Automated classification of convective downburst events in wind gust observations

Abstract

Wind observations near the ground are critical in assessing the impact of wind on structures. All wind climates comprise a mixture of several disjoint meteorological mechanisms that require separation before assessment. In this paper previous studies distinguishing between convective and non-convective gust events are reviewed. Classification by visual inspection of the gust speed timeseries is generally agreed to be easy and accurate, but it becomes impractical for very large datasets.  Recent automated approaches, using statistics, pattern recognition and neural networks, are calibrated against 4000 visually classified gust events from 20 locations across the USA over 22 years. The most promising method is developed to use only gust speed statistics to distinguish five classes of gust event: synoptic scale storms, deep convection, the forward flank and the rear flank of gust fronts, and downbursts from isolated thunderstorms.  A 6^th class collects non-meteorological artefacts in the data. The ensemble-averaged timeseries of each class form a distinctive hierarchy. The misclassification error rate against the visual classification is 7.8%, with most errors between adjacent classes. When applied to >10⁷ gust events >20kn from 450 locations across the USA, the class hierarchy remains stable. The method is implemented by open-source R scripts.