Surface pressure information retrieved from GPS radio occultation measurements

Authors


Abstract

Numerical weather prediction (NWP) experiments have been performed to investigate the impact of surface pressure information retrieved from Global Positioning System radio occultation (GPSRO) measurements. The GPSRO measurements are able to reduce the impact of removing all conventional surface pressure observations from the NWP system. The GPSRO measurements can constrain global mean surface pressure errors in analyses and short-range forecasts to the −1 to −1.5 hPa level over a 3-month period, whereas they can be as large as −4 hPa when both GPSRO observations and conventional surface pressure measurements are removed from the NWP system. The large mean errors that arise when both GPSRO and conventional surface pressure measurements are removed are caused by an interaction between the assimilation of the conventional upper-air measurements and the bias correction of radiances with the variational bias correction system (VARBC). In the absence of GPSRO, the VARBC amplifies and spreads errors that are introduced by the conventional upper-air data. However, the GPSRO measurements cannot fully compensate for the loss of the conventional data, particularly in the Northern Hemisphere. It is also shown that the surface pressure results obtained when GPSRO are assimilated are sensitive to biases introduced by other observations, such as aircraft temperature measurements, because these biases can be aliased into surface pressure increments. Experiments designed to compare the amount of information provided by other parts of the global observing system highlight the complementary nature of GPSRO and scatterometer wind retrievals, and show that reasonable surface pressure analyses, with an uncertainty of ∼ 1 hPa, can now be obtained by assimilating just satellite data, without using any conventional data to ‘anchor’ the NWP system. However, the full observing system, including both satellite and conventional observations, remains superior to assimilating just satellite observations.

1. Introduction

The possibility of retrieving useful surface pressure information from satellite measurements has been discussed since the early years of satellite meteorology (e.g. Smith et al., 1972). Much of the research in this area has been motivated by possible applications in operational numerical weather prediction (NWP), and it has been noted (e.g. O'Brien et al., 1998) that satellite-based surface pressure retrievals could augment the conventional surface pressure observation network, particularly in the Southern Hemisphere (SH). Scatterometers now routinely provide information on surface pressure gradients over oceans from surface wind retrievals. However, designing a satellite observing system capable of providing useful absolute surface pressure information based on the measurement of radiances has proved challenging. The surface pressure retrieval errors must be less than 1 hPa (∼ 0.1%) for NWP applications. This places significant constraints on the accuracy of the measurement instrumentation and processing, the knowledge of the spectroscopic parameters required to simulate and interpret measurements in the retrieval, and the a priori atmospheric information required to estimate the surface pressure information content. Nevertheless, a number of active and passive radiance-based observing systems designed to retrieve surface pressure have been proposed (O'Brien, 2002, provides a review), but none are yet used in NWP. Most research has focused on passive remote sensing of the O2 A band (759–771 nm) (Mitchell and O'Brien 1987). Promising results have been reported by O'Brien et al. (1998) who have demonstrated an accuracy of 0.1% (∼ 1 hPa) in experiments retrieving surface pressure from high-resolution A-band spectra of sunlight reflected from the sea surface, using a grating spectrograph directed towards sunglint from a research aircraft. The particular advantage of measuring high-resolution spectra is that it helps resolve an ambiguity between signals reflected by the surface and signals scattered by the atmosphere (Mitchell and O'Brien, 1987). However, the accuracy of surface pressure retrievals from the lower resolution O2 A-band measurement systems that are available, but cannot resolve the reflected/scattered ambiguity, is considerably poorer and these measurements are not suitable for NWP. For instance, van Diedenhoven et al. (2005) have found biases of 20 hPa in surface pressure values retrieved from Scanning Imaging Absorption spectrometer for Atmospheric Chartography (SCIAMACHY) measurements, and Lindstrot et al. (2009) have reported root-mean-square errors of between 10 and 15 hPa for surface pressure retrieved over land from the Medium-Resolution Imaging Spectrometer (MERIS) measurements, although this can be reduced to ∼ 3 hPa by spatial averaging.

Global Positioning System radio occultation (GPSRO) measurements (e.g. Kursinski et al., 1997) are an alternative source of surface pressure information, averaged over a ∼300–400 km limb path in the horizontal. These measurements are now assimilated at the major operational NWP centres (Healy and Thépaut, 2006; Cucurull et al., 2007; Aparicio and Deblonde, 2008; Poli et al., 2009; Rennie, 2010), and it has generally been found that they have a significant positive impact on upper-tropospheric and mid-stratospheric temperatures. This is primarily because the GPSRO measurements have superior vertical resolution to satellite radiances and they can be assimilated without bias correction to the NWP model. They are also reasonably simple to forward model and assimilate. It is perhaps less obvious that GPSRO measurements should also provide useful surface pressure information. This arises because the assimilated quantities, usually either bending angle (α) or refractivity (N) profiles, are assimilated as functions of a height variable, and the integration of the hydrostatic equation is a component of the GPSRO forward model, H, which maps the NWP model state, x, to observation space, y = H(x). This introduces a clear, physically based sensitivity of the simulated measurements with respect to the NWP model surface pressure. In general, increasing the model surface pressure increases the simulated bending angle or refractivity values, and reducing the surface pressure reduces them, and this sensitivity is sufficiently strong to enable the retrieval of surface pressure information.

Theoretical GPSRO information content studies using one-dimensional variational (1D-Var) retrieval techniques have been used to estimate the surface pressure retrieval errors when it is retrieved simultaneously with temperature and humidity profile information (Healy and Eyre, 2000; Palmer et al., 2000). For example, Healy and Eyre (2000, their Table 5) estimated that a priori (background) surface pressure errors of 1.25 hPa could be reduced by ∼16% to 1.05 hPa, but it has proved more difficult to demonstrate this accuracy with real data. Poli et al. (2002) found that biases in GPS/MET refractivity profiles in the troposphere introduced biased surface pressure increments in a 1D-Var retrieval, which degraded the fits to nearby radiosondes. However, it should be noted that the processing of observations has improved considerably since that time with the introduction of open-loop signal tracking in the troposphere (Sokolovskiy, 2001, gives a discussion of the signal tracking methods). Healy and Thépaut (2006) initially found it necessary to modify the relevant GPSRO tangent-linear and adjoint routines to remove the surface pressure sensitivity introduced by the hydrostatic integration when performing early experiments with Challenging Mini-satellite Payload (CHAMP) measurements, in order to remove large increments in the SH. Although this modification was subsequently reversed, prior to the operational assimilation of GPSRO measurements at ECMWF in December 2006, it has nevertheless remained difficult to conclusively demonstrate the value of GPSRO surface pressure information in an NWP system. This is partly a result of the GPSRO measurement numbers when compared with the number of conventional surface pressure measurements. The combined number of synoptic, METAR (hourly measurements at airports) and ship measurements assimilated per day is typically around 112 000, and there are also ∼13 500 drifting buoy observations per day. In contrast, there are generally around 2200 globally distributed GPSRO bending-angle profiles actively assimilated per day. Nevertheless, given that GPSRO observation numbers may increase in the future with the introduction of new constellations such as the Constellation Observing System for Meteorology, Ionosphere and Climate (COSMIC-2), and that new receivers will be capable of detecting more than one GNSS signal, it remains interesting to explore the impact and possible limitations of the surface pressure information provided by GPSRO.

We have investigated the ability of GPSRO measurements to constrain the surface pressure field in a degraded NWP system, where all conventional surface pressure observations are removed. These experiments are similar to those conducted by Thépaut and Kelly (2004) and Dumelow (2004), although both the NWP systems and data availability have changed significantly since those studies. It can be demonstrated that the GPSRO measurements contain useful surface pressure information. It is shown that, although the GPSRO measurements are not able to fully compensate for the loss of all surface pressure observations (particularly in the Northern Hemisphere, NH), they are able to constrain the global mean surface pressure field, and mitigate the evolution of large systematic surface pressure errors. However, we also show that the results with GPSRO are sensitive to small changes in the forward model, and to both NWP model bias and biases introduced by other observations, which can alias into erroneous surface pressure increments. We have also compared the impact of surface pressure provided by GPSRO with that provided by other parts of the Global Observing System (GOS). These results may be useful when designing the future GOS. In particular, it is important to understand the role of conventional observations as the satellite component of the GOS evolves.

The assimilation of GPS bending-angle profiles at ECMWF is described in section 2. The assimilation experiments are described and the main results are presented in section 3. The discussion and conclusions are given in section 4.

2. Assimilation of GPS measurements

The GPSRO measurement technique is described in detail by Kursinski et al. (1997). ECMWF assimilates ionospheric corrected bending-angle profiles, α, as a function of impact parameter, a, with a one-dimensional observation operator. The approach is an updated version of the observation operator described by Healy and Thépaut (2006), and it is reviewed briefly here to highlight the main differences. The forward model essentially contains three major components. The first is the integration of the hydrostatic equation, to compute the height of the NWP model levels. The second is the evaluation of refractive index, n, (or equivalently refractivity, N = 106(n − 1)) on the model levels, and the third is the evaluation of the bending-angle integral.

The GPSRO sensitivity to surface pressure arises because the variables are assimilated as a function of a height variable, a, so the computation and transformation between height variables is important in the context of retrieving surface pressure information. In the observation operator, the geopotential heights of the model levels are evaluated using standard routines which are part of the NWP system (Simmons and Burridge, 1981), but following the work of Aparicio et al. (2009) these geopotential heights are now adjusted to account for non-ideal gas effects (Healy, 2011). This adjustment generally reduces the model geopotential heights by around ∼0.05%. The geopotential heights are converted to geometric height, h, with a transform given by List (1984), but we use g0 = 9.80665 m s−2, rather than the value 9.8 m s−2 used by List in the original formula. The accuracy of this transform has been considered by comparing it with the EGM96 geoid model (Lemoine et al., 1998). We have found that it can introduce errors of ∼(−)4.7 m at a height of 30 km (e.g. latitude = 20°, longitude = −65°), and the differences tend to be largest where the geoid undulation is greatest. However, more generally the differences are smaller, and the root-mean-square differences between the List transform and EGM96 averaged over the globe are only 0.64 m and 0.93 m at 20 and 30 km, respectively. Forward model errors of this magnitude will introduce errors of order 0.1 hPa, are therefore do not have a significant impact on the retrieval of surface pressure from GPSRO measurements. The geometric heights, h, are mapped to radius values, r, by adding the radius of curvature, Rc, and geoid undulation, u, provided with the measurements (r = h + Rc + u).

The refractivity, N, on model levels is given by a three-term formulation (Thayer, 1974),

equation image(1)

where Pd is the pressure of dry air, e is the vapour pressure, T is the temperature, and Zd and Zw are the dry air and water-vapour compressibilities, respectively. The numbers are the ‘best average’ refractivity coefficients proposed by Rüeger (2002), but the k1 = 77.643 K hPa−1 coefficient has been revised to account for its use in a formula that includes non-ideal gas compressibility (Healy, 2011). The values of the refractivity coefficients, and the formulation of the refractivity equation, are still areas of active research, and they are both a potential source of systematic forward model error. For example, recent work by Aparicio and Laroche (2011) suggests a reduction of the k1 coefficient to equation image K hPa−1, based on latest measurements of the polarisibilty of oxygen in the physics literature, which is ∼ 0.09% lower than the k1 value used here.

The observation operator evaluates total ionospheric corrected bending angle, α, as a function of impact parameter, a, by integrating

equation image(2)

where N is the refractivity derived from the model and x = (1 + 10−6N)r, where r is the radius value of a point on the ray-path. This simplification of the bending-angle integral assumes equation image and equation image. It is assumed that the refractivity varies exponentially with x between model levels, leading to a computation of the integral in terms of a sum of Gaussian error functions (Healy and Thépaut, 2006).

The bending-angle operator used operationally at ECMWF now includes ‘tangent point drift’ (e.g. Cucurull, 2012), but most results presented here do not include this effect, unless stated otherwise in the text.

3. Assimilation experiments

The surface pressure information content of GPSRO measurements has been investigated in a series of assimilation experiments, covering the period 15 October 2009 to 31 January 2010. The experiments use the IFS CYCLE 36R1, which was operational at ECMWF from 26 January 2010 to 22 June 2010. The experiments are run at T511 resolution and use incremental 4D-Var assimilation (Rabier et al., 2000; Klinker et al., 2000) with a 12 h assimilation window. The satellite radiances are bias corrected using VARBC (Dee, 2005), where the bias correction coefficients are estimated as part of the analysis. Both COSMIC and GRAS (Global Navigation Satellite System Receiver for Atmospheric Sounding) GPSRO measurements are assimilated without bias correction. The COSMIC measurements are assimilated between the surface and 50 km, whilst the GRAS measurements between 8 and 50 km in the NH and SH Extratropics, and between 10 and 50 km in the Tropics. A single combined observation plus forward model error covariance matrix is used globally for all instruments. The standard deviations of the combined errors are assumed to vary with impact height, which is defined as h = aRc. The percentage error is assumed to be 20% of the observed value at h = 0, falling linearly with h to 1% at h = 10 km. Above 10 km, the error is assumed to be 1% of the observed value until this reaches a lower limit of 3×10−6 radians. The errors are assumed to be uncorrelated in the vertical.

3.1. Removing conventional surface pressure observations from the NWP system

The experiments have been designed to illustrate the surface pressure information content of GPSRO measurements by selectively removing (‘blacklisting’) different combinations of observations from the 4D-Var system. The three principal experiments are:

  • The full observing system assimilated operationally at ECMWF (The appendix gives details): ‘CTL’.

  • The full observing system minus all conventional surface pressure observations: ‘NOPS’.

  • The full observing system minus all conventional surface pressure observations and all GPSRO measurements: ‘NO(PS+RO)’.

A number of additional experiments–generally of shorter duration–have also been performed to investigate different aspects of the results obtained from the three main experiments. These shorter experiments are described in the text. The verification scores are generally against the operational ECMWF analyses calculated with a 12 h assimilation window, taken as a proxy for truth, which is reasonable since we are looking at significantly degraded observing systems.

Figure 1 shows the spatial variation of the mean and standard deviation of the (NOPS minus ECMWF operations) and (NO(PS+RO) minus ECMWF operations) surface pressure analysis differences, computed at both 0000 and 1200 UTC for the period 1 November 2009 to 31 January 2010. The assimilation of the GPSRO measurements in the NOPS experiment clearly produces better agreement with the ECMWF operational surface pressure analyses, in terms of both the mean and standard deviation values.

Figure 1.

The (a) mean and (b) standard deviation of the (NOPS minus ECMWF operations) surface pressure analysis differences (hPa), and the (c) mean and (d) standard deviation of the (NO(PS+RO) minus ECMWF operations) surface pressure analysis differences (hPa), for the period 1 November 2009 to 31 January 2010.

Figures 2 and 3 show the time series of the mean and standard deviation of the errors in 24 h forecasts of pressure at mean sea level (PMSL) in the NH and SH Extratropics. We generally show 24 h forecast scores, but the corresponding results for the 12 h and 36 h forecast ranges are very similar and their use would not change the main results presented here. In addition, the general characteristics of the results for the Tropics are similar to the NH and SH Extratropics, so the results for the Tropics are not shown here. These results clearly demonstrate that the GPSRO measurements provide a useful constraint on the surface pressure when comparing the NOPS and NO(PS+RO) experiments, with both the standard deviation and mean of the forecast errors being reduced throughout the period of the experiment. In particular, the mean errors when the GPSRO data are assimilated are reasonably stable in time, between −1 to −1.5 hPa, whereas they can be as large as −4 hPa when the GPSRO measurements are not assimilated. The mean errors in each experiment are reasonably stable over the day-1 to day-6 forecast range, as shown for the NH in Figure 4, where the statistics are computed for the period 1 November 2009 to 31 January 2010. Large global systematic surface pressure errors in experiments where all conventional surface pressure measurements are blacklisted have been reported previously by Thépaut and Kelly (2004) and Dumelow (2004), although the size of the errors found here are larger than those in Thépaut and Kelly (2004), which were ∼3–4 m in the 1000 hPa geopotential height field, globally. Thépaut and Kelly (2004) emphasised that the conventional surface pressure observations are essential in order to ‘anchor’ the NWP systems. ‘Anchoring’ is a general concept used here and elsewhere to describe measurements that prevent–or at least constrain–a gradual drift of the NWP system away from the true state. Clearly, in this general sense, GPSRO measurements are also able to constrain the mean state of the surface pressure field in a similar way, albeit with a residual mean error.

Figure 2.

The time series of the (a) mean and (b) standard deviation of the errors in the 24 h PMSL forecasts in the Northern Hemisphere for the CTL (black dot-dashed), NOPS (black solid) and NO(PS+RO) (grey solid) experiments, over the period 15 October 2009 to 31 January 2010.

Figure 3.

As Figure 2, but for the Southern Hemisphere.

Figure 4.

The mean errors in PMSL in the northern hemisphere as a function of forecast range for the CTL (black dot-dashed), NOPS (black solid) and NO(PS+RO) (grey solid) experiments, for the period 1 November 2009 to 31 January 2010.

The GPSRO measurements have the largest impact on the standard deviation of the errors in the SH with a reduction of around 0.06 hPa between 1 November 2009 and 31 January 2010. However it is also clear that, at the short range, the GPSRO measurements are not able to fully compensate for the loss of all the conventional surface pressures observations, particularly in the NH, where the standard deviation of the 24 h forecast error with the full observing system is about 0.14 hPa smaller. This is perhaps not surprising given the number and the spatial distribution of conventional surface pressure measurements that have been removed from the 4D-Var. The impact on the standard deviation of the surface pressure errors as a function of forecast range is shown in Figure 5. The NH scores are clearly degraded throughout the entire forecast range as a result of removing the conventional surface pressure observations, and the GPSRO measurements clearly cannot compensate for this loss. This degradation is statistically significant at the 95% level from day 1 to day 6. However in the SH, where the number of conventional surface pressure observations is lower, the GPSRO measurements have some impact. The differences between the NOPS experiment and CTL experiment are slightly smaller, but there is a statistically significant degradation between day 1 to day 3, with neutral results after day 4. The standard deviations of the surface pressure errors are generally smaller in the NOPS experiment than in the NO(PS+RO), and these improvements are statistically significant at day 1 and day 2 in the NH, and from day 1 to day 5 in the SH. Qualitatively similar results are found for the 500 hPa geopotential height scores in the NH and SH.

Figure 5.

The standard deviation of the errors in PMSL in the (a) Northern and (b) Southern Hemispheres, as a function of forecast range for the CTL (black dot-dashed), NOPS (black solid) and NO(PS+RO) experiments, over the period 1 November 2009 to 31 January 2010.

The large mean PMSL errors in the NO(PS+RO) experiment shown in Figures 2 and 3, and their similarity in both hemispheres, have been investigated. The mean PMSL errors are caused by an interaction between the assimilation of the conventional upper-air observations in the NH, and the bias correction of radiances using VARBC. The mean errors can be reduced if the VARBC coefficients are fixed to their value at the start of the experiment (0000 UTC, 15 October 2009) and do not evolve as the experiment progresses. Figure 6 shows the mean 24 h PMSL errors in the SH with and without the VARBC coefficients evolving as a function of time, in the NO(PS+RO) experimental configuration. Similar results are found in the NH and Tropics. The mean error values gradually diverge, and they differ by ∼1.5 hPa after 28 days of experimentation, so assimilating radiances with VARBC clearly has a significant impact on the mean errors shown in Figures 2 and 3. This illustrates how the conventional surface pressure observations anchor the bias corrections applied to satellite radiances. It may also partly explain why the mean errors found in this work are larger than seen by Thépaut and Kelly (2004), who employed a static bias correction scheme, which implicitly uses surface pressure information from the period when the bias coefficients are calculated from the radiance departures. Similar arguments apply to the differences shown in Figure 6. However, we have also found that the satellite radiances do not initiate, or drive, the surface pressure errors in the early stages of the experiment, by performing two experiments which split the observation dataset used in the NO(PS+RO) experiment into (i) satellite minus GPSRO, and (ii) conventional-only subsets. Surprisingly, in the first experiment, running with no anchor measurements, the mean and standard deviation of the PMSL errors are small and consistent in time, indicating that this subset of observations does not cause the systematic PMSL errors.

Figure 6.

The time series of the mean errors in 24 h PMSL forecasts (hPa) in the Southern Hemisphere in the NO(PS+RO) experiment when the VARBC coefficients used to correct satellite radiances are fixed (black line) and when they are allowed to evolve (grey line).

In contrast, large PMSL errors quickly develop over North America and Northern Europe in the conventional-only experiment. Mean differences of –1 hPa relative to both the CTL experiment and the operational surface pressure analysis are apparent after the first assimilation cycle (Figure 7). They arise as a result of correlations in the background-error covariance matrix because there is a negative correlation between surface pressure errors and temperature errors near the surface and around the tropopause. Any systematic positive temperature increments at these levels result in a reduction in the surface pressure. Aircraft temperature measurements appear to be the main cause (Ballish and Kumar, 2008). As the experiment progresses, the negative surface pressure biases gradually increase and spread across the NH and into the Tropics, but large positive errors develop in the SH south of latitude −30°, which is inconsistent with Figure 3. Furthermore, negative mean surface pressure errors in the SH do not occur in the (satellite minus GPSRO) experiment either. Therefore, the SH errors in the NO(PS+RO) experiment must be caused by the combination of the two subsets of observations: the errors are initially introduced through the assimilation of conventional upper-air observations in the NH, but then amplified and spread across the globe through the VARBC because of the anchoring to these conventional observations. This accounts for the similar behaviour in both hemispheres. This explanation in terms of VARBC spreading the information also appears to be consistent with the work of Dee and Uppala (2009), who have shown how a drift in the bias corrections applied to tropospheric Advanced Microwave Sounding Unit (AMSU-A) channels globally is caused by the increasing numbers of biased aircraft temperature measurements (which are predominantly located in the NH) assimilated in ERA-Interim.

Figure 7.

The PMSL analysis differences (hPa, shading) of the ‘conventional only’ (excluding surface pressure measurements) minus CTL experiments after the first assimilation cycle at 0000 UTC on 15 October 2009. The surface pressure analysis of the CTL experiment is also shown (contours).

The GPSRO measurements in the NOPS experiment are able to constrain the mean PMSL errors globally to within −1.5 hPa, and this is fairly constant in time. We have found that the magnitude of this residual mean PMSL error is directly related to biases of order −0.1% to −0.2% in the ‘observed minus background’ (o–b) bending-angle departures, between ∼10 and 30 km, where the GPSRO measurements have high information content and are given the most weight in the assimilation process. Biases in the bending angle (o–b) departures of this size in this height interval is a robust feature for all GPSRO instruments, including both COSMIC and GRAS which have been processed at different centres. In the CTL experiment, the (o–b) departures remain biased at the −0.1% to −0.2% level between 10 and 30 km throughout the three-month period of the experiment. However, in the NOPS experiment, where the conventional surface pressure observations are removed from the assimilation system, the biases in the (o–b) bending-angle departures are reduced, and essentially removed, within a few assimilation cycles by reducing the surface pressure. This improvement in the GPSRO bending-angle departures is then retained throughout the experiment. The surface pressure is essentially being used as a sink variable in the first few cycles to remove a GPSRO (o–b) bias that is apparent in the upper troposphere and lower/middle stratosphere. We have found that the residual mean surface pressure errors are essentially unchanged when GPSRO measurements below 10 km are blacklisted. Furthermore, the sensitivity to small biases of order 0.1% has also been demonstrated in an experiment where all the forward modelled bending angles are artificially reduced by 0.1%, resulting in a reduction in the residual mean surface pressure error of ∼ 0.7 hPa globally. In this context, note that the k1 coefficient suggested in Aparicio and Laroche (2011) would reduce the forward modelled bending angles by ∼ 0.09% above 10 km, and therefore this could account for just under half of the −1.5 hPa mean error.

Although forward model biases may account for some of the residual −1.5 hPa mean PMSL error, this is not the complete picture. When considering this mean PMSL error, it must be recognised that GPSRO observations do not measure the surface pressure directly, and so it has to be estimated simultaneously with temperature and humidity profiles, using a priori forecast state and error estimates in order to partition the information appropriately. Linear estimation theory shows that biases in the analysis state, equation image, can be related to biases in the a priori state vector, equation image, and the combined observation and forward modelling biases, equation image, by the matrix equation

equation image(3)

where, I is the identity matrix, K is the gain matrix and H is the gradient of the observation operator with respect to the state vector. Therefore, biases in this a priori information and other observations can produce biases in the analyses, even if the GPSRO measurements in the vector of observations are unbiased. The bias in the bending angle (o–b) departures between 10 and 30 km is partly caused by a cold NWP model bias in the stratosphere, and a warm bias in the mid and upper troposphere. In the NH, the latter is a result of assimilating aircraft temperature measurements, which are known to be biased warm at cruise level (Ballish and Kumar, 2008). (Aircraft temperature measurements have been bias-corrected operationally at ECMWF since 15 November 2011, but these measurements are not bias-corrected in this work). The influence of this warm bias, which peaks around 200 hPa near the aircraft cruising altitude, is apparent well above the aircraft cruising altitude because of the shape of the bending-angle weighting functions which form the GPSRO rows of H (Eyre, 1994); the weighting functions have a long, positive hydrostatic tail, meaning that increasing the temperature below a ray tangent height of a given ray path will increase the ray bending. Therefore, a warm bias in the troposphere will result in stratospheric model levels which are too high, and this will increase the value of the simulated bending angles in the stratosphere, and contribute to the negative (o–b) bias in the 10–30 km vertical interval. In the CTL experiment, when all observations are assimilated, GPSRO measurements tend to pull the analysis away from the aircraft temperatures, and produce a closer fit to radiosonde temperature observations at 200 hPa. However, when the surface pressure observations are removed, this is no longer the case, and the fit to aircraft measurements is improved whilst the fit to radiosondes is degraded slightly. It appears that the GPSRO measurements reduce the surface pressure, rather than correct the temperature bias introduced by the aircraft measurements, when no conventional surface pressure measurements are assimilated. Removing the aircraft temperature measurements from the assimilation system gradually produces an improvement in the residual mean surface pressure error, as shown in Figure 8 for the NH, with a reduction of ∼0.7 hPa in the final month of the experiment (1–31 January 2010), with the largest differences being over North America. Some improvements at the end of the experiment are also found in the Tropics (∼0.6 hPa) and in the SH (∼0.5 hPa). This is an encouraging result showing that the residual mean errors at the −1 to −1.5 hPa level are not purely determined by the GPSRO measurement accuracy or the forward modelling. However, it also clearly highlights a limitation, since model biases, or biases introduced by other observations, can be aliased into surface pressure increments because the GPSRO measurements are sensitive to more than one variable, and the vertical weighting functions are complicated. This problem will arise with any surface pressure information retrieved from satellite observations, because the observations will be sensitive to other parameters, and potentially biased a priori information will be used to estimate the surface pressure component.

Figure 8.

The time series of the mean errors in 24 h PMSL forecasts (hPa) in the Northern Hemisphere when the aircraft temperature measurements are removed from the NO(PS+RO) experiment (black line) and when they are included (grey line).

3.2. Contribution to the surface pressure analysis from various observing systems

As part of this study, we have also investigated the impact on the surface pressure analyses and forecasts of assimilating just GPSRO measurements (‘GPSRO-only’), and compared it with following configurations:

  • GPSRO and all scatterometer measurements assimilated operationally: ‘GPSRO+SCAT’.

  • All conventional observations, including all available surface pressure observations: ‘CONV-only’.

  • All satellite observations, including all radiances, GPSRO, atmospheric motion vectors and scatterometers, and no conventional observations, with VARBC evolving with time: ‘ALL-SAT’.

  • The control experiment defined in section 3.1: ‘CTL’.

The GPSRO+SCAT has been included because these observing systems should complement each other, since the scatterometer will provide information on the gradients of surface pressure over the oceans on scales which are difficult to detect with GPSRO, because of the width of the GPSRO horizontal weighting function. The SCAT measurements were provided by ERS, ASCAT and QUIKSCAT until 22 November 2009. The ‘ALL-SAT’ experiment demonstrates how much information the satellite component of the GOS contributes to the surface pressure analyses, in the absence of any conventional measurements. Note that, in the GPSRO-only, GPSRO+SCAT and ALL-SAT experiments the GPSRO measurements have been assimilated with tangent point drift included in the observation operator, so they are not completely consistent with the CTL experiment, which assumes a fixed tangent point location. However, this improvement in the GPSRO operator does not affect the main results presented here.

The mean and standard deviation of PMSL errors in the NH and SH are shown in Figure 9, computed for the 61-day period, from 1 November to 31 December 2009. There are a number of interesting points. The mean PMSL forecast errors are largest when GPSRO measurements are assimilated, probably as a result of systematic NWP model errors being mapped into surface pressure increments, and biases introduced in the processing of observations to bending angle and the forward modelling. However, note that the biases are largest in the ALL-SAT experiment, again suggesting an amplification of the biases introduced by GPSRO through the bias correction of the satellite radiances.

Figure 9.

The (a) mean and (b) standard deviation of the PMSL errors in the Northern Hemisphere for the day-1 to day-6 forecast range, in the GPSRO-only (grey solid), GPSRO+SCAT (black solid), ALL-SAT (black dots), CONV-only (grey dashed) and CTL (black dashed) experiments, for the period 1 November to 31 December 2009. (c, d) are as (a, b), but for the Southern Hemisphere.

The standard deviations of the PMSL errors are smaller in the GPSRO+SCAT experiment than in the GPSRO-only experiment throughout the forecast range. We have found that the inclusion of scatterometer measurements significantly improves the PMSL analyses over the oceans, both in terms of the positioning and structure of synoptic systems, even in some cases introducing low pressure systems which are absent in the GPSRO-only analyses. For example, we have found cases where large analysis errors of order ∼22 hPa are reduced to ∼2 hPa as a result of assimilating the scatterometer measurements. It has also been verified that assimilating only scatterometer measurements produces much larger errors than assimilating just GPSRO measurements; the standard deviations of the 24 h forecast surface pressure errors reach ∼7 hPa and 3 hPa in the northern and southern Extratropics, respectively, after 10 days. The standard deviations of the PMSL forecast errors in the ALL-SAT experiment are lower than in both the GPSRO and GPSRO+SCAT configurations, demonstrating that the 4D-Var data assimilation system is able to retrieve useful surface pressure information from the full range of satellite data currently assimilated, and that most of this information is from sources other than GPSRO. The impact of the ALL-SAT and CONV-only configurations on the standard deviations is similar in the NH, but in the SH the CONV-only experiment is considerably poorer. Figure 10 shows the standard deviations of the (CONV-only minus ECMWF operations) and (ALL-SAT minus ECMWF operations) surface pressure analysis differences computed for the period 1 November to 31 December 2009. The (ALL-SAT minus ECMWF operations) differences tend to be larger over land than sea, and are largest in the NH. This is because fewer satellite measurements are assimilated over land. The CONV-only experiment is reasonably consistent with the operational ECMWF analyses in the NH, but the differences are large in the SH polar region. It is found that the standard deviation of forecast errors in the SH grow much more rapidly in the CONV-only configuration as a result of the analysis errors south of −60°. The standard deviation of the surface pressure analysis differences with respect to ECMWF operational analyses can be larger than 5 hPa over ocean in this region (e.g. latitude = −70°, longitude = −159°). This illustrates that the satellite measurements are largely determining the accuracy of the surface pressure analysis south of −60°.

Figure 10.

The standard deviation of the (a) (CONV-only minus ECMWF operations) and (b) (ALL-SAT minus ECMWF operations) surface pressure analysis differences (hPa), for the period 1 November to 31 December 2009.

It is also notable that the CONV-only error standard deviation scores are poorer than the GPSRO-only after day-2, but this difference is not significant at the 95% level. The difference between the GPSRO+SCAT and CONV-only is statistically significant throughout the forecast range.

It is not a surprising result to find that satellite data are more important than conventional data in the SH, because this has been demonstrated a number of times by removing all satellite data from the data assimilation system. However, it is surprising to see that the satellite data performs so well in the absence of any conventional data, because it is generally assumed that some conventional data is required to anchor the NWP system when satellite information is being assimilated.

4. Discussion and conclusions

We have performed data assimilation experiments to investigate the surface pressure information content of GPSRO measurements, and to see how this compares with different components of the global observing system. A number of experiments have been conducted with all of the conventional surface pressure observations removed, in order to study the impact of GPSRO. We have demonstrated in the NOPS experiment that the GPSRO measurements can constrain the mean surface pressure errors to be ∼ −1.5 hPa in these degraded NWP systems, and this is fairly stable in time. In contrast, the mean errors vary considerably when neither the conventional surface pressure nor GPSRO measurements are assimilated (NO(PS+RO)), and they can be as large as −4 hPa. The residual mean errors in the NOPS experiment can be related to biases (o–b) bending angles departures between 10 and 30 km. We have also found that reducing the forward modelled bending angles by 0.1% can improve the surface pressure biases by 0.7 hPa, and this may suggest that forward model bias arising because of uncertainty in the refractivity coefficients is causing some of the biases. More specifically, the analysis of Aparicio and Laroche (2011) suggests equation image K hPa−1 compared to the value k1 = 77.643 K hPa−1 used in this work. The implementation of this revised value requires careful evaluation, but clearly it would have an impact on the mean surface pressure errors shown here. However, the bending angle departure biases between 10 and 30 km are also related to ECMWF NWP model biases, and a temperature bias introduced by aircraft temperature measurements. (Aircraft temperature measurements are now bias-corrected at ECMWF.) The mean surface pressure errors can be improved by ∼0.5–0.7 hPa by removing aircraft temperature measurements from the analysis, with the largest improvements in the NH. This sensitivity of the surface pressure results to NWP model biases or biases introduced by other observations will be a problem for any surface pressure information retrieved from satellite measurements, because the measurements will also be sensitive to the temperature and humidity information. This limitation emphasises the importance of conventional observations that specifically measure surface pressure.

The large surface pressure errors found in the NO(PS+RO) experiment are caused by an interaction between the assimilation of the conventional upper-air observations and the bias correction of radiances with VARBC. The assimilation of the conventional observations introduces systematic surface pressure increments in the data-rich regions in the NH. Surprisingly, we have found that biases do not develop when just satellite data (excluding GPSRO) are assimilated, and both the mean and standard deviation of the PMSL errors are small and stable in time. However, the satellite radiance data cannot constrain the impact of biases introduced by the conventional observations, and in fact the anchoring of the radiance bias corrections to these conventional observations through the VARBC amplifies and spreads the large systematic errors across the globe. This also accounts for the similarity of the time series of mean errors in the NH and SH, shown in Figures 2 and 3.

The impact of various components of the GOS on the surface pressure analyses and forecasts has been investigated. It has been shown that the combination of GPSRO and scatterometer clearly reduces the standard deviation of the PMSL errors when compared with assimilating just GPSRO, illustrating the complementary nature of these measurement systems. However, the ALL-SAT experiments are better than both of these configurations in terms of the standard deviation of the PMSL errors, showing that the 4D-Var assimilation system is able to retrieve useful surface pressure information from all of the satellite observations, and also indicating that the contribution of the GPSRO is currently relatively modest. However, as noted above, in the absence of GPSRO, the satellite radiance data is unable to correct or constrain any biases introduced by other sources, and indeed the impact of these biases tends to be amplified. It has also been shown that the impacts of the CONV-only and ALL-SAT are similar in the NH, but the ALL-SAT is clearly superior in the SH. The results in the SH are consistent with previous impact studies, where satellite data have been removed from a full system, but it is surprising that reasonable results can be obtained with no conventional measurements anchoring the NWP system.

The experiments assimilating just GPSRO measurements are characterised by mean errors with a magnitude of less than 0.4 hPa. These are caused by both NWP model biases, and biases introduced in the preprocessing and forward modelling of the GPSRO measurements. As noted above, in the context of the observation/forward modelling errors, we have found that artificially reducing the forward modelled bending angles by 0.1% increased the surface pressure values by +0.7 hPa. This level of sensitivity suggests the observation/forward model biases of order 0.01% are required for the retrieval of accurate, essentially bias-free surface pressure information in this NWP system. Since the change in the operational processing of COSMIC data on 12 October 2009, there has been very good consistency in the bias characteristics of all of the GPSRO instruments in the 10–30 km vertical interval. (The statistics of matched occultations are produced operationally by the Radio Occultation Meteorology Satellite Applications Facility (ROM SAF); http://www.romsaf.org/monitoring/matched.php). Although this is not proof that the measurements are unbiased, because different processing centres may be making the same errors, it provides some confidence in the quality of the data.

One area of future work relates to the experiment performed to understand the biases in the NO(PS+RO) experiment, where only satellite data (excluding GPSRO) were assimilated. This experiment did not contain any conventional or GPSRO anchor measurements, but both the mean and standard deviation of the PMSL errors remained small and stable with time. This result requires further investigation.

In summary, we have shown that it is possible to retrieve useful surface pressure information from satellite measurements, and use it to constrain an NWP system. A number of experiments have been performed to specifically investigate the surface pressure information content of GPSRO measurements, and it has been shown that the measurements can constrain the development of large mean surface pressure errors in the analyses and forecasts, and reduce the standard deviation of the forecast errors. However, the accuracy of the surface pressure results is sensitive to forward model bias, NWP model biases, and biases introduced by other observations. The sensitivity to NWP model error and biases introduced by other observations is because the GPSRO does not measure surface pressure directly. This limitation is a general problem with any surface pressure information retrieved from satellite measurements, highlighting the intrinsic value of conventional observations. Nevertheless, it has been demonstrated that a 4D-Var system can produce accurate surface pressure analyses and forecasts assimilating just satellite data.

Appendix

The control experiment used all conventional and satellite observations assimilated operationally at ECMWF during the period October 2009 to January 2010. Table A1 summarises the measurements that were assimilated, and includes the major ‘blacklisting’ changes during the period, and when instruments were added, removed or became passive in the NWP system.

Table A1. Summary of observations assimilated in the control experiment.
Conventional measurements
Radiosondes 
Pilot winds 
Drifting buoys 
SynopticLAND, SHIP, SHRED, METAR
AircraftAIREP, AMDAR, ACARS
Wind profilersAmerican, European and Japanese
  1. BL = blacklisted

Satellite sounder/imager measurements
PlatformSensor
METOP-AHIRS, AMSU-A, MHS, IASI
METEOSAT-7METEOSAT
METEOSAT-9MSG
ENVISATSCHIAMACHY (from 16 Sep 2009),
 MERIS
MTSATMTSATIMG
NOAA-15AMSUA
NOAA-16AMSUA, AMSUB
NOAA-17SBUV, HIRS,
 AMSUB (BL from 22 Dec 2009)
NOAA-18AMSUA (from 12 Jan 2010),
 AMSUB
NOAA-19HIRS, MHS,
 AMSUA (from 12 Jan 2010)
DMSP-15SSMI (BL from 24 Nov 2009)
GOES-11GOESIMG
GOES-12GOESIMG
AQUAAIRS (BL from 9 Jan 2010),
 AMSUA, AMSRE
AURAOMI
Other satellite measurements
ScatterometersERS, ASCAT,
 QUIKSCAT (to 22 Nov 2009)
GPSROMETOP-A GRAS, COSMIC
AtmosphericMETEOSAT-7, METEOSAT-9,
motion vectorsMTSAT-1R
 (BL 23 Nov to 8 Dec 2009),
 GOES-10, GOES-11,
 TERRA, AQUA

Acknowledgements

This work was conducted as part of the Radio Occultation Meteorology Satellite Applications Facility (ROM SAF), which is a decentralised opexsrational Radio Occultation processing centre under EUMETSAT.

Ancillary