Machine learning shows a limit to rain-snow partitioning accuracy when using near-surface meteorology

Partitioning precipitation into rain and snow with near-surface meteorology is a well-known challenge. However, whether a limit exists to its potential performance remains unknown. Here, we evaluate this possibility by applying a set of benchmark precipitation phase partitioning methods plus three machine learning (ML) models (an artificial neural network, random forest, and XGBoost) to two independent datasets: 38.5 thousand crowdsourced observations and 17.8 million synoptic meteorology reports. The ML methods provide negligible improvements over the best benchmarks, increasing accuracy only by up to 0.6% and reducing rain and snow biases by up to -4.7%. ML methods fail to identify mixed precipitation and sub-freezing rainfall events, while expressing their worst accuracy values from 1.0 °C-2.5 °C. A potential cause of these shortcomings is the air temperature overlap in rain and snow distributions (peaking between 1.0 °C-1.6 °C), which expresses a significant negative relationship (p < 0.0005) with partitioning accuracy. Thus, the meteorological characteristics of rain and snow are similar at air temperatures slightly above freezing with increasing overlap associated with decreasing performance. We suggest researchers switch their focus from marginally improving inherently limited precipitation phase partitioning methods using near-surface meteorology to creating new methods that assimilate novel data sources-e.g., crowdsourced precipitation phase observations.