SR/SSD 98-23

6-1-98

Technical Attachment

WSR-88D DOPPLER RADAR ADAPTABLE PARAMETER OPTIMIZATION OF THE MESO/TVS ALGORITHM

David Matson

National Weather Service, Little Rock, Arkansas

1. INTRODUCTION

An important mission of the National Weather Service (NWS) is to provide advance notification of severe weather to the public. This goal is achieved, in part, by using the WSR-88D Doppler radar to evaluate the potential for tornadogenesis. Doppler velocity fields are evaluated by the MESO/ TVS (Mesocyclone/Tornado Vortex Signature) algorithm (MTA) to produce decision aids for radar operators who identify areas primed for tornadic formation.

Several tornadic episodes (F0 - F4, some of which contained multiple track events) that affected the National Weather Service Forecast Office Little Rock (NWSFO LIT) county warning area (CWA) were collectively studied with the aim to improve MTA-generated TVS recognition relative to ground truth. Archive level II data from the KLZK (Little Rock) WSR-88D radar data acquisition (RDA) site, available since OCT 95 from the National Climatological Data Center (NCDC), was replayed using WSR-88D Algorithm Testing and Display Systems (WATADS) Build 9 software system. Times and locations of tornadoes were established from local storm damage surveys and Storm Data (NOAA,1995-97).

Algorithm performance statistics tests were calculated for various combinations of two adaptable parameters: Threshold Pattern Vector (TPV) and Threshold TVS Shear (TTS). The goal of this study is to see if algorithm performance can be increased by decreasing these two parameters from their default values (10 and 72 hr-1, respectively). Because of beam broadening and radar horizon problems, this study concentrates on the MTA's ability to detect a TVS within 60 nm of the KLZK RDA site and mesocyclone features out to a range of 120 nm. If this study shows that algorithm performance can be increased by changing these two parameters, WSR-88D field sites can take advantage of a Unit Radar Committee's (URC's) level of change authority to implement an improvement.

2. MESO/TVS ALGORITHM (MTA)

In many tornadic storms, an important precursor of tornadogenesis is the mesocyclone, tersely defined as an area of thunderstorm-scale rotation (usually cyclonic) associated with deep moist convection. On average, nearly 30% of mesocyclones produce tornadoes while around 90% are associated with some type of severe weather. The MTA is modeled after a Rankine-combined vortex which assumes solid body rotation, tangential velocities increase linearly with distance from the center, surrounded by a region of potential flow in which velocities fall off with distance from the center of circulation. The MTA identifies mesocyclones by finding three dimensional cyclonic shear regions that also meet a momentum criterion.

The TPV adaptable parameter defines the horizontal size of the smallest mesocyclone that the MTA searches for. After MTA identifies a mesocyclone, the algorithm checks to see if a Tornado Vortex is present. This process includes an assessment of the highest value of shear in, and in close proximity to, the mesocyclone. If the highest value of shear found is greater than the threshold value (TTS) then a TVS is identified.

2A. How The MTA Works

MESO/TVS Algorithm processing utilizes dealiased base velocity data sampled in .13 nm gates. The base velocity product encodes the first of four gates and assigns the base vector in .54 nm 1 resolution bins. For clockwise antenna rotation, the algorithm searches for an increase in azimuthal Doppler velocities at a fixed range that pass shear and momentum (Eqs. 1 and 2) thresholds, producing pattern vectors. For a set of areal pattern vectors in close proximity (2D processing), a feature is identified by the MTA if a minimum number of pattern vectors are present (TPV). Next, a symmetry evaluation is performed. If a 2D feature passes this condition, it is identified as uncorrelated shear. In the MTA's 3D processing stage, if two or more symmetric features are closely aligned in the vertical, the feature is identified as a mesocyclone (MESO). When the symmetry condition applies to only one elevation angle in a 3D feature, it is categorized as 3D correlated shear (3DC).

SHEAR = V / L (1)

MOMENTUM = V * L (2)

where V Vin - Vout and L is the length of V

If a MESO is detected by the MTA, the highest shear regions are searched for at each elevation in and near to (an additional 5% areal coverage) the MESO. If shear is greater than or equal to the TTS setting, a potential TVS is identified for cyclonically rotating storms. If two or more potential TVSs are vertically linked, a TVS is then generated.

2B. The Significance of TPV And TTS Change

The result of lowering the TPV parameter will reduce the minimum number of pattern vectors for the MTA to detect features, enabling smaller-scale recognition while increasing the potential for more TVS identifications. As a result of lowering the TTS benchmark, an increase in MTA-generated TVSs will be realized.

3. METHODOLOGY

In this MTA performance evaluation, TVS scoring statistics include the use of probability of detection (POD), false alarm ratio (FAR) and critical success index (CSI), defined in Eqs. 3-5. Algorithm output was evaluated for volume scans: (1) that contained tornadic circulations within 60 nm; (2) that occurred at least 30 minutes before tornadoes were on the ground; and (3) that occurred up to 30 minutes after tornadoes were on the ground. An algorithm hit was counted if a TVS was identified with a confirmed tornadic circulation and when a TVS detection occurred upstream of tornadoes within 60 nm, even if beyond 60 nm (lead hit). An algorithm miss was counted if a tornadic circulation was not accompanied during archive data playback by a TVS detection. A false alarm was counted when the MTA identified a TVS on a nontornadic cell.

POD = HITS / { HITS + MISSES} (3)

FAR = FALSE ALARMS / { HITS + FALSE ALARMS} (4)

CSI = HITS / { HITS + MISSES + FALSE ALARMS} (5)

The Archive II data sets used in this study for TVS detections were determined by the tornadic episodes listed in Table 2 (Appendix A). The highest multiple track tornadic event (15) producing the highest fatality count (25), appropriately referred to as the 1 March Tornadic Outbreak (1997), is the largest subset in Table 2. For this scenario, TVS statistics were computed from Tornado tracks #1-5, #10 within the 60 nm radial boundary from the RDA site (range ring) in Figure 1. Tornado tracks #6-9 and #11-16 will be used for MESO/3DC shear detections later in this study. More information pertaining to this specific event can be accessed via the Internet from the NWSFO Little Rock homepage at: http://www.srh.noaa.gov/ftproot/LZK/HTML.

For all tornadic data sets listed in Table 2, the optimal MTA performance was determined objectively by comparing POD and FAR values obtained using default values of TPV and TTS with POD and FAR values obtained using modified values of TPV and TTS. A skill ratio index (SRI), shown in Eq. 6, is proportional to the change in POD and inversely proportional to the change of FAR.

SRI = POD / FAR = [(POD (NONDEF) - POD (DEF)) / (FAR (NONDEF) - FAR (DEF))] (6)

TPV values were lowered (from TPV=10) to 4 in increments of 3. Following suit, the TTS default threshold was reduced to 45, 30 and 20 hr-1 at each TPV setting. The possibility of grouping the data for multiple MTA adaptable parameter optimization was also explored in this study, relating the degree of convective organization from the environment (sounding data) and the F-scale magnitudes in each event, though no meaningful groups could result. However, a common thread in each of the tornadic events did occur in that pre-storm environments were significant enough for the Storm Prediction Center (SPC) to have tornado watches in effect.

An account was also kept for MESO/3DC shear hits beyond 60 nm from the RDA site using the same time window for TVS detections, including lead time hits up to 30 minutes before tornadoes (even if 60nm), listed in Table 3 (Appendix B). Here, the statistics were simplified by calculating POD, measured solely as a function of TPV=10, 7 and 4.

4. RESULTS

Algorithm performance statistics (POD, FAR and CSI) are presented in Figure 2 for the 1 March Tornadic Outbreak. As TPV and TTS are lowered, POD increased; however, so did FAR. The best MTA performance for this particular event occurred at the highest POD with null false alarms (TPV=7 and TTS=45 hr-1). A better MTA performance evaluation may have resulted for a TPV of 7 and TTS of 36 hr-1. Nevertheless, this adaptable parameter combination was not examined in this study.

Performance values shown in Figure 2 are significantly different than a WSR-88D TVS optimal mode study reported by Mitchell (1997). Differences between time window scoring methods may account for some of the deviation. In addition, many of the hits for the 1 March event in Table 2 were associated with Tornado tracks #2 and #4, well within 40nm of the RDA site where the radar's beam width is more likely to resolve and adequately sample vortices capable of producing tornado-like shear.

Performance statistics, for all cases combined in Table 2, are shown in Figure 3. POD statistics, as a function of the TPV and TTS settings, reveal comparable skill levels to those in the 1 March Tornadic Outbreak in Figure 2. In sharp contrast, the performance of the MTA is compromised by significant increases in the FAR, narrowing the FAR gap between the 1 March Tornadic Outbreak to the findings reported in Mitchell (1997). In order to objectively determine the optimal MTA settings for all tornadic events in Table 2, trials for each combination of non-default MTA adaptable parameters are run and displayed with the associated SRI number in Table 1. Values of SRI and average TVS lead time are presented in Figure 4. The combination of TPV=7 and TTS=45 hr-1 provided the highest SRI value. Therefore, the optimal TPV and TTS settings for highest MTA performance are defined as Trial #4 as shown in Figure 4.

Reducing TPV and TTS increased average lead times (Fig. 4). Compared to MTA default settings, Trial #9 provided the best average lead time (about 12 minutes). However, a high number of false alarms decreased SRI, indicating that Trial #9 did not provide the best performance. SRI values suggest that Trial #4 provided the best performance with a lead time of about 6 minutes.

SRI VALUES FOR TVS DETECTIONS

Trial #1 #2 #3 #4 #5 #6 #7 #8 #9
TPV 10 10 10 7 7 7 4 4 4
TTS 45 30 20 45 30 20 45 30 20
SRI 3.63 3.00 2.31 3.79 2.37 2.48 2.45 2.61 2.37

Table 1. MTA performance trial runs of TPV, TTS and SRI values.

MTA-generated MESO and 3DC shear hits and misses are used to measure tornadic skill performance beyond 60 nm from tornadic data in Table 3 (Appendix B). Values of POD as a function of TPV are shown in Figure 5, revealing that TPV=7 offers the largest TPV increase (TPV=3, starting from TPV=10), matching the optimal TPV setting for TVS detections. The net increases in TPV=7, 4 POD from the default TPV threshold were found to be 12.9% and 15.8%, respectively.

5. OPERATIONAL IMPACTS

As a result of lowering the MTA adaptable parameter thresholds for the KLZK WSR-88D, more TVSs will be identified in an operational mode. That is, smaller and weaker circulations from mini-supercells, bow echoes, bookend vortices, frontal shear zones, etc., will be subject to TVS identification, whether tornadic or not. This limitation stems from the fact that the current algorithm cannot distinguish between highly sheared rotational signatures and tornadoes. The MTA optimization will amplify the skill level for TVS identified tornadic circulations yet increase the realization of false alarms. Thereby, where MTA-generated products identify areas of concern, an increased emphasis is placed upon the radar/warning meteorologist to analyze for traditional storm structures and evolutions using WSR-88D reflectivity products, storm-relative velocity products, alphanumeric MESO products at the Principal User Processor Applications Terminal (displays MDA features and classifications such as depth, diameter, and shear), etc., as follow up steps in the decision to warn process for tornadoes. A non-conventional WSR-88D derived product, combined shear, may provide some insight into echo pattern configurations and shear strength in weakly sheared events (Wilken, 1997).

Furthermore, the optimal MTA settings in this study imply a twofold increase in average lead time between a TVS and tornado (from 3 to 6 minutes). As a result, the difference may provide a "real" lead time for tornado warning notification to the public factoring in dissemination time lag effects, albeit slightly based on the TVS predictor alone if not detected sooner by other means. Beyond 60 nm, it is of additive benefit that the optimized MTA TPV adaptable parameter will identify more vertically correlated features in tornadic circulations.

6. SUMMARY AND CONCLUSIONS

In this MTA performance evaluation for tornadic data sets from the KLZK WSR-88D Archive II database using WATADS 9, POD, FAR and CSI statistics for TVS detections are generated as a function of various TPV/TTS combinations. The paradigm that the default MTA settings for TVSs produce a good basis for tornado warnings is not substantiated in the composited analysis (POD ~ 10% and FAR ~ 42%, Fig. 3), further motivating the need for an MTA adjustment. This is likely the result of selecting short-lived and lower end F-scale magnitude tornadoes as part of the data compositing mix. Indeed, on a case by case basis, a measure of statistical variability will occur, as demonstrated with the 1 March Tornadic Outbreak (Fig. 2).

Radar beam propagation limitations confine TVS detections up to 60 nm from the RDA site. When all data were combined, TVS skill levels in detecting tornadic circulations, as well as false alarms, increased when TPV/TTS default values were lowered. Thus, to objectively determine an optimal combination of MTA adaptable parameters, an SRI statistic is derived, representing the change in skill versus false alarms from default settings. The MTA was optimized with Trial #4 (TPV=7, TTS=45 hr-1). Tornado detection lead time increased from 3 to 6 minutes when adaptable parameters under study were lowered from their default values.

Beyond the 60 nm limit and up to the radar's maximum unambiguous range, tornado detections are compared to MESO and 3DC shear hits and misses. At TPV=7, a 12.9% increase in POD occurred from the default TPV setting.

The optimized MTA components will provide added utility for the radar/warning meteorologist in the operational setting by providing more information on tornadic storms across the F-scale spectrum. However, since the MTA does not factor in synoptic and mesoscale conditions, radar reflectivity signatures, convective-scale interactions (such as boundary enhancement and collisions) that may affect storms with tornadic potential, it is paramount for the radar/warning meteorologist anticipating or working a tornadic event to be knowledgeable of the environment and keep alert to changing conditions.

Finally, with regard to pending changes in related algorithms that will be part of the WSR-88D Build 10 software, this WSR-88D Build 9 TVS optimization study will prepare radar operators with more frequent, and accurate, TVS detections. Fundamental differences exist between the NSSL tornado detection algorithm (TDA) for WSR-88D Build 10 software implementation (scheduled for release in August 1998) and the current WSR-88D Build 9 TVS. The primary differences between the two algorithms are that the NSSL TDA-generated TVSs can occur without the existence of an algorithm-identified mesocyclone and examines only gate-to-gate velocity differences, the WSR-88D Build 9 TVS examines shear between maximum and minimum velocities within an MTA-generated MESO. Preliminary results indicate that the NSSL TDA at the default settings for Build 10, using the local database in Appendix A, yielded similar POD and a lower FAR compared to the optimized adaptable parameters statistics for the current TVS algorithm.

7. ACKNOWLEDGEMENTS

My thanks to George Wilken, SOO at the NWSFO Little Rock for his suggestion of this study, providing the Archive II WSR-88D tapes, installing and assisting me in the use of WATADS 9. Also, thanks to Robert Lee of NSSL for a technical review of this paper.

8. REFERENCES

NOAA, 1995-1997: Storm data and unusual weather phenomena. National Climatic Data Center.

NOAA, 1996: WATADS (WSR-88D Algorithm Testing and Display System) Reference Guide for Version 9.0, 6-66. (Available from Stormscale Research and Applications Division, National Severe Storms Laboratory, 1313 Halley Circle, Norman, OK 73069)

Mitchell, E.D., 1997: A Performance Evaluation and Comparison of the NSSL Tornado Detection Algorithm and the WSR-88D Tornadic Vortex Signature Algorithm, Preprints, 28th Conference of Radar Meteorology, AMS, Boston, MA, 351-352.

Wilken, G.R., 1997: Using WSR-88D Shear Products During Severe Storm Events. (Technical Attachment SR/SSD 97-41. NWS Southern Region Headquarters, Fort Worth, TX, 16 pp.)


Figure 1. 1 March Tornadic Outbreak (in tracks).
Range ring is approximately 60 nm from the
KLZK WSR-88D RDA site just north of Little Rock.


Figure 2. TVS statistics (POD, FAR, and CSI) for 1 March
Tornadic Outbreak in 1997 as a function of TPV/TTS settings.


Figure 3. TVS statistics for all cases combined
(POD, FAR, and CSI) as a function of TPV/TTS.


Figure 4. TVS and SRI and average lead time statistics
as a function of TPV/TTS settings.


Figure 5. MTA MESO/3 DC shear POD as a function of
TPV settings. Hits and misses were associated
with tornadoes greater than 60nm from the
RDA site. Lead hits may be, and occasionally were, closer in.

8. APPENDIX A

ARKANSAS TORNADO TABLE FOR TVS STATISTICS

DATE (UTC) ARKANSAS COUNTY (CITY), DISTANCE IN MILES. [TRACK # FOR 1 MARCH EVENT IN FIGURE 1) TIME (UTC) HIGHEST

F-SCALE MAGNITUDE

LARGEST PATH

WIDTH (YARDS)

PATH LENGTH (MILES) DEFAULT MTA HITS / MISSES / FALSE ALARMS
10/27/95 SALINE (SARDIS) 0622-0624 F2 50 0.5 0 / 1 / 0
11/11/95 PRAIRIE TO WOODRUFF (4 NNW DES ARC TO 3 SW McCLELLAND) 0502-0512 F2 100 8.5 0 / 2 / 2

03/06/96 GARLAND (PERCY TO PLEASANT HILL) 0150-0202 F1 50 8 1 / 2 / 0
05/27/96 DALLAS (1.3 SSW DALARK TO 2 SE TULIP) 2015-2040 F3 440 15 0 / 5 / 0
05/27/96 SALINE (1 E BENTON)

SALINE (BRYANT)

2035-2036

2041-2042

F0

F0

20

20

0.1

0.1

0 / 1 / 0

0 / 1 / 1

03/01/97 LONOKE TO WHITE (5.5 NW CABOT TO 8 SW SEARCY) [# 4] 2037-2055 F3 100 18 1 / 3 / 0
03/01/97

CLARK TO HOT SPRING (3.5 NE ARKADELPHIA TO 6.5 E MALVERN)

[# 1]

2050-2110 F4 600 24 0 / 4 / 0
03/01/97 WHITE (10 NE SEARCY TO NEAR VELVET RIDGE) [# 5] 2115-2130 F2 150 13 0 / 3 / 0
03/01/97 SALINE TO PULASKI (5 SE BENTON TO 1 S PROTHO JUNCTION) [#2] 2125-2150 F4 1408 17 2 / 6 / 0
03/01/97 POPE TO VAN BUREN (1 S OAK GROVE TO THE OZARK NATIONAL FOREST) [#10] 2130-2150 F2 880 15 0 / 4 / 0
03/01/97 LONOKE (NEAR FURLOW) [#3] 2202-2204 F2 100 2.3 0 / 1 / 0

Table 2. Arkansas tornadic events for TVS detections.

9. APPENDIX B

ARKANSAS TORNADO TABLE FOR MESOCYCLONE/3DC SHEAR POD

DATE (UTC) ARKANSAS COUNTY (CITY), DISTANCE IN MILES. [TRACK # FOR 1 MARCH EVENT IN FIG. 1) TIME (UTC) HIGHEST F-SCALE MAGNITUDE LARGEST PATH WIDTH (YARDS) PATH LENGTH (MILES) DEFAULT MTA TPV HITS / MISSES
03/06/96 IZARD TO SHARP (7 SW MELBOURNE

TO 2.8 ENE CALAMINE)

0145 - 0245 F3 150 35.5 10 / 3
04/15/96 STONE TO IZARD (4 NW FOX TO 2 NE HORSESHOE BEND) 0010 - 0122 F4 880 44 10 / 5
05/27/96 CLARK TO DALLAS (2.5 SSW OKOLONA TO 2 SE TULIP) 1925 - 2040 F3 440 41 10 / 2
03/01/97 HEMPSTEAD TO CLARK (NEAR HOPE TO 3.5 NE ARKADELPHIA)

[#1]

1950 - 2050 F4 1056 38 6 / 5
03/01/97 BAXTER (3 N NORFOLK TO 3 N JORDAN

[#15]

2010 - 2020 F1 200 0.5 0 / 2
03/01/97 YELL (NEAR BELLEVILLE)

[#8]

2050 - 2100 F1 50 1 1 / 1
03/01/97 YELL (NEAR CHICKALAH)

[#9]

2105 - 2115 F1 50 1 2 / 1
03/01/97 WHITE TO GREENE (NEAR VELVET RIDGE TO NEAR PARAGOULD)

[#5]

2130 - 2310 F2 150 65 15 / 12

03/01/97 VAN BUREN (4 SW SHIRLEY TO THE OZARK NATURAL FOREST)

[#11]

2212 - 2214 F0 20 2.5 2 / 1
03/01/97 STONE (0.5 E RUSHING)

[#12]

2222 - 2223 F0 20 0.5 0 / 1
03/01/97 STONE TO INDEPENDENCE (NEAR MARCELLO TO NEAR BETHESDA)

[#13]

2250 - 2255 F1 25 5 8 / 0
03/01/97 WOODRUFF TO POINSETT (1.5 W PATTERSON TO 5 NE HICKORY RIDGE)

[#7]

2255 - 2330 F2 880 20 12 / 0
03/01/97 SHARP (NEAR CAVE CITY)

[#14]

2315 - 2316 F1 40 1 1 / 1

Table 3. Arkansas tornadic events for MESO/3DC shear detections.