| description abstract | In this work, a new approach to weather model output postprocessing is presented. The adaptive boosting algorithm is used to train a set of simple base classifiers with historical data from weather model output, surface synoptic observation (SYNOP) messages, and lightning data. The resulting overall method then can be used to classify weather model output to identify potential thunderstorms. The method generates a certainty measure between ?1 and 1, describing how likely a thunderstorm is to occur. Using a threshold, the measure can be converted to a binary decision. When compared to a linear discriminant and a method currently employed in an expert system from the German Weather Service, boosting achieves the best validation scores. A substantial improvement of the probability of detection of up to 72% and a decrease of the false alarm rate down to 34% can be achieved for the identification of thunderstorms in model analysis. Independent of the verification results, the method has several useful properties: good cross-validation results, short learning time (≤10 min sequential run time for the experiments on a standard PC), comprehensible inner values of the underlying statistical analysis, and the simplicity of adding predictors to a running system. This paper concludes with a set of possible other applications and extensions to the presented example of thunderstorm detection. | |
