The Use of TOVS Cloud-Cleared Radiances in the NCEP SSI Analysis System

John C. Derber Environmental Modeling Center, National Centers for Environmental Prediction, Washington, D.C.

Search for other papers by John C. Derber in
Current site
Google Scholar
PubMed
Close
and
Wan-Shu Wu General Sciences Corporation, Laurel, Maryland

Search for other papers by Wan-Shu Wu in
Current site
Google Scholar
PubMed
Close
Full access

We are aware of a technical issue preventing figures and tables from showing in some newly published articles in the full-text HTML view.
While we are resolving the problem, please use the online PDF version of these articles to view figures and tables.

Abstract

With improved assimilation techniques, it is now possible to directly assimilate cloud-cleared radiances, rather than temperature and moisture retrievals, in objective analyses. The direct use of the cloud-cleared radiances became the operational technique for using satellite sounding data at the National Centers for Environmental Prediction (NCEP) in October 1995. The methodology for using the data (including bias correction, ozone analysis, skin temperature analysis, and quality control) are described in this paper. The impact of the direct use of the radiances compared to the previously operational use of satellite sounding data shows considerable improvement in NCEP’s forecast skill, especially in the Southern Hemisphere. It is anticipated that additional positive impacts will occur with application of the technique to other remotely sensed data.

Corresponding author address: Dr. John C. Derber, National Centers for Environmental Prediction, W/NP23, World Weather Building, Washington, DC 20233.

Abstract

With improved assimilation techniques, it is now possible to directly assimilate cloud-cleared radiances, rather than temperature and moisture retrievals, in objective analyses. The direct use of the cloud-cleared radiances became the operational technique for using satellite sounding data at the National Centers for Environmental Prediction (NCEP) in October 1995. The methodology for using the data (including bias correction, ozone analysis, skin temperature analysis, and quality control) are described in this paper. The impact of the direct use of the radiances compared to the previously operational use of satellite sounding data shows considerable improvement in NCEP’s forecast skill, especially in the Southern Hemisphere. It is anticipated that additional positive impacts will occur with application of the technique to other remotely sensed data.

Corresponding author address: Dr. John C. Derber, National Centers for Environmental Prediction, W/NP23, World Weather Building, Washington, DC 20233.

1. Introduction

The impact of conventional satellite retrievals on the quality of numerical weather prediction forecasts has been mixed. Although positive impact has been found in the Southern Hemisphere, it has not been possible to show a consistent positive impact in the Northern Hemisphere (e.g., Tracton et al. 1980; Halem et al. 1982; Andersson et al. 1991; Mo et al. 1995). With the recent improvement in both the quality of the analysis and the forecast model, it has become even more difficult to show a positive impact on the forecasts. Substantial improvements in the use of the data have become necessary to enhance the impact from the satellite sounding data.

The theory of data assimilation allows the inclusion of observational information if one can transform the analysis variables into the same form as the observations. This procedure differs from the traditional practice of transforming the observations into analysis variables. To include sounding information from satellites, radiances can be used directly rather than attempting to retrieve temperature and moisture profiles before use in the analysis.

With the development of analysis systems such as the National Centers for Environmental Prediction’s (NCEP, formerly the National Meteorological Center) spectral statistical interpolation (SSI, Parrish and Derber 1992; Derber et al. 1991) and European Centre for Medium-Range Weather Forecasts’ (ECMWF) three-dimensional variational (Courtier et al. 1993) analysis systems, the incorporation of the radiances directly in an analysis and assimilation system has become practical in an operational environment. The analysis then is essentially a 3D retrieval of the mass, momentum, and moisture fields derived from all available data including the radiances. In the 3D analysis (or retrieval), a very accurate background field (a short-term forecast), and all other data (surface temperatures, radiosonde data, radiances from different satellites, etc.) are available. Modern 3D variational analysis systems blend this information naturally and extend the influence of the data in both the horizontal and the vertical.

The 3D analysis problem and the temperature retrieval problem are both underdetermined with the data insufficient to determine all the degrees of freedom. Both problems have been made well posed by introducing a background (or guess) field, which is used to infer those degrees of freedom not present in the data. One of the problems with the traditional technique of producing a 1D retrieval and then incorporating the temperature and moisture profile in the analysis is that often different background fields and background error statistics were used. Because of the inconsistency of the background fields, information was being incorporated in the analysis that came not from the observations but from the background used in the retrieval. Since the background field and complete observational error information usually were not available for the retrievals, a degradation of the analysis would result. Several institutions have attempted to get around this particular problem by using the analysis background as the retrieval background. However, this scheme [known as interactive retrievals or 1D variational retrievals (Baker et al. 1995; Eyre et al. 1993)] still does not properly account for the 3D error structure in the background errors and can result in difficulties because of correlated errors between the retrievals and the analysis background fields. Direct use of the radiances overcomes these difficulties. It is still necessary however to properly define the observational and representativeness errors for the radiances.

On 25 October 1995, the direct use of radiances replaced the use of National Environmental Satellite, Data and Information Service (NESDIS) retrievals (Southern Hemisphere) and interactive retrievals (Northern Hemisphere) in the operational NCEP global analysis and assimilation systems. A similar system for using radiance data became operational at the ECMWF when their 3D variational analysis system (Courtier et al. 1993) was implemented on 30 January 1996. In this paper, a description of the operational use of cloud-cleared radiances in the NCEP analysis and assimilation system will be described. The basic theory for the use of the data will be presented in section 2. In the following four sections, four important components in the use of the data (bias correction, the ozone analysis, skin temperature analysis, and radiance quality control) will be discussed. This will be followed by some results from the direct use of radiances in section 7.

2. Theory

The SSI analysis system (Parrish and Derber 1992; Derber et al. 1991) produces an analysis through the minimization of an objective function given by
i1520-0493-126-8-2287-e1
where x is the analysis variable, xb is the background field (a 6-h forecast), y is a vector of all the observations, O is the observational and representativeness error covariance matrix (Lorenc 1986), and K is a transformation operator from the analysis variable to the space of the observation vector. The Jc term is a dynamical constraint term included in the procedure to increase the balance in the analysis increment. Since this term is small and has no impact on the basic discussion, for clarity’s sake we will eliminate it in the following discussion. However, the Jc term is included in the analysis used to create all results.

The analysis variable (x) is defined in a space where the background error covariance is the identity matrix. If this definition were not true, the first term in (1) would be given by [xxb]TB−1[xxb], where B is the background error covariance matrix. However, by incorporating the background error covariance in the definition of x, the conditioning is considerably improved and the convergence of the minimization algorithm is enhanced. For this reason, the analysis variable (x) is defined such that the following series of operations will result in the model variables (vorticity, divergence, virtual temperature, surface pressure, and specific humidity) on the model’s Gaussian grid.

  1. Horizontally, the analysis variable (x) is defined in spectral space and in globally defined vertical modes. The spectral amplitudes of x are multiplied by the inverse of the square root of the background error variances. The background error variances, which vary by wavenumber and vertical mode, are fixed in time and estimated from scaled differences between 45 24-h and 48-h forecasts valid at the same time (see Parrish and Derber 1992; Rabier 1997, personal communication).

  2. The analysis variable is projected from vertical mode space to the model-level space. The vertical mode space was defined from global vertical error covariance matrices estimated from the same 45 24–48-h forecast differences.

  3. The solution from 2 is transformed into the model variables through a statistically modified linear balance equation. The balance equation implies a balanced part of the mass field (temperature and surface pressure) and divergence from the streamfunction. The coefficients in the balance equation use statistics from the 45 24–48-h forecast difference fields.

  4. A spectral to grid transformation to the model’s Gaussian grid.

The transformation from the analysis variables to the model variables described above represents the first half of the K operator, which is applied for all types of data. The second half is the creation of pseudo-observations at the observation location for each type of data. For many observation types, this creation of pseudo-observation requires only a bilinear or trilinear interpolation to the observation locations. For the radiance data, the transformation is more complex. The temperature and moisture on the Gaussian grid are bilinearly interpolated in the horizontal to the observation location to create a temperature and moisture profile. The radiative transfer code currently being used (RTTOV, Eyre 1991) requires the vertical temperature and moisture profile on fixed pressure surfaces, which results in an additional interpolation of the profiles in the vertical to 40 pressure surfaces for the temperature and 15 pressure surfaces for the moisture. Above the top of the model (about 2.7 mb), the NESDIS operational retrieval is used for the background profile.

In addition to the vertical profile, the integration of the radiative transfer equation also requires values for the surface skin temperature and total ozone. The surface skin temperature predicted by the forecast model over land and the global SST analysis (Reynolds and Smith 1994) over oceans are used as the background surface skin temperature. The total ozone is taken from the analysis described in section 4. The surface skin temperature, the total ozone, the 40-level temperature and 15-level moisture profiles are used in the RTTOV radiative transfer code described in Eyre (1991) to produce a simulated observation. Note that since the radiances provided by NESDIS are cloud cleared, limb corrected, and surface emissivity corrected, no information on the cloudiness, local zenith angle, or surface emissivity have to be supplied to the radiative transfer code even though these degrees of freedom are allowed.

Thus, the final components of the K operator (after transforming the analysis variables to the model variables on the grid) include interpolation of the model variables to the 40 levels of temperature, 15 levels of moisture, surface pressure, skin temperature, and ozone at the observation locations, and integrating the radiative transfer equation. To find the solution to the analysis equation, (1) is differentiated with respect to x, resulting in
i1520-0493-126-8-2287-e2
where KT is the adjoint of the linearization of the K operator. Note again the Jc term in (1) has been dropped for clarity. For the radiances, K includes all of the previously mentioned components of the K operator except we have assumed the surface pressure at the sounding location does not change in the analysis. We believe the effect of this approximation is quite small. It is often more convenient to work in terms of differences from the background field. Defining x′ = xxb (2) becomes
i1520-0493-126-8-2287-e3
If the analysis corrections are small and the radiative transfer is not very nonlinear, then the linear approximation that K(x)K(xb)Kx′ is valid where K is a linearization around xb. This approximation appears to be valid for the High-resolution Infrared Soundor (HIRS) temperature channels but not completely for the more nonlinear water vapor channels. However, without this approximation, it is necessary to update K during each iteration of the solution procedure. At NCEP, the operational requirements allow only 30 min for the full resolution analysis (including quality control and other initial data preprocessing). Currently, the recalculation of K in each iteration would require too much computer time to allow it to vary in every iteration and still meet operational requirements unless the number of soundings were reduced. For this reason, the approximation was assumed to hold and the water vapor channels were given reduced weight (observation error increased) to account for the errors resulting from the approximation. Further experimentation is being performed to quantify and improve this approximation.
Applying this approximation, (3) becomes
i1520-0493-126-8-2287-e4
At the minimum to J, the derivative of J with respect to x is equal to zero. The minimum to (1) could be found by setting the derivative (4) equal to zero and performing a very large matrix inversion. However, this is not feasible on current computers. Therefore (4) was solved using a simple linear conjugate gradient scheme (e.g., Gill et al. 1981). Note that the third term on the right-hand side is not a function of x′ and thus represents a forcing of the solution away from x′ = 0 by the differences of the observations from the background (the short-term forecast from 6 h previous). The K operator for the radiance data is produced from the previously mentioned RTTOV code and is saved for use in Eq. (4).

The error covariance matrix, O, plays a vital role in (4), specifying the degree of confidence in the various components of the observations. This matrix should not only contain information on the observational error but also errors in representativeness (Lorenc 1986). Thus, this matrix includes the error in the radiative transfer modeling. The specification of this matrix is difficult. It is clear that the errors are probably correlated spatially because of the errors in the radiative transfer, instrument errors and errors arising from imperfect cloud clearing, emissivity correction, and other components. However, these correlations are probably quite different from the spatial correlations found in the temperature and moisture retrievals and are currently not well known. For this reason, we have chosen these errors to be spatially uncorrelated. Note that the horizontally correlated error is very closely related to the bias correction discussed in section 3. Also, because the interchannel error correlations are not known, we have set them equal to zero.

The error standard deviations (in terms of brightness temperatures) currently being used in the NCEP operational assimilation system and in this paper for the NOAA-12 and NOAA-14 are given in Table 1. These errors were estimated by first finding the fit of the radiances calculated from the background field to the observations and then reducing the values of these errors until the quality of the resulting forecast field is maximized. These observation errors are modified empirically before use in the analysis in several situations. First, because we are using the full 120-km resolution data, the observational errors are increased by a factor of 2 to ensure the radiance data does not overwhelm the other data. Then in cloudy situations, the errors assigned to HIRS channels 4–19 are increased to infinity (partly cloudy by a factor of 2). If the observation is over ice or snow, tropospheric channels increase by 2 and if over land by a factor of 3. Other factors such as high topography, a misfit of the model, and the real topography or the presence of coastlines also result in small modifications of the observational variances.

3. The cloud-cleared radiance data and bias correction

The cloud-cleared radiance data, supplied by NESDIS, is identical to that used in the NESDIS operational retrievals. The radiance information is supplied and used in terms of brightness temperature with units of degrees kelvin. These data have undergone substantial preprocessing by NESDIS before becoming available for usage (Smith et al. 1979; McMillin and Dean 1982; Reale et al. 1986). The data is from three instruments, the HIRS, the microwave sounding unit (MSU), and the stratospheric sounding unit (SSU). The MSU and SSU field of views are interpolated to the HIRS field of views by NESDIS. The data has been statistically limb corrected and (adjusted to nadir) surface emissivity corrected in the microwave channels and cloud cleared in the tropospheric HIRS channels. Also, HIRS channel 8 has been corrected in an attempt to remove the moisture signal. We have used the full 120-km resolution data using all clear, cloudy, and partly cloudy soundings that pass the quality control (section 6). Note that the radiance data was used over both the ocean and land. However, the quality control over the land was made more strict and the observational errors increased.

When the observed quantities are compared to the model-simulated values, substantial biases are noted [Eyre et al. (1993) and references therein]. The source of the biases could be from instrument calibration problems, the ground processing of the data (e.g., cloud clearing, correction to nadir, etc.), inadequacies in the forward modeling (e.g., interpolation or radiative transfer errors), or biases in the model fields. Closer examination of the data indicates that most of the errors are coming from the ground processing and the radiative transfer errors. The resulting biases are not constant for each channel, but rather spatially varying and dependent on other satellite parameters.

For example, a comparison between the observations and the model-simulated values as a function of the local zenith angle (the distance for nadir as the instrument scans across the path) is shown in Fig. 1. (Note that in the dataset currently available from NESDIS, the true local zenith angle is not available, but rather only the bin number with 18 bins, 9 on either side of nadir.) Since the brightness temperatures are corrected to nadir (local zenith angle = 0) and there is no reason the model simulated values should have a local zenith angle dependence, it is clear that this error must be introduced in NESDIS’s limb correction procedure. In some channels, the magnitude of the local zenith angle bias is comparable or larger than the signal in many of the channels. Of course, the best way to account for the bias is to remove all satellite calibration problems, remove all ground processing problems, and improve the radiative transfer. However, this is not feasible in the short term. Practically, the effects of the spatially dependent biases in the observations can be controlled in two different ways. First, by including a spatially dependent correlation in the observational (and representativeness) error covariance matrix. Second, the bias can be estimated at the observation locations and removed from the data. We have chosen the second of these options and describe the scheme in the remaining part of this section.

The scheme for predicting the bias is a simple linear equation for each satellite and each channel based partially on Eyre (1992). As predictors of the bias, we have chosen scaled values of a constant term, MSU channels 2–4, HIRS channel 1, the solar zenith angle, the approximate local zenith angle (the mean of the angle for the bin), and the square of the approximate local zenith angle. These predictors are multiplied by a set of coefficients to produce the bias correction. The coefficients are created by augmenting the analysis vector [x in (1)] with the bias correction coefficients and solving for them along with the rest of the analysis variables. By specifying the coefficients in this manner, the imperfect specification of the coefficients does not override the information contained in the other data used in the analysis and the system can adjust to changes in the data or processing fairly quickly. Also, the system is greatly simplified, eliminating the need for creating large files containing collocated radiosondes and radiances. The disadvantage of this method is the possibility that the model bias will become incorporated in the coefficients and feed back into the system and amplify the model bias. We do not believe such a feedback is likely since the coefficients are global, dependent only on the channel and the predictors. No evidence of this feedback has been noted to date in the operational implementation of the system since October 1995.

By making the coefficients analysis variables, it is also necessary to specify the background error variances. Initially these background error variances have been specified in a simple manner by scaling the predictors in a manner that makes the variance of the various predictors approximately equal (except the constant predictor) and then applying a constant diagonal matrix for the background errors. The constant used in the background errors gives fairly large weight to the values from the previous analysis. Figure 2 shows a time sequence of the coefficients for the square of the local zenith angle predictor (the strongest predictor of the bias) for a few channels of NOAA-14. The coefficients are fairly stable with time but occasionally show a diurnal cycle. The source of this diurnal cycle is currently unknown.

Figure 3 shows a comparison of the standard deviations and biases for the same global set of observations with and without bias correction. The standard deviation and the bias are both reduced since the bias correction varies spatially. However, the reduction in the standard deviation is small. The standard deviation for most channels is quite small, less than 1° in terms of brightness temperature. (No quality control has been performed on this data except the removal of tropospheric HIRS channels when specified as cloudy profiles by NESDIS and the removal of sounding locations over land and over ice. With the inclusion of the quality control described in section 6, the variance was further reduced.) The importance of including the local zenith angle dependence as a predictor can be seen by comparing the results in Fig. 3 to those in Fig. 1. The local zenith angle–dependent component of the error is a significant fraction of the signal after bias correction. Without the inclusion of the local zenith angle predictors, the analysis would extrapolate the local zenith angle–dependent biases between the satellite passes producing significant synoptic-scale errors in the analysis.

4. A simple ozone analysis

The HIRS longwave channels are affected by the total ozone (e.g., Cox 1994). The effect is not trivial with the change in the HIRS channel 6 calculated brightness temperature reaching 0.6°C in tropical profiles per 100 Du (Dobson units). Since local differences from the global mean can easily reach 100 Du, comparing this signal to the values in Table 1 for channel 6 shows this is not a trivial component of the error. Other groups (e.g., Andersson et al. 1994) have chosen to use a constant ozone value and allow the bias correction scheme to account for the resulting error in the radiative transfer calculation. We have chosen instead to provide a simple total ozone analysis for the radiative transfer calculations.

In theory, the ozone analysis could be created by including the total ozone in the control variable, x. Then, with the inclusion of HIRS channel 9 in the observations, and the inclusion of appropriate statistics and bias correction, the ozone analysis would be created with the rest of the analysis. Unfortunately, at the time of this work, the ozone was not properly incorporated in our version of the radiative transfer for channel 9. Therefore, we have instead created a simple independent analysis of the total ozone. This analysis is probably less accurate than that which could be created by analyzing everything together. Because the proper radiative transfer has been developed since the completion of this work, a 3D ozone analysis will be completely integrated with the temperature and moisture analysis.

The current total ozone analysis is essentially a 2D univariate version of the SSI analysis system. The ozone values are supplied by the operational NESDIS TIROS-N Operational Verticle Sounder total ozone retrievals. Note that the total ozone retrievals are created only in locations thought to be clear.

The 2D analysis is produced independently from the mass, momentum, and moisture analysis by minimizing the objective function given by
i1520-0493-126-8-2287-e5
where Z represents the total ozone analysis, Zb is the background field, yz is the vector of ozone observations, Fz is the observation error covariance matrix, Bz is the background error covariance matrix, and K is a linear transformation of the observations to the observation location. The background field, Zb, is provided by the previous analysis. Since the Z and Zb vectors are defined in spectral space, the background error covariance matrix is also defined in spectral space. The background error covariance matrix is defined as a first-order autoregressive function with a length scale of 400 km and a variance equal to the observational error variances (10 Du2).

Also included in the analysis is a simple bias correction, which attempts to make the mean correction between the two polar satellites (currently NOAA-12 and NOAA-14) approximately the same and to account for a local zenith angle–dependent error. There is also a simple quality control that includes a gross check, a maximum skin temperature check, and a retrieved-model predicted skin temperature check. The maximum skin temperature check is intended to remove ozone values over very hot deserts where retrieval errors are common. The retrieved-model predicted skin temperature check is intended to remove cloud contaminated retrievals.

The ozone analysis is performed before the rest of the analysis. Therefore at those locations where there are retrieved values of ozone, the analysis should be most accurate. Since clear locations are where the HIRS longwave channels are most important, the analysis should be most accurate where it is needed the most. For other applications, further enhancement of the analysis procedure will be necessary.

After allowing the ozone analysis spinup over a sufficiently long time (greater than two weeks), the standard deviation of the fit of the background field to the data stabilized at about 12 Du. An example of the analysis after more than six months of assimilation is shown in Fig. 4. Note the lower values in the Tropics, the synoptically correlated values in the midlatitudes and the ozone hole around the South Pole.

5. Skin temperature analysis

The accurate specification of the (apparent) surface skin temperature can be extremely important in the estimation of the simulated brightness temperatures. Many of the lower tropospheric channels have a significant contribution from the surface. If this contribution from the surface is not properly included, there will be an aliasing of the error in the vertical, which can result in a small-scale oscillation in the resultant temperature profile. For this reason, we have included a surface skin temperature variable in the control variable, x.

Note that the surface skin temperature is not necessarily representative of the actual skin temperature, since the HIRS channels that observe the ground will be usable only when it is clear. Also, errors in the correction of the microwave surface emissivity could result in a change in the apparent surface skin temperature. Finally, when some of the HIRS channels are slightly cloud contaminated and the data makes it through the cloud checks and quality control, this field can act to absorb the cloud contamination without strongly affecting the atmospheric analysis. For all of these reasons, it was decided that this field currently should not be introduced into the forecast model as a surface skin temperature, but would remain only an analysis variable.

The surface skin temperature component of the control variable was defined spectrally as the difference from the background field. The background field is produced by the forecast model over land and ice and the sea surface temperature analysis (Reynolds and Smith 1994) over the ocean. The covariance structure was defined as the same first-order autoregressive function used for the ozone and with an error variance equal to 6°C2. The background error is assumed independent from the other analysis variables. This covariance function is far from perfect. It is clear that the error variance should vary substantially over the globe and with significant contrasts between land and sea. An example of the difference between the analysis and background skin temperatures is given in Fig. 5.

6. Quality control

Quality control is vital for the use of any type of data. The presence of a single data point containing large errors can result in substantial degradation of the analysis and subsequent forecast. A rather simple quality control has been developed and the various parameters adjusted empircally for use with the brightness temperature data. Without a doubt, the quality control routines are rejecting some good data. However, it is better to err on the side of rejecting too many observations as opposed to too few.

The quality control is performed by a combination of two tests, a gross check and a check against the predicted values from nearby observations. For each observation location, an observation quality parameter is set based on the expected observational error variance for that channel. This quality control parameter is modified by the position across the track of the scan, whether it is over land, sea, snow, sea ice or a transition region, the elevation, the difference between the model and the real orography, and the latitude (the criterion is made tighter in the Tropics). For the HIRS channels, it is made more strict if it is designated partly cloudy by NESDIS, the differences between the window channel observations or the simulated window channels are too large, and for the shortwave channels if the solar zenith angle is small. Note that some of the modifications are to eliminate observations that are contaminated, and some are to eliminate situations where the simulated observations are deficient.

The observation quality parameter is then compared to the difference between the true and simulated observations. At the same time, it is compared to a simple interpolation of nearby observational increments to the observation location. The observation is rejected if the difference from the simulated observation is greater than three times the observation quality parameter or if the difference from the simulated observation and the difference between the interpolated increments and the observation increment are both greater than the observation quality parameter. To minimize the number of observations that are rejected due to errors in nearby observations, the checking is repeated three times with only the observations that have passed the previous time through the quality control check being used for the interpolation of increments. This procedure allows the reacceptance of some observations that were rejected in the previous passes through the data.

Note that the rejections were performed independently for each channel. Thus, the rejection of one lower tropospheric HIRS channel because of possible cloud contamination does not necessarily result in the rejection of other tropospheric HIRS channels at the same location. This quality control procedure has some known deficiencies and probably rejects some good data in order to ensure the removal of all the bad data. Further improvement of the quality control procedures is planned.

7. Results

The direct use of radiances has been extensively tested in the NCEP data assimilation system. The results were quite good, outperforming the previously operational use of satellite data (Baker et al. 1995) by a substantial amount in all tests. At the end of these experiments, the direct use of radiance became the scheme for using TOVS information in the NCEP operational data assimilation system (Caplan et al. 1997). Since the final tests prior to implementation included other changes (e.g., a change in the analysis of the divergent component of the wind and alterations to the model physics), we have rerun a three-week period with only those changes described in this paper. These reruns were produced using the direct use of cloud-cleared radiances (RAD), NESDIS operational retrievals in the Northern Hemisphere and interactive retrievals in the Southern Hemisphere (RET, the previous operational configuration), doubling the weight (dividing the observational error variance by the square root of 2) given to the retrievals (RET2), and with no satellite sounding data (NOSAT).

Note that it is impossible to produce a clean comparison between the use of retrievals and radiances. It is not possible to make the observational errors equivalent, since the radiance errors are convoluted through the radiative transfer equation and the guess profiles used by NESDIS are not available. The simple specification of the observational errors in terms of the radiances is one of the major reasons for using the radiances directly. Also, it is difficult to use exactly the same datasets because different quality control decisions are being made for the two sets of data.

To examine the effect of the use of RAD and RET on the analyses, the RAD analysis and the RET analyses are differenced from the NOSAT analysis for the last analysis time of the assimilation period (Figs. 6 and 7). For the purpose of these figures only, all three analyses were produced using the same background field (from the RET assimilation). Thus, the results differ only because of the different use of the data and the different quality control decisions. Figures 6 and 7 show the temperature differences for σ levels 1 and 13 (approximately the surface and 500 mb). Note the larger differences in the radiance assimilation than when the retrievals were used. Larger changes can be introduced using the retrievals by decreasing the observational errors. However, experiments performed in which the observational error was decreased showed only small changes in the forecast skill (see below).

An indication of the quality of the analysis can also be found from the fit to the data. In Table 2, the fits to the brightness temperatures for microwave channels 2 and HIRS channel 6 for the background field (6-h forecast) from the RET, RAD, and NOSAT assimilation, and analyses produced from the RET background field using the radiances, retrievals, retrievals with double the weight, and no satellite data are shown. From the comparisons to the analyses, the improvement to the background is small except when the radiances are used directly. The fit to the data does not improve significantly when additional weight is given to the retrievals. By comparing fits of the background fields to the data, it can be seen that the experiment directly using the radiances is also retaining the information in the radiances resulting in a better fit to the radiance observations.

Last, the final measure of the quality of the system can be evaluated by comparing forecasts from each system. Five-day forecasts were produced for each day of the three-week period. Each set of forecasts were verified against their own analyses. Little difference could be seen after day 3 whether the forecasts were verified against the RET or RAD assimilation. The average 500-mb anomaly correlations and rms errors are shown in Fig. 8 for the Northern Hemisphere and in the Southern Hemisphere in Fig. 9. Similar results can be found at other levels. The Northern Hemisphere results are slightly improved by the use of the retrievals with an additional improvement of the same or slightly larger magnitude by using the radiances directly. In the Southern Hemisphere, the improvement is much larger by the introduction of the retrievals and radiances. By five days, the forecast skill from the radiances is approximately equal to a 4.5-day forecast from the retrievals. The improvement from using radiances instead of retrievals is of at least the same magnitude as that produced from using retrievals instead of no satellite sounding data. In the short range, the verification of the forecasts produced using radiances is slightly worse. This is probably due to increased spatial variance in the analyses and forecasts. Increasing the weight given the retrievals results in a degradation of the results to the level of the NOSAT case in the Northern Hemisphere and no improvement in the Southern Hemisphere.

8. Conclusions

Advances in data assimilation now allow many observation types to be incorporated directly into the analysis system. The direct use of TOVS cloud-cleared radiances described in this paper is one example. By directly using the radiances, the ill-posed retrieval step before the analysis and inconsistencies between the retrieval and analysis can be eliminated. This more consistent use of the data allows the extraction of more useful information from the data and results in improved numerical predictions.

Many of the requirements for the preprocessing of the data (e.g., limb correction and emissivity correction) are no longer necessary within the context of the direct use of radiances. By bypassing many of the preprocessing steps, some of the errors introduced by the preprocessing can be eliminated and more information extracted from the data. The direct use of the raw data is currently under development at NCEP and other numerical weather prediction centers.

The use of cloud-cleared radiance data required the development of adequate radiative transfer models, quality control procedures, bias correction procedures, and skin temperature and ozone analyses. Each of these components and the cloud-cleared radiance data could be further improved. The radiative transfer contains significant errors including biases in the moisture channels. The quality control procedure is probably rejecting good data and requires much improvement. The bias correction appears to be working quite well on large-scale biases, but it would certainly be better if the biases were eliminated from the data, radiative transfer, and interpolations. Finally, the skin temperature analysis and the ozone analysis are currently quite crude and could certainly be enhanced. The skin temperature analysis is occasionally a sink for improperly cloud-cleared radiances. The statistics used in the skin temperature analysis certainly could be enhanced with an inclusion of a surface type–dependent background error covariance. In the future, additional information from other satellite- and ground-based instruments will be included in the ozone analysis to produce a 3D ozone analysis. This will require considerable enhancement of the radiative transfer and background error statistics.

The results from this study are very encouraging for using satellite observations directly in the analysis system. The impact of directly using the radiances on the NCEP forecasts was one of the largest in recent history (Caplan et al. 1997). There are many additional sources of data that could be incorporated in this manner (e.g., radiances from the DMSP SSM/I, T and T2 instruments, radiances from GOES-8 and -9 sounders, and imagery from various satellites). The most important difficulties in using these data are the availability of a fast radiative transfer code and its gradient; the biases in the data and radiative transfer calculation; the presence of many unrepresentative observations, which must be removed by quality control; and the availability of the data in a consistent, easily accessible manner. Addressing each of these difficulties for each set of satellite data requires considerable effort and resources. However, given the substantial resources expended on developing, launching, and maintaining the instruments and satellites, the additional resources required to use the information in the satellite radiance data to its fullest potential is small.

Acknowledgments

The authors would like to thank R. Saunders, T. McNally, E. Kalnay, and the anonymous reviewers for their useful suggestions concerning this paper. The authors would also like to thank M. Goldberg and L. McMillin for their input during the course of this research. Finally we would like to acknowledge J. Eyre and ECMWF for making the essential RTTOV radiative transfer software package available for use at NCEP.

REFERENCES

  • Andersson, E., A. Hollingsworth, G. Kelly, P. Lönnberg, J. Pailleux, and Z. Zhang, 1991: Global observing system experiments on operational statistical retrievals of satellite sounding data. Mon. Wea. Rev.,119, 1851–1864.

  • ——, J. Pailleux, J.-N. Thépaut, J. R. Eyre, A. P. McNally, G. A. Kelly, and P. Courtier, 1994: Use of cloud-cleared radiances in three/four-dimensional variational data assimilation. Quart. J. Roy. Meteor. Soc.,120, 627–653.

  • Baker, W., H. Fleming, M. Goldberg, B. Katz, and J. Derber, 1995: Development and implementation of satellite temperature soundings produced interactively. NWS Tech. Procedures Bulletin 422, 8 pp. [Available from Office of Meteorology, National Weather Service, 1325 East–West Highway, Silver Spring, MD 20910.].

  • Caplan, P., J. Derber, W. Gemmill, S. Y. Hong, H.-L. Pan, and D. Parrish, 1997: Changes to the 1995 NCEP operational MRF model analysis/forecast system. Wea. Forecasting,12, 581–594.

  • Courtier, P., and Coauthors, 1993: Variational assimilation at ECMWF. ECMWF Tech. Memo. 194, 84 pp. [Available from European Centre for Medium-Range Weather Forecasts, Shinfield Park, Reading, Berkshire R62 9AX, United Kingdom.].

  • Cox, S. J., 1994: Prediciton of ozone from meteorological parameters for use in GLOSS and UV index forecasting. UKMO Forecasting Research Division Tech. Rep. 85, 23 pp. [Available from Forecasting Research Division, Meteorological Office, London Road, Bracknell, Berkshire RG12 2SZ, United Kingdom.].

  • Derber, J. C., D. F. Parrish, and S. J. Lord, 1991: The new global operational analysis system at the National Meteorological Center. Wea. Forecasting,6, 538–547.

  • Eyre, J. R., 1991: A fast radiative transfer model for satellite sounding systems. ECMWF Tech. Memo. 176, 28 pp. [Available from European Centre for Medium-Range Weather Forecasts, Shinfield Park, Reading, Berkshire R62 9AX, United Kingdom.].

  • ——, 1992: A bias correction scheme for simulated TOVS brightness temperatures. ECMWF Tech. Memo. 186, 28 pp. [Available from European Centre for Medium-Range Weather Forecasts, Shinfield Park, Reading, Berkshire R62 9AX, United Kingdom.].

  • ——, G. Kelly, A. P. McNally, E. Andersson, and A. Persson, 1993:Assimilation of TOVS radiances through one dimensional variational analysis. Quart. J. Roy. Meteor. Soc.,119, 1427–1463.

  • Gill, P. E., W. Murray, and M. H. Wright, 1981: Practical Optimization. Academic Press, 401 pp.

  • Halem, M., E. Kalnay, W. E. Baker, and R. Atlas, 1982: An assessment of the FGGE satellite observing system during SOP-1. Bull. Amer. Meteor. Soc.,63, 407–429.

  • Lorenc, A. C., 1986: Analysis methods for numerical weather prediction. Quart. J. Roy. Meteor. Soc.,112, 1177–1194.

  • McMillin, L. M., and C. Dean, 1982: Evaluation of a new operational technique for producing clear radiances. J. Appl. Meteor.,21, 1005–1014.

  • Mo, K. C., X. L. Wang, R. Kistler, M. Kanamitsu, and E. Kalnay, 1995: Impact of satellite data on the CDAS–reanalysis system. Mon. Wea. Rev.,123, 124–139.

  • Parrish, D. F., and J. C. Derber, 1992: The National Meteorological Center’s spectral statistical interpolation analysis system. Mon. Wea. Rev.,120, 1747–1763.

  • Reale, A. L., D. G. Gray, M. W. Chalfant, A. Swaroop, and A. Nappi, 1986: Higher resolution operational satellite retrievals. Preprints, Second Conf. on Satellite Meteorology/Remote Sensing and Applications, Williamsburg, VA, Amer. Meteor. Soc., 16–19.

  • Reynolds, R. W., and T. M. Smith, 1994: Improved global sea surface temperature analyses. J. Climate,7, 929–948.

  • Smith, W. L., H. M. Woolf, C. M. Hayden, D. Q. Wark, and L. M. McMillin, 1979: The TIROS-N Operational Vertical Sounder. Bull. Amer. Meteor. Soc.,60, 1177–1187.

  • Tracton, M. S., A. J. Desmarais, R. J. van Haaren, and R. D. McPherson, 1980: The impact of satellite soundings on the National Meteorological Center’s analysis and forecast system—The Data Systems Test results. Mon. Wea. Rev.,108, 543–586.

Fig. 1.
Fig. 1.

Mean background fit to brightness temperatures as a function of bin number for local zenith angle. Only clear brightness temperatures not over land or ice are used in statistics.

Citation: Monthly Weather Review 126, 8; 10.1175/1520-0493(1998)126<2287:TUOTCC>2.0.CO;2

Fig. 2.
Fig. 2.

Coefficient of square of local zenith angle as estimated by the analysis system as a function of time (1200 UTC 18 September 1996–0600 UTC 28 September 1996) for NOAA-14.

Citation: Monthly Weather Review 126, 8; 10.1175/1520-0493(1998)126<2287:TUOTCC>2.0.CO;2

Fig. 3.
Fig. 3.

(a) Standard deviation and (b) bias with and without bias correction for NOAA-14. Only clear brightness temperatures not over land or ice are used in statistics.

Citation: Monthly Weather Review 126, 8; 10.1175/1520-0493(1998)126<2287:TUOTCC>2.0.CO;2

Fig. 4.
Fig. 4.

Total ozone analysis (in Dobson units) for 16 September 1996.

Citation: Monthly Weather Review 126, 8; 10.1175/1520-0493(1998)126<2287:TUOTCC>2.0.CO;2

Fig. 5.
Fig. 5.

Skin temperature analysis increments for 23 August 1995.

Citation: Monthly Weather Review 126, 8; 10.1175/1520-0493(1998)126<2287:TUOTCC>2.0.CO;2

Fig. 6.
Fig. 6.

RAD-NOSAT (a) and RET-NOSAT (b) analysis differences for the analysis level closest to the surface from 23 August 1995.

Citation: Monthly Weather Review 126, 8; 10.1175/1520-0493(1998)126<2287:TUOTCC>2.0.CO;2

Fig. 7.
Fig. 7.

Same as Fig. 6 except for sigma level near 500 mb.

Citation: Monthly Weather Review 126, 8; 10.1175/1520-0493(1998)126<2287:TUOTCC>2.0.CO;2

Fig. 8.
Fig. 8.

Northern Hemisphere anomaly correlations (a) and rms errors (b) for 5-day forecasts from NOSAT, RAD, RET, and RET2 experiments. Results are averages over 21 forecasts (29 July 1995∼18 August 1995).

Citation: Monthly Weather Review 126, 8; 10.1175/1520-0493(1998)126<2287:TUOTCC>2.0.CO;2

Fig. 9.
Fig. 9.

Same as in Fig. 8 except for Southern Hemisphere.

Citation: Monthly Weather Review 126, 8; 10.1175/1520-0493(1998)126<2287:TUOTCC>2.0.CO;2

Table 1.

Basic observational error standard deviations assigned to observed brightness temperatures (in °C). Note that these errors are modified based on various factors before being used in analysis (see text).

Table 1.
Table 2.

(a) Fit of observations to background (°C) after 3 weeks assimilation. Fit of observations to analysis (°C) using same background field (from RET assimilation) (b). Only clear brightness temperatures not over land or ice are used in statistics.

Table 2.
Save
  • Andersson, E., A. Hollingsworth, G. Kelly, P. Lönnberg, J. Pailleux, and Z. Zhang, 1991: Global observing system experiments on operational statistical retrievals of satellite sounding data. Mon. Wea. Rev.,119, 1851–1864.

  • ——, J. Pailleux, J.-N. Thépaut, J. R. Eyre, A. P. McNally, G. A. Kelly, and P. Courtier, 1994: Use of cloud-cleared radiances in three/four-dimensional variational data assimilation. Quart. J. Roy. Meteor. Soc.,120, 627–653.

  • Baker, W., H. Fleming, M. Goldberg, B. Katz, and J. Derber, 1995: Development and implementation of satellite temperature soundings produced interactively. NWS Tech. Procedures Bulletin 422, 8 pp. [Available from Office of Meteorology, National Weather Service, 1325 East–West Highway, Silver Spring, MD 20910.].

  • Caplan, P., J. Derber, W. Gemmill, S. Y. Hong, H.-L. Pan, and D. Parrish, 1997: Changes to the 1995 NCEP operational MRF model analysis/forecast system. Wea. Forecasting,12, 581–594.

  • Courtier, P., and Coauthors, 1993: Variational assimilation at ECMWF. ECMWF Tech. Memo. 194, 84 pp. [Available from European Centre for Medium-Range Weather Forecasts, Shinfield Park, Reading, Berkshire R62 9AX, United Kingdom.].

  • Cox, S. J., 1994: Prediciton of ozone from meteorological parameters for use in GLOSS and UV index forecasting. UKMO Forecasting Research Division Tech. Rep. 85, 23 pp. [Available from Forecasting Research Division, Meteorological Office, London Road, Bracknell, Berkshire RG12 2SZ, United Kingdom.].

  • Derber, J. C., D. F. Parrish, and S. J. Lord, 1991: The new global operational analysis system at the National Meteorological Center. Wea. Forecasting,6, 538–547.

  • Eyre, J. R., 1991: A fast radiative transfer model for satellite sounding systems. ECMWF Tech. Memo. 176, 28 pp. [Available from European Centre for Medium-Range Weather Forecasts, Shinfield Park, Reading, Berkshire R62 9AX, United Kingdom.].

  • ——, 1992: A bias correction scheme for simulated TOVS brightness temperatures. ECMWF Tech. Memo. 186, 28 pp. [Available from European Centre for Medium-Range Weather Forecasts, Shinfield Park, Reading, Berkshire R62 9AX, United Kingdom.].

  • ——, G. Kelly, A. P. McNally, E. Andersson, and A. Persson, 1993:Assimilation of TOVS radiances through one dimensional variational analysis. Quart. J. Roy. Meteor. Soc.,119, 1427–1463.

  • Gill, P. E., W. Murray, and M. H. Wright, 1981: Practical Optimization. Academic Press, 401 pp.

  • Halem, M., E. Kalnay, W. E. Baker, and R. Atlas, 1982: An assessment of the FGGE satellite observing system during SOP-1. Bull. Amer. Meteor. Soc.,63, 407–429.

  • Lorenc, A. C., 1986: Analysis methods for numerical weather prediction. Quart. J. Roy. Meteor. Soc.,112, 1177–1194.

  • McMillin, L. M., and C. Dean, 1982: Evaluation of a new operational technique for producing clear radiances. J. Appl. Meteor.,21, 1005–1014.

  • Mo, K. C., X. L. Wang, R. Kistler, M. Kanamitsu, and E. Kalnay, 1995: Impact of satellite data on the CDAS–reanalysis system. Mon. Wea. Rev.,123, 124–139.

  • Parrish, D. F., and J. C. Derber, 1992: The National Meteorological Center’s spectral statistical interpolation analysis system. Mon. Wea. Rev.,120, 1747–1763.

  • Reale, A. L., D. G. Gray, M. W. Chalfant, A. Swaroop, and A. Nappi, 1986: Higher resolution operational satellite retrievals. Preprints, Second Conf. on Satellite Meteorology/Remote Sensing and Applications, Williamsburg, VA, Amer. Meteor. Soc., 16–19.

  • Reynolds, R. W., and T. M. Smith, 1994: Improved global sea surface temperature analyses. J. Climate,7, 929–948.

  • Smith, W. L., H. M. Woolf, C. M. Hayden, D. Q. Wark, and L. M. McMillin, 1979: The TIROS-N Operational Vertical Sounder. Bull. Amer. Meteor. Soc.,60, 1177–1187.

  • Tracton, M. S., A. J. Desmarais, R. J. van Haaren, and R. D. McPherson, 1980: The impact of satellite soundings on the National Meteorological Center’s analysis and forecast system—The Data Systems Test results. Mon. Wea. Rev.,108, 543–586.

  • Fig. 1.

    Mean background fit to brightness temperatures as a function of bin number for local zenith angle. Only clear brightness temperatures not over land or ice are used in statistics.

  • Fig. 2.

    Coefficient of square of local zenith angle as estimated by the analysis system as a function of time (1200 UTC 18 September 1996–0600 UTC 28 September 1996) for NOAA-14.

  • Fig. 3.

    (a) Standard deviation and (b) bias with and without bias correction for NOAA-14. Only clear brightness temperatures not over land or ice are used in statistics.

  • Fig. 4.

    Total ozone analysis (in Dobson units) for 16 September 1996.

  • Fig. 5.

    Skin temperature analysis increments for 23 August 1995.

  • Fig. 6.

    RAD-NOSAT (a) and RET-NOSAT (b) analysis differences for the analysis level closest to the surface from 23 August 1995.

  • Fig. 7.

    Same as Fig. 6 except for sigma level near 500 mb.

  • Fig. 8.

    Northern Hemisphere anomaly correlations (a) and rms errors (b) for 5-day forecasts from NOSAT, RAD, RET, and RET2 experiments. Results are averages over 21 forecasts (29 July 1995∼18 August 1995).

  • Fig. 9.

    Same as in Fig. 8 except for Southern Hemisphere.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 1997 1089 270
PDF Downloads 691 198 19