Assessing Observation Impacts in NWP: Atmospheric Motion Vector Winds
In the last three decades, the number of atmospheric observations available to scientists has increased so dramatically that dealing efficiently with the sheer volume of data has become a challenge. In the early 1970s, fewer than 100,000 observations were available to scientists for use. Today, more than 4 million observations are taken with various instruments and observing platforms in any given six-hour period, and the number is still increasing.
The Goddard Earth Observing System Atmospheric Data Assimilation System, version 5, (GEOS-5 ADAS) receives several million of these observations every six hours, and by thinning the data (e.g., selecting only certain observations because of data density) and employing quality control measures, roughly 2.5 million of those observations are assimilated into GEOS-5 real-time analyses four times a day. Historically, quantifying the impact that a specific observation or type of observation has on a model forecast required performing observing system experiments (OSEs), or running the model with and without assimilating the observed data in question, but this approach is rather computationally expensive, and is not a feasible way to assess the impact of all observations on a model's forecast.
An adjoint of the GEOS-5 assimilation system is used instead to compute the impact of all observations on forecast error simultaneously (see Errico, 2007). The result catalogues the effect that each type of assimilated observation has on a forecast, and is helpful in assessing which observations add the most value to a forecast. Knowing this, it's easier to identify which new instruments could be developed and deployed to best complement the existing observational network, and to be of most use for numerical weather prediction.
The impact a type of observation has on a model forecast depends highly on the observations themselves, on the manner in which they are assimilated, and even on the other observations being input into the assimilation system. A recent investigation into the impact of atmospheric motion vectors (AMVs), or satellite winds, on forecasts revealed just how sensitive impacts were to these factors.
Satellite winds have a much greater impact in the Navy's Fleet Numerical Meteorology and Oceanography Center (FNMOC) forecasts than in GMAO's GEOS-5 forecasts (Gelaro, 2012). One possible reason for this is that the FNMOC assimilates many more AMVs than GMAO does, and reduces them differently (the GMAO uses data thinning to assign observed values to the model gird, while the FNMOC uses a weighted regional averaging of observations). When the GMAO substitutes FNMOC's satellite winds for those used in the operational version of GEOS-5 (and also at NCEP), overall forecast skill increases, the global impact of AMV observations on the forecast nearly doubles, and all sources of satellite winds have an increased (positive) impact. This increased impact also comes from those satellite winds that are usually assimilated by GEOS-5, but are typically found to have a negative impact on GEOS-5 forecast results. The volume of data assimilated clearly appears to be important to observation impacts, as do the methods used in assimilation.
Though assimilating FNMOC AMV data into GEOS-5 as a substitute for its standard AMV data increases the impact that satellite winds have on the forecast, the overall observation impacts for all other data types appear to decrease (see Figure 1). The benefit of one type of observation to a forecast clearly depends on the other data being input (i.e., certain types of data may complement or take away from each other). For instance, the relatively small impact of satellite winds on GMAO forecasts versus FNMOC forecasts may be due to the fact that the GMAO assimilates far more satellite radiance observations than the FNMOC system does.
Figure 1: Fractional observation impacts for forecasts run from December 10, 2010 to January 31, 2011. The control runs (black) made use of the standard GEOS-5 data set, while the NRLAMV runs (magenta) substitute FNMOC AMVs for those normally used in GEOS-5.
References:
Errico, Ronald M., 2007: Interpretations of an adjoint-derived observational impact measure. Tellus, 59A, 273—276.
Gelaro, R., 2012: The Impact of Satellite Atmospheric Motion Vectors in the GMAO GEOS-5 Global Data Assimilation System. 5th WMO Workshop on the Impact of Various Observing Systems on Numerical Weather Prediction. Sedona, AZ, World Meteorological Organization. (Available online at www.wmo.int/pages/prog/www/OSY/Meetings/NWP5_Sedona2012/3b6_Gelaro.pdf .)
