In fulfillment of the CY 1997 NSSL/OSF MOU Deliverable D5.1.3 Modifications to the NSSL Storm Cell Identification and Tracking (SCIT) Algorithm and Their Effects on the NSSL Damaging Downburst Prediction and Detection Algorithm (DDPDA)

1.0 Purpose of modifications

The DDPDA uses the output of other NSSL algorithms such as the SCIT algorithm as input. In order to provide the best possible data set for the DDPDA's input stream, some modifications were made to the SCIT algorithm. Specifically, the DDPDA requires the best possible cell tracking and time-height trends of reflectivity and radial velocity data.

After changes were made to the SCIT algorithm code, the code was tested using cases from the NSSL Damaging Wind Events database.

2.0 Coding changes

In order to improve the identification and tracking of cells, two changes were made in the SCIT algorithm code. First, the vertical association of two-dimensional (2D) cell components was made more restrictive. This has the effect of reducing 2D component "switching" between cells that are in close proximity to each other. Secondly, the number of reflectivity thresholds used to identify two-dimensional cell components was increased. It was anticipated that this would allow additional reflectivity peaks to be identified and create higher-resolution 2D cell components.

2.1 Vertical association

Previously, cell components were associated in the vertical by a search routine that associated cell components on one elevation tilt with the first component found within 5 km quasi-horizontally on the next elevation tilt. If no component was found, the routine stepped outward in 2.5 km increments and continued the search until either a component, which hadn't already been assigned to a cell, is found or the search was completed for a 10 km radius from the current component's center.

The changes made to the vertical association part of the code now allow the user to specify the search area and the increment at which the routine steps though the search area. The parameters used to test this change caused the search to begin by first finding a component at the next highest elevation tilt within 0.75 km quasi-horizontally of the component from which the search originates (rather than 5 km). The steps are also 0.75 km, rather than the previous 2.5 km. The search is restricted to a maximum 7.5 km radius, instead of 10 km.

2.2 Reflectivity thresholds for SCIT components

In earlier versions of the SCIT algorithm, users were confined to 7 reflectivity thresholds, which by default were spaced in 5 dBZ increments from 30 dBZ to 60 dBZ. The changes now allow users to select any number of reflectivity thresholds from one to twenty. Those who wish to track large areas of convection may be more interested in using a lower number of reflectivity thresholds, while those interested in attaining high-resolution reflectivity core information (such as needed by the DDPDA) should use a larger number of thresholds spaced at smaller increments.

3.0 Performance comparison

Several tests were performed to determine the effects of the code changes on the SCIT algorithm and the DDPDA. Cells from the Damaging Wind Events Database (DWED) for "wet downbursts" were used to score the algorithms both before and after code and adaptable parameter changes were made. Most cells in the DWED "wet downburst" data set occurred over at least moderately-populated areas (25 people/km²) and typically reached 50 dBZ during some part of their life cycle. Each cell in the database contains the coordinates of the storm centroid as determined by a radar meteorologist at the beginning of each volume scan of its life-cycle.

The SCIT algorithm was run using three sets of adaptable parameters:

Method 1: 15 thresholds, every 2 dBZ from 26 dBZ to 54 dBZ; 0.75 km steps when increasing the vertical association search area.
Method 2 (WSR-88D default): 7 thresholds, every 5 dBZ from 30 dBZ to 60 dBZ; 2.5 km steps when increasing the vertical association search area (original configuration)
Method 3: 7 thresholds, every 5 dBZ from 30 dBZ to 60 dBZ; 0.75 km steps when increasing the vertical association search area.

3.1 Cell Tracking

Table 1 shows the effects of the changes on cell tracking. Out of 232 total cells, the total number of cell ID changes as compared to the "true" track was only 49 (21% of cells) using 15 reflectivity thresholds, as compared to 97 (42% of cells) using the original thresholds and search step interval. This was an overall decrease of 49% in the number of misidentified cells. When using seven reflectivity thresholds and a small search step interval (method 3), the results are about the same (or slightly worse) in most cases as using the original (method 2).

ID change

Radar Date Cells method 1 method 2
(default)
method 3

KFFC 7/17/95 35 4 10 9

KIND 7/15/95 24 7 12 16

KIWA 8/5/93 14 2 2 4

KIWA 8/20/93 19 1 9 8

KLOT 7/15/95 21 0 5 6

KMKX 7/15/95 76 22 25 27

KMLB 6/26/96 9 2 3 3

KMLB 8/17/96 23 11 22 20

KTBW 7/11/95 11 0 9 9

Totals 232 49 97 102

Table 1: Misidentified cells using the three methods listed in section 3.0.

**Table 1:** Misidentified cells using the three methods listed in section 3.0.
Radar	Date	Cells	method 1	method 2 (default)	method 3
KFFC	7/17/95	35	4	10	9
KIND	7/15/95	24	7	12	16
KIWA	8/5/93	14	2	2	4
KIWA	8/20/93	19	1	9	8
KLOT	7/15/95	21	0	5	6
KMKX	7/15/95	76	22	25	27
KMLB	6/26/96	9	2	3	3
KMLB	8/17/96	23	11	22	20
KTBW	7/11/95	11	0	9	9
Totals		232	49	97	102

3.2 Cell time-height trends

Table 2 shows the effects of the changes on 3D cell detections. Out of 1838 cell detections, 183 (10%) had missing 2D components when using method 1, as compared to 290 (16%) using method 2. This was an overall decrease of 37% in the number of cells with missing 2D components. Again, method #3 had slightly worst results than the other two tests.

3D cells with missing components

Radar Date 3D Cells method 1 method 2
(default)
method 3

KFFC 7/17/95 228 6 13 14

KIND 7/15/95 228 31 44 45

KIWA 8/5/93 154 25 47 50

KIWA 8/20/93 158 28 37 33

KLOT 7/15/95 170 12 18 19

KMKX 7/15/95 537 43 71 72

KMLB 6/26/96 52 4 5 6

KMLB 8/17/96 188 32 47 48

KTBW 7/11/95 133 2 8 8

Totals 1838 183 290 295

Table 2: Cell identifications with missing 2D components.

**Table 2**: Cell identifications with missing 2D components.
Radar	Date	3D Cells	method 1	method 2 (default)	method 3
KFFC	7/17/95	228	6	13	14
KIND	7/15/95	228	31	44	45
KIWA	8/5/93	154	25	47	50
KIWA	8/20/93	158	28	37	33
KLOT	7/15/95	170	12	18	19
KMKX	7/15/95	537	43	71	72
KMLB	6/26/96	52	4	5	6
KMLB	8/17/96	188	32	47	48
KTBW	7/11/95	133	2	8	8
Totals		1838	183	290	295

3.3 Downburst Prediction

Table 3 shows the performance of the DDPDA using the two best sets of SCIT input. The data were scored based on the three volume scans preceding a downburst event. The symbols along the top of the table include:

H = hit (correct prediction of a downburst event)
M = miss (incorrect non-prediction of a downburst event)
FA = false alarm (incorrect prediction of a non-event)
CN = correct null (correct non-prediction of a non-event)
POP = probability of prediction [H/(H+M)]
FAR = false alarm ratio [FA/(H+FA)]
CSI = Critical Success Index [H/(H+M+FA)]

In addition, the column marked "missing" contains the number of cell truth points for which there were no corresponding cell detections or for which the cell detection was incorrectly identified as a cell other than that in the truth database. In fact, 37% of the points in the truth database were misidentified using method 2, while only 9% were misidentified using method 1.

Method 1 showed an improvement over method 2 in the area of downburst prediction. POP (from 0.338 to 0.450), CSI (from 0.188 to 0.223), and FAR (from 0.703 to 0.692) were all improved by using 15 SCIT reflectivity thresholds and a smaller search area. Although the number of false alarms increased from 64 (method 2) to 81 (method 1), this is misleading due to the large number of missing 3D cell detections when using method 2.

H M FA CN missing POP FAR CSI

method 1 36 44 81 1146 129 0.450 0.692 0.223

method 2 27 53 64 760 521 0.338 0.703 0.188

Table 3: DDPDA performance using method #1 ("'hires' SCIT") and Method #2 ("'original' SCIT").

**Table 3:** DDPDA performance using method #1 ("'hires' SCIT") and Method #2 ("'original' SCIT").
	H	M	FA	CN	missing	POP	FAR	CSI
method 1	36	44	81	1146	129	0.450	0.692	0.223
method 2	27	53	64	760	521	0.338	0.703	0.188

4.0 Discussion

Based on the performance statistics generated during this study, it is recommended that, in cases where downbursts from "pulse"-type thunderstorms are favorable, an increased number of reflectivity thresholds be used and the search radius for building 3D cells should be more finely tuned.

In addition, using a slightly lower minimum reflectivity threshold such as 26 dBZ (instead of 30 dBZ) may assist in the early identification of cells and allow weaker components to be identified. This lower minimum threshold, in turn, helps improve cell tracking by preventing many occurrences of misidentification that occur when a new cell forms in close proximity to where an older cell is decaying.

No comparison was made between using "small" (0.75 km) and "large" (2.5 km) steps in the vertical association routines for the high-resolution (15 reflectivity threshold) SCIT data. In theory, using a smaller step size should improve the chances of correct vertical association of 2D cell components, especially when two or more cells are located within a few km of each other horizontally.

The use of additional SCIT reflectivity thresholds may take more computer processing power in large-scale precipitation events, as more 2D cell components and 3D cells are identified. In such cases, it may be desirable to use the original SCIT adaptable parameters.

By using the higher-resolution SCIT reflectivity thresholds and decreasing the search step criteria, cell tracking misidentifications were reduced by 49% and the number of cells with missing 2D components was reduced by 37%. In addition, the probability of downburst prediction increased from 0.338 to 0.488. These changes to the SCIT algorithm should also improve other output such as the trends of Probability of Severe Hail and cell-based Vertically Integrated Liquid.