The Influence of Assimilated Targeted Observations Upon Ensemble Forecasts of Convection Initiation During the Mesoscale Predictability Experiment

The influence of assimilating targeted meso-α- to synoptic-scale observations collected in the upstream, pre-convective environment upon subsequent short-range ensemble forecasts of convection initiation (CI) across the central United States for the fifteen aircraft missions conducted by the Mesoscale Predictability Experiment (MPEX) in May and June 2013 is evaluated in this study. Utilizing the ensemble Kalman filter implementation within the Data Assimilation Research Testbed software package as coupled to version 3.4.1 of the Advanced Research version of the Weather Research and Forecasting model, two nearly-identical thirty member ensembles of short-range forecasts are conducted for each mission. Initial conditions for one ensemble are generated through a cycled data assimilation process that incorporates the targeted MPEX dropsonde observations from that day's mission, and initial conditions for the other ensemble are generated through a cycled data assimilation process that excludes the targeted MPEX dropsonde observations. All forecasts for a given mission begin at 1500 UTC, extend forward 15 h, and are conducted on a domain encompassing the conterminous United States with 3 km horizontal grid spacing and 40 vertical levels. Verification is conducted over spatiotemporal thresholds of 50 km/0.5 h, 100 km/1 h, and 200 km/2 h of an observed CI event to assess the skill of probabilistic forecasts and quantify the influence that assimilating targeted observations has upon forecast skill for the events considered. Forecasts without the targeted observations have high probabilities of detection but also greatly overproduce CI, and the inclusion of targeted observations minimally improves some forecasts and minimally degrades other forecasts. Within the 100 km/1 h threshold, the targeted observations on average reduce distance errors between matched modeled and observed objects by 0.22 km while adding a time bias of 0.24 minutes. The forecast performance for three specific cases as well as implications for CI predictability are discussed.