Extended Abstract (540K)P1.15
Developing a statistical scheme to predict the occurrence of lightning in south Florida
Justin M. Winarchick, Florida State University, Tallahassee, FL; and H. E. Fuelberg
Statistical guidance equations are developed to determine the probability of noon to midnight lightning activity (the occurrence or non occurrence of at least one flash) in eastern Miami-Dade and Broward Counties of South Florida during the warm season (May-September). The guidance assumes that the sea breeze provides the dominant forcing for afternoon convective and lightning activity.
Fourteen years (1989-2002) of warm season lightning data from the National Lightning Detection Network and 1200 UTC Miami radiosonde data are used to develop and test the guidance equations. The lightning data document whether lightning occurred within the areas of interest during the noon to midnight time period. The radiosonde data are used to calculate fifty-four potential predictors, including wind, moisture, stability and temperature parameters. Two persistence variables (the previous day’s afternoon activity and the current day’s morning activity) also are included as potential predictors.
Binary logistic regression is used to relate the noon to midnight lightning activity to the pool of potential predictors. A stepwise screening procedure is used to build separate models for each month during the warm season for both counties. Deriving separate monthly models is found to improve forecast skill compared to a single warm season model. Each monthly model generally contains persistence and the wind, moisture, and stability parameters which are known to influence the strength and movement of the sea breeze and convective development.
A cross-validation procedure is used to test the models on independent data and to determine the stability of the models. The cross-validation process reveals that the models are statistically stable and perform well when tested on independent data. The probability of detection, calculated from the independent testing, ranges from ~67% during May to ~90% in August, while false alarm rates range from ~30% during May to only 15% in August.
Results from independent testing of the models also show that they improve on forecasts based solely on persistence. For example, the threat score for the guidance equations is ~69% versus ~61% for persistence alone. Furthermore, the hit rate improves from ~71% to ~77%. Although persistence is a powerful predictor of lightning activity in South Florida during the warm season, the guidance equations provide superior results.
Days when the models produced an incorrect forecast are examined. When no lightning was forecast but occurred anyway, quartiles of lightning activity are considered. The percent of incorrect forecasts decreases from low activity days (i.e., 1st quartile days, 1-7 afternoon flashes) to high activity days (4th quartile days, >125 afternoon flashes). Thus, incorrect forecasts are least likely on the most important days. Fewer days with lightning occurrence are incorrectly forecast during July and August than during May. A similar trend is observed on days when no noon to midnight lightning was observed but had been forecast.
This paper is a companion to one being submitted by Schafer and Fuelberg and optimally should be placed in the same session.
Supplementary URL: http://bertha.met.fsu.edu
Poster Session 1, Poster Session 1
Monday, 10 January 2005, 2:30 PM-4:00 PM
Previous paper Next paper