Using DBSCAN-RF Algorithm and Multi-source Features to Model Deep Convection in Hubei, China

Abstract

AbstractThe realistic representation of convection in atmospheric models is paramount for skillful predictions of hazardous weather as well as climate. In order to accurately describe mechanism of deep convection initiation, the forecasting models of thunderstorm occurrence are established from the perspectives of “point to face” and “integration of air and ground”, based on Cloud-to-Ground (CG) lightning and convection inducing factors. Firstly, we use the DBSCAN density clustering method to preprocess the discrete CG strokes, eliminating weak convection or noise; then we combine ERA5 with other valuable data sources and use machine learning to predict the probability of thunderstorms. Up to 49 input variables are used, representing, for example, instability, humidity, topography, land-cover. Feature importance derived from random forest (RF) models emphasize the high importance of conditional instability for deep convection. Topographic features accounts for 3%~4% of the total feature contribution, in which geographical position and elevation play a major role. In the comparison experiment of thunderstorm prediction with and without topographic factors, the former can make thunderstorm events and non-events tend to be predicted correctly, reduce false alarm ratio, and improve the overall skill of models. On the 2013–15 independent test, the 2013–15 RF model has a hit rate of 0.79, false alarm ratio is 0.65, and threat score is 0.32. Combining mesoscale reanalysis data with small-scale underlying surface data, the DBSCAN-RF can be used to further study climate trends in convective storms.KeywordsDeep convectionCG stroke clusteringMulti-source features