1.0 Introduction

This document acts as a final report for the OSF/NSSL MOU item 6.1 and deliverable D6.1.1. It describes changes made to the NSSL Damaging Downburst Prediction and Detection Algorithm (DDPDA) during 1997 as well as the current performance of the algorithm.
 

2.0 Algorithm Enhancements

During 1997, the DDPDA was completely rewritten in the C language in order to replace inefficient code and reduce the code's complexity. The algorithm's input streams were improved, and environmental data were added to the algorithm for the first time. The re-coding of the algorithm also allowed the cell time-height trend information to be improved and for gridded convergence information to be added as output.

2.1 Enhancements to Algorithm Input

Since the DDPDA uses the output from other algorithms such as the NSSL Storm-Cell Identification and Tracking Algorithm (SCIT) as input, it is necessary that these data be as accurate as possible. Modifications to the SCIT algorithm during 1997 improved both cell tracking and time-height trend information, which led to improved DDPDA performance (see deliverable report D5.1.3).

Mid-altitude convergence and low-altitude divergence data were also improved. Previous versions of the algorithm used separate convergence/divergence detection routines, where two- or three-dimensional (3D) convergence features were associated with nearby 3D cell features. Using this method, large convergence features would frequently be incorrectly associated with only one cell when they should have been associated with several. This led to much missing data in the time-height trends of convergence and divergence.

The new convergence detection routines are cell-based rather than convergence feature-based. At each radar elevation sweep of each 3D cell, the convergence is calculated for each range gate within a user-defined radius of each 2D cell component. A least-squares shear estimation method described by Elmore et al. (1993) was utilized to make the convergence calculations. This method also produces a rotational shear value for each range gate as a by-product of the calculation, so these data are stored as well. After the least-squares calculations are performed, radial velocity "pattern vectors" are built and radial convergence is also calculated as a radial velocity difference. Figure 1 shows an example of these data for one volume scan of a downburst-producing cell.
 

2.2 Near-Storm Environment Data

Near-storm environment data have been added to the algorithm in an attempt to improve predictions. So far, parameters have been tested which combine radar and environmental data. Future work may focus on using environmental data as a filter to reduce false alarms. A list of the data which may be used as algorithm input can be found in Table 1. Presently, these parameters are read from a text file modified by the user, but the process should soon be automated to use NSSL Near-Storm Environment (NSE) algorithm. Additionally, many of these parameters have not yet been tested in the algorithm, but may be used in future version which predict high-based "dry" downbursts or strong straight-line winds produced by bow echoes.

The parameters that were tested include convergence at the height of cloud base and at the height of minimum equivalent potential temperature, and maximum reflectivity at and above the minimum equivalent potential temperature height.

2.3 Downburst Prediction

One hundred seventy-eight different variables which could be calculated from radar data were examined to determine which had the best potential for predicting downbursts. These parameters were typically variations on ways to detect the primary downburst precursors described by Smith and Eilts (1997). The parameters include data from the SCIT algorithm Hail Detection Algorithm (HDA), as well as the convergence and rotation information discussed above.

The current equation uses three parameters which are optimized based on the data in the Damaging Wind Events Database. These include the maximum reflectivity above the height of the minimum equivalent potential temperature, the depth of the convergence greater than .004 s^-1, and the maximum convergent velocity difference between 2 and 6 km above radar level (ARL).

Unlike earlier versions of the algorithm, the prediction equation can be changed based on the range from the radar. Analysis of downburst events in the database indicate that velocity-based precursors are more prevalent at close ranges (less than 70 km from the radar) than at long ranges.

2.4 Other Algorithm Output

Other improvements to the algorithm output have been made as well. The cell time-height trends of convergence and divergence have been improved to ensure data points at each level a 2D cell component is detected. A gridded field of convergence and rotation is also produced by the algorithm, and may be used in future work with strong straight-line winds produced by bow echoes.

3.0 Performance Evaluation

The algorithm's performance tested using events in the NSSL Damaging Wind Events Database. Additional results based on the low-resolution version of the SCIT algorithm can be found in deliverable report D5.1.3.

3.1 Database

The database used for testing included 223 cells which were identified over moderately-populated areas (greater than 25 people per km²). According to Storm Data, 26 of these produced severe downbursts. Only cells over populated areas were used in order to increase confidence that cells with no damage reports did not, in fact, produce any damage. Cells were identified by a scientist and added to the database. Data were from eight different radar data sets, and included data from Arizona, Florida, Georgia, Illinois, Indiana, and Wisconsin.

Approximately 5000 lines of code were written to assist with development of the ground truth database and scoring of the algorithm. This should be of considerable benefit in future algorithm development, especially as the DDPDA is optimized for different geographical regions and environments, as it drastically speeds up the cell identification and algorithm scoring process.

3.2 Results

Evaluation of the algorithm was based on these 223 cells and 26 downburst events. Algorithm "hits", "misses", "false alarms", and "correct nulls" were calculated based on the following criteria:
 

• Hit (H) - A correct prediction of a downburst which occurred on one of the three volume scans preceding the event.
• Miss (M) - A non-prediction of a downburst one to three volumes preceding the event.
• False Alarm (FA) - A prediction of a downburst event at any other time than that defined as a "hit".
• Correct Null (CN) - A correct non-prediction of a non-event.
• Probability of Prediction (POP) = H/(H+M)
• False Alarm Ratio (FAR) = FA/(H+FA)
• Critical Success Index (CSI) = H/(H+M+FA)
• Heidke Skill Score (HSS) = [2*(CN*H-FA*M)]/[(CN+FA)*(FA+M)+(CN+M)*(M+H)]

At this point of its development, the algorithm suffers from a moderately high False Alarm Ratio (0.692), which may be correctable though the use of environmental data. These environmental data have been added to the algorithm, but most have not been tested in detail yet. Additionally, the algorithm may be expected to have improved performance when optimized seasonally and regionally. The probability of downburst prediction (0.450) should also improve once these changes are made.

In addition, the reflectivity-based wind gust prediction technique described by Stewart (1996) was scored based on the same criteria. For scoring purposes, a severe event was classified as a cell with a predicted gust of 25 ms^-1 or greater. These results are given in Table 3. This technique suffers from an even greater false alarm problem than the DDPDA, but may be useful when combined with other DDPDA parameters in future statistical studies.