Forecasting lightning around the Korean Peninsula by postprocessing ECMWF data using SVMs and undersampling

Abstract

We use machine learning to generate binary forecasts of the occurrence of lightning within a particular location and time interval. The training data is weather variables found in the forecasts from the European Centre for Medium-range Weather Forecasts, correlated against subsequent lightning reports, for a region containing the Korean Peninsula. Lightning is uncommon, so the amount of data which does not involve lightning tends to swamp the training process. Thus we only consider spatial locations at which lightning frequently occurs, and we also undersample the subset of the remaining data-points which are not associated with lightning. Results from support vector machines and random forests had equitable threat scores of 0.0885 and 0.0828, respectively. The ETS of results from SVMs can be increased to 0.1241 if temporal resolution is reduced by a factor of 2, and 0.1499 if spatial resolution is reduced by a factor of 3.