366 WEATHER AND FORECASTING VOLUME 29 An Objective Analysis of Tornado Risk in the United States TIMOTHY A. COLEMAN University of Alabama in Huntsville, Huntsville, Alabama P. GRADY DIXON Mississippi State University, Mississippi State, Mississippi (Manuscript received 10 May 2013, in final form 12 September 2013) ABSTRACT In this paper, an objective analysis of spatial tornado risk in the United States is performed, using a somewhat different dataset than in some previous tornado climatologies. The focus is on significant tornadoes because their reporting frequency has remained fairly stable for several decades. Also, data before 1973 are excluded, since those tornadoes were rated after the fact and were often overrated. Tornado pathlength within the vicinity of a grid point is used to show tornado risk, as opposed to tornado days or the total number of reported tornadoes. The possibility that many tornadoes in the Great Plains were underrated due to the lack of damage indicators, causing a low bias in the number of signifi- cant tornadoes there, is mostly discounted through several analyses. The kernel density analysis of 1973– 2011 significant tornadoes performed herein shows that the area of highest risk for tornadoes in the United States extends roughly from Oklahoma to Tennessee and northwestern Georgia, with the highest risk in the southeastern United States, from central Arkansas across most of Mississippi and northern Alabama. 1. Introduction Edwards 1999), Kelly et al. (1978) reported that the area of the United States with the highest tornado frequency Debates have been ongoing about the location of the was the Great Plains, or the classic tornado alley [as most tornado-prone regions of the United States over mapped by Schaefer et al. (1980) and from McNulty the past 10 yr or so. For many years, ‘‘tornado alley,’’ (1995); Fig. 1a]. Brooks et al. (2003), using tornado days a region in the Great Plains extending from northern and limiting the risk areas to those with tornado seasons Texas to Iowa, was considered to be the area with the that are consistent by time of year, suggested that the greatest tornado risk in the country. However, the idea primary tornado alley in the United States extends from of another local maximum of tornadoes in the south- Texas to North Dakota, similar to and slightly northwest eastern United States was first mentioned by Alan of the classic tornado alley (Fig. 1b). Concannon et al. Pearson in 1971 (Gagan et al. 2010), and further shown to (2000) examined days with significant tornadoes [those some degree by Concannon et al. (2000), Ashley (2007), rated on the Fujita scale as category 2 (F2) or higher] Dixon et al. (2011), Smith et al. (2012),andKellogg and from 1921 through 1995, and found an L shaped region Forbes (2013). of maximum tornado risk that included most of the Researchers have also used various ways of defining classic tornado alley from Schaefer et al. (1980), then tornado-prone areas and, by implication, tornado risk. extended eastward through Arkansas, Mississippi, and Using data from what eventually became the Storm Pre- Alabama (Fig. 1c). Smith et al. (2012),usingamore diction Center (SPC) tornado database (e.g., Schaefer and limited dataset (2003–11) including tornadoes of all intensities, suggest that the southern part of the tornado-proneareashownbyConcannon et al. (2000), Corresponding author address: Tim Coleman, Dept. of Atmo- spheric Science, University of Alabama in Huntsville, NSSTC, 320 from Oklahoma to Alabama, has the highest tornado Sparkman Dr., Huntsville, AL 35805. risk, with a maximum over parts of Mississippi and E-mail: [email protected] Alabama. However, this small dataset was probably DOI: 10.1175/WAF-D-13-00057.1 Ó 2014 American Meteorological Society Unauthenticated | Downloaded 09/30/21 02:21 PM UTC APRIL 2014 C O L E M A N A N D D I X O N 367 FIG. 1. (a) Annual tornado frequency for all tornadoes from Schaefer et al. (1980), shown in McNulty (1995), based on records before 1980. (b) Tornado occurrence based on frequency of all tornadoes, 1980–99, combined with a measure of seasonality of tornadoes (Brooks et al. 2003). (c) Number of days per century with one or more sigtors 1921–95 (Concannon et al. 2000). (d) Count of all tornado events, 2003–11 (Smith et al. 2012). (e) Number of sigtors 1950–2004 (Ashley 2007). (f) Number of days with sigtor per century, 1961–2010 (Carbin et al. 2012). overly influenced by the 27 April 2011 outbreak (Fig. 1d). There are several issues that have not been fully Ashley (2007), using only significant tornadoes during addressed in prior tornado climatology studies. First of the years 1950–2004, and Carbin et al. (2012), examining all, the Fujita scale was not published until 1971 (Fujita the return frequency of significant tornadoes from 1961 1971; Edwards et al. 2010), and was not adopted for through 2010, also showed that the region from Okla- rating tornadoes in their near-immediate aftermath in the homa to Alabama has the highest tornado risk (Figs. 1e United States until 1973. So, tornadoes before this time and 1f). were rated years after the fact in the SPC database, using Unauthenticated | Downloaded 09/30/21 02:21 PM UTC 368 WEATHER AND FORECASTING VOLUME 29 newspaper stories and photographs that likely over- tropical cyclone–related tornadoes, which tend to be emphasized the damage, causing overrating of some weak [EF0 or EF1; e.g., Schultz and Cecil (2009)]. Maps tornadoes prior to 1973 (e.g., Schaefer and Edwards showing the smoothed average annual pathlength of 1999; Brooks and Craven 2002; Schaefer and Schneider tornadoes within 40 km (25 mi) of a point were gener- 2002; Anderson et al. 2007). Also, there has been ated using kernel density estimation (KDE, e.g., a steady increase in annual reported tornadoes in the O’Sullivan and Unwin 2003). United States, even since 1973, due to better spotting KDE is a very common method used in spatial anal- techniques, National Weather Service (NWS) warning ysis and has been applied to tornado risk in several verification procedures, and improved detection by ra- previous papers (e.g., Brooks et al. 2003; Dixon et al. dar. This overall increase has primarily been due to an 2011; Smith et al. 2012; Marsh and Brooks 2012). As increase in weak tornadoes (F0 or F1, hereinafter re- discussed in detail by Dixon et al. (2011), KDE analysis ferred to as weaktors). However, the reporting of sig- implies that spatial patterns have magnitudes at every nificant tornadoes (F2 or greater, hereafter referred to given point, as opposed to only the places directly af- as sigtors) has remained relatively stable since 1973. fected (in this case, by tornadoes) (e.g., O’Sullivan and Additionally, as seen from the results of various re- Unwin 2003). The KDE method used in this study em- searchers in Fig. 1, most tornado climatologies have ploys the Epanechnikov quadratic kernel probability used either tornado days or the number of tornadoes as density function as calculated by ESRI ArcGIS Spatial the basis for their studies, not utilizing pathlength. Analyst software (Silverman 1986; de Smith et al. 2007; However, pathlength seems more desirable because it is Dixon et al. 2011; Marsh and Brooks 2012; Smith et al. more proportional to the area impacted by tornadoes 2012). There are several other types (i.e., shapes) of than tornado numbers or tornado days. kernel functions, but Dixon and Mercer (2012) shows The purpose of this research is to further our under- that the differences in spatial patterns due to kernel standing of the regions of greatest tornado risk in the shape are negligible compared to those related to the United States by addressing the above issues in an ob- radius of the kernel, which is consistent with previous re- jective analysis. Section 2 will discuss the data and search (Silverman 1986; de Smith et al. 2007). Through- methodology used in the study. Section 3 will examine out this study, a kernel radius of 250 km is used because of some trends in tornado reporting and discuss the issues the relatively short study period of the current analysis. mentioned above, including the importance of path- The steps we used in performing the KDE analysis were length in assessing tornado risk. This information will as follow: 1) The 250-km probability density function help remove some biases that are present in other studies, (KDE) was applied to each tornado path on a 1-km2 producing an objective analysis of tornado risk in the grid. This results in each 1-km2 grid cell within 250 km of United States in section 4. Section 5 contains a summary the tornado path in question receiving a portion of the and conclusions. total tornado path (according to its distance from the path). 2) The sum of each point’s fractional pathlengths (measured in kilometers of tornado path per square kilo- 2. Data collection and methodology meter) for all tornadoes to affect that grid cell was calcu- The primary data source for this study is the SPC lated on a 5-km output grid. So, while only 1 out of every 25 tornado database, which provides the date and time, lo- grids is displayed, this coarser output grid changes the cation, intensity [Fujita or enhanced Fujita (EF) scale], resolution of the map and does not change the numerical pathlength, and other statistics on nearly every tornado results (Dixon et al. 2011). 3) Every cell is then multiplied in the United States for 1950–2011. We focus primarily by the approximate area of a 40-km-radius circle (5024 km2) on data from 1973 through 2011, to eliminate any bias to represent a meaningful risk metric (in kilometers of from tornadoes rated after the fact.