The algorithmic detection of pulse thunderstorms within a large, mostly non‐severe sample

The accurate differentiation of pulse thunderstorms from benign weakly forced thunderstorms (WFTs) is both a historical and contemporary forecasting challenge. Little research has been directed toward WFTs, and the few existing efforts are characterized by small sample sizes and inflated proportions of pulse thunderstorms. The purpose of this study was to determine whether pulse thunderstorms can be successfully differentiated from non‐severe WFTs within a large, mostly non‐severe sample, a more operationally realistic scenario. Random forests, a decision‐tree‐based machine learning technique, was applied to radar, total lightning and environmental parameters of >84,000 WFTs, of which <1% were pulse thunderstorms. In particular, differential reflectivity ( Z DR ) fields were mined for occurrences of Z DR troughs, suggested by recent work as a reliable indication of impending downbursts. The random forest identified pulse thunderstorms from the 84,000 storm sample with a critical success index (CSI) of 0.29. The CSI improved to 0.50 in a subset of the most active geographical regions and convective environments, outperforming the US National Weather Service (CSI = 0.35) for warnings issued on pulse thunderstorms. Unfortunately, the presence of Z DR troughs contributed little skill to the random forest. Although forming at higher rates in pulse thunderstorms (61%) than non‐severe WFTs (5.1%), trough‐possessing non‐severe WFTs were roughly 9× more common than pulse thunderstorms with the same feature. Overall, the random forest shows promise as a potential decision aid for identifying pulse thunderstorms. Performance may be further improved if collinearity amongst the radar parameters and probable under‐reporting of severe weather can be overcome.