Deep Learning for Short-Term Characterization of Tornado Intensity

Tornadoes pose substantial risk to human life and property in the United States, causing upwards of 50 fatalities and $100 million of property damage annually. Previous work on short-term predictions of tornado intensity has discovered impacts on peak tornado intensity from radar-identifiable characteristics such as updraft area and low-level rotational velocity. Additionally, recent studies have suggested that there is some forecasting skill in predicting tornado intensity from near-storm environment data, particularly when it is combined with radar inputs. However, the tremendous amount of data available (e.g. radar, satellite, sounding) presents significant challenges for human operators when attempting to predict tornado intensity, particularly in near-real-time. This work demonstrates a preliminary system for characterizing short-term tornado intensity using a deep learning architecture. Input data consists of the Multi-Radar Multi-Sensor dataset (MRMS) and the Gridded NEXRAD WSR-88D Radar dataset (GridRad), both of which provide composite radar imagery across CONUS. Predictors for the deep learning model include real-time radar imagery and near-storm environment data. We will train separate convolutional neural networks (CNNs) on the MRMS and GridRad datasets to generate predictions of tornado intensity in the form of maximum azimuthal wind shear. We anticipate that these CNNs will be capable of producing physically realistic short-term predictions of tornado intensity. Such predictions could eventually be used by forecasters to characterize tornadoes in real time, enabling higher precision when providing life-saving information in future tornado warnings.