Factors Influencing Ensemble Sensitivity-Based Targeted Observing Predictions at Convection-Allowing Resolutions

Intelligent methods to target additional observations to improve forecasts have been available for decades, which take into account fast growing errors (i.e., singular vectors) and gradients of tangent linear models (i.e., adjoint sensitivity). Ensemble sensitivity analysis (ESA) is a more recently developed targeting method that combines information regarding the largest uncertainty of a forecast state along with areas ripe for error growth to estimate where observations could be gathered to improve forecasts. ESA relates a scalar forecast metric to prior model states through a regression of an ensemble of non-linear model forecasts and can be carried out during forecast post-processing. Incorporating uncertainty from the initial state, and an estimate of observation error, ESA predicts the impact of a new observation on the chosen forecast response distribution. Less-than-ideal results have been realized when ESA has been utilized for mesoscale convection/precipitation forecasts. Recent literature has noted only modest forecast improvements from new observations targeted with ESA. Given the desire to exploit advantages of ESA over other targeting methodologies (e.g., execution time), it is imperative to understand factors that influence the prediction of observation impacts and the impacts after assimilation.

A set of experiments is designed to evaluate the relative impact of nonlinearity, data assimilation procedures, and numerical noise on ESA-based targeting at convective-allowing resolutions. A 50-member ensemble is generated over ten cases of severe storms along the dryline with the Advanced Research core of the Weather Research and Forecasting model (WRF) and Data Assimilation Research Testbed (DART) software. Observing system simulation experiments (OSSEs) under the assumption of a perfect model are utilized to assimilate targeted observations at varying lead times and under different assimilation configurations. It is determined that localizing observations during assimilation negatively impacts the correlation of ESA-based predictions and actual impacts of the observation measured by changes to the forecast metric distributions of reflectivity and accumulated rainfall. Additionally, more accurate predictions of observations impacts are realized at shorter lead times (i.e., time before valid forecast metric), a function of non-linear effects that propagate through the moist dynamics. Moreover, numerical noise significantly contributes to poor observation-impact predictions, and observations that induce large changes to the forecast relative to the control forecast are more predictable. Implications of these results in regards to ESA applications for higher resolution simulations (e.g., storm scales) will also be discussed.