NCEP/Climate Prediction Center ATLAS No. 2

Intercomparison of the NCEP/NCAR and the NASA/DAO Reanalyses (1985-1993)

2. Overview of the NCEP/NCAR Reanalysis System

The NCEP/NCAR reanalysis data assimilation system consists of the NCEP Medium Range Forecasting (MRF) spectral model (Kanamitsu 1989; Kanamitsu et al. 1991) and the operational NCEP spectral statistical interpolation (SSI; Parish and Derber 1992) with the latest improvements (Kalnay et al. 1996). For convenience, the data assimilation system is briefly described below; more detail is found in Kalnay et al. (1996). For a discussion of the processing system see Kalnay et al. (1996), section 5a.

The analysis module of the data assimilation system incorporates the spectral statistical interpolation (SSI), a three-dimensional variational analysis scheme (Parrish and Derber 1992; Derber et al. 1991). This scheme, which replaced an earlier optimal interpolation (OI) analysis scheme, has led to major improvements in analyses and forecasts, especially in the Tropics, including a substantial reduction in the precipitation spinup. Because the balance imposed on the analysis by the SSI is valid over the globe, it is unnecessary to use an initialization procedure. Recent updates to the SSI scheme, including improved error statistics and the introduction of a divergence tendency constraint, have also been included (Derber et al. 1994; Parrish et al. 1996). The SSI used in the reanalysis is the same as the system implemented operationally at NCEP in January 1995.

The NCEP/NCAR assimilation system employs the T62/28-level version of the global spectral model as implemented in the NCEP operational system in December 1994. The vertical coordinate system is the sigma coordinate of Phillips (1957). This version of the model has 7 levels below 850-hPa and 7 levels above 100-hPa. The lowest model level is about 5 hPa from the surface while the top level is at about 3 hPa. The sigma levels are distributed in the vertical to provide enhanced resolution in the planetary boundary layer and at upper levels (e.g. minimal impact of the top boundary on the stratospheric analysis at 10-hPa). The enhanced orography (Mesinger et al. 1988) used in earlier versions of the model has been replaced by mean orography (see Kanamitsu et al. 1991) in the reanalysis model.

The model includes parameterizations of all major physical processes, that is, radiation (including the diurnal cycle and interaction with clouds), convection, large-scale precipitation, shallow convection, boundary layer physics, interactive surface hydrology, gravity wave drag, and vertical and horizontal diffusion. The details of the model dynamics and physics are described in NOAA/NMC Development Division (1988), Kanamitsu (1989) and Kanamitsu et al. (1991). A major difference in the model as described by Kanamitsu et al. (1991) is the use of a simplified Arakawa-Schubert convective parameterization scheme developed by Pan and Wu (1994) based on Grell (1993). This scheme results in much better prediction of precipitation than the previous Kuo (1965,1974) scheme over the United States (as measured by equitable threat scores) and has resulted in more realistic precipitation patterns over the Tropics. A recent improvement to the model is a diagnostic cloud scheme (Campana et al. 1994), which has resulted in better model generated outgoing longwave radiation (OLR). A new soil model, based on Pan and Mahrt (1987), has yielded better surface temperatures and more skillful predictions of precipitation.

SST: Reynolds reanalysis (OI SST approach of Reynolds and Smith 1994) for 1982 on and the UKMO GISST for earlier periods.

Snow cover: NESDIS weekly analyses and climatology, updated weekly (D. Garrett 1995, personal communication).

Sea Ice: Derived from SMMR/SSMI and quality controlled by B. Nomura for the period 1979-1993 (this is the same data used in the ECMWF reanalysis). After 1993, a similar algorithm developed by R. Grumbine is used. For earlier periods, analyses from the Joint Ice Center are used, when available, and GISST analyses otherwise.

Soil Wetness: Updated during the analysis cycle using the soil model of Pan and Mahrt (1987)

Complex Quality Control (CQC) is used to check the rawinsonde heights and temperatures for consistency (Gandin 1988). CQC computes residuals from several independent checks (i.e. it computes the difference between an observation and the expected value for that observation from each check), and then uses these residuals to accept, reject or correct data (Collins and Gandin 1992). The CQC includes a number of checks used operationally, including a hydrostatic check, increment check and horizontal and vertical interpolation check. A baseline check is used to determine errors and/or changes in station location. A temporal interpolation check, not used operationally, is also done in the reanalysis; statistics show that this check is comparable to the increment check, and is particularly useful when the first guess (hence incremental check) is not available. Kalnay et al. (1996) report that, at present, about 7% of the rawinsonde observations are found to have at least one error. Roughly 75% of the hydrostatically detectable errors are corrected, while 60% of the baseline errors are corrected. While the number of corrections for early years in the reanalysis is likely to be smaller (i.e. depending on data density), there may be a higher percentage of data that needs correcting.

Optimal Interpolation Quality Control (OIQC) is a final screening procedure for observations used in the assimilation (Woollen 1991; Woollen et al. 1994). OIQC detects data with gross errors generated by instruments, humans or communications and withholds them from the assimilation. OIQC also checks for observations with large representativeness errors, i.e. observations that appear accurate but that represent spatial and temporal scales impossible to resolve in the assimilation system. Kalnay et al. (1996) state 3 principles that guide the OIQC: (1) use of multivariate 3-d statistical interpolation for "buddy" checks (i.e. to obtain comparison values for each observation from nearby neighbors), (2) use of independent quality control components which collectively suggest whether errors exist in an observation, and (3) use of a "nonhierarchical" decision making algorithm in which final decisions on data acceptance/rejection are made only after all checks are completed. Other details concerning the OIQC are given in Kalnay et al. (1993), Woollen (1991) and Woollen et al. (1994).

Before the reanalysis module (consisting of the analysis, model and quality control discussed above) is executed, the data preprocessor reformats the data coming from many different sources into the BUFR format. This allows detection of major data format problems. The preprocessor includes a special satellite TOVS soundings data monitoring system, which is intended to quality control the NESDIS archive. The Complex Quality Control (CQC) system described above is also used in the preprocessor but without the use of a first guess. An automatic monitoring system has been developed to check the reanalysis output each month using climatological tests and statistics for each month.

The Data Assimilation Office at NASA's Goddard Space Flight Center has completed a reanalysis for the period 1985-1993. Documentation of the reanalysis dataset and the assimilation system used to produce it is found in Schubert et al. (1993). An atlas which serves as a users guide to the DAO assimilation and which provides an overview of the output has been produced by Schubert et al. (1995).

The DAO reanalysis data were produced with a fixed assimilation system consisting of the Goddard Earth Observing System (GEOS-1) GCM (Takacs et al. 1994) and an optimal interpolation scheme (Pfaendtner et al. 1995). The assimilation was performed at a horizontal resolution of 2° latitude by 2.5° longitude and 20 sigma levels in the vertical, including four levels below 850-hPa.

It is important to emphasize that the convective parameterizations are different in the DAO and the NCEP models. The GEOS-1 model uses the relaxed Arakawa-Schubert scheme (Moorthi and Suarez 1992) while the NCEP model uses a simplified Arakawa-Schubert convective parameterization (Pan and Wu 1994; Grell 1993). The soil moisture in GEOS-1 is computed off-line based on a simple bucket model using monthly mean observed surface air temperature and precipitation (Schemm et al. 1992) while the NCEP model uses model-generated precipitation, a 3-layer soil model and a modified simple SiB model for the evaporation.

We examine assimilated model prognostic variables (e.g. winds and specific humidity) which are instantaneous fields available every 6 hours and various diagnostic fields (e.g. precipitation, evaporation) that were generated by the physical parameterizations. The DAO assimilation system includes an incremental analysis update (IAU) procedure (Bloom et al. 1996) which reduces initial imbalances and spinup. Prior to the IAU, fields such as precipitation were dominated by shocks due to the insertion of observations. In the DAO system, surface fields such as precipitation and evaporation were accumulated every 3 hours during the assimilation cycle, while in the NCEP system, they are based on a 6-hour forecast.

In addition to the model differences mentioned above, there are many differences in the input data (see Kalnay et al. 1996 and Schubert et al. 1993 for details). For example, the SST data set used in the DAO incorporates the monthly mean blended sea surface temperature (SST) analyses of Reynolds (1988) and Reynolds and Marsico (1993), while NCEP incorporates the improved weekly global optimum interpolation SST analyses of Reynolds and Smith (1994). The major differences between the blended and OI analyses are in the regions of high SST gradients such as the eastern Pacific (Reynolds and Smith 1994).