Forecast Sensitivity to Observations using Data Denial and Ensemble-based Methods over the Dallas-Fort Worth Testbed
The research testbed known as the Dallas – Fort Worth (DFW) Urban Demonstration network was created to experiment with many kinds of mesoscale observations that could be used in a data assimilation system, in order to identify observational systems that are most impactful on high-resolution forecasts. Many observation systems have been implemented for the DFW testbed, including Earth Networks Weather Bug surface stations, Citizen Weather Observer Program (CWOP) amateur surface stations, Global Science and Technology (GST) mobile truck observations, CASA X-band radars, SODARs, and radiometers. These “non-conventional” observations are combined with conventional operational data from METARs, mesonet, aircraft, rawindsondes, profilers, and operational radars to form the testbed network.
In this study, the GSI-based EnKF data assimilation system was used together with the WRF-ARW model to examine impacts of observations assimilated for the 3 April 2014 convection initiation case. Data denial experiments were conducted testing the impact of high-frequency (5-min) assimilation of nonconventional data on the timing and location of CI, as well as storm development as they progress through the testbed domain. Results show nonconventional observations were necessary to capture local details in the dryline structure causing localized enhanced convergence and leading to CI. It was found that most of this impact came from the assimilation of thermodynamic observations. Accurate metadata is crucial to the application of nonconventional observations in high-resolution assimilation and forecasts systems.
The second part of this study was exploring the application of the ensemble-based forecast sensitivity to observations (EFSO). First, tests using a global two-layer model were performed to identify areas of improvement in the localization methods needed to make EFSO estimates accurate. Due to the time-forecast component, localization of the EFSO metric is more complicated than during assimilation because as forecast time increases the error correlation structures evolve with the flow. Application of an adaptive localization method – regression confidence factors based on a Monte Carlo “group filter” technique – led to insights into what localization should look like with the time-forecast component included, and subsequently improved EFSO estimates. Next, the EFSO method was applied to the high-resolution CI case and evaluated for accuracy in terms of several verification metrics, including energy norms partitioned by variable and composite reflectivity. Static and advected localization were applied to EFSO and compared for accuracy to the actual forecast error reduction.