Thunderstorm predictors and their forecast skill for the Netherlands

Thirty-two different thunderstorm predictors, derived from rawinsonde observations, have been evaluated specifically for the Netherlands. For each of the 32 thunderstorm predictors, forecast skill as a function of the chosen threshold was determined, based on at least 10280 six-hourly rawinsonde observations at De Bilt. Thunderstorm activity was monitored by the Arrival Time Difference (ATD) lightning detection and location system from the UK Met Office. Confidence was gained in the ATD data by comparing them with hourly surface observations (thunder heard) for 4015 six-hour time intervals and six different detection radii around De Bilt. As an aside, we found that a detection radius of 20 km (the distance up to which thunder can usually be heard) yielded an optimum in the correlation between the observation and the detection of lightning activity. The dichotomous predictand was chosen to be any detected lightning activity within 100 km from De Bilt during the 6 h following a rawinsonde observation. According to the comparison of ATD data with present weather data, 95.5% of the observed thunderstorms at De Bilt were also detected within 100 km. By using verification parameters such as the True Skill Statistic (TSS) and the Heidke Skill Score (Heidke), optimal thresholds and relative forecast skill for all thunderstorm predictors have been evaluated. It was found that Heidke reaches a maximum for more thundery index values than the TSS. In order to arrive at a single optimal threshold value, the TSS and Heidke were combined to form the Normalized Skill Score (NSS). When comparing forecast skill in a dichotomous forecasting scheme, the Lowest 100 hPa Lifted Index scores best, although other versions of the Lifted Index have relatively good performance as well. Even though the Boyden Index does not account for any moisture, it serves surprisingly well as a dichotomous thunderstorm predictor. The Rank Sum Score (RSS) has been used to assess relative forecast skill without the use of a dichotomous forecasting scheme. Again, the Lowest 100 hPa Lifted Index scored best. Finally, we have estimated the 'thundery case' probability as a function of the various thunderstorm predictors. We found that thunderstorm probability depends most on latent instability (especially near the surface), next on potential instability, and least on conditional instability. (C) 2003 Elsevier B.V. All rights reserved.