Analyses & Numerical Prediction

Operations Bulletin No. 52
Operational Implementation of 1DVAR GASP in NMOC
18 August 2000

INTRODUCTION

A major change was made to the analysis of satellite sounding data on 1 August 2000 when the analysis component of the operational Global Analysis and Prediction (GASP) system was upgraded to use a one-dimensional variational radiance (1DVAR) assimilation scheme developed in the Data Assimilation Group (formerly the Medium Range Prediction Group) of BMRC over the past year. Instead of using retrievals produced by NESDIS, the new scheme analyses the cloud-cleared radiances from the TOVS (NOAA-14) and ATOVS (NOAA-15) instruments. The radiances estimated from the first guess profile of moisture and temperature are compared with the observed values and the profile adjusted using a variational scheme to give a best fit to all the observed radiances. This adjusted profile is then used as the 'retrieval' in the analysis scheme with a weight adjusted for the fact that the first guess has already been used in deriving it.

This analysis upgrade has coincided with the switch of the operational version of GASP to the new NEC SX-5 supercomputer, replacing the previous version which was run on the SX-4 and has now been discontinued.

DESCRIPTION

The new 1DVAR assimilation scheme (Harris and Steinle, 1999) has been extensively tested in BMRC for over a year and found to lead to significantly improved forecasts over the global domain. The system is based on the ECMWF 1DVAR scheme (Eyre et. al. 1993) and modified by Dr Brett Harris to improve the original formulation. A brief summary of the scheme follows.

The NESDIS TOVS data currently received via GTS have already undergone considerable preprocessing. However, there are still residual biases in the data relative to the forward radiances. These can be categorized into scan bias and air-mass bias. The scheme corrects the radiances for these biases before further processing produces retrievals.

To correct for the scan bias, radiances from the previous two weeks of data are grouped into eighteen latitude bands and mean values are found for each scan position within each band. A scan correction is calculated from these data.

To correct the air-mass bias in each channel a regression scheme is used to predict the bias using a number of predictors from the model's first guess fields. The regression coefficients for the scheme are computed using a least-squares fit on a large sample, usually around two weeks of data. Only soundings near radiosondes are used to prevent the procedure from becoming unstable, as it is assumed that the model is relatively unbiased at these locations. The procedure used to construct the bias-correction dataset ensures that the coefficients are always up to date and continually adjust to changes in instrument characteristics.

After both types of radiance bias have been corrected, an iterative retrieval of temperature and moisture profiles is performed. The background and retrieved profiles are then projected to a set of thick layers to reflect the limited information content of the radiances. These layers are 1000-700, 700-500, 500-300, 300-100, 100-50, 50-30 and 30-10 hPa. Layer thickness is computed for each layer, and precipitable water is calculated for the lowest three layers only. Background and retrieval error covariances are projected to the same layers and the lot is passed to the assimilation system where scaling factors are calculated for the retrieval errors. These vary with location and retrieval type.

The use of variable scaling factors in the assimilation scheme has a number of advantages compared to the fixed parameter approaches used elsewhere. Among these is the fact that the retrieval error will automatically adjust to the information content of the radiances. Therefore, if fewer channels are used, as is the case for a cloudy sounding, the scaling factors increase for the relevant layers, thereby giving the retrieval less weight in the analysis. Conversely, when many more channels are used from an instrument such as AMSU-A from NOAA-15, the scaling factors will automatically decrease to account for the higher information content.

With the present scheme, retrievals have been thinned to approximately 250 km to reduce problems associated with the correlated background component of the retrievals. If the data is used at full 120 km density and default radiance errors, then the first guess is given too much weight and it is found that, particularly in the tropics, the analysis runs away, increasing the lower tropospheric temperature and introducing excessive moisture in the upper troposphere.

In the present scheme, NESDIS retrievals are substituted for the 1DVAR retrieval above 100 hPa. No retrievals are used over the land below 100 hPa, so that only the NESDIS product is used over these points at present.

Apart from the move to the new supercomputer the forecast component of this new system is essentially unchanged from the previous GASP.

PERFORMANCE OF THE SYSTEM IN PARALLEL OPERATIONAL TRIALS

The 1DVAR GASP system was run on the SX-5 in parallel to operations until 31 July 2000. A comparison of a range of quantitative verification statistics of the GASP system for the period 1 June 2000 until 31 July 2000 (61 cases) is shown in Figures 1 - 4. Here the verifications for each system are relative to their own assimilation analyses and only use the 1200 UTC based forecasts for this period. Virtually all results show significant improvements for the entire forecast period over Australia, as well as over the southern hemisphere, northern hemisphere and tropical regions in general.

PRODUCT AVAILABILITY

With the changeover to this new system few obvious changes should be evident except better forecasts. Product availability remains similar to that described in Analysis and Prediction Bulletin No 45 of 11 December 1998. In recent months however, the number and frequency of GASP fields ingested into the real time database have been increased in response to user requests. These changes include:

• An increase in the frequency of the 1 x 1 high-resolution database surface fields (geometry name=global_181x360) from 6-hourly to 3-hourly.
• Addition of various extra screen level surface fields to the 2.5 x 2.5 database (geometry name=global2_73x144). These include air_temp, mix_rat, dwpt and rel_hum, all at level ht_sfc=2m,
• An increase the number and levels of upper-level fields saved at six-hourly intervals in the 2.5 x 2.5 database for the full model run (previously only saved at levels other than 1000 hPa for 24-hourly intervals past day two). These include air_temp, mix_rat, geop_ht, wind_ucmp, and wind_vcmp at isbr_lvl = 900, 850, 700, 500, 300, 200 hPa.

COMPUTATIONAL ASPECTS

The analysis code has been significantly optimized so that, in spite of the extra computation required for the 1DVAR scheme, the assimilation cycle runs considerably faster than the old code without the 1DVAR calculations. Consequently, although the assimilation cycle runs on only 4 CPUs on the SX-5 compared to 8 on the SX-4, the running time is reduced to about 30 minutes compared to the 45 minutes required on the SX-4.

FUTURE DEVELOPMENTS

Some enhancements to the GASP system currently being investigated in BMRC include:

• Increasing the number of vertical model levels to 50.
• Raising the top model level to improve the representation of the stratosphere and provide a better vertical profile for use in physical retrievals from TOVS radiances.
• Use of scatterometer winds within the GASP analysis system with the lowest model and analysis level located nearer the surface.
• Refined radiative transfer code with more sophisticated cloud optical properties.
• Development of an ensemble prediction scheme.
• 3DVAR, where a set of radiances is analysed simultaneously in a three-dimensional sense.

REFERENCES

Eyre, J.R., Kelly, G.A., McNally, A.P., Andersson, E., and Persson, A., 1993, Assimilation of TOVS radiance information through one-dimensional variational analysis. Q. J. R. Met. Soc., 119, 1427-1463

Harris, B. and Steinle, P. 1999, Variational TOVS radiance Assimilation in the GASP model at BMRC, Technical Proceedings of the tenth International TOVS Study Conference, Boulder, Colorado 27 Jan 2 Feb 1999.

Figure 1. Comparison of S1 skill scores over the Australian region for the previous operational GASP model which used NESDIS retrievals only (dashed line) and the new 1DVAR model (solid line). The comparison is based on 61 cases comprising the 1200 UTC predictions for the period 20000601 to 20000731.

a) MSLP

b) 500 hPa geopotential height

c) 200 hPa geopotential height

Figure 2 As for fig. 1 but for anomaly correlations in the Southern Hemisphere annulus

(60S-20S, 0E-360E)

a) MSLP

b) 500 hPa geopotential height

c) 200 hPa geopotential height

Figure 3. As for fig. 1 but for anomaly correlations in the Northern Hemisphere annulus

(20N-60N, 0E-360E)

a) MSLP

b) 500 hPa geopotential height

c) 200 hPa geopotential height

Figure 4. As for fig. 1 but for RMS errors in the tropics region (30S-30N, 0E-360E)

a) MSLP

b) 500 hPa geopotential height

c) 200 hPa geopotential height