How Radar-Derived Divergence can Improve Very-Short-Term Forecasting of Convective Initiation

This study presents a technique for very-short-term forecasting of convective initiation (CI) using radar-derived divergence fields. Convective nowcasting is largely dominated by spatiotemporal tracking of rainfall features, usually in radar reflectivity fields. Algorithms identify existing objects in the fields and derive a motion vector field by tracking the objects in past radar imagery. Various nowcasting methods then use the motion vectors to extrapolate and evolve features forward in time and space. Save for computationally expensive techniques, there is little operational forecasters can do to forecast CI in the very-near term. This project utilized radar-derived divergence (convergence) fields, vertically integrated in accordance with the mass continuity equation, to isolate areas of interest where CI was likely to occur in the very-near future. We limited cases to weakly-forced, ordinary convection typical of summertime, pop-up thunderstorms. Little to no tilt of the updrafts and downdrafts associated with this convective mode ensured that a simple vertical integration captured these features well. If objects appeared in the integrated divergence fields before CI was evident on reflectivity, that instance was classified as a “hit”. The method used a reflectivity threshold of >=35 dBZ to classify initiation. Data from the GOES East Geostationary Lightning Mapper provided a separate baseline to assess hits and misses. We present the results in the context of hits, misses, probability of detection (POD), false alarm rate (FAR), and critical success index (CSI). Across 13 ordinary convective cases, we observed 189 areas of interest that either did or did not lead to CI. We classified 40 as false alarms (FAR=0.21) and 137 hits (POD=0.92, CSI=0.72) with respect to reflectivity. We discuss how the process could be fully automated using machine learning and what the output could look like for an operational product.