Using Deep Learning with Polarimetric Radar Data to Predict Hail Size

During the summers from 2006 through 2015, CIMMS and NSSL conducted the Severe Hazards Analysis and Verification Experiment (SHAVE). The goal for SHAVE was to primarily collect hail reports at high spatial resolution following thunderstorms. This was completed by using student telephone operators who monitored Multi-Radar, Multi-Sensor (MRMS) outputs on a web-enabled data entry page. The project collected over 54,000 hail reports during the 10 years of operations. Cases exhibiting good spatial resolution have been collected and radar data, including MRMS data, have been processed. In total, 735 cases were processed, with 427 polarimetric cases yielding around 20,000 hail reports, and SHAVE reports were paired to both single-radar and MRMS data.

The Maximum Estimated Size of Hail (MESH) algorithm is an operational hail size estimation algorithm used with both for single-radar and MRMS data; it uses a vertical integration of reflectivity weighted by both height relative to the melting level and reflectivity magnitude. However, MESH, despite its name, is not meant to be a 1-to-1 estimate of the absolute maximum hail size expected. Rather, the MESH output is meant to represent the 75th percentile of observed hail diameters (i.e., ~75% of hail stones will be smaller than the MESH value). This 75% threshold is roughly accurate for both single-radar and MRMS applications. However, anecdotally, MESH is often used as a 1-to-1 estimate of maximum hail size, which then leads to mixed impressions regarding the accuracy of the algorithm when verification information is available. There is a need for a 1-to-1 hail sizing algorithm, however, vertical integrations of reflectivity and curve fitting have shown their limits in providing accurate size estimates in large part due to significant overlap of reflectivity distributions for even general hail size categories.

Recently, an effort has started at CIMMS and NSSL to apply a range of machine learning techniques to the SHAVE data in an effort to improve hail size estimation, both in terms of detection and prediction. This presentation will focus on the study's neural networks being developed. Characteristics of polarimetric variables in small neighborhoods surrounding reports can be used to train a fully-connected, deep neural network to predict hail size. The spatial density of SHAVE reports allows for the reports to be gridded to verification images, thus fully convolutional neural networks can be trained to provide spatial predictions of hail size, where the polarimetric variables and other information, such as distance from the radar and height of the melting level, can serve as different channels of an image for use in the network. Comparisons to other machine learning models will be included.