Machine Learning Techniques for Radar-Based Hail Size Prediction

Previous work has investigated the usefulness of radar-based reflectivity profiles with machine learning models to predict hail size. These investigations have examined one or two similar datasets and a limited selection of machine learning models to find the best solution. This work takes several datasets of reflectivity profiles with derived radar products to find the best subset of data to be used for training. The data used includes verification from the Severe Hazards Analysis and Verification Experiment (SHAVE) with different sources of reflectivity profiles. The reflectivity profiles used are derived from a single radar or using merged radar data completed with the Multi-Radar Multi-Sensor (MRMS) framework. These merged profiles contain reflectivity values at specific isotherms, the height of the isotherms, and reflectivity at heights of 1 to 20 km MSL. In addition to the reflectivity profiles, radar derived variables including low- and mid-level azimuthal shear, vertically integrated liquid (VIL) and maximum expected size of hail (MESH) are incorporated as well near storm environmental data. For the single radar profiles, polarimetric variables are included in some training to find any additional skill in hail size prediction.

Machine learning models investigated include simple fully connected neural networks, deep neural networks, deep neural networks with skip layers, gradient boosted decision trees, one-dimensional convolutional neural network, and categorical classification using these networks. For each dataset, the best performing hyperparameters were chosen for each model and the best were compared to the other datasets. This presentation shows using similarly structured machine learning models and how different datasets used for training the models perform in predicting hail sizes. This allows the best MRMS model to be implemented on larger datasets such as the Multi-Year Reanalysis of Remotely Sensed Storms (MYRORSS), an MRMS database for 1998-2011, and can be used to a new radar-based hail climatology.