Ensemble Sensitivity-Based Observation Targeting Experiments for Southern Plains Dryline Convection

A variety of observation targeting methods exist (e.g., adjoint sensitivity, singular vectors, ensemble transform) to estimate the observation types and locations that would improve a specific forecast. Observation targeting methods have primarily been tested and applied on synoptic scales for lead times of 24 hours or greater using airborne platforms and dropwindsondes. In this study, the authors test the ensemble-based sensitivity analysis (ESA) targeting methodology for forecasts of severe convection along drylines in the southern plains using observing system simulation experiments (OSSEs). ESA is a computationally inexpensive technique that relates changes in forecast variables to changes in an initial state through linear regression. Forecast changes from targeted observations may be estimated by comparing forecasts that assimilate the targeted observations and those that assimilate observations from random locations. The primary goals of this study are to determine the quantitative impact observations may have and whether enhanced predictability may be gained through targeted observing for mesoscale convection forecasts in spite of the linear assumptions underpinning ESA.

Ten cases of dryline-initiated convection are simulated from 2011 to 2013 within an OSSE framework. Ensemble simulations are produced from a cycling system that utilizes the Weather Research and Forecasting (WRF) model v3.3.1 within the Data Assimilation Research Testbed (DART). A one-way nested configuration with three domains at 36, 12, and 4-km grid spacing is setup using the EAKF data assimilation procedure. The ensemble is cycled for 48 hours prior to forecast initialization, with adaptive inflation and covariance localization utilized. A “truth” (nature) simulation is produced by supplying a 4 km WRF run with GFS analyses and integrating the model forward ~78 hours, from the beginning of ensemble initialization through the end of the forecast. Each nested domain extracts observations from the nature run, with random error from a normal distribution added, and assimilates the observations every six hours during cycling. Target locations for surface and radiosonde observations are computed six hours into the forecast based on a chosen scalar forecast metric (e.g., maximum reflectivity at convection initiation). A new forecast is initialized six hours after the prior forecast was initialized, assimilating observations based on the following three experiments: (1) Only the targeted observations are assimilated; (2) A swath of observations outside of target areas are assimilated (i.e., random observations); (3) Both targeted and random observations are assimilated. Using these methods, a proper analysis of the impact from targeted observations is accomplished for dryline convection forecasts.