Classification and prediction of hail using self-organizing neural networks

The occurrence of severe hailstorms in Switzerland is a relatively frequent event. Therefore reliable predictions of hail would be extremely important for the protection of human lives and property. Quite different developments of hail types with a different duration time can be observed. The main problem predicting hail lies in the need of a prediction after a short observation period. That means, that after typically 5 minutes, an estimate should be given on what will happen 30 minutes in the future of a hailstorm cell observed by the maximal radar reflectivity. Using self-organizing feature maps (SOM) for data analysis in its original form leads to a misspecification of the prediction. We developed an extended version of the SOM such that a prediction of hail becomes possible. The main idea of our approach is first to use an extended SOM for the classification of a “typical” hail development. After the classification a prediction using the extended vector can be made. We compared our results to classical approaches, as used in meteorology, and to a regression model. For an increasing prediction time we achieved an significant improvement of the prediction error in comparison to the classical and statistical models