Evaluation of Machine Learning Techniques for Precipitation Type Forecasting

In this study, we use an artificial convolutional neural network to produce probabilistic precipitation categorizations based on atmospheric profiles from the Global Ensemble Forecast System (GFS). Traditional precipitation classification techniques often rely on a derived set of physical parameters which summarize an atmospheric profile, and classifications are made following a rule-based algorithm. In our approach, advanced machine learning techniques discover natural summary features by learning from an example set where observed precipitation types are matched with the GFS model profiles that produced them. The neural network considers location-dependent climatological precipitation probabilities and model-generated vertical profiles of temperature, wet-bulb temperature, and specific humidity as predictors. Following a series of mathematical variable transformations, it maps those profile inputs into an independent probability distribution over precipitation type labels. Even though the limited spatial and temporal resolution of the GFS allows only a coarse representation of the temperature and specific humidity profiles, the trained network performs very well on rain and snow categories (RA, SN), but also shows skill for traditionally difficult categories such as ice pellets (IP) and freezing rain (FRZA). Finally, model peformance is compared to traditional precipitation type algorithms to assess relative changes in predictability.