• Anthes, R. A., and Coauthors, 2008: The COSMIC/FORMOSAT-3 Mission: Early results. Bull. Amer. Meteor. Soc., 89, 313333.

  • Aparicio, J. M., , and G. Deblonde, 2008: Impact of the assimilation of CHAMP refractivity profiles in Environment Canada global forecasts. Mon. Wea. Rev., 136, 257275.

    • Search Google Scholar
    • Export Citation
  • Auligné, T., , A. P. McNally, , and D. P. Dee, 2007: Adaptive bias correction for satellite data in a numerical weather prediction system. Quart. J. Roy. Meteor. Soc., 133, 631642.

    • Search Google Scholar
    • Export Citation
  • Bauer, P., , and G. Radnóti, 2009: Study on Observing System Experiments (OSEs) for the evaluation of degraded EPS/Post-EPS instrument scenarios. ECMWF Rep., EUMETSAT Contract EUM/CO/07/4600000454/PS, ECMWF, Reading, United Kingdom, 99 pp.

  • Bechtold, P., and Coauthors, 2012: Tropical errors and convection. ECMWF Tech. Memo. 685, ECMWF, Reading, United Kingdom, 82 pp.

  • Cardinali, C., , and S. Healy, 2013: Evaluation of the assimilation of GPS-RO observations at ECMWF. Quart. J. Roy. Meteor. Soc., in press.

    • Search Google Scholar
    • Export Citation
  • Cardinali, C., , S. Pezzulli, , and E. Andersson, 2004: Influence-matrix diagnostic of a data assimilation system. Quart. J. Roy. Meteor. Soc., 130, 27672785.

    • Search Google Scholar
    • Export Citation
  • Cucurull, L., , J. C. Derber, , R. Treadon, , and R. J. Purser, 2007: Assimilation of Global Positioning System Radio Occultation Observations into NCEP’s Global Data Assimilation System. Mon. Wea. Rev., 135, 31743193.

    • Search Google Scholar
    • Export Citation
  • Dee, D., 2005: Bias and data assimilation. Quart. J. Roy. Meteor. Soc., 131, 33233343.

  • Dee, D., , and S. Uppala, 2009: Variational bias correction of satellite radiance data in the ERA-Interim reanalysis. Quart. J. Roy. Meteor. Soc., 135, 18301841.

    • Search Google Scholar
    • Export Citation
  • Eyre, J. R., 1994: Assimilation of radio occultation measurements into a numerical weather prediction system. Tech. Memo. 199, ECMWF, Reading, United Kingdom, 22 pp.

  • Haseler, J., 2004: Early delivery suite. ECMWF Tech. Memo. 454, ECMWF, Reading, United Kingdom, 26 pp.

  • Healy, S. B., , and J.-N. Thépaut, 2006: Assimilation experiments with CHAMP GPS radio occultation measurements. Quart. J. Roy. Meteor. Soc., 132, 605623.

    • Search Google Scholar
    • Export Citation
  • Kursinski, E. R., and Coauthors, 1996: Initial results of radio occultation observations of earth’s atmosphere using the Global Positioning System. Science, 271, 11071110.

    • Search Google Scholar
    • Export Citation
  • Poli, P., , S. B. Healy, , F. Rabier, , and J. Pailleux, 2008: Preliminary assessment of the scalability of GPS radio occultations impact in numerical weather prediction. Geophys. Res. Lett., 35, L23811, doi:10.1029/2008GL035873.

    • Search Google Scholar
    • Export Citation
  • Poli, P., , P. Moll, , D. Puech, , F. Rabier, , and S. B. Healy, 2009: Quality control, error analysis, and impact assessment of FORMOSAT-3/COSMIC in numerical weather prediction. Terr. Atmos. Ocean, 20, 101113.

    • Search Google Scholar
    • Export Citation
  • Poli, P., , S. B. Healy, , and D. P. Dee, 2010: Assimilation of Global Positioning System radio occultation data in the ECMWF ERA-interim reanalysis. Quart. J. Roy. Meteor. Soc., 136, 19721990, doi:10.1002/qj.722.

    • Search Google Scholar
    • Export Citation
  • Rennie, M. P., 2010: The impact of GPS radio occultation assimilation at the Met Office. Quart. J. Roy. Meteor. Soc., 136, 116131, doi:10.1002/qj.521.

    • Search Google Scholar
    • Export Citation
  • Rocken, C., and Coauthors, 1997: Analysis and validation of GPS/MET data in the neutral atmosphere. J. Geophys. Res.,102 (D25), 29 849–29 866.

  • Wickert, J., and Coauthors, 2001: Atmosphere sounding by GPS radio occultation: First results from CHAMP. Geophys. Res. Lett., 28, 32633266.

    • Search Google Scholar
    • Export Citation
  • Wu, D., , A. Mannucci, , F. Xie, , C. Ao, , D. Diner, , and J. Teixeira, 2011: Climate and weather sensors on iridium-NEXT: A combined GPS-RO and WindCam system for PBL remote sensing. Proc. GEOScan Workshop, Annapolis, MD, NASA Jet Propulsion Laboratory, 21 pp.

  • View in gallery

    Horizontal distribution of all assimilated GPS-RO data for a given 12-h 4D-Var assimilation window.

  • View in gallery

    Mean June–August 2008 day-1 model temperature and meridional wind error in (a) logarithmic and (b) linear scale to emphasize stratosphere and troposphere, respectively. Bold colors indicate errors that are statistically significant.

  • View in gallery

    Mean temperature analysis difference between experiments with 100% and (a) 0%, (b) 5%, (c) 33%, and (d) 67% of the total observation number available in July 2008. (left) 100- and (right) 200-hPa levels. Results are from the 1200 UTC analyses only.

  • View in gallery

    As in Fig. 3, but for experiments without radiosondes.

  • View in gallery

    Vertical profile of global GPS-RO observation DFS loss when only 67% (black bars), 33% (dark gray bars), and 5% (light gray bars) of the data are present compared to the full constellation (100%). Results obtained from 16 Aug 2008. Units are bits.

  • View in gallery

    Departure between 9-h model forecast (first guess) and radiosonde observations of temperatures in the tropical stratosphere (10–100 hPa). (a) Standard deviations and (b) biases for GPS-RO OSEs with 100% (dash–triple dotted), 67% (dash–dotted), 33% (dashed), 5% (dotted), and 0% (solid) of the total observation number available in 15 Jul–15 Aug 2008. Statistics based on 50 000–80 000 data per level.

  • View in gallery

    As in Fig. 6, but for troposphere (100–1000 hPa).

  • View in gallery

    Model fit to AMSU-A channels 5–14 radiance observations in Southern Hemisphere. Standard deviations of (a) first-guess departures normalized to 0% GPS-RO constellation, and (b) mean first-guess departures (thin lines) and bias corrections (thick lines).

  • View in gallery

    As in Fig. 8, but for Northern Hemisphere. 33% GPS-RO constellation (dashed), if aircraft (dotted), or radiosonde (solid) temperature data are removed.

  • View in gallery

    Normalized RMS temperature error difference between GPS-RO denial experiments (numbers denote the percentage used) and control for temperature. Positive values indicate positive impact of GPS-RO data. (from left to right) Southern Hemisphere, tropics, and Northern Hemisphere. (from top to bottom) 200, 500, 700, and 1000 hPa. Forecast verification is against operational analysis, verification period is 1 Jul–30 Sep 2008.

  • View in gallery

    Dependence of 24-h temperature forecast error relative to 0%-GPS-RO experiment (%) at (a) 50, (b) 100, (c) 150, (d) 200, (e) 500, and (f) 1000 hPa. Solid, dotted, and dashed lines denote scores for Northern Hemisphere, Southern Hemisphere, and tropics, respectively, over the period 15 Jul–15 Aug 2008.

  • View in gallery

    Temperature forecast root-mean-square error reduction if GPS-RO (red) or IASI (black) data are added to a baseline observing system (see text for details). (from left to right) Southern Hemisphere, tropics, and Northern Hemisphere. (from top to bottom) 200, 500, 850, and 1000 hPa. Forecast verification is against operational ECMWF analysis, verification period is 7 Jul–31 Aug 2008.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 93 93 58
PDF Downloads 53 53 35

GNSS Radio Occultation Constellation Observing System Experiments

View More View Less
  • 1 European Centre for Medium-Range Weather Forecasts, Reading, United Kingdom
© Get Permissions
Full access

Abstract

Observing system experiments within the operational ECMWF data assimilation framework have been performed for summer 2008 when the largest recorded number of Global Navigation Satellite System (GNSS) radio occultation observations from both operational and experimental satellites were available. Constellations with 0%, 5%, 33%, 67%, and 100% data volume were assimilated to quantify the sensitivity of analysis and forecast quality to radio occultation data volume. These observations mostly constrain upper-tropospheric and stratospheric temperatures and correct an apparent model bias that changes sign across the upper-troposphere–lower-stratosphere boundary. This correction effect does not saturate with increasing data volume, even if more data are assimilated than available in today’s analyses. Another important function of radio occultation data, namely, the anchoring of variational radiance bias corrections, is demonstrated in this study. This effect also does not saturate with increasing data volume. In the stratosphere, the anchoring by radio occultation data is stronger than provided by radiosonde and aircraft observations.

Corresponding author address: Peter Bauer, ECMWF, Shinfield Park, Reading RG2 9AX, United Kingdom. E-mail: peter.bauer@ecmwf.int

Abstract

Observing system experiments within the operational ECMWF data assimilation framework have been performed for summer 2008 when the largest recorded number of Global Navigation Satellite System (GNSS) radio occultation observations from both operational and experimental satellites were available. Constellations with 0%, 5%, 33%, 67%, and 100% data volume were assimilated to quantify the sensitivity of analysis and forecast quality to radio occultation data volume. These observations mostly constrain upper-tropospheric and stratospheric temperatures and correct an apparent model bias that changes sign across the upper-troposphere–lower-stratosphere boundary. This correction effect does not saturate with increasing data volume, even if more data are assimilated than available in today’s analyses. Another important function of radio occultation data, namely, the anchoring of variational radiance bias corrections, is demonstrated in this study. This effect also does not saturate with increasing data volume. In the stratosphere, the anchoring by radio occultation data is stronger than provided by radiosonde and aircraft observations.

Corresponding author address: Peter Bauer, ECMWF, Shinfield Park, Reading RG2 9AX, United Kingdom. E-mail: peter.bauer@ecmwf.int

1. Introduction

The feasibility of making radio occultation (RO) measurements using Global Navigation Satellite System (GNSS-RO) transmissions was demonstrated in the global positioning system/meteorology program (GPS/MET; Kursinski et al. 1996; Rocken et al. 1997) and the Challenging Minisatellite Payload (CHAMP; Wickert et al. 2001) missions. GNSS-RO measurements have subsequently proven to be a very valuable addition to the operational numerical weather prediction (NWP) observing system, since the launch of the joint Taiwan–U.S. Constellation Observing System for Meteorology, Ionosphere, and Climate (COSMIC)/Formosa Satellite Mission 3 (COSMIC/FORMOSAT-3) constellation of receivers in 2006 (Anthes et al. 2008). The measurements are particularly useful in NWP because they can be assimilated without bias correction, and they have strong sensitivity to small-scale temperature structures in the mid/upper troposphere and stratosphere, an area that is otherwise only weakly constrained by other observations in the analysis, and that is prone to large model uncertainties. Many operational NWP centers have now reported a significant impact on upper-tropospheric and stratospheric temperatures in their NWP systems with the current number of observations (Healy and Thépaut 2006; Cucurull et al. 2007; Aparicio and Deblonde 2008; Poli et al. 2009; Rennie 2010), and they also have a significant impact in climate reanalyses (Poli et al. 2010).

To date, only receivers of the GPS signals have been deployed. Hereafter, these will be referred to as GPS-RO observations. In the future, it is expected to also have multiple receivers using the European Galileo and other systems, such as Globalnaya Navigatsionnaya Sputnikovaya Sistema (GLONASS), which will potentially greatly increase the number of occultations available for assimilation into NWP systems. However, it is an open question if more occultation measurements are required for NWP applications, or whether the impact of GPS-RO is already close to saturation with the current observation numbers.

In this context, the term “saturation” refers to the decrease of additional forecast improvement when adding more data. In theory this will not happen in a perfectly constrained data assimilation system because, even if the same location and time are sampled multiple times by observations with uncorrelated measurement errors, the analysis error should be reduced as a result of the decrease of observational noise. However, the analysis system’s limitations and forecast error metrics do not resolve further improvement indefinitely and therefore saturation of impact is observed.

The objective of this paper is the quantification of the rate of radio occultation observational impact with increasing observation number, in both the analyses and forecasts. This work extends a preliminary investigation presented by Poli et al. (2008), who demonstrated that removing 50% of the GPS-RO data in the Météo-France NWP system degraded the analysis and forecast quality. However, the GPS-RO measurements have been used more conservatively in their system than we do here. The present study aims at supporting current constellation assessment and future mission planning, for example in the framework of the joint University Corporation for Atmospheric Research (UCAR)–Taiwanese project COSMIC-2, the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT) Polar System Second Generation (EPS-SG), or considerations to add GNSS receivers on research missions of opportunity or commercial programs (Wu et al. 2011). The need for enhanced GNSS radio-occultation observing systems is also highlighted in the World Meteorological Organization’s Vision for the Global Observing System (http://www.wmo.int/pages/prog/www/OSY/GOS-redesign.html).

The impact quantification is based on observing system experiments (OSEs) within the European Centre for Medium-Range Weather Forecasts (ECMWF) data assimilation framework and by varying the number of assimilated GPS-RO observations. Section 2 describes which OSEs have been performed employing increasing numbers of observations from both operational [the GNSS Receiver for Atmospheric Sounding (GRAS)] and experimental missions [COSMIC, CHAMP, Gravity Recovery and Climate Experiment (GRACE-A), TerraSAR-X, etc.], for the period July–September 2008, which represents the most data-rich time for GNSS observations so far. Given this maximum constellation, reduced scenarios can be simulated by data denial and the incremental reduction of observation impact on analyses and forecasts can be tested. This assessment is presented in sections 3 and 4 where the main question of whether impact saturation has been reached is also addressed.

Another aspect of this study is the importance of GNSS observations to provide anchoring points for variational bias corrections that are widely applied to satellite radiances. These represent the bulk of the observational data volume in operational NWP systems. The presence of model biases, particularly in the upper troposphere–lower stratosphere (UTLS) and in the stratosphere are likely to be absorbed by radiance bias corrections. Anchoring observations, such as radiosondes and GPS-RO, reduce this effect and produce a more consistent weight given to observational data (Dee and Uppala 2009).

The study is summarized and conclusions are drawn in section 5. The scenarios employed in this study also provide information to aid future constellation design. Furthermore, the results can be used for the “calibration” of assimilation experiments with simulated GPS-RO data, designed to estimate the impact of increasing observation numbers significantly beyond those currently available (e.g., Harnisch et al. 2013, manuscript submitted to Mon. Wea. Rev.).

2. Configuration

a. Model data assimilation

The experiments have been run with model cycle CY36R4 of the Integrated Forecast System (IFS) that became operational at ECMWF on 11 September 2010. This choice ensured that very recent versions of model physics, data assimilation, and observation treatment were incorporated in the OSEs. The experiments have been run at reduced horizontal resolution, namely with a T511 wavenumber truncation (40 km versus the current operational resolution of T1279, i.e., 16 km). The vertical resolution of the experiments has been kept at 91 levels with a model top level pressure of 0.01 hPa. Previous OSEs have indicated that a reduced horizontal resolution produced sufficiently accurate results due to the large-scale structure of increments produced by temperature-related observations (Bauer and Radnóti 2009).

The system has been run in the global ECMWF four-dimensional variational data assimilation (4D-Var) configuration that produces two analyses per day (valid at 0000 and 1200 UTC) with 12-h assimilation windows, also known as delayed cutoff data assimilation (DCDA; Haseler 2004). Only one medium-range forecast has been run per day as opposed to two forecasts in the operational configuration. Despite the different data coverage in the 0000 and 1200 UTC window, this is not expected to have affected the results and it provides sufficient statistical significance due to large enough samples. The main experimentation periods were chosen to be July–September 2008, the latter having been chosen because of the additional experimental GPS-RO data available for this period.

All experiments have been initialized with the operational suite on the first day of the period. A 14-day spinup phase was included allowing for the system to adjust to the enhanced GPS-RO observing system that was activated on day one. Given that the global observation influence of the observations is about 0.2 in the ECMWF system (Cardinali et al. 2004), it is expected that by the end of the spinup period, the system has lost memory of the operational system used to initialize the first analysis. A global observation influence of 0.2 means that 20% of the information content in a given analysis is provided by new observations while 80% are contributed from previous analyses, and thus observations, integrated forward in time by the model.

The experiment evaluation has been restricted to the remaining part of the period. The evaluation has been performed based on standard observation consistency statistics to evaluate the impact of adding or denying selected observations on the fit of the analysis and short-range forecast to all other observations. A generally improved (degraded) system will draw closer to (away from) all other observations. This tool provides a robust means of impact verification due to the large quantities of assimilated data from a wide range of instruments. Further, standard forecast skill scores for key parameters (geopotential heights, temperatures, vector wind at 1000, 500, 200 hPa; global and regional) including tests of statistical significance are computed. Additional evaluation has been perfomed using the diagnostic tools of analysis sensitivity to observations (Cardinali et al. 2004).

The variational bias correction (Dee 2005; Auligné et al. 2007) for the experimentation period was initialized with the operational system output on the initial date and kept active throughout the experimentation periods. This ensured that, as in the operational system, a trade-off between analysis increments and the evolution of biases as a function of model state is performed.

b. Background observing system

Since the experiments are meant to represent current and near-future conditions, the background, which is the non–GPS-RO observing system, should mimic a configuration that will be available during the next, say, 10 years. The set of satellite sounding instruments that has been available over the past 5 years has been rather rich since many early satellites such as National Oceanic and Atmospheric Administration-15 (NOAA-15) are still operational but more recent systems such as the National Aeronautics and Space Administration (NASA)–NOAA Suomi National Polar-orbiting partnership (Suomi-NPP) and the EUMETSAT Meteorological Operational (MetOp-A/B) have been added. For reference, the 2009 radiance observing system is compiled in Table 1. In addition to radiances, clear-sky radiance (CSR) and atmospheric motion vector (AMV) products are assimilated from (Meteorological Satellite-7 and -9) Meteosat-7 and -9, (Geostationary Operational Environmental Satellite-11 and -12) GOES-11 and -12, and Multifunctional Transport Satellite-1R (MTSAT-1R) as well as AMVs from Aqua and Terra Moderate Resolution Imaging Spectroradiometer (MODIS), GPS-RO observations from COSMIC 1–6 and MetOp-A GRAS as well as scatterometer wind vectors from MetOp-A Advanced Scatterometer (ASCAT), the European Remote Sensing satellite-2 (ERS-2) scatterometer and Quick Scatterometer (QuikSCAT) SeaWinds, total column ozone products from NOAA-17 and -18 Solar Backscatter Ultraviolet Instrument (SBUV), Aura Ozone Monitoring Instrument (OMI), while MetOp-A Global Ozone Monitoring Experiment-2 (GOME-2), Environmental Satellite (Envisat) Scanning Imaging Absorption Spectrometer for Atmospheric Chartography (SCIAMACHY), Meteosat Second Generation (MSG) Spinning Enhanced Visible and Infrared Imager (SEVIRI), and Envisat Global Ozone Monitoring by Occultation of Stars on the Michaelson Interferometer for Passive Atmospheric Sounding (GOMOS/MIPAS) data are only monitored.

Table 1.

Satellite radiance observing system used at ECMWF (status March 2009, monitored instruments in italic).

Table 1.

Over the next 10 years, we do not expect an increase of available observations since several aging satellites are expected to become decomissioned. A basic estimate of availability is one sounder system [High Resolution Infrared Radiation Sounder (HIRS), Advanced Microwave Sounding Unit-A (AMSU-A), Microwave Humidity Sounder (MHS), and the Infrared Atmospheric Sounding Interferometer (IASI)] on board the prime MetOp-A satellite in the midmorning orbit, two conventional sounder systems (HIRS, AMSU-A, MHS) and one advanced sounder (CrIS) from NOAA satellites [Suomi-NPP, Joint Polar Satellite System-1 (JPSS-1)] in the afternoon orbit, two microwave imaging systems in morning orbits [Defense Meteorological Satellite Program (DMSP) satellites F-17 or F-18 Special Sensor Microwave Imager/Sounder (SSMIS) and future Department of Defense satellites], two scatterometers on board MetOp-A and Oceansat-2 (and follow-on), 3–5 GPS receivers (COSMIC, COSMIC-2, GRAS), and three total column ozone sensing instruments (SBUV on board NOAA-18 or -19, GOME-2 on board MetOp-A, and OMPS on board Suomi-NPP and JPSS-1). This set is expected to be complemented by at least five satellites in geosynchronous orbits (two European, two U.S., and one Japanese satellite), and altimetry data from research satellites. The number of HIRS instruments may be revised given the fact that this series will be discontinued in the future since its observing capabilities are fully covered by advanced sounders. Chinese instruments are not accounted for in this context because the expected observation quality and data availability is not sufficiently known.

In this study, the reference observing system has been defined as a subset of the operationally used set of observations described before. With respect to the operational satellite observation usage (Table 1) the following systems remained:

  • for conventional (HIRS, AMSU-A/B, MHS) and advanced (AIRS, IASI) soundings only those from NOAA-18, Aqua, and MetOp-A;
  • for microwave imager radiances only SSM/I data from DMSP-13;
  • for total column ozone products only those from SBUV instruments on board NOAA-17 and NOAA-18;
  • for scatterometry only ERS-2, QuikSCAT and MetOp-A ASCAT data;
  • only AMVs from geostationary satellites (GOES-East/West, Meteosat-7/9, MTSAT-1).
This forms the background observing system to which different GPS-RO constellations have been added.

c. GPS-RO constellation

The experiments with GPS-RO data have been performed for 3 months, July–September 2008, a period for which experimental GPS-RO data from CHAMP, GRACE-A, TerraSAR-X, and SAC-C have been acquired in addition to the COSMIC and GRAS data that are operationally used at ECMWF. The COSMIC data used in this study was a reprocessed dataset, using the latest operational processing code employed at the UCAR. The reprocessed COSMIC data is thus more consistent with GRAS data, in terms of bias characteristics of the bending angles in the lower/midstratosphere, although the biases with respect to ECMWF short-range forecasts have increased. The SAC-C and TerraSAR-X datasets were also processed at UCAR using the latest operational code, specifically for this study. The combined number of bending angle profiles available per day, averaged over the 3-month period, is 2940 which is about 20% more data than used in the 2009 operational system. This number is comparable to the number of occultations that are currently available (as of January 2013), since the successful launch of MetOp-B on 17 September 2012, which includes one GRAS instrument.

In the control experiment all available GPS-RO data have been assimilated and three denial experiments have been run assimilating only 5%, 33%, and 67% of the GPSRO data, respectively. The 5% experiment (i.e., about 150 profiles) is representative of the data numbers from a single instrument like a mission of opportunity (e.g., CHAMP). The 33% and 67% experiments roughly represent a half and a full COSMIC constellation, respectively. The total number of different bending angle observations in a randomly chosen 4D-Var cycle can be seen in Table 2 and the horizontal distribution of all the assimilated GPSRO data in the same cycle is shown in Fig. 1.

Table 2.

Total number of GPS-RO bending angle observations from different receivers for a typical 12-h 4D-Var assimilation window.

Table 2.
Fig. 1.
Fig. 1.

Horizontal distribution of all assimilated GPS-RO data for a given 12-h 4D-Var assimilation window.

Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00130.1

3. Analysis impact

Since GPS-RO observations mainly constrain temperature in the analysis with excellent vertical resolution across the mid- to upper troposphere and throughout the stratosphere it is worth showing the mean ECMWF model errors as zonal cross section (Fig. 2). The mean errors are defined as the mean difference between 24-h forecasts and operational analyses verifying at the same time. Figure 2a suggests that across the stratosphere the model has a cold bias of up to 0.5 K that broadens in altitude toward the summer hemisphere and that breaks up and becomes less pronounced toward the winter hemisphere. Below the tropopause and mostly in the summer hemisphere, the bias changes sign with similar magnitude and extends down to 700–800 hPa. The positive bias is collocated with areas of summertime convection over land and deep convection over warm oceans and the ITCZ. The cold bias in the lower stratosphere is also most pronounced in areas with convection. These structures suggest that the biases have a dependence on radiation and its interaction with water vapor, trace gases and clouds as well as on the troposphere–stratosphere exchange and the model’s representation of the tropopause.

Fig. 2.
Fig. 2.

Mean June–August 2008 day-1 model temperature and meridional wind error in (a) logarithmic and (b) linear scale to emphasize stratosphere and troposphere, respectively. Bold colors indicate errors that are statistically significant.

Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00130.1

The meridional wind errors in Fig. 2 are largest in the tropics and suggest that the Hadley circulation is generally too weak. However, the model exhibits quite substantial regional variability of lower-level divergence errors (Bechtold et al. 2012) depending on land–sea distribution and global/regional circulation. The forecast errors beyond day 1 tend to amplify with forecast range, particularly in the stratosphere. Thus observations correct for a systematic model error in the analysis and both GPS-RO and infrared/microwave radiance data contribute most strongly to this correction in upper troposphere and stratosphere.

In theory, 4D-Var assumes a perfect and thus unbiased model. In reality, the observation bias correction does absorb some of the model biases but the radiance bias correction is set up per channel and has rather large-scale predictors. These do not resolve finer-scale structures that greatly vary between latitudes and heights. Consequently, even bias-corrected radiance observations do correct mean model errors. GPS-RO observations are not bias corrected and hence act even stronger on correcting mean model errors.

a. Mean state

Figure 3 shows the impact of successively removing GPS-RO observations on mean temperature analysis at 100 and 200 hPa (i.e., the height region in which the mean model-minus-observation difference changes sign in the UTLS; see Figs. 6 and 7). The differences have been calculated with reference to the full constellation that is 100% of GPS-RO data being available. The sign change between these heights is clearly visible and the decreasing relative impact of GPS-RO data going toward larger observation volumes. The strongest effect is seen in the tropics and over oceans where fewer radiosonde (and aircraft) observations are available. Therefore, the weakest impact is seen over the northern midlatitude continents where the highest-density radiosonde networks are found. The GPS-RO data impact there is nearly zero and thus does not much depend on the number of available occultations. However, over oceanic areas the difference between analysis from the 67% and 100% GPS-RO coverage experiments is rather significant and suggests that even more observations will produce also stronger impact.

Fig. 3.
Fig. 3.

Mean temperature analysis difference between experiments with 100% and (a) 0%, (b) 5%, (c) 33%, and (d) 67% of the total observation number available in July 2008. (left) 100- and (right) 200-hPa levels. Results are from the 1200 UTC analyses only.

Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00130.1

If the same set of experiments is run and also radiosonde data are withdrawn (Fig. 4) the main, above described features are still visible and the areas of impact are deepened. At 100 hPa the overall effect of GPS-RO data on the analysis is now even larger and now also noticeable over midlatitude land areas in the Northern Hemisphere. The impact at 200 hPa is quite similar to the experiments in which radiosonde data are included, most probably reflecting the importance of aircraft temperature measurements at this level in the Northern Hemisphere. It is evident that radiosonde data, similar to GPS-RO, perform similar work on adjusting the model bias in the analysis in terms of magnitude and sign. Because of the approximately homogeneous coverage of occultations over the globe, the mean analysis increments are distributed more evenly in space (and time) and will thus produce a geographically more balanced analysis state that is not available from radiosondes alone.

Fig. 4.
Fig. 4.

As in Fig. 3, but for experiments without radiosondes.

Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00130.1

b. DFS

The analysis impact of individual observation types can be further illustrated using the degree of freedom of signal (DFS; Cardinali et al. 2004). The DFS measures the observational influence in the data assimilation scheme as a function of data volume, observation, and background error variances as well as the linearized observation/model operator. The value of mean DFS or (mean observation influence) is expected to be between 0 and 1, where mean DFS = 0 means that the observation has no influence and DFS = 1 means that the analysis at the given point is fully driven by that observation and not the model background. For gaining computational efficiency, the actual calculation is performed by a numerical approximation so that values outside this range can occur. Of importance is that the DFS provides an analysis impact assessment without the need for performing OSEs and with the entire observing system being present.

Note that the DFS denotes the information content of the observations if they are used optimally in the analysis and if the assumptions the analysis is based on are correct; for example that the system is linear, unbiased, and the error characteristics of background and observations have Gaussian distributions.

The loss of total DFS as a function of GPS-RO data denial for both GPS-RO and all observations with respect to the full constellation (100% data) increases linearly with data loss to about 15% when all GPS-RO data have been withdrawn. This number emphasizes the fundamental importance of occultation measurements in the analysis given the fact that about 10 times more radiance observations are assimilated. The DFS loss can also be displayed as a function of height (Cardinali and Healy 2013) and is largest where the mean observation influence is largest. This is shown as the vertical loss profile of DFS for the 67%, 33%, and 5% OSEs in Fig. 5. The figure suggests that the loss of information follows the mean observation influence profile and that it changes linearly with data volume. The total loss of information is therefore largest where the information provided by this observation type is largest, between 15 and 25 km (i.e., also where the largest impact of GPS-RO data on mean analysis state is seen), where the model errors change sign and the high vertical resolution of GPS-RO data is thus most beneficial. This is also the height range in which the observation errors that are assigned to GPS-RO observations have their smallest values.

Fig. 5.
Fig. 5.

Vertical profile of global GPS-RO observation DFS loss when only 67% (black bars), 33% (dark gray bars), and 5% (light gray bars) of the data are present compared to the full constellation (100%). Results obtained from 16 Aug 2008. Units are bits.

Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00130.1

c. Fit analysis-observations

Better analyses are expected to produce a consistently better fit of the model fields when compared against all observations. This assumption applies to both the analysis and first guess (i.e., the short-range forecast that produces a first estimate of the actual state and that has been initialized from the previous analysis). This method is very stable and generally considered unambiguous.

Figure 6 shows the fit of short-range temperature forecasts with respect to tropical radiosonde observations from the experiments using 0%, 5%, 33%, 67%, and 100% of the GPS-RO observations as listed in Table 2, respectively. The effect of changing GPS-RO constellation on standard deviations is shown relative to the full constellation (i.e., 100% data; Fig. 6a). Withdrawing all GPS-RO bending angles increases standard deviations by 3%–4% and this change is fairly uniform between 10 and 100 hPa. Even reducing the number to 67% (i.e., reducing the maximum possible constellation to what is operationally available today) is noticeable and decreases the fits by 0.7%. The impact on biases is shown in Fig. 6b and reveals a much larger height dependence. The short-range forecast errors without GPS-RO observations are near zero between 50 and 70 hPa and increase significantly above and below this layer. Adding GPS-RO observations reduces this height dependence until a fairly constant mean difference of 0.2–0.3 K is reached with the full constellation.

Fig. 6.
Fig. 6.

Departure between 9-h model forecast (first guess) and radiosonde observations of temperatures in the tropical stratosphere (10–100 hPa). (a) Standard deviations and (b) biases for GPS-RO OSEs with 100% (dash–triple dotted), 67% (dash–dotted), 33% (dashed), 5% (dotted), and 0% (solid) of the total observation number available in 15 Jul–15 Aug 2008. Statistics based on 50 000–80 000 data per level.

Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00130.1

Between 50 and 70 hPa in the tropics, differences between the model and radiosondes in the absence of GPS-RO observations tend to be very small. This feature is also visible in Fig. 2a near 25°N latitude and is the result of compensating positive errors over northern Africa and negative errors elsewhere at this latitude and thus not an expression of generally small errors. Adding GPS-RO observations therefore reduces regional differences and produces model-observation departures that are more homogeneous across latitudes and height.

In the troposphere, the interpretation of the GPS-RO impact becomes slightly more complicated (Fig. 7). Again, the standard deviations are reduced with increasing observation number in the mid- to upper troposphere. At lower levels, the impact is rather neutral and the 67% constellation even produces a slightly better fit between the model and radiosondes below 700 hPa. The mean differences show that the GPS-RO observations basically make the model colder throughout the troposphere. This cooling counteracts the model’s warm bias in the upper troposphere but increases the model’s warm bias in the lower troposphere. This is interpreted as an integrated effect of these observations on the atmospheric column from altitudes where the bending angles produce very large increments to levels where the model bias switches sign from the upper troposphere to the lower stratosphere. The increments propagate downward since the weighting function of the bending angle operator has a long hydrostatic tail toward the surface (Eyre 1994), because the computed bending angles are sensitive to the height of the model levels. This effect also has a regional pattern depending on the presence of other assimilated temperature observations as from radiosondes and aircrafts.

Fig. 7.
Fig. 7.

As in Fig. 6, but for troposphere (100–1000 hPa).

Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00130.1

d. Bias correction anchoring

The analysis system employs a variational bias correction Dee (2005) that is applied to the majority of satellite data and selected conventional observations. This bias correction accounts for fluctuations in instrument calibration and enhances the consistency between the rather diverse observation types, but it is also prone to absorb model bias. It is therefore useful to investigate the change of bias corrections with respect to changes in the number of observations that do not require bias correction when assimilated.

Figure 8 shows the difference between the fit of the analysis to AMSU-A channels 5–14 radiance observations between experiments, and the mean bias corrections applied to these observations in the Southern Hemisphere. The impact of changing GPS-RO observation number on the standard deviations is visible as was the case for radiosondes but with values of 1%–2% or O(0.01 K) is not significant. The impact on mean differences is nearly zero but the impact on mean bias correction is more obvious (up to 0.15 K; Fig. 8b). This demonstrates that the GPS-RO observations anchor the bias correction without affecting the analysis and first-guess statistics much. The degree of anchoring changes with the number of GPS-RO observations. Bias corrections become smaller with increasing numbers for channels 6–12. This is a good sign as adding unbiased anchoring GPS-RO observations should reduce the biases across a consistent analysis system and therefore the (uncorrected) GPS-RO data counteracts model bias, and so smaller radiance bias corrections are needed.

Fig. 8.
Fig. 8.

Model fit to AMSU-A channels 5–14 radiance observations in Southern Hemisphere. Standard deviations of (a) first-guess departures normalized to 0% GPS-RO constellation, and (b) mean first-guess departures (thin lines) and bias corrections (thick lines).

Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00130.1

The weighting functions of AMSU-A channels 6–12 peak between 400 and 10 hPa and covering those altitudes at which GPS-RO observation error variances are the smallest and, therefore, their effect on the analysis the strongest. Note that the weighting functions of radiance observations are fairly broad and do not resolve the more detailed model bias structure seen in Fig. 2. Thus, the change of sign for bias corrections as a function of channel reflects the combined effect of model, radiometer calibration, and radiative transfer biases.

As with the radiosonde temperatures (Fig. 7), the increasing bias corrections with increasing GPS-RO data volume for levels below 400 hPa (AMSU-A channel 5) relates to the downward propagation of GPS-RO temperature increments due to their long hydrostatic weighting function tail. Again, the different curves are just laterally displaced. Above 10 hPa a different mechanism comes into play. The bias correction of AMSU-A channel 14 is fixed to zero to avoid it being too strongly determined by significant model biases. This also has an impact on the bias correction of channels 12–13 due to their overlapping weighting functions. It was only possible to fix the channel 14 bias correction to zero when GPS-RO measurements were also being assimilated (A. P. McNally 2007, personal communication).

Removing alternative temperature observations from the background observing system should change the above statistics in a similar way as shown in Fig. 7. Figure 9 shows this effect based on the 33% GPS-RO constellation for the Northern Hemisphere where the highest density of complementary temperature observations exists. Removing radiosonde (solid) or aircraft (dotted) temperature data increases first-guess error standard deviations by a small amount and slightly increases AMSU-A bias corrections, most significantly for tropospheric channels. Again, for altitudes where GPS-RO observations receive the highest weight (see Fig. 5), the withdrawal of data is least visible indicating the dominant anchoring effect by GPS-RO data. Even at levels where the assimilation of aircraft observations is most effective (e.g., at 200–250 hPa over the North Atlantic and Pacific), little impact of these data on both infrared and microwave sounder bias corrections is found. These results represent large-scale statistics and a more detailed investigation into the impact of bias correction anchoring by conventional observations on smaller spatial and temporal scales is required.

Fig. 9.
Fig. 9.

As in Fig. 8, but for Northern Hemisphere. 33% GPS-RO constellation (dashed), if aircraft (dotted), or radiosonde (solid) temperature data are removed.

Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00130.1

4. Forecast impact

Forecast evaluation is more difficult because different reference standards can produce rather different scores. Bauer and Radnóti (2009) have shown that an evaluation with the experiment’s own analysis or the operational analysis can produce inconsistent results, in some cases with opposite signs in the short range. The argument for using the experiment’s own analyses for evaluation is that if an additional observation is expected to change the mean analysis state, only the own analysis provides a fair reference while any other would produce a poorer analysis state. This, however, can also produce problems in case of a very poor observing system and simply due to the fact that additional observations can add systematically larger increments to the analysis and therefore increase the root-mean-square difference between (short range) forecasts and analyses.

Here, standard forecast evaluation is performed with the operational analyses. The operational observing system contains more data in the above experimentation periods since the background system of the experiments was designed to represent poorer conditions as expected in the forthcoming 10 years. Additionally, the operational system is run at higher spatial resolution.

Figure 10 shows the normalized RMS forecast error difference between the different GPS-RO denial experiments (0%–67%) and the control experiment (100%) for tropospheric temperatures. Down to 500 hPa the more GPS-RO data are used the better the scores become. The impact lasts 2–3 days at 200 and 700 hPa but into the medium range at 500 hPa. The error reduction at 200 hPa is rather dramatic and reaches 50% at day 1 in the tropics and Southern Hemisphere.

Fig. 10.
Fig. 10.

Normalized RMS temperature error difference between GPS-RO denial experiments (numbers denote the percentage used) and control for temperature. Positive values indicate positive impact of GPS-RO data. (from left to right) Southern Hemisphere, tropics, and Northern Hemisphere. (from top to bottom) 200, 500, 700, and 1000 hPa. Forecast verification is against operational analysis, verification period is 1 Jul–30 Sep 2008.

Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00130.1

The negative effect of adding GPS-RO observations at 1000 hPa, especially in the tropics, matches the radiosonde statistics in Fig. 7. The GPS-RO data produce a cooling in the mid- to upper troposphere that propagates down toward the surface due to the above-mentioned long weighting function tail. The switch of sign from a cold to a warm model bias in the boundary layer cannot be resolved by the GPS-RO observations. There, it is also affected by increased noise and is thus assimilated with increased observation error variances. Therefore, the upper-tropospheric signal pushes the analysis away from other observations at lower levels and thus increases the analysis and forecast errors near the surface.

Another feature in Fig. 10 is a negative impact between days 2 and 4 at 200 hPa in the Northern Hemisphere and at 500 hPa in the Southern Hemisphere. Again, this deterioration increases with GPS-RO data volume. This degradation is most likely an artifact from the differences in GPS-RO data preprocessing between the denial experiments and the operational system, the latter being used for verification in this figure. The UCAR processing (see section 2c) translates to a small but noticeable change of observation-minus-model biases that produces suboptimal analyses. Areas where the model bias changes sign are particularly sensitive to this difference.

Figure 11 presents the reduction of temperature forecast errors across stratosphere and troposphere relative to the OSE in which all GPS-RO have been withdrawn. Note that 24-h forecast errors are between 0.5 and 1.5 K, and that the largest errors are found near 100 hPa and in the tropics. The relative improvement of scores by adding GPS-RO observations is significant when 5% and 33% of the totally available data are added and tapers slightly off by adding more data. Error reduction reaches more than 20% at 50, 100, and 150 hPa and becomes weaker in the mid- to lower troposphere. An important observation is that—given the shape of the curves—impact saturation has not been reached and that we can expect further improvement if more data were available.

Fig. 11.
Fig. 11.

Dependence of 24-h temperature forecast error relative to 0%-GPS-RO experiment (%) at (a) 50, (b) 100, (c) 150, (d) 200, (e) 500, and (f) 1000 hPa. Solid, dotted, and dashed lines denote scores for Northern Hemisphere, Southern Hemisphere, and tropics, respectively, over the period 15 Jul–15 Aug 2008.

Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00130.1

Within the context of this study it was not tested whether reducing observation errors and thus increasing GPS-RO observation weight could produce such a saturation. The retuning of observation errors is rather complex and should not be performed for one observing system in isolation. This is due to the fact that the balance between observation and background error formulation as well as the design of the bias correction employing a perfect model assumption requires tuning and cannot be obtained from first principles alone.

Last, the general capability of GPS-RO data to constrain temperature analyses has been compared with other satellite data, especially from advanced sounders. This has been investigated with another set of experiments in a poor-baseline context, in which no satellite radiance and aircraft observations have been assimilated. This setup helps to amplify the impact of individual observing systems and thus to distinguish individual contributions from the nonlinear superposition of all components in an operational framework. Here, the experiments have also been verified with the operational ECMWF analysis.

Figure 12 shows the temperature RMS forecast error reduction at various levels and latitude bands if GPS-RO (100%; red curves) or IASI radiance data (black curves) are added to the baseline, respectively. Note that these experiments reflect that typically 3 million IASI radiances and 300 000 bending angles are being assimilated per 12-h cycle. In addition, GPS-RO data are assimilated with less weight in the lower troposphere and the upper stratosphere and thus receives the biggest weight between heights of 8–40 km.

Fig. 12.
Fig. 12.

Temperature forecast root-mean-square error reduction if GPS-RO (red) or IASI (black) data are added to a baseline observing system (see text for details). (from left to right) Southern Hemisphere, tropics, and Northern Hemisphere. (from top to bottom) 200, 500, 850, and 1000 hPa. Forecast verification is against operational ECMWF analysis, verification period is 7 Jul–31 Aug 2008.

Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00130.1

At 200 hPa in the Southern Hemisphere and at 500 hPa in the tropics, both observing systems produce almost identical impact. At 200 hPa in the tropics, GPS-RO data even outperform IASI data for the first 2 days. At the remaining levels and areas, IASI performs slightly better. In the Northern Hemisphere, the difference between all experiments is smaller due to the large conventional observation density. While these experiments do not reflect the individual instrument impact in a fully operational framework, they highlight how powerful GPS-RO data are in current systems and how this can compare to other rather sophisticated instruments.

5. Summary and conclusions

GPS-RO data represent a very important part of the observing system because they do not require bias correction and produce accurate observations of (mainly) temperature with very good vertical resolution in the upper troposphere and stratosphere where radiance data are more sparse and NWP model errors are large. For the experiments in this study, CHAMP, GRACE-A, TerraSAR-X, and SAC-C data in addition to the currently assimilated receivers onboard MetOp-A GRAS and the COSMIC/FORMOSAT-3 constellation have been acquired to make available as many GPS-RO observations as possible. All data have been assimilated and three denial experiments have been run using 5%, 33%, and 67% of the GPS-RO data, respectively. The 5% experiment (i.e., about 150 profiles) is representative of the data numbers from a single receiving instrument while the 67% experiment corresponds to the number available from six satellite constellation like COSMIC.

The results from the 5% experiment indicate that the impact on temperature analysis is significant but not yet sufficient to remove the entire upper-atmospheric temperature bias. An important aspect is that model bias changes sign across the tropopause (too warm upper troposphere, too cold lower stratosphere) and that this bias has a seasonal cycle, being more pronounced in the summer hemisphere. GPS-RO observations have high vertical resolution but are assimilated in concert with other, lower-resolution observations such as radiances (with similarly broad bias corrections), and with background error covariance structures that are not well resolved at these levels as well. Thus, the net effect of adding GPS-RO observations is filtered. This also points at areas with large potential for future impact enhancement.

Gradually increasing the number of observations further corrects the model bias, and the analysis fits to radiosonde observations clearly indicate that more observations would produce even more benefit. The radiosonde fits also suggest that the impact has a vertical structure that follows the model bias, namely, that GPS-RO data warm the analysis in the stratosphere and cool it in the troposphere. The impact is also affected by other observations in the system and running further experiments in which radiosonde and aircraft data have been withdrawn as well demonstrated a significant amplification of the GPS-RO contribution. Evaluating the information content of the observations reveals that the information loss increases linearly with the number of observations and follows the vertical profile of mean observation influence. This means that losing GPS-RO observations affects most strongly those heights at which the observations have the largest impact and that coincide with the levels of best vertical resolution, smallest observation error variances and yet model errors with the largest vertical gradients.

The significant role of GPS-RO observations as anchors in variational radiance bias-correction schemes has been demonstrated. Bias corrections become smaller the more GPS-RO data are available. The radiance bias corrections also show sensitivity to removing radiosonde and aircraft temperature observations; however, this is mostly visible in the troposphere whereas GPS-RO data dominate the anchoring throughout the stratosphere. Again, the anchoring effect scales with available data volume.

RMS temperature forecast errors and geopotential height anomaly correlations are positively affected by GPS-RO data, especially in the upper troposphere and lower to midstratosphere. Above 200 hPa the impact remains visible even for 7 days. A surprising negative effect is found, however, at 1000 hPa, especially over the tropics for temperature. This degradation is very likely due to large increments in the upper troposphere propagating downward and simply shifting the entire profile toward lower temperatures. Since the model bias switches sign again in the lower troposphere, the increments produce enhanced errors. This is in agreement with boundary layer radiosonde statistics from the analyses. The main reason for this effect is a downward propagation of increments through the tails of the GPS-RO weighting function.

At higher levels the strongest forecast improvement from adding GPS-RO observation is seen in the Southern Hemisphere, where fewer radiosonde and aircraft temperature observations are available than elsewhere. From the available observations, no clear saturation effect could be found. Even with 50% more observations than assimilated in today’s operational systems (i.e., 100%/67%), further improvement in short-range scores is seen. The actual level of saturation can only be derived from experiments that are based on simulated observations. Given the importance of GPS-RO observations for NWP and the current decline of available satellite instruments, it is worth investigating the sensitivity of NWP analyses and forecasts to larger than existing constellations to produce recommendations for space agencies.

Finally, baseline experiments have been run to compare the impact of IASI radiance data with that of GPS-RO observations. Up to 500 hPa the IASI impact is clearly stronger. At 200 hPa over the Southern Hemisphere IASI and GPS-RO seem to be of equal impact. Over the tropics GPS-RO observations appear to bring slightly more improvement than IASI. Over the Northern Hemisphere and in the first two days GPS-RO data improve the forecast slightly stronger, but afterward the IASI impact overtakes. These results have to be viewed keeping in mind that about one order of magnitude more IASI radiances than bending angles are assimilated at ECMWF.

Acknowledgments

Gábor Radnóti was funded by EUMETSAT Contract EUM/CO/09/4600000644 and EUCOS Contract EUCOS-ADM-2010-001. Sean Healy is funded by the EUMETSAT Radio Occultation Meteorology Satellite Application Facility (ROM SAF). The reprocessing of the COSMIC, TerraSAR-X, GRACE-A, CHAMP, and SAC-C data was performed by Bill Schreiner, Doug Hunt, and Sergey Sokolovskiy (UCAR) and was crucial for performing this study. This work is greatly acknowledged. The authors are grateful for the valuable comments and suggestions from Tony McNally and Florian Harnisch.

REFERENCES

  • Anthes, R. A., and Coauthors, 2008: The COSMIC/FORMOSAT-3 Mission: Early results. Bull. Amer. Meteor. Soc., 89, 313333.

  • Aparicio, J. M., , and G. Deblonde, 2008: Impact of the assimilation of CHAMP refractivity profiles in Environment Canada global forecasts. Mon. Wea. Rev., 136, 257275.

    • Search Google Scholar
    • Export Citation
  • Auligné, T., , A. P. McNally, , and D. P. Dee, 2007: Adaptive bias correction for satellite data in a numerical weather prediction system. Quart. J. Roy. Meteor. Soc., 133, 631642.

    • Search Google Scholar
    • Export Citation
  • Bauer, P., , and G. Radnóti, 2009: Study on Observing System Experiments (OSEs) for the evaluation of degraded EPS/Post-EPS instrument scenarios. ECMWF Rep., EUMETSAT Contract EUM/CO/07/4600000454/PS, ECMWF, Reading, United Kingdom, 99 pp.

  • Bechtold, P., and Coauthors, 2012: Tropical errors and convection. ECMWF Tech. Memo. 685, ECMWF, Reading, United Kingdom, 82 pp.

  • Cardinali, C., , and S. Healy, 2013: Evaluation of the assimilation of GPS-RO observations at ECMWF. Quart. J. Roy. Meteor. Soc., in press.

    • Search Google Scholar
    • Export Citation
  • Cardinali, C., , S. Pezzulli, , and E. Andersson, 2004: Influence-matrix diagnostic of a data assimilation system. Quart. J. Roy. Meteor. Soc., 130, 27672785.

    • Search Google Scholar
    • Export Citation
  • Cucurull, L., , J. C. Derber, , R. Treadon, , and R. J. Purser, 2007: Assimilation of Global Positioning System Radio Occultation Observations into NCEP’s Global Data Assimilation System. Mon. Wea. Rev., 135, 31743193.

    • Search Google Scholar
    • Export Citation
  • Dee, D., 2005: Bias and data assimilation. Quart. J. Roy. Meteor. Soc., 131, 33233343.

  • Dee, D., , and S. Uppala, 2009: Variational bias correction of satellite radiance data in the ERA-Interim reanalysis. Quart. J. Roy. Meteor. Soc., 135, 18301841.

    • Search Google Scholar
    • Export Citation
  • Eyre, J. R., 1994: Assimilation of radio occultation measurements into a numerical weather prediction system. Tech. Memo. 199, ECMWF, Reading, United Kingdom, 22 pp.

  • Haseler, J., 2004: Early delivery suite. ECMWF Tech. Memo. 454, ECMWF, Reading, United Kingdom, 26 pp.

  • Healy, S. B., , and J.-N. Thépaut, 2006: Assimilation experiments with CHAMP GPS radio occultation measurements. Quart. J. Roy. Meteor. Soc., 132, 605623.

    • Search Google Scholar
    • Export Citation
  • Kursinski, E. R., and Coauthors, 1996: Initial results of radio occultation observations of earth’s atmosphere using the Global Positioning System. Science, 271, 11071110.

    • Search Google Scholar
    • Export Citation
  • Poli, P., , S. B. Healy, , F. Rabier, , and J. Pailleux, 2008: Preliminary assessment of the scalability of GPS radio occultations impact in numerical weather prediction. Geophys. Res. Lett., 35, L23811, doi:10.1029/2008GL035873.

    • Search Google Scholar
    • Export Citation
  • Poli, P., , P. Moll, , D. Puech, , F. Rabier, , and S. B. Healy, 2009: Quality control, error analysis, and impact assessment of FORMOSAT-3/COSMIC in numerical weather prediction. Terr. Atmos. Ocean, 20, 101113.

    • Search Google Scholar
    • Export Citation
  • Poli, P., , S. B. Healy, , and D. P. Dee, 2010: Assimilation of Global Positioning System radio occultation data in the ECMWF ERA-interim reanalysis. Quart. J. Roy. Meteor. Soc., 136, 19721990, doi:10.1002/qj.722.

    • Search Google Scholar
    • Export Citation
  • Rennie, M. P., 2010: The impact of GPS radio occultation assimilation at the Met Office. Quart. J. Roy. Meteor. Soc., 136, 116131, doi:10.1002/qj.521.

    • Search Google Scholar
    • Export Citation
  • Rocken, C., and Coauthors, 1997: Analysis and validation of GPS/MET data in the neutral atmosphere. J. Geophys. Res.,102 (D25), 29 849–29 866.

  • Wickert, J., and Coauthors, 2001: Atmosphere sounding by GPS radio occultation: First results from CHAMP. Geophys. Res. Lett., 28, 32633266.

    • Search Google Scholar
    • Export Citation
  • Wu, D., , A. Mannucci, , F. Xie, , C. Ao, , D. Diner, , and J. Teixeira, 2011: Climate and weather sensors on iridium-NEXT: A combined GPS-RO and WindCam system for PBL remote sensing. Proc. GEOScan Workshop, Annapolis, MD, NASA Jet Propulsion Laboratory, 21 pp.

Save