Applications of EFSO to Improve NWP
Ensemble Forecast Sensitivity to Observations (EFSO) was introduced by Kalnay et al. (2012) inspired by the adjoint Forecast Sensitivity to Observations of Langland and Baker (2004). EFSO estimates the forecast error change generated by each observation. Since it is based on computations already available from an ensemble analysis, it is economical and efficient for evaluating the impact of each observation. EFSO finds whether each observation is beneficial (it decreases the total energy of the forecast errors) or detrimental (it increases it). EFSO has two major applications: 1) Data Monitoring and Selection: just one month of EFSO impacts reveals detrimental subsets of observations, and allows designing fixed QC for new instruments (Lien et al., 2017). 2) Proactive QC: This is a fully flow dependent QC designed to avoid forecast skill dropouts (Hotta et al., 2017, Ota et al., 2013). It rejects observations at each DA cycle that EFSO finds very detrimental, using the next 6hr analysis as truth. The “final analysis” (known as GDAS at NCEP) is corrected using the EFSO-estimated impacts of the detrimental observations, and these corrections are “cycled”, i.e., accumulated with time. Experiments with realistic GFS forecasts show that cycled PQC improves the 1 day forecasts by ~10-20%, and that the improvement remains significant (~5%) even after 5 days. Since the cycled PQ is performed on the “final analysis”, not on the “operational analysis”, it is not using “future data”, and it is both efficient and operationally feasible.
Prof. Eugenia Kalnay University of Maryland