The Application of a Clustering Algorithm to Identify and Measure Tornado Swarms

Tornado clusters and outbreaks are some of the most well-known and destructive weather phenomenon that wreak havoc across the United States. More widespread than any single tornado, such tornado clusters, or swarms, can impact anywhere from a vast section of the country to a single isolated region. The de-facto standard for measuring individual tornadoes is based on the Enhanced Fujita (EF) Scale, but a similar standard for measuring tornado clusters has not yet been established.

In this study an algorithmic approach to identify and measure tornado clusters is introduced. Tornado segments from NOAA’s Storm Data database are broken down into points based on tornado longevity. These points are then clustered using DBSCAN, a density-based clustering algorithm, adapted for spatial-temporal data. The measure of each tornado cluster is quantified by the total tornado time in each cluster.

Results shown include those from the tornado outbreaks on 27 April 2011 and 3 May 1999. The algorithm successfully extracts the three distinct rounds of tornadoes in the mid-south throughout the day on 27 April 2011. On 3 May 1999 the algorithm separates the tornadoes in Oklahoma from other, less well-known tornadoes that touched down in Kansas and Nebraska. In these examples, the method has shown utility in not just measuring tornado clusters, but can also highlight different clusters that occur within the context of a larger tornado outbreak.

By measuring and highlighting tornado clusters, it is hoped a climatology of clusters can be developed independently of the established climatology of individual tornado touchdowns. This new climatology and metric for tornado clusters can then be used to more accurately determine synoptic-scale or seasonal patterns that favor instances of tornado swarms in various regions of the United States.