The Southern rain–snow transition zone as a major source of uncertainty in precipitation-phase partitioning across China

Study region China, encompassing strong climatic and topographic gradients from the cold and arid northwestern and high-elevation regions to the warm and humid rain–snow transition zones of southern China. Study focus Reliable rain–snow partitioning is essential for representing snowfall accumulation, snowmelt, and precipitation inputs in hydrological modelling. Using surface observations from 1951 to 1979, we evaluated 12 conventional (Ta_0, Td_0, Tw_1, Trh, Dai08, KS98, Jenn_Uni, Jenn_Bi, Jenn_Tri, Ding14, NaquSig, and Yamazaki) threshold-based methods and six machine-learning models (RF, XGBoost, CatBoost, LightGBM, ExtraTrees, and HGB) across China using the snowfall F1 score (F1_snow), Heidke skill score (HSS), overall accuracy (OA) and macro-averaged F1 score (F1_macro). Their sensitivities to temperature (−10–10 °C), relative humidity (0–1), and elevation (0–5 km) were examined. Station-network thinning experiments were further conducted to assess sensitivity to uneven observational coverage, and model transferability was independently evaluated against NOAA Global Surface Summary of the Day (GSOD) observations for 1980–2024 without retraining. New hydrological insights for the region Differences among precipitation-phase methods were expressed much more clearly in snowfall identification than in overall accuracy. Nationwide F1_snow ranged from 0.837 to 0.891 for conventional methods and 0.921–0.929 for machine-learning models, whereas OA remained within 0.970–0.982 and 0.987–0.989, respectively. The largest classification uncertainty occurred within the major rain–snow transition belt at approximately 22°–32°N and under three critical environmental conditions: temperatures of 0–5 °C, RH < 0.4, and elevations of 2500–4000 m. Under these conditions, machine-learning models achieved F1_snow values of 0.587–0.619, 0.832–0.890, and 0.940–0.952, compared with conventional-method means of 0.421, 0.753, and 0.750, respectively. Station thinning produced mostly < ±2% bias in F1_snow, indicating limited sensitivity of the main spatial patterns to uneven station coverage. When transferred directly to GSOD for 1980–2024, the machine-learning models retained F1_snow values of approximately 0.90, demonstrating useful cross-period and cross-dataset applicability, although their relative advantage over conventional methods was reduced.