquid water path.