Abstract: Herein, a dataset is constructed using the data from 158 daily meteorological observation stations of Sichuan Province and road meteorological disaster data from 2018 to 2020. Using Pearson’s correlation coefficient, correlation analysis is performed on 14 meteorological factors, including temperature, humidity, precipitation, and their daily variation compound features, to examine the differences in road icing patterns between plain and mountainous terrains. Logistic icing-warning models are built for various terrains using the dataset. Dimensionality reduction is employed to enhance the models by employing three methods: method of least absolute shrinkage and selection operator (LASSO) with the L1 penalty term in the linear module, principal component analysis (PCA) based on orthogonal transformation to reduce data dimensions, and stepwise regression (SR) based on significance for variable screening. Random errors are reduced using five-fold cross-validation. The research results can provide a reference for road icing warnings and forecasting for complex terrains. The following conclusions can be drawn. (1) Low temperature is the primary condition for the occurrence of road icing, and the difference in correlation coefficients between relative humidity factors and icing demonstrates different patterns of water vapor supply in different terrains. (2) The LASSO-logistic model exhibits the best prediction performance for both plain and mountainous terrains, with average accuracies of 87.67% and 82.62%, respectively; the average accuracies of the SR-logistic model are 85.77% and 80.64%, respectively, while those of the PCA-logistic model are 84.09% and 77.44%, respectively. (3) Introducing compound features of meteorological factors changing over time can enhance the model accuracy, and the average accuracies of the Lasso-, SR-, and PCA-logistic models with added compound features are increased by 3.68%, 3.00%, and 3.00%, respectively. (4) All models performed better for plain areas than for mountainous areas, with an average accuracy of 5.06%, 5.13%, and 6.65% higher for Lasso-, SR-, and PCA-Logistic models.