Results and Conclusions

The performance of the BWER detection scheme on the training and on the test cases is shown in Table 1. The success index is high by either performance measure. The incidence of false alarms (element c) is low, indicating that the scheme possesses a very good discrimination capability. Hence, by using a fuzzy logic scheme, the uncertainties associated with the radar profile and the 3D structure of a BWER can be factored in and BWERs reliably detected.

Table 1: Performance of the BWER Detection Scheme: The letters b-i correspond to the elements of the table in Equation 4. The CSIs are computed using Equation 6 (the traditional measure) and Equation 5 (the measure that incorporates marginal BWERs).

Radar Location	Date	b	c	d	e	f	g	h	i	CSI( 5)	CSI( 6)
Fort Worth, TX	May 7, '95	2	0	1	1	0	6	0	4	-	-
Lubbock, TX	Jun 2, '95	3	3	4	1	3	15	1	16	-	-
Norman, OK	May 11, '92	7	5	0	4	3	0	2	6	-	-
Training cases		12	8	5	6	6	21	3	26	0.41	0.47
Oklahoma City, OK	Apr 21, '96	1	1	0	7	8	3	1	12	-	-
Oklahoma City, OK	Jun 1, '95	2	3	0	0	0	0	0	0	-	-
Test cases: Fuzzy logic classifier		3	4	0	7	8	3	1	12	0.54	0.63

Two of the training cases (the Lubbock and Norman cases) had several multi-storm clusters, greatly increasing the complexity of the storm structure. The false alarms that are found by the scheme are usually complex multicellular structures that give the appearance of weak BWERs. These multicellular structures usually lie in the vicinity of the storm's main updraft, i.e., in the region of a storm where we would expect a BWER to occur. Other false alarms are caused by extremely strong inflow notches that lie in the vicinity of the storm's main updraft. More information is needed to eliminate those false alarms that still are found by the scheme.

Unfortunately, a formal, quantitative definition of BWER structure is lacking and it is not known whether a scheme developed on cases in one part of the country will perform as well elsewhere. Since all the cases studied in this paper pertain to radar data collected in the Southern Plains during spring, a more exhaustive study, using data from different several regions of the country and on numerous storm days, should be conducted.

Finally, the BWER detection scheme described above depends crucially on output from the Mesocyclone Detection Algorithm (MDA). The MDA does not detect circulations in regions where the velocity data from the Doppler weather radar is range-folded (aliased). Consequently, the BWER detection scheme also fails in such areas. Unfortunately, removing the link to the MDA increases the number of false alarms to an unacceptable level.

When we remove the link to the MDA, the number of candidate regions greatly outnumbers the number of BWERs that we have in our sample cases. The problem then becomes one of classifying skewed distributions and identifying rare signatures. Our current research is geared towards overcoming these problems. To that end, we are developing a fuzzy aggregation operator that performs well in aggregating variables that are distributed in a skewed fashion [10]. To account for the rarity of BWER signatures, we use genetic algorithms where the fitness function is a logical combination of the CSIs found over the various sample cases [11].