Super-Channel Selection for IASI Retrievals

Transcription

1 Super-Channel Selection for IASI Retrievals Peter Schlüssel EUMETSAT, A Kavalleriesand 31, Darstadt, Gerany Abstract The Infrared Atospheric Sounding Interferoeter (IASI), to be flown on Metop as part of the EUMETSAT Polar Syste (EPS), will be used to derive a nuber of atospheric paraeters. A challenging task will be to ake proper use of the 8461 spectral radiance saples provided by IASI. Despite the developent of fast radiative transfer odels it will not be possible to ake direct use of all saples in a variational retrieval schee or to assiilate the all in nuerical weather forecasts because of the huge aount of data. As any of the spectral radiances are well correlated with each other it sees straightforward to cobine highly correlated ones to so-called super channel clusters. The advantages are reduced noise of the super channels, when copared to that of easured single spectral saples, and the possibility to chose only one of the saples to represent each cluster in radiative transfer calculations. The coposition of the super channels and their usefulness for the retrieval of teperature and water vapour profiles in diverse atospheric situations is studied by eans of RTIASI-5 siulations for globally distributed sets of atospheric and surface situations. A variational retrieval that akes use of super channels has been ipleented in the core ground segent of the EPS for the generation of level 2 products. Introduction The Infrared Atospheric Sounding Interferoeter (IASI), that has been built under EUMETSAT- CNES co-operation, will be flown on the Metop satellites as part of the EUMETSAT Polar Syste (EPS) fro 2006 onwards. Details of the IASI instruent are described by Cayla (1993). IASI will deliver highresolution radiance spectra that allow to retrieve atospheric teperature and huidity profiles for nuerical weather prediction and cliate research at accuracies of 1 K and 5%, respectively, at high vertical resolution. Trace gases to be derived fro IASI include ozone colunar aounts in deep layers, and colunar aounts of carbon onoxide, nitrous oxide, ethane, and carbon dioxide. Cloud paraeters easured fro IASI are cloud fraction, cloud top teperature, cloud height, and cloud phase. The IASI Level 2 processor (Schlüssel et al., 2005) akes use of different retrieval techniques. These include statistically based ethods like regression on principal coponent scores (Schlüssel and Goldberg, 2002; Goldberg et al., 2003) and artificial neural networks (Turquety et al., 2004). Alternatively, a variational retrieval (e.g. Rodgers, 2000) can be included either as stand-alone retrieval or be used in cobination with a statistical retrieval, where the latter provides the first guess for the forer

2 one. The variational retrieval ipleented is a siultaneous iterative retrieval seeking the axiu posterior probability solution for the iniisation of a cost function J = ( y ( x) y ) E 1 ( y( x) y ) T + ( x x b ) C 1 ( x x b ) T, (1) where x is the atospheric state vector as calculated iteratively, x b is the background atospheric state, C is the covariance atrix associated with the background, y is the easureent vector, y(x) is the forward odel operator at a given state x, and E is the cobined easureent and forward odel error covariance. The iniisation is achieved by the Marquardt-Levenberg ethod. However, despite the availability of fast radiative transfer odels, it is not possible to perfor a variational retrieval that includes separately all 8461 spectral saples in y and y. Constraints of the near-real tie processing with today's coputer capabilities require that the nuber of radiances included in the iterative retrieval is restricted to about 500. Although this is a strong liitation, the fact that the IASI spectru contains uch redundant inforation and any of the spectral saples are highly correlated with each other, enables the inclusion of the entire IASI spectru in the variational retrieval by the use of super-channel clusters. Super-Channel Coposition Redundancy in the IASI spectru allows the coposition of super-channel clusters, in which highly correlated spectral saples are cobined. Single clusters can be populated by one to any (often hundreds) single saples, of which one in each cluster, so-called lead saple, is chosen to represent the cluster in the calculation of the forward odel operators y L (x) and the related Jacobians. The easured radiance of the clusters are given by a weighted average of the contributing saples, where the average has to represent the radiance of the lead channel. The ore saples contribute to a cluster the lower the effective easureent noise will be in the averaged radiance of the cluster. The weighted average is best achieved by linear regression of the lead channel radiance against all radiances in the sae cluster. The weights consist of regression coefficients, taking into account the correlation and the noise of the respective saples. The "easured" radiance of a lead channel is calculated as N 1 yl = ai + bi yi, (2) N i= 1 where N is the nuber of spectral saples in a cluster, y i is the easured radiance of spectral saple i, and ai, b i are regression coefficients (for the lead channel we have a 1 =0 and b 1 =1). The errors in the easured radiance of the lead channel will include easureent and regression errors. y L Data Set and radiative Transfer Calculations For the generation and analysis of super-channel clusters IASI spectra have been siulated for a globally representative set of atospheric and surface situations. The basis for this set is the sub-sapled

3 ERA-40 (40 years ECMWF Re-Analysis) as described by Chevallier (2001). The rather sooth teperature profiles have been randoly perturbed with 1K/1k double-dirac dipoles in order to represent fine structure in the profiles. Land surface types have been randoly coposed of up to three types (out of 14 possible ones) and corresponding rando fractions. According to the surface types eissivity spectra have been coposed according to Snyder et al. (1998), where the ean eissivity spectra have been randoly varied between the possible liits of each type. Trace gas profiles have been taken fro cliatology, they have been randoly varied between 50 and 150% of the respective trace gas aounts. Aerosol types have been randoly selected, and their cliatological profiles have been randoly varied between 0 and 200%. About half of the data set has been set up to siulate oceanic conditions with sea surface conditions where the surface eissivity is paraeterised according to the wind-roughened sea surface. For the entire set of cloud-free situations siulations have been carried out with the fast radiative transfer odel RTIASI-5 (Matricardi, 2004a, 2004b). The instruent viewing angles for each siulation was chosen randoly aong the possible values. The resulting Gaussian apodised radiance spectra are noralised with the standard deviation of the respective noise spectra before entering the correlation analysis. Correlation and Regression Analysis A correlation analysis has been done between spectral saples. Starting at the low wavenuber boundary of 645 c -1 the spectru has been scanned for saples with a high correlation. All pairs with a correlation higher than a threshold are retained in the respective cluster, irrespective of the spectral distance between the saples, and each saple ust be a eber of exactly one cluster. The assued thresholds for the iniu correlation vary between 0.95 and The nuber of super-channel clusters obtained varies according to the threshold and generally decreases with rising threshold as seen in table 1. At a low correlation of 0.95 the nuber of super-channel cluster is as low as 47 and it is clear that the spectral inforation which was originally in the IASI spectru is lost in a sense coparable to a broadband filter radioeter. The nuber of saples within single clusters can reach several hundreds, but can also be as low as one. At higher correlation thresholds the clusters are ore sparsely populated and often contain only single spectral saples with pronounced, individual spectral signatures (Figure 1). Table 1: Nuber of super-channel clusters for given correlation thresholds. Correlation Nuber of Clusters

4 Fig. 1: Population of super-channel clusters at correlation thresholds of 0.99 (top), (iddle), and (botto). The spectral range covered by single clusters is shown in figure 2. It is seen that at the lower correlation a spectral range of alost 1000 c -1 can be covered by single clusters. At increasing correlation the spectral range in a single cluster is liited to about 300 c -1. As entioned above, the population of the superchannel clusters also reduces the easureent noise entering the variational retrieval. The noise figures to be considered for the lead channels consist of easureent noise and regression errors. At lower correlation the latter prevail, while the easureent noise becoes the ajor part at high correlation iplying a low regression error. Figure 3 illustrates for a correlation of 0.99 (222 super-channel clusters) the effective noise on the lead channel together with the population of the respective super channel. The noise (shown for a noise-noralised radiance) is reduced for any clusters by a factor greater than 5. For clusters populated with single saples only the noralised noise stays at 1, corresponding to the plain easureent noise.

5 Fig. 2: Spectral range covered by single super-channel clusters at correlation thresholds of 0.99 (top), (iddle), and (botto). Two exaples for super-channel populations at a correlation threshold are shown in figure 4 together with the portion of a IASI radiance spectru. They deonstrate the big differences that are possible between super channels and the variety of spectral spacing between different saples within single clusters. Conclusion and Discussion IASI spectral saples can be efficiently cobined to super-channel clusters in order to reduce the nuber of easureents entering a variational retrieval or an assiilation of IASI spectra into nuerical weather forecast odels, whilst retaining redundant inforation to reduce the easureent noise of the radiance saples entering the retrieval or assiilation schee.

6 Fig. 3: Noralised noise (blue) of the lead channel radiance after weighted averaging and the corresponding population (red) of super channels for a correlation of Lead channels are selected to represent the clusters in the radiative transfer siulations used as forward odel operators. The choice of lead saples ade in this study was based on a plain search along the wavenuber scale for the next saple that was not already eber of a cluster. Likewise, the lead saple selection could be optiised with respect to forward odel errors, selecting aong those saples of a cluster the one as lead saple which has the lowest forward odel error. The nuber of clusters can range fro as low as 47 to ore than 1633, assuing saple correlations ranging fro 0.95 to higher than The right choice for the nuber of clusters is to be optiised, trading off the noise reduction against spectral inforation. Operational retrieval or assiilation systes clearly require an upper liit of clusters that is suitable for near real-tie processing. A nuber of 222 to 417 clusters, corresponding to correlations of 0.99 and 0.995, respectively sees to be realistic for such systes at present coputing capabilities. The super-channel clusters derived for high correlation thresholds contain any clusters which are populated with single saples only. This is desirable when spectral inforation is needed that is unique in single saples and which is not to be seared by the creation of a highly-populated super-channel cluster. Likewise, it is always possible to reove single saples fro a cluster and include it as single saple in

7 Fig. 4: Exaples of super-channel clusters at a correlation threshold of Top: cluster 6 with lead saple at 647 c -1 including 15 saples; botto: cluster 3 with lead saple at c -1 including 36 saples. The blue curve is an atospheric reference spectru, the red lines indicate the positions of the spectral saples. the retrieval or assiilation, if the particular spectral inforation of this saple is needed (e.g. retrieval of trace gases). The spectral range covered by single clusters is not liited to narrow spectral regions, but can cover a ajor part of the IASI spectru. The drastic change in order of agnitude of the radiances across the spectru is handled efficiently by using noise-noralised radiances. For a odest nuber of 222 super channel clusters the noise is reduced substantially, a factor of 5 is achieved for any clusters. A variational retrieval that akes use of this set of super channel clusters has been ipleented and tested in the EPS core ground segent at EUMETSAT, clearly deonstrating the usefulness of this approach.

8 References Cayla, F.-R., 1993: IASI infrared interferoeter for operations and research, in: Chedin, A., Chahine, M.T., Scott, N.A. (Eds.), High Spectral Resolution Infrared Reote Sensing for Earth's Weather and Cliate Studies. NATO ASI Series, I 9, Springer Verlag, Berlin-Heidelberg, Chevallier, F., 2001: Sapled Databases of 60-level Atospheric Profiles fro the ECMWF Analyses. Research Report 4, EUMETSAT/ECMWF. Goldberg, M.D., Qu, Y., McMillin, L.M., Wolf, W., Zhou, L., Divakarla, M., 2003: AIRS Near-Real- Tie Products and Algoriths in Support of Nuerical Weather Prediction. IEEE Trans. Geosci. Reote Sens., 41, Matricardi, M., 2004a: The inclusion of Aerosols and Clouds in RTIASI, Report EUMETSAT Contract EUM/CO/02/989/PS, ECMWF, Reading, 57 pp. Matricardi, M., 2004b: RTIASI-5 User's Guide, Report EUMETSAT Conract EUM/CO/02/989/PS, ECMWF, Reading, 47 pp. Rodgers, C.D., 2000: Inverse Methods for Atospheric Sounding - Theory and Practice. World Scientific, Singapore, pp 238. Schlüssel, P. and Goldberg, M. 2002: Retrieval of atospheric teperature and water vapour fro IASI easureents in partly cloudy situations, Adv. Space. Res., 29(11), Schlüssel, P., Hultberg, T.H., Phillips, P.L., August, T., Calbet, C., 2005: The operational IASI level 2 processor. Adv. Space Res., in press. Snyder, W.C., Wan, Z., Zhang, Y., and Feng, Y-Z., 1998: Classification-based eissivity for land surface teperature easureent fro space, Int. J. Reote Sensing, 19, Turquety, S., Hadji-Lazaro, J., Clerbaux, C. Hauglustaine, D.A., Clough, S.A., Cassé, V., Schlüssel, P., Mégie, G., 2004: Operational trace gas retrieval algorith for the Infrared Atospheric Sounding Interferoeter, J. Geophys. Res., 109, D21301.