Developing a Machine Learning–Based Hail Climatology Using the SHAVE and MYRORSS Databases

Hail climatologies have traditionally been created using hail reports collected by the National Weather Service (i.e., Storm Data), however, previous work has shown that hail reports contain inaccuracies in reporting locations and hail diameters, additionally areas with low population densities will have inaccurate climatologies due to a lack of reporting. Remote sensing-based products, like the Multi-Radar Multi-Sensor (MRMS) Maximum Expected Size of Hail (MESH), give the ability to generate climatologies in a more objective manner. A previous climatology using MESH simply used MESH thresholds to define areas of any-sized hail and severe-sized hail. This work will explore using additional MRMS output, such as the azimuthal shear product, and using 1-hour hail swaths as inputs to neural networks to generate a CONUS-wide hail climatology.

To train the networks, 735 cases from the Severe Hazards Analysis and Verification Experiment (SHAVE) have been analyzed. For each case, nearby radar data has been processed through the MRMS framework. SHAVE hail reports have spacings typically near 2 km and include all hail sizes, including reports of no hail and of non-severe sizes that are typically unavailable from Storm Data. The more complete hail size range and the dense spacing of the SHAVE reports allow for the reports to be gridded and a verification image, rather than just simply a few points, can be generated. Deep, fully-connected neural networks and fully convolutional neural networks will be explored. Network predictions for exact hail size and hail size category will be explored. The results of the networks will be applied to the Multi-Year Reanalysis of Remotely Sensed Storms (MYRORSS), a database of MRMS data that has been completed for the years 1998 through 2011, to develop a hail climatology, which can be compared to previous MRMS-based hail climatologies that primarily only used MESH as the input.