Mapping Rain-Snow Partitioning Accuracy Across China's Diverse Climate Zones: Conventional and Machine Learning Methods

Based on surface observation data from China for the period 1961–1979, this study systematically evaluates the rain/snow partitioning performance of eight conventional physical threshold schemes (Ta, Td, Tw, Ding, Trh, Dai, Sigmoid, and Yamazaki) and six machine learning models (Logistic Regression (LR), Random Forest (RF), Multi-Layer Perceptron (MLP), XGBoost, LightGBM, and CatBoost) across diverse geographical settings and within specific ranges of temperature (−10 to 10 °C), relative humidity (0 to 1), and elevation (0 to 5 km), employing a comprehensive multi-metric evaluation framework that includes the F1 score, Heidke Skill Score (HSS), overall accuracy (OA), and overestimation error (OE).本研究基于1961年至1979年中国地面观测数据，系统评估了八种常规物理阈值方案（Ta、Td、Tw、Ding、Trh、Dai、Sigmoid及Yamazaki）与六种机器学习模型——逻辑回归（Logistic Regression，LR）、随机森林（Random Forest，RF）、多层感知机（Multi-Layer Perceptron，MLP）、XGBoost、LightGBM及CatBoost——在多样地理环境下，以及温度（-10℃至10℃）、相对湿度（0至1）和海拔（0至5千米）的特定区间内的雨雪分离性能，并采用涵盖F1分数、海德克技能得分（Heidke Skill Score，HSS）、总体准确率（Overall Accuracy，OA）与高估误差（Overestimation Error，OE）的多指标综合评估框架开展研究。