4.3A Anticipating the formation of tornadoes through data mining Amy McGovern Derek H. Rosendahl School of Computer Science School of Meteorology University of Oklahoma University of Oklahoma Norman, OK Norman, OK [email protected] [email protected] Adrianna Kruger Meredith G. Beaton School of Computer Science School of Computer Science University of Oklahoma University of Oklahoma Norman, OK Norman, OK [email protected] [email protected] Rodger A. Brown Kelvin K. Droegemeier NOAA School of Meteorology National Severe Storms Laboratory University of Oklahoma Norman, OK Norman, OK [email protected] [email protected] 1. Introduction we develop new data mining techniques (computer science) to understand the data and analyze the re- sults (meteorology). This becomes a cycle where Severe weather phenomena such as tornados, thun- the results inform new techniques and new tech- derstorms, hail, and floods, annually cause signifi- niques produce new results. The results presented cant loss of life, property and crop destruction, and in this paper represent the beginning of this re- disruption of the transportation systems. The annual search. economic impact of these phenomena is estimated to be greater than 13 billion dollars (Pielke and Car- bone 2002). Any mitigation of the effects of these storms would be beneficial. 2. Meteorological Data We propose to enhance our understanding of the formation of severe weather events, specif- With our goal of improving the detection and antic- ically focusing on tornadoes, through data min- ipation of tornadoes, we are not taking the tradi- ing/knowledge discovery techniques. The process of tional route of examining radar reflectivity and ra- knowledge discovery is about making sense of data. dial velocity gathered by a specific radar system. Generally, the data are too complex for humans to These radar systems may be limited in their ability quickly identify and understand the important pat- to detect and anticipate tornadoes due to inherent terns. Instead, knowledge discovery techniques can radar characteristics such as a beam increasing in be used to highlight salient patterns. We are devel- altitude and spreading as it travels away from the oping new data mining techniques for use on storm- radar causing the resolution volume to be too large scale weather data. to accurately identify tornadic circulations (e.g. Don- The long-term goals of our work are to improve aldson 1970; Brown et al. 1978; Wood and Brown the understanding of tornado formation to the point 1997). Additionally, analyzing only radar reflectiv- where refined detection and prediction algorithms ity and radial velocity provides a limited number of can be created. We will do this by engaging in an in- variables to use in depicting the true state of the at- terdisciplinary knowledge discovery process where mosphere. We therefore make use of new research 1 on data assimilation that will eventually enable us through 4 km. We also used a sounding from the to predict and detect severe weather on a real-time 20 May 1977 Del City, Oklahoma tornado (Ray et al. three-dimensional gridded data set containing all the 1981) fundamental and derived meteorological quantities. Examples of the fundamental quantities include the 2a. Data Extraction three components of air motion, temperature, pres- sure, and precipitation. Examples of the derived vari- ables include divergence, temperature gradient, ver- Our eventual goal is to examine the data using high- tical vorticity, and the pressure gradient force. The level features such as a hook echo, gust front, rear ability to have all fundamental and derived meteo- flank downdraft, occlusion downdrafts, etc. However, rological quantities at all grid points results in an im- automated identification and tracking of these high- proved and significantly broadened representation of level features is computationally difficult. Identifying the atmosphere. This new representation necessi- them requires creating a description that a majority tates the development of more sophisticated detec- of meteorologists will agree on or at least creating a tion and anticipation techniques. large enough labeled data set such that a machine learning algorithm could learn to extract these fea- Figure 1 gives a sample of the gridded data cre- tures. We will be addressing these issues in future ated using an ensemble Kalman filter assimilation of work. real observations (Tong and Xue 2005). The top two panels display the observed reflectivity and Doppler For the results in this paper, we chose to extract wind measurements in the May 29, 2004 tornado a set of 24 fundamental and derived meteorologi- in Oklahoma City. The remaining panels show as- cal quantities. These quantities are listed in Table similated radar reflectivity and retrieved temperature, 1. These represent the most important meteorolog- pressure, and vertical vorticity. ical quantities and enable us to observe each storm Because the technology to create real-time three- with a significant reduction in data size over examin- dimensional gridded data from actual observations ing each variable at each grid point. is currently under development, we are using simu- Any given storm simulation may generate several lated storm data produced from the Advanced Re- separate storm cells. We define a storm cell based gional Prediction System (ARPS), which is a three- on the maximum updraft. The algorithm for identi- dimensional, nonhydrostatic model that is one of the fying and tracking individual storms cells is given in top weather forecasting systems for mesoscale data Table 2. At each time step, we identify the domain- (Xue et al. 2000, 2001, 2003). The computational wide maximum vertical wind speed at 4 km height. grid for our study has a horizontal spacing of 0.5 km Once a single storm is being tracked, new storms within a 100 km by 100 km by 20 km domain. There will not be identified until their maximum vertical wind are 49 levels in the vertical with a stretched grid from speed exceeds that of the main storm. A 20 km by 50 m at the ground to 750 m near the top of the do- 20 km full height box is drawn around the location main. The model is run for three hours with history of the maximum vertical wind speed at 4 km height files saved every 30 seconds. for each storm. We then measure the maximum and Soundings are used to initialize horizontally ho- minimum of each quantity listed in Table 1 within the mogeneous base state environments wherein a box. We measure the maximums and minimums for thermal bubble initiates convection. The thermo- each variable for the surface to 2 km in height and dynamic profiles of the soundings are analytically then from 2 km to the top. For some variables, we constructed using the Weisman and Klemp (1982) also store the maximum and minimum values at the method with variations in the surface mixing ratios. surface. This allows us to identify whether a maxi- The hodographs have variations in magnitude and mum or minimum value is associated with a surface, shape similar to Adlerman and Droegemeier (2005) low, or mid to upper level feature. (e.g. half circle, quarter circle, quarter circle with tail, and straight). Only supercell storms are studied In current work, we are examining the use of in this research and therefore indices such as 0-6 a modified Storm Cell Identification and Tracking km Bulk Richardson Number (Weisman and Klemp (SCIT) algorithm (Johnson et al. 1998) for improved 1982) and storm-relative helicity (Davies-Jones et al. storm identification and tracking. The original SCIT 1990) are used to identify suitable soundings. Our algorithm used reflectivity to identify and track each preliminary results used soundings having surface storm. Because we do not have to depend only on mixing ratios of 13, 14, 15, 16, and 17 g kg−1 with reflectivity, we instead use discrete areas of signifi- a half circle hodograph of radius 10 m s−1 turning cant updrafts and track around each updraft. 2 Observed Radar Reflectivity Observed Radar Radial Velocity Assimilated Radar Reflectivity Retrieved Pressure Retrieved Temperature Retrieved Vorticity Figure 1: An example of real data (top panels) being used to create gridded data (center and bottom panels). This example is from the May 29, 2004 tornado in Oklahoma City and is courtesy of Fritchie and Droegemeier at the University of Oklahoma. 3 Table 1: Maximum and minimum quantities extracted for each storm cell. The bars represent averages. Variable Equation Units vertical velocity w m s−1 r 2 2 ∆hw ∆hw −1 vertical velocity horizontal gradient ∆x + ∆y s ∆hv ∆hu −1 vertical vorticity ∆x − ∆y s rainwater mixing ratio qr kg kg−1 r 2 2 ∆hqr ∆hqr −1 −1 rainwater mixing ratio horizontal gradient ∆x + ∆y kg kg m r 2 ∆v qr −1 −1 rainwater mixing ratio vertical gradient ∆z kg kg m perturbation potential temperature pt - ptbar K pressure perturbation (p’) p - pbar Pa 0 −1 ∆v p 2 vertical perturbation pressure gradient force rho ∗ ∆z m s ∆hu ∆hv − horizontal divergence ∆x + ∆y s 1 hail mixing ratio qh kg kg−1 r 2 2 ∆hqh ∆hqh −1 −1 hail mixing ratio horizontal gradient ∆x + ∆y kg kg m r 2 ∆hqh −1 −1 hail mixing ratio vertical gradient ∆z kg kg m √ horizontal wind speed u2 + v2 m s−1 ∆hv ∆hu ∆hu ∆hv −1 vertical stretching − ∆x − ∆y ∆x + ∆y s ∆hw ∆hu ∆hw ∆hv −1 tilting term ∆y ∗ ∆z − ∆x ∗ ∆z s 1 ∆h(rho) ∆hp ∆h(rho) ∆hp −1 baroclinic generation (rho)2 ∗ ∆x ∗ ∆y − ∆y ∗ ∆x s Σ(wij −ww>1)(ζij −ζw>1) vertical velocity (w) and vertical vorticity (ζ) √ p correlation Σ(wij −ww>1) Σ(ζij −ζw>1) r 2 2 ∆hpt ∆hpt −1 horizontal potential temperature gradient ∆x + ∆y K m radar reflectivity ref = 10 ∗ LOG10 (Zerain + Zesnow + Zehail) dBZ r 2 2 ∆h(ref) ∆h(ref) −1 radar reflectivity horizontal gradient ∆x + ∆y dBZ m r 2 ∆v (ref) −1 radar reflectivity vertical gradient ∆z dBZ m horizontal Laplacian of radar reflectivity ∇2(ref) dBZ m−2 2 2 Σ(wij −ww>1)(∇ (ref)ij −∇ (ref) ) vertical velocity and horizontal Laplacian of w>1 √ q 2 2 2 2 radar reflectivity correlation Σ(wij −ww>1) Σ(∇ (ref)ij −∇ (ref)w>1) 4 Table 2: Summary of how we identify individual storm cells and extract the maximum and minimum quanti- ties from each storm.