An Investigation of Two Machine Learning Radar-Based Hail Discrimination Algorithms

Machine Learning (ML) is being relied upon more heavily as software packages to develop ML algorithms become more readily available. Often, different ML approaches perform similarly, but depend upon different predictors to do the job. This paper focuses on two ML hail size discrimination algorithms: one based on neural nets and one based on a random forest. Both models use the same set of predictors derived from WSR-88D Dual-Pol radar data only; all cases include polarimetric variables, but no environmental data are used. The performance between the two models are compared but so are the importance of the predictors within each model. This is done using some of the newer techniques recently developed within the machine learning community. Implications about how well we can expect to do with existing data sets are made and additional information sources proposed.