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

Area of Science:

• Atmospheric Science
• Meteorology
• Climatology

Background:

• Partitioning precipitation into rain and snow using near-surface meteorology is a persistent challenge.
• The performance limits of current precipitation phase partitioning methods are not well understood.

Purpose of the Study:

• To evaluate the performance limits of precipitation phase partitioning methods.
• To compare benchmark methods with machine learning (ML) models using extensive datasets.
• To identify factors limiting the accuracy of precipitation phase determination.

Main Methods:

• Applied benchmark precipitation phase partitioning methods and three ML models (artificial neural network, random forest, XGBoost).
• Utilized two independent datasets: 38.5 thousand crowdsourced observations and 17.8 million synoptic meteorology reports.
• Analyzed the relationship between air temperature overlap and partitioning accuracy.

Main Results:

• ML methods provided negligible improvements (up to 0.6% accuracy increase) over best benchmarks.
• ML models showed limited ability to identify mixed precipitation and sub-freezing rain, with worst performance between 1.0°C-2.5°C.
• Significant negative relationship found between air temperature overlap (1.0°C-1.6°C) and partitioning accuracy (p < 0.0005).

Conclusions:

• Near-surface meteorology imposes inherent limitations on precipitation phase partitioning accuracy.
• Focus should shift from improving existing methods to incorporating novel data sources like crowdsourced observations.
• New approaches are needed to overcome the challenges posed by similar meteorological conditions for rain and snow near freezing temperatures.