Analysis of Evidential Deep Learning for Winter Precipitation Prediction

Winter precipitation hazards, including rain, snow, freezing rain, and sleet, significantly impact human safety and transportation. The winter precipitation type (p-type) machine learning model aims to improve the classification of these p-types, which have historically been inaccurately predicted. This evidential model incorporates uncertainty estimation directly into its predictions, unlike traditional methods that derive uncertainty from ensembles of models. This provides both forecasters and researchers insight into the confidence of the model’s prediction and allows for more informed decision making. Forecasters currently rely on extensive model guidance for weather predictions. This project aims to understand how they evaluate and trust new machine learning-based guidance, addressing the need for transparency in processes and inputs. This work involves comprehensive analysis of the model's outputs, comparing its performance against Numerical Weather Prediction models, and investigating failure modes through case studies and real-time data comparisons. By enhancing the p-type model with hyperparameter optimization and quality control of mPING crowdsourced training data, we also aim to refine the model's accuracy and reliability.