Doppler Weather Radar Research and Development
NOAA/NWS/CSTAR – Improving Tornado Detection with WSR-88D Data using Spectral Analysis
Funding Type: CIMMS Task III (Program Manager – Sam Contorno)
Objectives
Develop a novel algorithm to provide accurate tornado detection and extend
the detection range.
Accomplishments
Current tornado detection algorithms (TDA) that employ WSR-88D data search
for strong and localized azimuth shear in the velocity field. However,
the shear signature deteriorates with increasing range due to the smoothing
effect of radar weighting functions. Previous work by others, and our
recent analysis, has shown that a tornado vortex has a distinct spectral
signature that deviates from the typical Gaussian shape obtained from
non-tornadic regions.
We have developed and refined the Neuro-Fuzzy Tornado Detection Algorithm (NFTDA) that integrates both shear and spectral signatures. A comprehensive statistical analysis of NFTDA is performed using numerical simulations. The probability of detection (POD) as a function of the normalized range is presented on the left panel of the figure for the two tornado sizes of rt=100 and rt=200 m, where rt is radius of the tornado. The normalized range is defined as r0=θb/rt, where r0 is the range from the tornado to the center of the radar resolution volume and θb represents the radar beam width in radians. Each data point represents the mean of POD from 50 realizations, and each one has a different noise sequence added to the time series data. For the purposes of comparison, a tornado detection solely based on the thresholding of azimuthal velocity difference is also implemented and is termed “thresholding tornado detection” (TTD). The POD from the TTD using a threshold of 20 m/s, one of the thresholds used in the NSSL's TDA, is provided in the figure. For the larger tornado of rt=200 m both NFTDA and TTD have PODs of approximately 100% when the normalized distance is smaller than 8.7. At larger ranges, NFTDA still has high PODs even though the shear signature is diminishing with increasing range. For the case of rt=100 m NFTDA has a POD of approximately 80% at a range of 14 (80 km) while the TTD has a POD of 10%. It is evident that NFTDA provides higher PODs than TTD especially at far ranges for the two tornado sizes. The improvement is caused by the fact that even when a tornado is far from the radar, wide and flat spectra are still obtained from the vortex, but the shear signature is degraded significantly. Although the performance of the TTD method can be improved by lowering the threshold, the false detections will be increased.
It has been demonstrated that NFTDA can improve the detection from conventional TDA to extend the detection range and to detect small-size tornadoes. However, the spectral products are not readily available from operational WSR-88Ds. On the other hand, polarimetry products are expected to be available in the near future. Therefore, it is of interest to evaluate the performance of NFTDA with only operationally available products (i.e., mean Doppler velocity and spectrum width) and further to investigate the impact of polarimetric products on the tornado detection. Note that NFTDA is extremely flexible to vary the combinations of input parameters with minimum training from the existing data. The evaluation result is shown on the right panel using KOUN data from the 10 May 2003 case, for which polarimetric products are also available from NSSL. The ground damage path is denoted by cyan-shaded area. Note that the tornadoes were traveling away from the radar and tornado of F0-scale was reported during 0405~0419 UTC. The detection results from conventional TDA are denoted by green downward triangles. Conventional TDA can provide detection results that are consistent with the damage path up to 0353 UTC. Three NFTDA with three different sets of inputs were developed, i.e., (1) spectrum width (σv) and azimuthal velocity difference (δv), (2) spectrum width, azimuthal velocity difference, and polarimetric products (cross correlation coefficient and differential reflectivity), and (3) spectrum width, azimuthal velocity difference, and tornado spectral signatures (TSS) developed earlier. The results show that all NFTDA produce similar detections during 0329-0353 UTC that agree well with the damage path and conventional TDA. For latter time (i.e., tornado was weaker and located far from the radar), the NFTDA with level II-derived data (blue squares) produce two accurate detections at 0405 UTC and 0411 UTC, but several false detections occur. The addition of polarimetric products (red circles) can help to eliminate most false detections and extend the detection to 0417 UTC. Note that the false detection at 0359 UTC is still presented, which can be eliminated using NFTDA with spectral signatures.
This project is ongoing.
Publications
Wang, Y., T.-Y. Yu, M. Yeary, A. Shapiro, N. Shamim, M. Foster,
D. Andra, and M. Jain, 2007, Tornado detection using a neurofuzzy System
to Integrate Shear and Spectral Signatures. J. Atmos.
Oceanic Technol.,
submitted.
Xue, M, S. Liu, and T.-Y. Yu, 2007: Variational analysis of over-sampled dual-Doppler radial velocity data and applications to the analysis of tornado circulation. J. Atmos. Oceanic Technol., 24, 403-414.
Yeary, M, Y. Zhai, T.-Y. Yu, S. Nematifar, and A. Shapiro, 2006: Spectral signature calculations and target tracking for remote sensing. IEEE Transactions on Instrumentation and Measurement, 55(4), 1430-1442.
Yeary, M, N. Shamim, T.-Y. Yu, Y. Wang, 2007: Tornadic time series detection using eigen analysis and a machine intelligencebased approach. IEEE Geosci. Remote Sens. Lett., 3, 334-339.
Yeary, M,, N. Shamim, T.-Y. Yu, Y. Wang, and Y. Zhai, 2007, Support vectors to simultaneously minimize classification error and maximize geometric margin for radar spectral signatures. IEEE Transactions on Instrumentation and Measurement, submitted.
Yu, T.-Y., G. Zhang, A. B. Chalamalasetti, R. J. Doviak, and D. Zrnic, 2006: Resolution enhancement technique using range oversampling. J. Atmos. Oceanic Technol., 23, 228-240.
Yu, T.-Y., Y. Wang, A. Shapiro, M. Yeary, D. Zrnic, and R. Doviak, 2007, Characterization of Tornado Spectral Signatures Using Higher Order Spectra. J. Atmos. Oceanic Technol., accepted.
Wang, Y., T.-Y. Yu, M. Yeary, A. Shapiro, S. Nematifar, M. Foster, D. L. Andra, Jr., 2006: Tornado identification using a neuro-fuzzy approach to integrate shear and spectral. 23rd Conference on Severe Local Storms, St. Louis, MO, Amer. Meteor. Soc.
Wang, Y., T.-Y. Yu, M. Yeary, A. Shapiro, N. Shamim, M. Foster, D. L. Andra, Jr., and M. Jain, 2007: A novel approach of tornado detection using a machine intelligence system based on shear and spectral signatures, 33rd Conference on Radar Meteorology, Caims, Australia, Amer. Meteor. Soc.
(Left panel) Statistical analysis of the performance of NFTDA as a function of normalized range for rt=50 and 200 m, where rt is the radius of the tornado. The NFTDA results are denoted by thick solid lines. The results from the detection based on a threshold of velocity difference of 20 m/s (termed TTD) are also provided for comparison and are denoted by the thin dashed lines. It is evident that NFTDA can provide more accurate tornado detection than the conventional shear-based detection. (Right panel) Tornado detection results from conventional TDA (green triangles), NFTDA with Level-II data as input (blue squares), NFTDA with Level-II and polarimetric data (red circles), and NFTDA with Level-II and spectral data (yellow asterisks). It is shown that NFTDA with even only Level-II data can improve the conventional TDA. The polarimetric products can further improve the detection.