A variational data assimilation system for ground-based GPS slant delays

A variational data assimilation system for ground-based GPS slant delays
of 12
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
  QUARTERLY JOURNAL OF THE ROYAL METEOROLOGICAL SOCIETY Q. J. R. Meteorol. Soc.  133 : 969–980 (2007)Published online in Wiley InterScience(www.interscience.wiley.com) DOI: 10.1002/qj.79 A variational data assimilation system for ground-based GPSslant delays Heikki J¨arvinen, a Reima Eresmaa, a * Henrik Vedel, b Kirsti Salonen, a Sami Niemel¨a a and John de Vries c a Finnish Meteorological Institute, Helsinki, Finland  b  Danish Meteorological Institute, Copenhagen, Denmark  c  Royal Netherlands Meteorological Institute, de Bilt, The Netherlands ABSTRACT:  Tropospheric delay affects the propagation of the microwave signals broadcast by the Global PositioningSystem (GPS) satellites. Geodetic processing software enable estimation of this effect on the slanted signal paths connectingthe satellites with the ground-based receivers. These estimates are called slant delay (SD) observations and they arepotentially of benefit in numerical weather prediction.The three-dimensional variational data assimilation system of the High-Resolution Limited-Area Model (HIRLAM 3D-Var) has been modified for data assimilation of the SD observations. This article describes the ground-based GPS observingsystem, the SD observation operator, the estimation of the observation- and background-error standard deviations, themethodology of accounting for the observation-error correlation, and the tuning of the background quality control for thisobservation type.The SD data assimilation scheme is evaluated with experiments utilizing hypothetical observations from a single receiverstation, as well as a single-case experiment utilizing real observations from a regional GPS receiver network. The ability of the data assimilation system to extract the asymmetric information from the SD observations is confirmed. In termsof analysis increment structure and magnitude, SD observations are found to be comparable with other observationtypes currently in use, provided that the observation-error correlation is taken into account. Copyright  ©  2007 RoyalMeteorological Society KEY WORDS  GPS meteorology; numerical weather prediction; observing system  Received 17 October 2006; Revised 1 March 2007; Accepted 12 March 2007  1. Introduction The geodetic applications of the Global Positioning Sys-tem (GPS) are based on measuring the carrier phasesof the microwave signals broadcast by the satellites andreceived by the ground-based receivers (e.g. Hofmann-Wellenhof   et al ., 2001). The measurements are affectedby a number of uncertainty factors. One of these is thetropospheric effect on the signal propagation, the so-called tropospheric delay. In the case of a sufficientlydense network of geodetic GPS receivers, the tropo-spheric delay can be estimated as a by-product of thegeodetic processing. The tropospheric delay estimates canbe considered as meteorological observations containingsignatures of the atmospheric humidity field, and theyform the basis for the concept known as GPS meteorol-ogy (Bevis  et al ., 1992). These scalar-valued observationshave dimension of length. For a near-zenith satellite anda receiver at mean sea level, the tropospheric delay variesbetween two and three metres. *Correspondence to: Reima Eresmaa, P.O. Box 503, FI-00101,Helsinki, Finland. E-mail: Reima.Eresmaa@fmi.fi GPS meteorology provides potential for both subjec-tive forecasting and nowcasting as well as for the numer-ical weather prediction (NWP). The latter application isstrongly coupled with the process of data assimilation.The data assimilation produces the maximum likelihoodestimate of the atmospheric state, which is used as the ini-tial condition for the numerical forecasts. Development of NWP systems towards fine-scale atmospheric modellingsets requirements for the resolution of the observing sys-tems. Ground-based GPS receivers are recognized as apotential source of high spatio-temporal resolution obser-vations of atmospheric humidity. The European meteo-rological and geodetic communities are collaborating toadvocate the use of the ground-based GPS receiver net-works as an observing system for future operational NWPsystems (e.g. Elgered  et al ., 2005; Vedel  et al ., 2006).Such a system is now being set up in the E-GVAPproject (EUMETNET GPS Water Vapour Programme,http://egvap.dmi.dk).It is not yet known precisely how the GPS measure-ments should be processed in order to provide the highestbenefit to NWP models. The currently used geodetic soft-ware support processing of zenith total delay (ZTD),which is an estimate of the tropospheric delay in a vertical Copyright  ©  2007 Royal Meteorological Society  970  H. J¨ARVINEN  ET AL . column above each receiver station (e.g. Elgered  et al .,2005). By using surface meteorological data of pressureand temperature, ZTD can be separated into hydrostaticand wet contributions (zenith hydrostatic delay, ZHD andzenith wet delay, ZWD), and the latter further convertedinto integrated water vapour (IWV) (Bevis  et al ., 1992).These estimates are essentially weighted means of thevertical projections of the corresponding quantities alongthe individual slanted signal paths.It is obvious that part of the meteorological informa-tion is lost in the ZTD processing. A single ZTD estimategives no indication about the azimuthal asymmetry, infor-mation that is contained in the raw GPS measurements.This asymmetry is typically most significant in the vicin-ity of strong horizontal humidity gradients, such as frontalregions, often accompanying severe weather phenomena(e.g. Koch  et al ., 1997). In such cases, the use of theZTD estimates can be considered suboptimal. It wouldbe more beneficial to process the raw GPS measurementsin a manner which conserves the asymmetry. One possi-bility to achieve this is to estimate the tropospheric delayalong the individual signal paths. Efficient meteorolog-ical use of such slant delay (SD) observations requiresfurther development of the data assimilation systems.So far, the majority of the ground-based GPS dataassimilation experiments have been conducted using theZTD, ZWD or IWV estimates (e.g. De Pondeca and Zou,2001; Pacione  et al ., 2001; Vedel and Huang, 2004; Vedeland Sattler, 2006 and references therein). In general itis found that the data assimilation of the ZTD has aweak, but mainly positive, impact on NWP forecastsof humidity and precipitation. Ha  et al . (2003) studieddata assimilation of slant wet delays with hypotheticaldata in a mesoscale NWP system. Hypothetical SDobservations were used also by MacDonald  et al . (2002)for diagnosing a three-dimensional humidity field. Thesestudies show promising results for retrieving the humidityfield on the basis of the SD measurements, even though ithas not yet been demonstrated that the data assimilationof real SD observations actually improves the NWPforecast scores.Methodology for data assimilation of SD observa-tions has been implemented in the High-ResolutionLimited-Area Model (HIRLAM; Und´en  et al ., 2002) aspreparation for impact studies. This article reviews theimplementation. The variational data assimilation in theHIRLAM framework is described in Section 2. Section 3discusses SD as an observation type. The data assimila-tion aspects specific to SD are covered in Section 4. Theexperiments evaluating the SD data assimilation systemare described in Section 5. A summary and the main con-clusions are given in Section 6. 2. The HIRLAM 3D-Var The hydrostatic HIRLAM limited-area NWP systemis developed in a co-operation between the nationalmeteorological services of Denmark, Finland, Iceland,Ireland, the Netherlands, Norway, Spain and Sweden(Und´en  et al ., 2002). There is ongoing co-operationbetween the HIRLAM and ALADIN (Bubnov´a  et al .,1993) consortia aiming at the joint development of anon-hydrostatic mesoscale NWP system.The analysis system of HIRLAM is based on the vari-ational data assimilation technique (Gustafsson  et al .,2001; Lindskog  et al ., 2001), in either three (3D-Var)or four dimensions (4D-Var). While the currently opera-tional implementations use 3D-Var, the 4D-Var set-up isbeing fine-tuned for operational use in the near future.The incremental formulation (Courtier  et al ., 1994) isapplied in the HIRLAM 3D-Var in order to reduce thecomputational load of the analysis problem. The back-ground field  x b  is updated with the analysis increment  δ x corresponding to the minimum of the incremental costfunction J(δ x ) = 12 δ x T B − 1 δ x + 12 (H  [ x b ] + H δ x − y ) T R − 1 (H  [ x b ] + H δ x − y ),( 1 ) where  y  represents the observations,  B  and  R  are theerror covariance matrices for the background and theobservations, respectively, and  H   is the observation oper-ator producing the model state in observation space.The incremental formulation allows one to update thehigh-resolution background with a low-resolution analy-sis increment, i.e. the dimensions of   x b  and  δ x  need notbe the same.The incremental cost function (1) is based on theassumption that  H   is linear for realistic  δ x  in theneighbourhood of   x b . The linearized version of theobservation operator  H  is thus applied for projecting theanalysis increment to observation space. Moreover, theiterative minimization of (1) makes use of the adjoint of the tangent-linear observation operator  H T , as seen fromthe cost function gradient ∇  δ x J   = B − 1 δ x + H T R − 1 (H  [ x b ] + H δ x − y ), ( 2 ) which is calculated at each iteration step.The analysis increment  δ x  of the HIRLAM 3D-Varconsists of the increments of horizontal wind components( u  and  v ), temperature ( T  ), specific humidity ( q ) andlogarithm of surface pressure (ln  p s ). Introduction of anew observation type requires the development of theobservation operator ( H  ,  H ,  H T ) and the estimation of the observation-error statistics for  R . 3. Processing of the SD observations The nominal constellation of the GPS system consistsof 24 satellites in circular orbits with a radius of about26600 km. The satellites broadcast a modulated signal ontwo carrier-wave frequencies in the microwave domain. Copyright  ©  2007 Royal Meteorological Society  Q. J. R. Meteorol. Soc.  133 : 969–980 (2007)DOI: 10.1002/qj  DATA ASSIMILATION OF GPS SLANT DELAYS  971The most accurate positioning applications make use of the phase pseudoranges, i.e. phases of the received car-rier waves. The charged particles at high atmosphericaltitudes affect the wave propagation causing the iono-spheric refraction. This frequency-dependent effect canbe largely eliminated by computational methods in casethe receiver has a capacity of measuring the carrier-wavephases at the two frequencies (e.g. Hofmann-Wellenhof  et al ., 2001).The GPS processing deals with the raw pseudo-range measurements. Within the processing, a number of unknown parameters are either estimated or eliminated.The tropospheric delay is one of the contributing effectsin the measurements. It is generally not possible todirectly estimate the SD along each individual signalpath. Therefore, SD processing is usually done as a two-step procedure (de Haan  et al ., 2002).In the first step, ZTD is estimated as part of the geode-tic network solution simultaneously with the satellite andreceiver positions and the clock errors. Accurate estima-tion of ZTD requires separation between ZHD and ZWD.The separation makes use of surface pressure data at thereceiver locations. Pre-defined hydrostatic and wet map-ping functions ( m h  and  m w ) are used for relating theZHD and ZWD to the corresponding contributions of slant delay at zenith angle  ζ  . Once the zenith delays aresolved, the  symmetric  part of SD,  SD s , is obtained as SD s = m h (ζ) ZHD + m w (ζ)ZWD. ( 3 ) In the second step, the least-squares residuals of thenetwork solution, obtained in the first step, are interpretedas the asymmetric part of SD,  SD a . The final SD is then SD  = SD s + SD a . ( 4 ) The undesirable slowly varying effects of (e.g.) signalmultipath propagation and antenna phase-centre varia-tions can be reduced by utilizing multipath maps that aregenerated by averaging the data over several days priorto the measurements. Due to the observing geometry, inhigh latitudes there are relatively more data available atlarge zenith angles.The SD processing method is still subject to criticaldiscussions among GPS scientists. The simulation studyperformed by Elosegui and Davis (2004) shows that alarge part of the asymmetric information, contained in theraw measurements, is lost during the processing of SD.This is a consequence of the least-squares fitting method,which aims at finding the parameter values that providea best fit to all observational data. Obviously, this is asevere limitation for the information content of the SDobservations and possibly also for their value to the NWPmodels.As the information content of the SD observationsis somewhat questionable, it is critical to assess theasymmetry properties of the observations before drawingconclusions from extensive assimilation experiments.Such an assessment has been conducted recently byEresmaa  et al . (2007) from the observation modellingperspective. The conclusions suggest that the asymmetriccontribution to the SD observations is of the order of a few parts per thousand of the total SD, and it ismost significant at large zenith angles. Moreover, theasymmetry contained in the SD observations is found tobe related to real atmospheric asymmetry in the vicinityof the GPS receiver stations.Another point of concern is the observation-error cor-relation, which is expected to be significant. The simul-taneous processing of all raw delay measurements makesthe ZTD observations in the receiver network mutuallydependent, implying correlated observation errors. Fur-thermore, several SD observations are generated from asingle ZTD observation. Consequently, the ZTD obser-vation errors propagate through the mapping functionsto the SD observations making the SD observation errorscorrelated at one time instant at one receiver station. Thisimplies that the corresponding SD observations are mutu-ally dependent. 4. SD data assimilation characteristics This section reviews the specific actions taken in orderto enable the SD data assimilation with the HIRLAM3D-Var system. Since the SD observation operator hasalready been discussed in detail by Eresmaa and J¨arvinen(2006), it is described only briefly here.4.1. Observation operatorThe observation operator is built upon the definition SD  =   s (n − 1 ) d s  = 10 − 6   s N   d s, ( 5 ) where  n  is the real part of the atmospheric refractiveindex,  N   ≡ 10 6 (n − 1 )  is refractivity,  s  is the signal pathfrom the satellite to the receiver and d s  is a signal pathelement in length units. This definition for troposphericdelay is commonly used in geodetic literature (e.g.Hofmann-Wellenhof   et al ., 2001).The observation operator consists essentially of threealgorithms. First, the slanted signal path,  s , through theNWP model grid is determined. The signal path is definedas a set of coordinates (latitude  ϕ , longitude  λ  and heightabove the mean sea level  h ) for the intersections withthe NWP model levels. The algorithm for the signal pathdetermination makes use of the concept of geometricalpath (GP) given by the satellite azimuth and zenith anglescorresponding to the satellite–receiver geometry. Therefractive bending of the signal path is taken into accountby an explicit correction on top of the GP approximation.(Initial tests with an alternative approach for the signalpath determination by ray tracing (Rodgers, 2000; Healy,2001) seem to produce similar results when applied toHIRLAM.)Second, the NWP model state is projected from themodel grid to the signal path intersections using bi-linear Copyright  ©  2007 Royal Meteorological Society  Q. J. R. Meteorol. Soc.  133 : 969–980 (2007)DOI: 10.1002/qj  972  H. J¨ARVINEN  ET AL . interpolation in the horizontal. Once the model variablesare interpolated, refractivity  N   at the intersections isobtained through N   = k 1 pT  + (k 2 − k 1 )qp( 0 . 622 + 0 . 378 q)T  + k 3 qp( 0 . 622 + 0 . 378 q)T  2 , ( 6 ) where  p  is pressure,  T   is temperature,  q  is spe-cific humidity and the empirical coefficients are  k 1  = 77 . 60 K hPa − 1 ,  k 2  = 70 . 4 K hPa − 1 and  k 3  = 3 . 739 × 10 − 5 K 2 hPa − 1 (Bevis  et al ., 1994). Relation (6) isderived from the ideal gas assumption.Third, the refractivity is integrated using the modelstate variables to yield SD. The refractivity integration isdone layer by layer in the vertical. For each layer of depth h , the parameters  a  and  b  of the exponential function N   = exp (a + bh) ( 7 ) are uniquely determined from values of   N   at the bottomand the top of the layer. Use of these parameters allowspiecewise analytical integration.The observation operator has been validated againstobservational SD data. The model counterparts are pro-duced for an SD dataset consisting of nearly 400000observations with satellite zenith angle cutoff of 80 ° . Theobservations are processed for 17 permanent receiver sta-tions in Northwest Europe and they cover the time period1–24 May 2003. The dataset is processed at the Techni-cal University of Delft, the Netherlands. The root-mean-square difference between the observed and the modelledSD is about 2 cm near the zenith and about 10 cm at thezenith angle of 80 ° . Also the tangent-linear observationoperator and its adjoint are implemented and evaluatedwith appropriate tests.One practical issue is recognized as a potential problemspecific for the SD observations. Due to the high demandson the computational efficiency, the current NWP sys-tems are usually implemented on parallel computers, suchthat each processor handles only one geographical area,i.e. subdomain, at a time. It can occasionally happen thatthe SD signal path at a large zenith angle intersects twoor more subdomains. For such observations, the forwardmodel and the data assimilation code will be significantlymore complicated. In the current HIRLAM implementa-tion, those observations are simply rejected in the dataassimilation. The drawback from this simple solution isthat the details of parallelization affect the selection of observations and further the resulting analysis. The par-allelization of the SD data assimilation in HIRLAM isstill an open technical question.4.2. Estimation of the error standard deviationsThe data assimilation algorithms are derived from theassumption that the observation- and the background-error distributions are Gaussian with expectation valuezero. The observation- and background-error standarddeviations ( σ  o  and  σ  b ) need to be explicitly specified.In case the assumption of unbiased observations cannotbe justified, a bias correction algorithm also needs tobe implemented. The HIRLAM data assimilation systemdoes not yet include a bias correction for SD.The statistical properties of the observation minusmodel background (OmB) departures provide the basisfor estimation of   σ  o  and  σ  b . The standard deviation of theOmB departure contains contributions of both the obser-vation and the background errors. The contribution of thebackground error is determinated by the randomizationmethod (Andersson  et al ., 2000). This method makes useof a sample of Gaussian random vectors (here 500), eachrepresenting the background error in the model grid-pointspace. The randomization method simulates the propaga-tion of the background errors from the model grid-pointspace to observation space by applying the tangent-linearobservation operator to each of the Gaussian random vec-tors. The estimation method assumes horizontal homo-geneity and independence of azimuth angle.The standard deviation of the OmB departure shows astrong dependence on the satellite zenith angle (Eresmaaand J¨arvinen, 2006). Therefore,  σ  o  and  σ  b  are estimatedseparately for nominal zenith angles 0 ° , 15 ° , 30 ° , 40 ° ,50 ° , 60 ° , 65 ° , 70 ° and 75 ° . The estimation procedure is asfollows:(1) The variance of the OmB departure is calculated ateach nominal zenith angle. Only the departures forwhich the zenith angle is within 0.5 ° of the nominalzenith angle are used.(2) The randomization method is applied in order toobtain the background-error variance at each nominalzenith angle.(3) At each nominal zenith angle, the background-errorvariance is subtracted from the OmB variance inorder to obtain the observation-error variance. Theobservation and background errors are assumed tobe uncorrelated.(4)  σ  o  and  σ  b  at the nominal zenith angles are calculatedas the square roots of the corresponding variances.Finally, functional models for  σ  o  and  σ  b  are constructedassuming the form f(ζ) = c cos ζ   + d ( 8 ) as a function of zenith angle  ζ  . The parameters  c  and  d  are determined separately for  σ  o  and  σ  b  using the least-squares fitting method. The input data for the fitting isobtained in the form of   σ  o  and  σ  b  at the nominal zenithangles.The results of the fitting are  c = 11 . 27 mm and  d   =− 5 . 669 × 10 − 2 for  σ  o , and  c = 7 . 550 mm and  d   = 2 . 654 × 10 − 3 for  σ  b . Figure 1 shows the fitted curves as afunction of zenith angle. For all zenith angles,  σ  o  exceeds σ  b . Moreover, the ratio  σ  o /σ  b  increases with increasing Copyright  ©  2007 Royal Meteorological Society  Q. J. R. Meteorol. Soc.  133 : 969–980 (2007)DOI: 10.1002/qj  DATA ASSIMILATION OF GPS SLANT DELAYS  973 0 20 40 60 80ZENITH ANGLE (DEG)0246810    S   T   A   N   D   A   R   D   D   E   V   I   A   T   I   O   N   (  c  m   ) OmBob errorbg error Figure 1. Standard deviations of SD OmB departure (solid line), SDbackground error (dashed line) and SD observation error (dash-dottedline) as functions of zenith angle. zenith angle. This is interpreted as follows. With increas-ing zenith angle, the SD observation becomes increas-ingly sensitive to various noise terms, such as the signalmultipath. Such sensitivity is absent from the SD modelcounterparts. Therefore, the increase with zenith angle isfaster for  σ  o  than for  σ  b . Consequently, the SD observa-tions receive relatively smaller weight in the data assim-ilation than the model background, especially at largezenith angles.It should be noted that the randomization method onlyprojects the background-error statistics from grid-pointspace to observation space. The obtained  σ  b  values aretherefore realistic only as far as the background-errorcovariance matrix is realistic. An alternative procedure tocompute the error standard deviations based on the NMCmethod by Parrish and Derber (1992) has been appliedby de Vries (2006) with generally supportive results.4.3. Observation-error correlationAs pointed out in section 3, the observation-error cor-relations of SD are likely to be significant. Observationerrors are assumed to be uncorrelated for most observa-tion types currently assimilated in the HIRLAM system,the only exception being the Advanced TIROS Opera-tional Vertical Sounder radiances (Schyberg  et al ., 2003).The observation-error covariance in the HIRLAM contextis estimated for the ZTD (Eresmaa and J¨arvinen, 2005). Acovariance model for the SD based on elevation angle andsite separation has been derived by de Vries (2006). TheSD data assimilation system allows specification of thelocal error correlation properties between the SD obser-vations at the same receiver station.The dimension of the correlated blocks in the  R  matrixis of the order of 20 to 200, depending on the assimilationtime-window length and the observation thinning details.The observation-error correlation is accounted for byfollowing the procedure introduced in J¨arvinen  et al .(1999).Let us define the departure  d  between the analysis andthe observations as d = H  [ x b ] + H δ x − y , and the effective departure   d  as  d = R − 1 d . The observation part of the incremental 3D-Var costfunction (1) then becomes J  o  = 12 d T R − 1 d = 12 d T  d , ( 9 ) and the observation part of the cost function gradient (2)becomes ∇  δ x J  o  = H T R − 1 d = H T  d . ( 10 ) Once the observation-error covariance matrix has beenspecified, the effective departure is obtained by solvingthe linear system R  d = d , ( 11 ) for  d . In the HIRLAM implementation of SD data assim-ilation, (11) is solved by making use of the Choleskydecomposition for  R , which is positive-definite. This pro-cedure allows calculation of the cost function and itsgradient in a general case of correlated observation errors.The sparseness of   R , due to (e.g.) the lack of temporalcorrelations in the 3D-Var context, makes the decompo-sition tractable even in the case of many hundreds if notthousands of observations.4.4. Observation quality controlThe background quality control (BgQC) aims at identifi-cation and rejection of observations contaminated withnon-Gaussian gross errors. The BgQC will reject theobservation  y i  if it does not satisfy the inequality (H  i [ x b ] − y i ) 2 σ  2o ,i + σ  2b ,i ≤ L, ( 12 ) where  H  i [ x b ] is the non-linear model counterpart to  y i , σ  o ,i  and  σ  b ,i  are the error standard deviations of the obser-vation and the model counterpart, respectively, and  L  isthe rejection limit (J¨arvinen and Und´en, 1997; Lindskog et al ., 2001). The BgQC is tuned in HIRLAM for the SDobservations by specifying  L  for this observation type.Andersson and J¨arvinen (1999) showed how the rejec-tion limit  L  can be tuned using a large number of OmBdepartures. First, a normalized departure  z i  is calculatedfor all observations  y i  as z i  = y i  − H  i [ x b ]   σ  2o ,i + σ  2b ,i . ( 13 ) Copyright  ©  2007 Royal Meteorological Society  Q. J. R. Meteorol. Soc.  133 : 969–980 (2007)DOI: 10.1002/qj
Similar documents
View more...
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks