APPLICATIONS OF ENSEMBLE FORECAST SENSITIVITY TO OBSERVATIONS FOR IMPROVING NUMERICAL WEATHER PREDICTION
MetadataShow full item record
Massive amounts of observations are assimilated every day into modern Numerical Weather Prediction (NWP) systems, and more are being deployed. The large volume of data prevents thorough monitoring and screening (QC) the impact of each assimilated observation using standard observing system experiments (OSEs). The presence of so many observations also makes very difficult to estimate the impact of a new observing system using OSEs. Forecast Sensitivity to Observation using adjoint formulation (AFSO, Langland and Baker, 2004) provides an efficient impact evaluation of each observation on forecasts. We propose 3 applications using the simpler ensemble formulation of FSO (EFSO, Kalnay et al., 2012) to improve NWP, namely (1) online monitoring tool, (2) data selection, and (3) proactive quality control (PQC). We first demonstrate PQC on a simple Lorenz (1996) model, showing that EFSO is able to identify artificially '`flawed" observations. We then show that PQC improves the quality of analysis and forecast of the system, even if the observations are flawless, and the improvement is robust against common sub-optimal of DA configurations in operation. A PQC update method reusing the original Kalman gain is found to be both accurate and computationally efficient. EFSO and PQC are then explored with realistic GFS systems. A close-to-operation GFS-GSI Hybrid En-Var system is used to examine the data monitoring and selection applications. The benefit of the online observation monitoring and data rejection based on EFSO is very apparent. Identifying and deleting detrimental radiance channels results in a forecast improvement. Results obtained on a lower resolution GFS system show that PQC significantly improves the quality of analysis and 5-day forecasts for all variables over the globe. Most of the improvement comes from "cycling" PQC, which accumulates improvements brought by deleting detrimental observations over many cycles, rather than from deleting detrimental observations in the current cycle. Thus we avoid using "future data" in PQC and its implementation is shown to be computationally feasible in NCEP operations.