2B.2 Real-Time and Climatological Storm Classification through Deep Learning

Monday, 7 January 2019: 11:00 AM
North 125AB (Phoenix Convention Center - West and North Buildings)
G. Eli Jergensen, University of Oklahoma, Norman, OK; and A. McGovern, H. Obermeier, and T. Smith

Classification of storm type assists in weather prediction, particularly in determining features such as storm duration, precipitation, and damage potential. Storm classification also enables the analysis of long-term trends in weather patterns. Using characteristics drawn from radar data provided by the Multi-Year Reanalysis of Remotely Sensed Storms (MYRORSS) and environmental data, our goal is to develop an automated storm classification algorithm using machine learning techniques (specifically deep learning) that can classify both in real-time to aid forecasters and to accurately build a climatology of storms across the continental United States.

This talk considers two primary models of interest, a fully connected Artificial Neural Network (ANN), and a Convolutional Neural Network (CNN). While both are trained over the environmental data, they differ in how they receive the MYRORSS. Since ANNs are ill-equipped to handle 2-dimensional data (like radar images), we train ours over the statistics of the radar images (min, mean, max, 5th-, 25th-, 75th-, and 95th-percentiles, skewness, and kurtosis), which are extracted beforehand. The CNN, which is particularly well-suited to 2D data, is provided with the raw radar images. To provide comparisions with other machine learning methods, we compare to previous work on Support Vector Machines (SVMs) and Decision Trees, both Random Forest and Gradient Boosted. All models are evaluated in terms of both accuracy and skill, measured with the Peirce Skill Score. Additionally, we provide insight into the model by determining the most important variables.

Preliminary results on Artificial Neural Networks are highly suggestive that ANNs perform at least as well as SVMs and Gradient Boosted Decision Trees, achieving accuracies of roughly 0.70 and Peirce scores of approximately 0.40. While we presently have no quantitative results for CNNs, the similarity of ANNs and CNNs in the last few layers is supportive that CNNs should at least perform as well as ANNs.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner