A Machine Learning Approach to Smoothing Tornado Climatology Maps

Tornadoes rank among the most dangerous weather events for human safety. In 2023 alone tornadoes caused 91 deaths and 955 injuries. As cities grow and population density continues to increase, a robust understanding of tornado frequency across the country is needed to assist with urban planning and risk management. The historical tornado record like many historical records for severe weather has several limitations. For example, spatial coverage is incomplete, and many tornadoes were missed at the beginning of the record because of large areas with low population density and limited radar coverage. To produce a more complete spatial map of frequency, this study implements a Gaussian Process Regression (GPR) model using observed tornado record from 1950 to 2022 as training data. Previously, studies have applied inflexible smoothing techniques such as Gaussian kernels of fixed bandwidth to produce smoothed risk maps. In contrast, GPR provides a methodology that allows both local and global structure to be preserved and helps to fill in missing data points by constraining Bayesian fitted functions using observed data. GPR is also specifically skilled at working with nonlinear functions like the tornado dataset where there are specific localized hotspots of high tornado activity. Finally, GPR provides estimates of the uncertainty of the model and provides low uncertainty at points of known data with higher uncertainty at points where data may be missing. The model is optimized through choice of both the internal covariance function and inclusion of a noise kernel, with hyperparameter tuning to accommodate influence scales of length. Through use environmental co-variates in addition to the original observations we are able to better constrain the uncertainty of observed frequency in lower certainty regions. These results highlight known hotspots of tornado frequency, while also revealing local scale structures that would be obscured using prior techniques. The smoothed counts will be used to develop a statistical model that defines tornado frequency across the country. This model will objectively determine the area’s most at risk, and when combined with environmental covariates, can produce nonstationary projections of future tornado risk.