Towards the assimilation of AIRS cloudy radiances

Transcription

1 Towards the assimilation of AIRS cloudy radiances N. FOURRIÉ 1, M. DAHOUI 1 * and F. RABIER 1 1 : National Center for Meteorological Research (CNRM, METEO FRANCE and CNRS) Numerical Weather Prediction Department (GMAP) 1 :42 av. Gaspard Coriolis, Toulouse Cedex, France *: Direction de la Météorologie Nationale, Casablanca, Morocco Abstract The assimilation of AIRS cloudy radiances is investigated inside the 4D-Var assimilation scheme of the French global model ARPEGE. Two approaches are used: the first one is based on a combination of a diagnostic cloud scheme together with the radiative transfer model RTTOVCLD to simulate AIRS cloudy radiances. With this diagnostic cloud scheme, only large scale processes are taken into account. Therefore, the situations ruled by convection, are discarded. Linearity and accuracy requirements for the observation operator significantly reduce the number of AIRS channels suitable for use in 4D-Var. The second approach makes use of the cloud top pressure and the cloud cover derived from the CO2-slicing technique. In a first experiment, the CO2-slicing outputs are directly used by RTTOV in the 4D-Var to simulate the cloud-affected spectrum. In a second experiment, the CO2-slicing outputs are adjusted by a 1D-VAR step before being used by RTTOV in the 4D-Var. Preliminary experiments have been conducted in a simplified framework over a ten day period. The number of channels from cloudy pixels assimilated in the 4D-Var is small, but larger with the CO2- slicing approach than with the diagnostic approach and in the reference experiment. The impact of the assimilation of cloudy radiances is neutral with respect to the fit of the assimilation to the observations, except for AMSU-B data in the case of the CO2-slicing approach. In this latter case, more AMSU- B observations are assimilated than in the reference experiment. A very slight and non-significant impact on the forecast is found for the first approach. With a slightly more positive impact, results with the CO2-slicing approach are encouraging. The positive impact is more pronounced with a prior adjustment by the 1D-Var of the cloud parameters. Results are thus significantly but weakly improved in the southern hemisphere. Finally, some points to be studied in the near future are presented. Introduction The assimilation of cloudy radiances is a challenge for numerical weather prediction centres as it is now well established that sensitive areas for the forecast of cyclogenesis are cloudy (McNally, 2002 and Fourrié and Rabier 2004). In a first step, Dahoui et al (2005) assessed various cloud detection schemes suitable for the assimilation of AIRS radiances with MODIS observations. They showed that the ECMWF and the CO2-slicing algorithms were very accurate. Both were implemented in the 321

2 Control variable T, Q, Ps Control variable δ T, δ Q, δ Ps T Jrad, q Jrad, Ps Jrad Cloud diagnostic scheme TL of cloud scheme Adjoint of cloud scheme Cloud Var Qi, Ql, CC RTTOVCLD Cloud Var δqi, δ Ql, δ CC RTTOVCLD TL Increment of cloud variables RTTOVCLD AD Radiances Radiances Radiance increment Computation of Jrad and Jrad Figure 1: Sketch of the observation operator for the screening (left-hand side) and for the minimisation (right-hand side) used with the simplified cloud diagnostic scheme. ARPEGE 4D-Var data assimilation scheme. In the second part of their paper, they tried to select, according to the cloud type, the AIRS channels for which the observation operator has a good linear behaviour. This observation operator consisted of the RTTOV radiative transfer model containing a cloud-processor module and a simplidied cloud diagnostic scheme modelling only the large-scale clouds. An attempt to assimilate cloudy radiances in the 4D-Var of ARPEGE is made in this paper. Two approaches have been chosen for the assimilation of cloudy observations. The first one uses a simplified diagnostic cloud scheme together with a radiative transfer model that can take into account multiple cloud layers. The second one is based on the CO2-slicing approach. In a second section, we describe both approaches and the experimental settings. Preliminary results are given in the ensuing section and in the last one, some conclusions are drawn and the future work is discussed. How to deal with clouds in the radiative transfer for the assimilation of cloudy radiances? In this section we describe the observation operators used for each approach in the 4D-Var assimilation of the AIRS cloudy radiances. Use of a simplified cloud diagnostic scheme In that case, the observation operator (Fig. 1) consists of a cloud diagnsotic scheme and of a radiative transfer model. For the radiation scheme, we rely on the RTTOVCLD model (Saunders et al, 2002, RTTOV model with a cloud-processor module). The cloud scheme used is a simplified version of the ARPEGE operational cloud scheme (Gérard, 2001). Only the large scale clouds are modelled. The cloud scheme diagnoses, from the temperature and humidity profiles and the surface pressure, the cloud cover, the cloud liquid and cloud ice water contents at the 41 vertical levels corresponding to those of the ARPEGE model. All these output parameters are supplied as input together with the atmospheric control variables T, q, P s to RTTOVCLD. For the clear observations, the cloud profiles are set to zero and the same observation operator as for the cloudy observations is used. 322

3 Screening Radiances T,Q,Ps CO2slicing Control variables δ T, δ Q, δ Ps RTTOV TL Radiances Cloud top pressure and effective emissivity Computation of Jrad and Jrad T Jrad, q Jrad, Ps Jrad RTTOV AD Radiance increments Figure 2: Sketch of the observation operator used with the CO2slicing method. Use of the CO2-slicing The CO2-slicing method (Chahine 1974, Menzel et al, 1983 and Lavanant, 2002) has been extensively used to retrieve the cloud-top pressure and the cloud effective emissivity. The algorithm, based on the radiative tranfer principles uses the measured radiances of a subset of 124 AIRS channels selected in the CO2 absorption band (between and cm 1 ), which is very sensitive to the presence of clouds. In that case, the cloud parameters are computed during the screening and do not change during the minimisation step of the 4D-Var (Fig. 2). Use of the CO2-slicing adjusted with 1D-Var The 1D-Var is able to improve the cloud characterization (in terms of spectral effective emissivity and top pressure), particularly for the low level clouds for which the CO2-slicing alone is less accurate. The input of the 1D-Var is the temperature and the humidity profiles, the surface pressure and the cloud parameters (cloud top and effective emissivity). The 1D-Var gives an adjustment of the temperature and humidity profiles and of the cloud parameters but only the latter are kept for the 4D-Var minimisation. The observation operator of the 1D-Var consists of RTTOV with the simulation of a single thin and opaque cloud layer. The 1D-Var is run during the screening step and the cloud parameters are then kept constant during the 4D-Var minimisation as in the previous approach. The observation operator in the 4D-Var is described in figure 3. Experimental settings 5 experiments have been conducted during 10 days. They have been performed to test the feasibility of the assimilation of AIRS cloudy channels. Only the low-level clouds for which the cloud top pressure is included in the hpa bracket are selected in order to compare more easily the results obtained with the diagnostic approach with the other ones. In addition, the effective emissivity 323

4 Screening Radiances T,Q,Ps CO2slicing Control variables δ T, δ Q, δ Ps RTTOV TL Cloud top pressure and effective emissivity T Jrad, q Jrad, Ps Jrad RTTOV AD Radiances Computation of Jrad and Jrad Radiance increments Figure 3: Sketch of the observation operator for the experiment with the CO2slicing adjusted together with an 1DVar of the cloudy observations has to be included between 0.3 and 0.99 to discard low-level cloud cases with small emissivity for which the CO2-slicing method is less accurate. Furthermore, to make the comparison of the cloudy observation impact easier, cloudy observations are assimilated only in the extra-tropics (beyond [-40 ; 40 ]) because the diagnostic method is only valid for stratiform clouds. Here are the different experiments studied: A control experiment with the assimilation of clear sky radiances detected with the CO2-slicing algorithm (called here-after control ). An experiment with the ECMWF cloud detection. In this case, clear channels are assimilated (called here-after ECMWF-like ). An experiment with the assimilation of clear pixels and cloudy pixels with the cloud diagnostic scheme (called here-after diagnostic ) An experiment with the assimilation of clear pixels and clouds pixels with the CO2-slicing (called here-after CO2-slicing ). An experiment with the assimilation of clear pixels and cloudy pixels with the cloud top pressure and the cloud cover adjusted with the 1D-Var (called here-after CO2+1D-Var ). The observation errors are the same for clear and cloudy channels. They vary between 0.5 and 2 K. The same bias correction is applied to all experiments. This bias correction is flat, as at the beginning of the experimention, no robust varying bias correction was available. In addition at the beginning of the experiments 102 channels were assimilated in a pre-operational suite. 324

5 Experiments observation numbers channel number clear cloudy total clear cloudy total Control Diagnostic CO2slicing CO2+1D-Var ECMWF-like Table 1: Number of pixels and of channels assimilated for the 5 experiments as a function of the cloudiness for the first analysis time 2006/06/08 00 UTC. Results Impact on the analysis Table 1 presents the different observations and channels assimilated in the 5 experiments for the first analysis cycle of the period. The first-guess is the same for this first analysis time for all the experiments and it facilitates the comparison between the experiments. With respect to the control, the number of clear observations is similar for the 3 cloudy experiments. The number of clear observations assimilated with the ECMWF experiment is larger than the control but the number of clear pixels is similar. For the cloudy observations, 50 % and 75% additional observations are assimilated. The CO2+1D- Var assimilates the largest number of cloudy pixels and the total number of assimilated AIRS observations (1700) is similar to the ECMWF experiment. The diagnostic experiment only takes 15% additional data into account. The number of additional cloudy observations is larger for the experiments based on the CO2-slicing (30% and 45% of addition channels for the CO2slicing and the CO2+1D-Var experiments respectively). Figure 4 shows the distribution of the cloud top pressures for the assimilated cloudy observations for the first analysis time. The diagnostic and CO2slicing methods exhibit a similar distribution. In contrast, the CO2-1DVar experiment has more cloudy channels with higher cloud top pressure and thus it assimilates more low-level cloud observations. Figures 5 and 6 show the root mean square differences between the observations and the similated brightness temperatures from the background and the analysis for the assimilated clear and the cloudy channels respectively for the first analysis time. The RMS for the clear observations are very similar for the control experiment and the 3 cloudy experiments. The RMS are lower for the ECMWF experiment. For the cloudy observations, the RMS are larger for the diagnostic channels even if few cloudy channels in the first band are assimilated. The two other methods have similar innovations but with some differences at the end of the spectrum. We have also looked at the impact of the other observation types on the assimilation. This impact is globally neutral in terms of number of assimilated observations and innovation statistics (not shown). For the whole period of the study, the more pronounced impact is for the AMSU-B observa- 325

6 CO2slicing CO2+1Dvar diagnostic % cloud top pressure (hpa) Figure 4: Distribution of the cloud top pressures of the assimilated observations for the diagnostic, the CO2slicing experiment and the CO2slicing adjusted with 1D-Var experiment for the first analysis time: 8 June 2005 at 00UTC. tions for which we have a small increase of the assimilated observations with the CO2slicing (+0.5%) and the CO21D-Var (+1.8%) experiments. This fact gives us an indirect indication of a slightly better quality of the background for the CO21D-Var experiment. Impact on the forecasts Figures 7, 8, 9 and 10 show the impact on the forecast of the different experiments versus the control run. The reference used here is the radiosondes. For the diagnostic experiment (Fig. 7), the impact of the cloudy radiances is very small, with a slight improvement in the northern hemisphere and a slight degradation on the southern hemisphere. The impact is neutral in the Tropics regions as no cloudy observation is assimilated in these regions in our study. This impact is non significant. This can be explained by the fact that the observation operator requires strong selection criteria for the assimilation of channels which leads to a very small number of additional assimilated data. For the CO2-slicing experiment (Fig. 8), the impact of the cloudy radiances on the geopotential forecast is slightly positive in the southern hemisphere and almost neutral in the northern hemisphere and in the Tropics. For the CO2-1D-Var experiment (Fig. 9), the positive impact is more pronounced in the southern hemisphere but the impact in the northern hemisphere is slightly negative. The adjustment of the cloudy parameters by the 1D-Var lead to a positive impact that becomes statistically significant even if it is small. 326

7 o g REF o a REF o g ECMWF o a ECMWF o g DIAG o a DIAG o g CO2 o a CO2 o g CO2 1D o a CO2 1D 1 rms (K) wave number (cm ¹) Figure 5: RMS difference between observations and background and analysis for the clear observations on 8 june 2005 at 00UTC. 1.5 o g DIAG o a DIAG o g CO2 o a CO2 o g CO2 1D o a CO2 1D 1 rms (K) wave number (cm ¹) Figure 6: RMS for cloudy observations For the ECMWF experiment (Fig. 10), the impact is positive at the end of the forecast range on the southern hemisphere and neutral elsewhere. It should be mentioned that these results are preliminary as only a 10-day period has been tested. The bias correction and the error specification should also be improved. Conclusion and discussion about the future work The goal of this study was to test the feasibility of the assimilation of AIRS cloudy observations in the ARPEGE model. To do so, two approaches have been used. The first one makes use of a simplified cloud diagnostic scheme and the second one uses the cloudy parameters derived from the observation 327

8 Figure 7: Root mean square error, standard deviation and bias of the difference for the geopotential field for the control experiment and the diagnostic experiment as a function of the forecast range for the 10 days of study (from 8 to 17 june 2005). Blue contours show improvements, red ones degradations. themselves in the radiative transfer model. For the second approach, an adjustment of the cloudy characteristics has also been tested. These three methods have been compared to a control experiment that assimilated only clear observations and with an ECMWF-like experiment that assimilated only clear channels over a ten days period. To make the comparison easier, only low-level clouds (for which the cloud top pressure is in the bracket hpa) have been used for all the experiments. We have shown that the more pronounced impact with respect to the control experiment is obtained for the experiment that uses the CO2-slicing method and a 1D-Var for the cloud characteristics. It enables to take into account the largest number of cloudy observations in the analysis. However, more effort should be devoted to these experiments. The bias correction and the observation errors should be specific to cloudy observations, the period of study has to be extended. In addition to optimize the impact of the cloud characteritics, the minimisation on the cloud parameters inside the 4D-Var minimisation should bring improvement through a better coherence between the atmospheric profile and the cloud characteristics. A hybrid approach combining both methods could also be tested : the CO2-slicing adjusted together with the 1D-Var could be used in the first 328

9 Figure 8: Root mean square error, standard deviation and bias of the difference for the geopotential field for the control experiment and the CO2slicing experiment as a function of the forecast range for the 10 days of study (from 8 to 17 june Blue contours show improvements, red ones degradations. minimisation of the 4D-Var and the diagnostic sheme could be used in the second one. Acknowledgments The Numerical Weather Prediction Satellite Application Facilities is acknowledged for having funded the scientific visits of M. Dahoui to the CNRM. Jean Antoine Maziejewski is also warmly thanked for his helpful comment on this paper. References Chahine, M. T., 1974: Remote sounding of cloudy atmospheres, I: The single cloud layer. J. Atmos. Sci.31, Dahoui, M., Lavanant, L., Rabier, F. and Auligné, T., 2005: Use of the MODIS imager to help with AIRS cloudy radiances. Quart. J. Roy. Meteor. Soc. (131),

10 Figure 9: Root mean square error, standard deviation and bias of the difference for the geopotential field for the control experiment and the CO2-1D-Var experiment as a function of the forecast range for the 10 days of study (from 8 to 17 june Blue contours show improvements, red ones degradations. Dahoui, M. L., 2006: Vers l assimilation des radiances nuageuses. Thse de l universit Paul Sabatier. PhD defended on the 19th of June 2006, Toulouse, France. Fourrié, N. and Rabier, F., 2004: Cloud characteristics and channels selection for IASI radiances in meteorologically sensitive areas Quart. J. Roy. Meteor. Soc. (130), Gérard L., 2001: Physical parametrizations in the ARPEGE ALADIN operational model. Internal documentation available from CNRMP/GMAP, Météo-France, Toulouse, France. Lavanant, L., 2002: Cloud processing in IASI context. Proceedings of the 12th international TOVS Study conference. 26/02 to 4/03/2002, Lorne, Australia. McNally, A. P., 2002: A note on the occurence of cloud in meteorologically sensitive areas and the implications for advanced infrared sounders. Quart. J. Roy. Meteor. Soc.(128), McNally A. P. and Watts P. D., 2003: A cloud detection algorithm for high-spectral resolution infrared sounders. Quart. J. Roy. Meteor. Soc., 129, Menzel, W., Smith T. and Stewart, T. R., 1983: Improved cloud motion vector and altitude assignment using VAS. J. Appl. Meteor., 22,

11 Figure 10: Root mean square error, standard deviation and bias of the difference for the geopotential field for the control experiment and the ECMWF-like experiment as a function of the forecast range for the 10 days of study (from 8 to 17 june Blue contours show improvements, red ones degradations. Saunders, R. W., Brunel, P. Chevallier, F., Deblonde G., English S.J., Matricardi M. and Rayer, P.J., 2002 : RTTOV-7 Science and Validation Report, NWP Forecasting Research Tech. Rep., 387, 51 pp. 331

12