Machine Learning Methods to Infer Precipitation Phase from Temperature and Moisture Profiles

The phase of falling precipitation can have a large societal impact for both hydrology (snow storage, rain-on-snow events), meteorology (snowstorms, freezing rain) and climate (snow albedo feedback). In Canada, many surface weather stations report precipitation information in the form of total precipitation (liquid-equivalent), but very few weather stations directly report snow. Thus, precipitation phase must be inferred from ancillary data such as temperature and moisture. Each scientific community has developed its own tool for the determination phase in the absence of direct observations: from simple rules based on air temperature, dew point temperature or wet bulb temperature to sophisticated microphysics schemes passing by methods based on the discrimination of features extracted from vertical temperature profiles. With the recent advances of machine learning, there is an opportunity to investigate another set of methods based on deep neural networks. Using ERA5 and ERA5-Land model re-analyses as the reference, we trained several recurrent neural networks (RNN) on vertical profiles of temperature and moisture to infer the snow fraction – the ratio of solid precipitation to total precipitation. Since precipitation phase (solid, liquid or mixed) was not directly available in the model re-analysis, we defined it using two thresholds: snow fraction of less than 5% for liquid, snow fraction of more than 95% for solid phase, and mixed phase for everything in between. The best performing neural network for regressing snow fraction is found to be a Gated Recurrent Unit (GRU) RNN using profiles up to 500 hPa above the surface of both temperature and relative humidity. A slight decrease in performance is observed if profiles up to 700 hPa are used instead. A feature experiment also reveals that the performance is significantly better when using both temperature and moisture profiles, but it does not really matter what type of moisture observations are used (either dew point spread, wet bulb temperature or relative humidity). For classifying precipitation phase, the balanced accuracy is over 90%, clearly outperforming the implementation of Bourgouin’s method used operationally in part of Canada. Compared with the K-Nearest Neighborhood (KNN) method trained on surface observations only, it is seen that the greatest gain in performance for GRU-RNN is when the surface temperature is close to zero degrees Celsius. These preliminary results indicate the great potential of the proposed algorithm for determining snow fraction and precipitation phase in the absence of direct observations. The proposed algorithm could potentially be used for inferring snow fraction and precipitation phase in several applications such as (1) precipitation analysis for forcing hydrological models, (2) weather nowcasting, (3) weather forecast post-processing and (4) climate change impact studies.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.