Exploring the Application of Machine Learning to Identification of Storm Objects

As new methodologies are sought to potentially improve the accuracy of the segmentation of independent storm objects from radar data, the wealth of current research into machine learning applications should be carefully considered to see if it offers insights or solutions to the problem. If applicable, machine learning algorithms hold the promise of providing storm segmentation results that are more efficiently computed and are more useful to researchers and decision makers than the current state of the art.

The current storm segmentation method developed as part of the Multi Radar Multi Sensor (MRMS) system at the National Severe Storms Laboratory has used a k-means clustering algorithm combined with watershed-based region selection to provide effective results for many years. However, this method, as well as other widely-used methods such as the dual thresholding approach used in the Thunderstorm Identification, Tracking, Analysis and Nowcasting (TITAN) software suite from the National Center for Atmospheric Research, can sometimes suffer from results that do not correspond well with human perception of storm objects. In the case of the segmentation method within MRMS, this is often due to the inconsistent identification of objects across time. This causes challenging problems for the broad range of activities that depend upon storm object identification, including tracking of the objects across time, analysis of storm cell characteristics, and intuitive display of storm information to end users.

This work investigates the feasibility of applying machine learning methods to the problem of segmenting meteorological data into independent storm objects. This involves consideration of how machine learning can be applied to the problem, which machine learning algorithms might be applicable, whether or not existing data sources are sufficient for the training and calibration of machine learning algorithms, and whether or not the required data processing will likely be possible in real-time operations. Also presented is a brief review of current state of the art methods as well as other new storm segmentation methods under consideration. The future work that will be required to compare these methods against each other and validate their results is discussed.