Abstract

Data from the Infrared Atmospheric Sounding Interferometer (IASI) onboard MetOp has been assimilated at the Met Office in both Global and North Atlantic and European (NAE) model configurations since November 2007. It has been a considerable challenge to reduce data volumes to a manageable level within the constraints of the forecast system and this paper will summarise the processing methodology employed. Pre-operational trials of IASI assimilation in the Global model delivered a positive impact on forecasts approximately twice as large as that shown by AIRS. A series of observing system experiments confirms the relative performance of IASI and AIRS, and shows that impact from IASI is equivalent to a single AMSU-A+MHS. Analysis of a second season Global model IASI assimilation trial indicates that the measurement of impact is strongly affected by variations in performance of the control forecast. Furthermore, the impact of IASI is strongly dependent on the variables and methods chosen for verification: although improvements to the large scale fields (e.g. mean sea-level pressure and geopotential height) were also seen in the NAE configuration, no forecast impact was seen for variables such as visibility and rain-rate. Selection, preprocessing and assimilation of IASI data Data selection For a six-hour assimilation cycle, 324,000 IASI spectra each consisting of 8461 channels are processed. This equates to 2.7x10 9 channel observations per cycle, in comparison with approximately 0.04x10 9 channel observations for ATOVS data from five satellites. One of the greatest difficulties in the use of IASI data is therefore to extract detailed atmospheric information as efficiently as possible to enable assimilation within the current computing constraints of our operational numerical weather prediction (NWP) system. The approach taken to enable fast delivery of benefit from IASI was to perform spatial and spectral reduction at the earliest possible opportunity in the data processing chain. The 8461 channels are reduced to 314 using the 300-channel selection proposed by Collard (2007) -which selects the channels with the greatest information content for a variety of meteorological situations -and adding in 14 extra monitoring channels chosen by the Centre National d'Etudes Spatiales (CNES; Blumstein, pers. comm.). The observations are thinned spatially to 1 pixel from 4 in each field of view using the IASI L1c AVHRR cluster information to select the most homogeneous field of view. This is carried out in AAPP International TOVS Study Conference-XVI Proceedings This spatial and spectral data reduction approach is a rather blunt instrument, but it is anticipated that in future we may be able to dispense with the spatial thinning and use principal component scores to compress information from the whole spectrum into a relatively few pieces of information which can be assimilated efficiently. Following data reduction, for the Global model, we process a maximum of 81,000 observations (0.025x10 9 channel observations) for each six-hour assimilation cycle. These are further reduced to around 3,000 observations for assimilation through rejection of observations during preprocessing, details of which follow in the next section. For the NAE model, the data coverage is highly variable, but between 200 and 800 observations are typically assimilated in each six-hour 4D-Var cycle. Preprocessing Before assimilation, the IASI data are passed through a 1D-Var and quality control stage known as the Observation Processing System (OPS) where the observations are compared with forward-modelled atmospheric columns constructed from model fields. The process consists of several steps including bias correction, surface type assignment, cloud detection, channel selection and 1D-Var. At the Met Office a static bias correction scheme is used rather than variational bias correction. This correction consists of a different constant offset for each channel and each scan angle plus a linear function of the 850-300 hPa thickness and 200-50 hPa thickness, with different coefficients for each channel. The Met Office model has a land/sea mask and also a sea ice fraction which are used to assign a surface type to the observation. The model surface type is used in conjunction with a surface type determined from AMSU observations which are mapped to the IASI footprints in AAPP (English et al., 1997). The AMSU data is tested for consistency with eight different surface types and an optimal estimation method is used to match each observation to the best-fitting surface type. Not all of the 314 channels are used during preprocessing and assimilation. In OPS, channels in the 8μm ozone band, all channels in Band 3, and several other channels which seem to cause problems in the 1D-Var minimisation (for example, the very highest peaking channels) 56 International TOVS Study Conference-XVI Proceedings are rejected. Only water vapour channels peaking below 520hPa (from their temperature Jacobians for the US standard atmosphere) are used -this leaves 31 water vapour channels in the minimisation. For observations over land or where the microwave cloud test is failed, channels peaking below 400hPa (for the US standard atmosphere) which may be sensitive to the surface emission are rejected. No water vapour channels are used over land. Assimilation Channels peaking above 50hPa are used in OPS but are not assimilated via 4D-Var because the increments they generate exhibit stratospheric ringing. In total, 138 channels can be assimilated depending on the cloud conditions and surface type (see International TOVS Study Conference-XVI Proceedings For 4D-Var, the errors are inflated in line with those for other radiance data to compensate for lack of off-diagonal elements in the covariance matrix: the standard deviations of error are inflated to 0.5K for the 15μm CO 2 band, 1K for window channels and 4K for water vapour channels. These values were chosen to maximise the weight given to IASI without allowing the fit to other satellite sounders to deteriorate significantly. In particular, the large errors assumed in the water vapour band help to compensate for lack of inclusion of any correlated error. Global Model Assimilation Trials June 2007 A series of assimilation trials where IASI was included on top of the current operational satellite data usage were run for the period 24 th May -24 th June 2007. The trials tested assimilation under a range of conditions: with and without water vapour channels; with 0.5K and 1K observation errors for the temperature sounding channels; at different horizontal resolutions; and with different model physics configurations. Forecast impact at the Met Office is measured by use of a score called the NWP Index (see Annex A for detail) which combines improvements in forecast skill in a number of atmospheric parameters. Assuming one knows the persistence error perfectly (a reasonable assumption in comparison with the forecast error), over a one-month trial where the impact is reasonably consistent in time and across the components of the Index (also reasonable in this case), using the standard error as a measure of forecast error, an improvement in the Index of 0.5 points is considered to be significant. The results of all trials run were very consistently positive, and the inclusion of water vapour channels and use of 0.5K observation errors for the temperature sounding channels produced the largest positive impact of 1.2 points verified against observations and 0.8 points against analyses -1.0 points overall. The largest impacts seen were in the tropical and northern hemisphere geopotential height forecasts against analyses; and against observations in the wind fields and 500hPa height in the tropics and southern hemisphere, and 500hPa height and PMSL in the northern hemisphere. The NWP Index was improved for most days of the forecast trial period against both observations and analyses except for a few days towards the end International TOVS Study Conference-XVI Proceedings The strong performance of IASI leading to marked improvements in forecast skill are particularly impressive considering that IASI has been added to a system already assimilating sounding data from ATOVS on three NOAA satellites and on MetOp (whose observations are coincident with IASI), AIRS on EOS Aqua, and SSMIS on DMSP F16. This result suggests that new information is being added by the assimilation of IASI. Bell et al. (2008) compare humidity increments for IASI and SSMIS and show that using the 31 water vapour channels that were tested in the trial, 1D-Var specific humidity increments from IASI span the troposphere where SSMIS increments are concentrated in the lower troposphere. IASI data from the assimilation system. Removing IASI had negative impact, but not as significant as over the summer period: a degradation of 0.4 verified against observations and 0.6 against analyses (0.5 overall). A time series of the trial period The difference in NWP Index scores can be attributed to the way the forecasts are verified and the change in performance of the control between the two periods. The skill scores used to calculate the NWP Index compare the performance of the trial relative to persistence with the performance of the control relative to persistence. In the northern hemisphere, persistence is much poorer as a forecast during winter than summer, and the control therefore exhibits much greater skill during the winter 2007 trial period 60 International TOVS Study Conference-XVI Proceedings summer and winter trial periods, the reduction in RMS error of the control is an order of magnitude larger. Figure 5 also indicates that during the summer trial IASI performed strongly for the 850hPa wind forecasts in the tropics, variables to which the NWP index calculation is particularly sensitive. Global Model Observing System Experiments A series of observing system experiments were run for a slightly shortened winter trial period of 12 th December 2007 to 4 th January 2008 where individual satellite sounding instruments were tested on top of a no-satellite-sounding baseline. Five experiments were run: • No satellite sounding -AMSU, MHS, HIRS, AIRS, IASI and SSMIS excluded The verification of these trials was performed against observations only. Verifying against radiosondes obviously has drawbacks, not least from the point of view of the spatial variability of the network -in particular the sparseness of the verifying observations in the southern hemisphere is of concern given that satellite data have traditionally shown greater 61 International TOVS Study Conference-XVI Proceedings impact here but at least the verification is consistent across the experiments. Verifying against analyses would probably favour either the control or the experiment for such a data-poor experimental set up. Figure 6: Percentage change in weighted skill relative to persistence of the control runs used for IASI assimilation experiments in December and June: a positive value indicates that the December control showed more skill relative to persistence than the June control. A time series of the daily Index change shows that in general the satellite sounding observations are all showing similar patterns of impact in terms of day-to-day variability 62 International TOVS Study Conference-XVI Proceedings It is extremely encouraging that assimilation of IASI in clear areas only is providing similar forecast benefit to the assimilation of ATOVS data in both clear and cloudy conditions. The relative performance of IASI and AIRS in the assimilation system demonstrates that more information can be extracted from hyperspectral infrared sounders when data with low radiometric noise are given appropriate weight in the assimilation system. 63 International TOVS Study Conference-XVI Proceedings North Atlantic and European Model Assimilation Trial An IASI assimilation trial for the period 24 th May to 24 th June was also run for the NAE model configuration. Despite extremely positive results in the global model, IASI showed neutral impact in the NAE, in common with previous satellite data assimilation trials for AIRS (Hilton et al., 2006) and AMSU (Candy et al., 2003). Verification for the NAE is very different from the global model. A UK NWP Index is calculated from six weighted forecast variables: surface visibility, six hour precipitation accumulations, total cloud amount, cloud base height, surface temperature and surface wind verified against observations. Annex B describes the calculation of the UK Index. Satellite radiances typically show little impact on these "weather" variables, despite providing benefit in the upper air fields for the NAE model. The highest impact (for the full NAE domain) was +0.12%. No objective significance testing has been applied to this result. However, the operational change procedure at the Met Office considers a change of 0.3% to be a small but significant impact, particularly if this impact was shown to be consistent in time and across different verification areas. Therefore, whilst this is a subjective significance test, 0.12% is unlikely to be a significant change. In contrast, The verification results for the upper air and large-scale fields are in keeping with the results from the global model trials. This suggests that the negligible impact seen when adding satellite data into Met Office limited area models may be partly due to the way in which the model is verified, but also that the improvements seen in the large-scale model fields are disconnected from changes in the surface weather variables which make up the UK Index. 64 International TOVS Study Conference-XVI Proceedings Conclusion Experiments have shown that assimilation of IASI in the Met Office Global model provides significant forecast benefit on top of a system which already makes use of ATOVS, AIRS and SSMIS data. Observing system experiments testing individual satellite instruments on top of a no-satellite-sounding baseline suggest that IASI provides a similar level of improvement in forecast skill to AMSU-A+MHS, and significantly more than current implementations of AIRS and HIRS. The benefit seen in the North Atlantic and European model is negligible for the forecast variables which are used for the verification of Met Office limited area models, suggesting that benefits in large-scale fields such as geopotential height and PMSL do not feed into surface weather variables. The considerable improvement in the global forecast was achieved with a simple assimilation scheme treating IASI in the same way as the previous generation of sounding instruments. There is much more information to be extracted from IASI, and future work will focus on using the spectral information more effectively to allow us to increase the amount of data assimilated and thus improve forecasts further. This could be through spectral compression, or using more channels over land and over cloud (Pavelin, 2006). 65 International TOVS Study Conference-XVI Proceedings