Abstract
Past binary logistic regression (BLR) and perfect prognosis schemes have shown skill for predicting the probability of one or more cloud-to-ground (CG) lightning flashes for regions such as Florida. This study examines the ability of the BLR and perfect prog techniques to forecast only those flashes at the daily onset of warm season convective activity. A statistical model is developed for two domains in the western United States in the vicinities of Amarillo, Texas and Albuquerque, New Mexico, to predict the probability of one or more CG flashes during the 1800 UTC to 2100 UTC time window. Warm season convection in these locales is influenced by factors such as the Southwest Monsoon, drylines, and topography. Each domain consisted of a 28 × 26 grid at a 10-km resolution. One equation for each domain was developed using 10-year climatology and predictors from the North American Regional Reanalysis (NARR) dataset for the dependent period May through September of 1994 – 2004. In our perfect prog scheme, the reanalysis data were treated as observations. Forecasts were made for the 2005 warm season, and evaluated using the Brier Score, Brier Skill Score, and reliability diagrams. Due to the relative rarity of the event being forecast, forecast probabilities tended to be small, rarely exceeding the 30% range in the Amarillo domain and 70% (with a majority below 50%) in the Albuquerque domain. Both domains typically exhibited skill over climatology and good reliability for low forecast probabilities, which constituted a majority of the forecasts. Sample forecasts for both domains are examined. The Amarillo model demonstrated competent performance on 10 June 2005, but poor performance on 1 September 2005. The Albuquerque model performed well on 1 September 2005, but performed poorly on 17 July 2005. Despite the low magnitude of probabilities forecast, both models showed promise, commonly placing their relative maxima in the vicinity of the verifying initiating flashes, with some obvious exceptions.
|
