Abstract: Utilizing ERA5 reanalysis data and surface meteorological observations from China for the period 2000–2019, this study examines the spatiotemporal distribution characteristics of winter precipitation phases across the complex terrain of Hunan and Hubei provinces. After optimizing key physical factors, several machine learning models were developed, including two lightweight gradient-boosting machine models with data resampling techniques (ADASYN-LGBM and Hybrid-LGBM) and a fully connected multilayer perceptron incorporating the Focal Loss function (FocalLoss-MLP). These individual models were then integrated into an ensemble classifier (Ensemble) using a soft-voting strategy to enhance classification accuracy and robustness under severe class imbalance. The results reveal that the spatial distribution of winter precipitation phases in Hunan and Hubei is strongly influenced by local topography, atmospheric circulation patterns, and regional climate conditions. Notably, the frequency of freezing rain in Hunan is significantly higher than in Hubei, primarily due to topographical blocking by the Nanling Mountains and the frequent confluence of cold and warm air masses. Diurnal variations also influence the distribution of precipitation phases. Feature importance analysis indicates that 2 m temperature, 0°C layer height, and atmospheric thickness are dominant factors in model discrimination, while the diurnal binary factor and latitude contribute substantially as auxiliary variables. All four models achieved the best discrimination performance for the dominant rain category, followed by snow, freezing rain, and sleet, suggesting that model performance is highly dependent on sample size. However, for sleet, which accounted for only 2.2% of the samples, the test set threat score values ranged from 8% to 18%. This relatively low performance is linked to the overlapping meteorological characteristics and blurred boundaries during the rain–snow phase transition. Specifically, the Ensemble model compensates for the limitations of single models in identifying specific precipitation phases, thereby further improving overall identification accuracy. Typical case studies confirm its superior performance in finely characterizing the rain–snow boundary and accurately identifying freezing rain.