Based on the daily meteorological observation data of 158 stations in Sichuan Province and the road meteorological disaster data from 2018 to 2020, the correlation analysis is used for a total of 14 factors including temperature, humidity and Precipitation Factors and their daily variation composite characteristics to study the difference of road icing modes under plain and mountainous terrain. The results show that the correlation coefficient between average relative humidity and icing is 0.451 and -0.451 in the two terrain respectively. The difference of icing mode in mountainous and plain terrain is mainly reflected in the different factors that meet the water vapor conditions required for icing. In addition, based on the dimensionality reduction of principal components, stepwise regression and lasso principle, the logistic icing prediction model is constructed, the accidental error is reduced by 50% cross validation, and the obtained model is evaluated by using the standards of accuracy, ROC curve and AUC value. The performance of the three models under the two terrains ranked the same. Lasso-Logistic model had the best effect, with an average accuracy of 86.98%, followed by SR-Logistic, with an average accuracy of 84.05%, and PCA-Logistic was the lowest, with an average accuracy of 84.09%. In the typical case test, the prediction accuracy of Lasso-Logistic in plain and mountainous areas is 86.15% and 83.70% respectively, which is the highest among the three models. At the same time, the introduction of the composite characteristics of meteorological factors over time can improve the prediction accuracy of the model. Among them, Lasso-Logistic, PCA-Logistic and SR-Logistic have improved the accuracy of the model by 3.68%, 3.00% and 3.00% compared with the model with only daily mean. The results show that lasso logistic icing prediction model can effectively predict road icing events under two kinds of terrain, has strong universality, and can provide reference for road icing disaster early warning.