Assimilation of COST 716 Near-Real Time GPS data in the nonhydrostatic limited area model used at...

Meteorol Atmos Phys 91, 149–164 (2006)DOI 10.1007/s00703-005-0110-6

1 Institute of Applied Physics, University of Bern, Switzerland2 Federal Office of Meteorology and Climatology, Zurich, Switzerland3 Swiss Federal Office of Topography, Wabern, Switzerland4 Institute of Applied Physics, University of Bern, Switzerland

Assimilation of COST 716 Near-Real Time GPS datain the nonhydrostatic limited area model used at MeteoSwiss

G. Guerova1, J.-M. Bettems2, E. Brockmann3, and Ch. Matzler4

With 9 Figures

Received August 13, 2003; revised August 31, 2003; accepted December 21, 2004Published online: June 30, 2005 # Springer-Verlag 2005

Summary

Application of the GPS derived water vapor intoNumerical Weather Prediction (NWP) models is one ofthe focuses of the COST Action 716 ‘‘Exploitation ofGround based GPS for climate and numerical weatherprediction applications’’. For this purpose the GPS datacovering Europe have been collected within the Near-RealTime (NRT) demonstration project and provided forObserving System Experiments (OSE). For the experimentspresented in this manuscript the operational NWP systemof MeteoSwiss is used. The limited area nonhydrostaticaLpine Model (aLMo) of MeteoSwiss covers most ofwestern Europe, has a horizontal resolution of 7 km, 45layers in the vertical, and uses a data assimilation schemebased on the Newtonian relaxation (nudging) method.In total 17 days analyses and two 30 hours daily forecastshave been computed, with 100 GPS sites assimilatedfor three selected periods in autumn 2001, winter andsummer 2002. It is to be noted that only in the last perioddata from 10 french sites, i.e. west of Switzerland areassimilated.

The GPS NRT data quality has been compared with thePost-Processed data. Agreement within 3 mm level ZenithTotal Delay bias and 8 mm standard deviation was found,corresponding to an Integrated Water Vapor (IWV) biasbelow 0.5 kg=m2. Most of the NRT data over aLMo domainare available within a prescribed time window of 1 h 45 min.In the nudging process the NRT data are successfully used bythe model to correct the IWV deficiencies present in thereference analysis; stronger forcing with a shorter time scale

could be however recommended. Comparing the GPSderived IWV with radiosonde observations, a dry radiosondebias has been found over northern Italy. Through GPS dataassimilation the aLMo analysis bias and standard deviationin the diurnal cycle has been reduced. The negative biasof �0.64 kg=m2 in the reference analysis has been reduc-ed to 0.34 kg=m2 in GPS analysis. However, the diurnalcycle statistic from the forecast does show the characteristicnegative bias only slightly reduced starting with the GPSanalysis.

The GPS IWV impact on aLMo is large in June 2002 andmoderate in September 2001 OSE. January OSE is incon-clusive due to inconsistent use of humidity data below thefreezing point. In June 2002 OSE, a substantial IWV impactis seen up to the end of the forecast. Over Switzerland thedry bias in the reference analysis has been successfully cor-rected and the 2 m temperature and dew point have beenslightly improved over the whole aLMo domain. The sub-jective verification of precipitation against radar data inautumn 2001 and summer 2002 gives mixed results. In theforecast the impact is limited to the first six hours and tostrong precipitation events. A missing precipitation patternhas been recovered via GPS assimilation in June 20 2002forecast. A negative impact on precipitation analysis on June23 has been observed.

The future operational use of GPS will depend on dataavailability; European GPS networks belong mainly to thegeodetic community. A further increase of GPS networkdensity in southern Europe is welcome. The GPS derivedgradient and Slant Path estimates could possibly improveefficiency of IWV assimilation via the nudging technique.

1. Introduction

The last decade brought a new generation ofnonhydrostatic meso-� scale models in opera-tional Numerical Weather Prediction (NWP).They run with a horizontal resolution below10 km (mesh size) and are aimed to betterresolve storm-scale phenomena, complex topo-graphy and to provide reliable short range fore-cast up to 2 or 3 days.

For the needs of high resolution NWP modelsa further extension of the existing MeteorologicalObserving System will be necessary. In particu-lar, Bettems (2002) reported that additional infor-mation about local structures in the humidityfield will be needed. Some promising candidatesare the weather radar and lidar networks. In thelast decade it was expected that satellite-borneinstruments like Special Sensor Microwave=Imager (SSM=I) and Advanced MicrowaveSounding Unit (AMSU) would deliver watervapor information with the high temporal andspatial resolution needed for the NWP models.However, the satellite retrievals are often limitedto the ocean regions and due to the frequencyemployed the satellite instruments are also dis-turbed by rain or rain and ice. MODIS, the Mod-erate Resolution Imaging Spectroradiometer, hasa spatial resolution of 1 km but a poor temporalresolution, i.e. daily averaged water vapor pro-ducts. Another source of water vapor informa-tion are the ground-based networks of GlobalPositioning System (GPS) receivers, currentlyoperated for geophysical purposes. With im-provements of the GPS accuracy, i.e., improvedmapping functions and antenna phase centremodels (Haase et al, 2002), it is possible toobtain reliable information of the atmosphericsignal delay in the zenith direction (ZTD). In1992, Bevis (Bevis et al, 1992) proposed to usethe ground-based GPS for retrieving the watervapor content of the atmosphere. In the follow-ing years several publications investigated theaccuracy of GPS in comparison with the conven-tional sources of water vapor information likeradiosondes and water vapor radiometers andin comparison with unconventional ones likesunphotometer, sunspectrometer and VLBI (VeryLong Baseline Interferometer). The findings, assummarized in Haase et al (2002), confirm theaccuracy of GPS derived Integrated Water Vapor(IWV) in the range of 1 kg=m2 or better, i.e., the

same level of accuracy as the conventionalobservations.

The validation of mesoscale NWP modelsusing GPS is presented in Cucurull et al (2000),Kopken (2001), Haase et al (2002), Tomassiniet al (2002), and Guerova et al (2003). Tomassiniet al (2002) studied the diurnal IWV cycle fromGPS and Local Model over Germany in summer2000. They found a systematic IWV underesti-mation, higher than 1 kg=m2, in the model anal-ysis for the hours between 06 and 18 UTC. TheHIRLAM model has been validated for the wes-tern Mediterranean area during MAGIC project(Haase et al, 2002). An IWV bias of 0.5 kg=m2

and standard deviation of 3 kg=m2 is reportedas well as a latitude dependence of the standarddeviation. The validation studies agree that adense GPS network provides a valuable addi-tional information for NWP models.

Assimilation experiments with GPS IWV arereported in several publications: Kuo et al(1996), Guo et al (2000), Smith et al (2000), DePondeca and Zou (2001), Vedel et al (2002),Falvey and Beavan (2002), Tomassini et al (2002),Gutman et al (2004), and Nakamura et al (2004).The GPS impact is mainly evaluated by analyzingthe model precipitation skills. The results pointtowards an overall neutral impact when assimilat-ing GPS. However, during active weather phasesthe GPS is reported to have a positive influence onthe location of the front boundaries, and a continu-ous improvement of precipitation forecast skillshas been obtained in a five year assimilation pe-riod in the USA (Gutman et al, 2004). Falvey andBeavan (2002) and Tomassini et al (2002) use anudging scheme for assimilation of the GPS data,which makes their results of particular interestfor the study presented here. Falvey and Beavan(2002) report that continuous GPS assimilationimproved the upwind total rainfall to 1% signifi-cance level only. They also discuss one casewhere the humidity profile adjustment deterio-rated the precipitation structure; the reason wasthe proportional moisture removal from all modellevels based on the GPS observation. In Tomassiniet al (2002) a mixed impact on precipitation anal-ysis is reported for a severe weather case.

In Europe the work on application of ground-based GPS in operational meteorology wasstarted with the MAGIC project (Haase et al,2002). The project objectives were to collect, vali-date and assimilate (HIRLAM model) the GPS

150 G. Guerova et al

data from the western Mediterranean. In 1999, theCOST Action 716 ‘‘Exploitation of ground basedGPS for climate and numerical weather predic-tion application’’ followed (Elgered, 2001). Incontrast to the MAGIC the COST 716 GPS net-work covers fully western Europe and six of theEuropean Meteorological Offices, operating theHIRLAM, LM and UK mesoscale models, wereinvolved in the data validation and assimilationwork. The Swiss contribution to COST 716 wasestablished as a collaboration between threeInstitutes namely, the Swiss Federal Office ofTopography (Swisstopo), the Federal Office ofMeteorology and Climatology (MeteoSwiss) andthe Institute of Applied Physics at the Universityof Bern (Brockmann et al, 2002). The first goal inevaluating the potential of GPS for meteorologi-cal purposes, was the verification of the opera-tional NWP models against GPS data fromthe Automated GPS Network of Switzerland(AGNES), reported in Guerova et al (2003;2005). The second goal, evaluating the impactof GPS data in the NWP system of MeteoSwiss,was initiated in 2001. A first GPS assimilationexperiment was calculated for two weeks of Sep-tember 2001. The precipitation verification overSwitzerland shows improvements of the scores aswell as an improved diurnal precipitation cycle. Inaddition we found out that the typical horizontalscale for spreading of the observation incrementswas too large, particularly in presence of strongwater vapor gradients; to avoid this problem thecorresponding model parameter was modified.The results obtained in this first experiment(Guerova et al, 2004) have been consideredencouraging, and three new Observing SystemExperiments (OSE) have been calculated in2002, which prime objective was to investigatethe model analysis and forecast sensitivity toGPS IWV in different seasons and a possibledegradation of low stratus in winter months. Inthis respect this work differs to the work ofTomassini et al (2002), which concerns only sin-gle case studies and Guerova et al (2004) whereno sensitivity in forecast has been reported.

This manuscript reports the results of theSeptember 2001 and June 2002 OSEs. In Sect. 2,the mesoscale model of MeteoSwiss is described.The COST 716 Near-Real Time (NRT) demonstra-tion project is outlined in Sect. 3. Section 4 reportsthe OSE set-up. The results are discussed in Sect. 5and summarized in Sect. 6.

2. The operational NWP systemat MeteoSwiss

2.1 The aLpine Model – an overview

Since April 2001, a nonhydrostatic mesoscalemodel named the aLpine Model (aLMo) is usedfor operational NWP at MeteoSwiss. The aLMois the Swiss configuration of the COSMO(COnsortium for Small-Scale MOdelling) limitedarea model (Doms et al, 2001) developed by theNational Weather Services of Switzerland, Italy,Poland and Greece under the lead of the NationalWeather Service of Germany (DWD). The Swissimplementation of the model has a horizontalresolution of about 7 km, and the domain coversmost of western Europe (Fig. 1). A terrain follow-ing vertical coordinate system is used, with 45vertical layers and about 100 m vertical resolutionin the lowest 2 km of the atmosphere (Bettems,2002), and a top level at 20 hPa. A filtered oro-graphy was introduced to produce more realisticprecipitation fields. The aLMo is nested in theECMWF global spectral model IFS. The initialconditions are obtained by the aLMo continuousassimilation cycle and only boundary data areprovided by the IFS model. Lateral boundaryformulation is by the Davies (1976) relaxationtechnique. The aLMo prognostic variables are:horizontal and vertical Cartesian wind compo-nents, temperature, perturbation pressure, specifichumidity and cloud water content.

The aLMo parameterization schemes take a va-riety of physical processes like grid-scale clouds

Fig. 1. aLMo domain and the location of the EuropeanGPS sites (red dots) used in our experiments. The colorscale presents the orography and the thin red line the coun-try boundaries. The black dots present the location ofPayerne (Switzerland) and Torino (Italy)

Assimilation of COST 716 Near-Real Time GPS data 151

and precipitation, subgrid-scale clouds, moist con-vection, radiation, vertical diffusion of sensibleheat and humidity, boundary layer and soil pro-cesses into account. The subgrid-scale cloudinessis a combination of cloudiness due to convectiveprocesses and cloudiness described as an empiri-cal function depending on relative humidity andheight. The grid scale clouds are described viacloud water saturation adjustment. For precipita-tion formation a Kessler type bulk parameteriza-tion scheme is used, and four categories of watersubstances are considered, namely: water vapor,cloud water, rain and snow. The aLMo hydrologi-cal cycle and microphysical processes are pre-sented in Fig. 2. As seen from Fig. 2, the cloudice phase is neglected, assuming a fast transfor-mation from cloud water to snow. A cloud icescheme has been developed (Doms, 2002) butwas not yet tested at the time of the experiments.

2.2 The data assimilation scheme

Data assimilation with aLMo is based on the nud-ging or Newtonian relaxation method (Schraff,1997), where the atmospheric fields are forcedtowards direct observations at the observationtime. The nudging is a continuous data assimila-tion scheme, similar in this respect to the 4D-Varmethod. A relaxation term is introduced in themodel equations, small enough to not signifi-cantly disturb the dynamic balance of the model.Balance terms are also included: (1) hydrostatic

temperature increments balancing near-surfacepressure analysis increments, (2) geostrophicwind increments balancing near-surface pressureanalysis increments, upper-air pressure incre-ments balancing total analysis increments hy-drostatically. A simple quality control usingobservation increments thresholds is in action.Besides GPS IWV, only conventional observa-tions are assimilated (Doms et al, 2004): synop,ship, buoy (surface pressure, 2 m humidity, 10 mwind for stations below 100 m above msl), temp,pilot (wind, temperature and humidity profiles)and airep, amdar (wind, temperature).

The nudging equation for the specific humid-ity – q and for a single observation has thefollowing form:

@

@tqðtÞ ¼ Qðq;u;v; . . . ; tÞþGq �Wq � ½qobs � qmod�;

ð1Þ

where Q denotes the model physics anddynamics, qobs and qmod are the observed andthe model specific humidity, Gq ¼ 6 � 10�4 s�1

is the coefficient defining the relaxation scaleand Wq consists of spatial and temporal weightsand of a quality factor. The second term in Eq. (1)is the nudging term and the part ½qobs � qmod� iscalled the observation increment. For the hori-zontal spreading of the observation incrementsan autoregressive weight function is used:

WqðxÞ ¼ ð1 þ x=sÞe�x=s; ð2Þ

where x is the distance between the model gridpoint and the observation and s is a correlationscale factor. The temporal weight function WqðtÞequals one at the observation time and decreaseslinearly to zero at 1.5 hours before and 0.5 hoursafter the observation time, i.e., a 2 hour asym-metric saw tooth shaped time window is used.The observation density is also taken intoaccount to set the value of Wq. The observationincrements are computed once every 6 timesteps, i.e., once every 240 s.

In the nudging scheme a direct assimilation ofGPS derived IWV is not possible as no prognos-tic variable of this type is available in the modelequations. Thus, an indirect assimilation proce-dure based on Kuo et al (1993) has been devel-oped at DWD (Tomassini et al, 2002). Thisprocedure is briefly summarized here. First, the

Fig. 2. Hydrological cycle and microphysical processes im-plemented in aLMo. Four types of hygrometeors are consid-ered namely: water vapor, cloud water, rain and snow


Zenith Total Delay (ZTD) from GPS is convertedinto IWV following Bevis et al (1992) using themodel temperature and surface pressure. Second,IWV ratio between the observation and themodel is calculated. Third, using this ratio themodel specific humidity profile is scaled fromsurface to the 500 hPa level. Specific humidityis set to saturation at any level where the specifichumidity exceeds its saturation value due to thescaling. The humidity increments are spread lat-erally using a correlation scale factor of 35 km (sin Eq. (2)). Only GPS stations with a differencebetween station height and model orography larg-er than �100 m have been assimilated. For GPSstations above the model orography, the modellevel closest to the GPS height has been usedas initial level in the profile scaling. This condi-tion about the height of the station is particularlyimportant for the Alpine areas in Switzerland,France, Austria and Italy. Due to this conditionabout 20 GPS stations have not been assimilated,eight of them from the Swiss GPS network.

3. GPS Near-Real Time demonstrationproject

The Near-Real Time (NRT) demonstrationexperiment started in May 2001 as a main activ-ity in Working Group 2 of the COST Action 716(COST-716, 2002). The objective of this experi-ment was to develop a dense NRT GPS networkcovering Europe and providing ZTD estimatessuited for operational NWP, i.e. data deliverywithin 1 h 45 min after the observation time.Seven regional processing centers are contribut-ing to the NRT project, delivering hourly ZTDfiles from about 250 stations in a predefinedCOST format (COST-716, 2001a) to the ftp ser-ver of the UK Met Office. An extensive overviewabout the NRT project is available in Van derMarel et al (2003).

The assimilation experiments reported in thismanuscript used data from about 100 GPS sta-tions, processed by three processing centers; LPT(Swisstopo,Wabern,Switzerland),GOP(GeodeticObservatory Pecny, Czech Republic) and GFZ(GeoForschungsZentrum, Potsdam, Germany).The selection of processing centers is based onoptimal domain coverage, data quality and avail-ability. For the period May 2001–December2002 the three centers delivered more then 90%

of the data within the selected time window of1 h 45 min (Van der Marel et al, 2003). The dataprovided by LPT and GOP have been processedusing the Bernese Normal Equation StackingSoftware with a time window of 7 and 12 hours,respectively, 30 s sampling rate and elevationcut-off angle of 10�. The GFZ data have beenprocessed with Precise Point PositioningSoftware – (EPOS) with data sampling of 120 sand elevation cut-off 7�. Hourly observations arereported by LPT and GOP, at Hþ 300, and twoobservations per hour are reported by GFZ, atHþ 150 and Hþ 450. As seen in Fig. 1, the over-all GPS coverage over Switzerland and north ofSwitzerland is pretty good, but is poor in thesouthern part of the aLMo domain.

4. The observing system experimentdesign

All assimilation experiments reported here havebeen performed by nesting aLMo in the ECMWFglobal model, the boundary conditions beingobtained from the main ECMWF 4D – VARassimilation cycle with a 6-hour update fre-quency. All standard meteorological observationsare assimilated (SYNOP, BUOY, TEMP, aircraftwind and temperature), but no satellite data areused. Two daily 30 hour forecasts starting at 00and 12 UTC have been calculated without (refer-ence experiment) and with GPS data assimilation(GPS experiment). Initial conditions for theseforecasts were obtained from the correspondingaLMo assimilation cycle. In the forecast runs theobservations are still assimilated during the firsttwo hours. One should note, that the French GPSstations have only been available for the June2002 OSE.

Three weather regimes have been selectedfor these OSE, namely: an advective period inSeptember 2001, a winter low stratus case inJanuary 2002 and a summer convection periodin June 2002. During the five day period from9 to 13 September 2001 the weather was drivenby a cyclone located over the Baltic Sea and acold front moving slowly eastward, passing overSwitzerland on September 9 and initiating cyclo-genesis in the Gulf of Genoa in the night ofSeptember 10. The five days period between 10and 14 of January 2002 is characterized by lowstratus over the Swiss Plateau and Southern


Bavaria, induced by an anticyclone with weakpressure gradients located over Hungary – a typi-cal winter situation. The third experiment, fromthe 18 to the 24 of June 2002, was an activeperiod with intense precipitation events and frontpassages (Table 3).

5. Experimental results

A comprehensive verification of the two Obser-ving System Experiments has been performed.The vertical structure and the near – surfaceparameters have been verified for the entiremodel domain. The precipitation observationswere collected for the western and central partof the domain. The diurnal IWV cycle andZTD are studied over Switzerland.

5.1 Diurnal IWV cycle and ZTDverification for Switzerland

The diurnal cycle of water vapor is an importantpart of the model verification. The diurnal IWVcycle for nine Swiss GPS stations at altitude below800 m is plotted in Fig. 3 for the analysis and theforecast of June 2002 OSE. The model bias andstandard deviation are computed to the NRT GPSdata. In Fig. 3a, one could see that the referenceanalyses show a bias of �0.64 kg=m2 which indi-cates water vapor underestimation. Note that the

day-time IWV underestimation is of the samemagnitude as reported in Tomassini et al (2002).A significant negative daily analysis bias isreported on June 21, 22 and 24 reaching �2.0,�3.8 and �2.5 kg=m2, respectively, and a positiveone on June 18 and 19 reaching 2.6 and 1.5 kg=m2.The GPS analyses are successfully correcting theday-time IWV underestimation present in thereference analysis. This resulted in a positive biasreduced to 0.34 kg=m2 as well as reduced stand-ard deviation from 0.87 kg=m2 in reference to0.48 kg=m2 in the GPS experiment. The visualinvestigation of the GPS analysis and the NRTobservation show satisfactory agreement, exceptfor the early morning hours; however, one shouldnote that NRT observations were missing in theearly morning of June 24. In Fig. 3b, the diurnalcycle from the 00 UTC forecast is presented. Boththe reference and GPS forecast underestimatethe IWV with the bias being negative equal to�0.79 and �0.54 kg=m2 correspondingly. Thusthe reported IWVunderestimation in the referenceanalysis is even strengthened in the reference fore-cast and could not be corrected by starting withGPS analysis. This suggests about model weak-ness to correctly predict the water vapor contentvery possibly related to overestimation of lightprecipitation, i.e. a known model deficiency.

In order to investigate the efficiency of the as-similation scheme and the accuracy of the NRT

Fig. 3. June OSE diurnal IWV cycle for nineSwiss GPS sites below 800 msl; (a) The refer-ence analysis – black line, the GPS analysis –red line, Post – Processed GPS green line andNRT GPS – blue line are shown; (b) same as(a) but for the 00 UTC forecast


GPS data, the IWV and ZTD for the GPS sitesPayerne (PAYE) are plotted in Fig. 4. In the panel,one sees that the IWV differences between the ref-erence and GPS experiment are up to 7–8 kg=m2

and that most of the large errors are corrected.However, it is also seen that the assimilation ofGPS does not always drive the model to theobserved value: for example at 07 UTC on June18 the difference between model and observationremains in the range of 3 kg=m2 and the mini-mum of IWV is reached with a three hours delay.

A possible explanation is the scale of thenudging term employed in aLMo, which is notoptimal for the fast temporal variations of thewater vapor; the nudging coefficient (G) usedhere corresponds to a time scale of about 30 min.

In Fig. 4a and b, one notes the strong discre-pancy between the NRT and PP GPS at midday ofJune 20. A sharp peak of 35 kg=m2 is seen in thePP data, which is not well pronounced in the NRTone. This peak has been verified against indepen-dent measurements from a sunphotometer, whichsupports the rapid jump and drop seen in the PPdata; this peak is in fact associated with the pas-sage of a warm front (Fig. 8d). Thus it is to beconcluded that NRT processing tends to smoothout the rapid variations of ZTD. This can beexplained with the Near-Real Time processing

strategy, which incorporates past observationsin a 7 hour window to compute the actual ZTDvalue (processing time window described inSect. 3). Note however, that NRT data smoothingdepends on processing strategy applied. LPTdata, presented in this section, and GOP applythe same smoothing strategy, while GFZ hasone hour time processing window with nosmoothing applied. Except for this single obser-vation, a surprisingly good agreement betweenthe two processing schemes is observed. Table 1

Fig. 4. IWV and ZTD for station Payerne inJune 2002 OSE; (a) IWV amount in kg=m2

from the NRT (blue line) and Post-Processed(green line) GPS data and the aLMo reference(black line) and GPS (red line) analysis; (b)ZTD in m at station Payerne from NRT (bluedots) and the Post-Processed (green dots) data

Table 1. Comparison of Near-Real Time and Post-Pro-cessed solutions. ZTD bias and std for nine Swiss GPS sitesfor the period 18 to 24 June 2002

GPS site ZTD [m]

BIAS NRTminus PP

STD NRTminus PP

EPFL (Lausane) 0.0028 0.0076ETHZ (Zurich) 0.0030 0.0081FHBB (Basel) 0.0004 0.0075GENE (Geneva) 0.0037 0.0072LUZE (Luzern) 0.0026 0.0088NEUC (Neuchatel) 0.0020 0.0087PAYE (Payerne) 0.0018 0.0078STGA (St. Gallen) 0.0025 0.0079UZNA (Uznach) 0.0020 0.0086


displays the bias of the Near-Real Time solutioncompared with the PP solution for nine Swisssites, which is in the range 0.4 to 3.7 mm. Theaveraged ZTD bias is 2.3 mm and the std is inthe 8 mm range. This is in agreement with thework reported in Haase et al (2002) about inter-

comparison of the MAGIC data set. A ZTD biasof 3 mm corresponds to a IWV bias lower than0.5 kg=m2. This result is certainly very encourag-ing for NWP, as the period under considerationis characterized by rapid changes in atmosphericconditions.

Fig. 5. Upper – air verification(aLMo minus TEMP) of relativehumidity bias (left) and std(right) for September, Januaryand June OSEs. The blue dashedline presents the GPS forecastand the red line the referenceforecast at: (a) þ00 h, and (b)þ06 h


5.2 Upper-air and near-surface verification

The vertical structure of the model atmospherehas been verified against 28 radiosonde stations(TEMP) regularly distributed over aLModomain, using the operational package ofMeteoSwiss. Bias and standard deviation (std)for temperature, relative humidity, wind direc-tion, wind speed and the geopotential are calcu-lated for the forecast times: þ00, þ06, þ12,þ18, þ24 and þ30 h. The temperature, windand geopotential (bias and std) do not differbetween the reference and GPS forecast. InFig. 5, the bias and standard deviation (aLMo –TEMP) of the relative humidity are plotted forforecast times at þ00 h and þ06 h, for the 17days period. A small increase of the bias of theorder of 3 to 4% is observed in the GPS experi-ment for the layers above 800 hPa while the std isslightly improved; the assimilation of GPS dataresulted in a global increase of the model humid-ity content. This is specially pronounced overnorthern Italy (Milan and Udine radiosondes)where the positive model humidity bias reaches10–12% in June 2002 OSE. The bias from Milanand Udine radiosondes is not limited only to thefirst þ06 h forecast, but is also present up toþ30 h forecast. This is a sign of a possiblesystematic underestimation in the radiosondehumidity profiles. Studies in Japan (Ohtani andNaito, 2000) and from MAGIC project in thewestern Mediterranean (Haase et al, 2002) reporta dry radiosonde bias in mid day observations.

Near-surface parameters have been verifiedagainst about 1000 surface stations from theSYNOP network; the pressure reduced to meansea level (PMSL), dry bulb temperature (T_2M)and dew point temperature (TD_2M) at 2 m have

been considered. Table 2 summarizes the biasand the std of the reference and the GPS analysesfor the two OSE periods. Overall T_2M andTD_2M biases are moderately improved in Sep-tember and June OSEs; impact of GPS on thestandard deviation shows the same trend, butwith a much smaller magnitude. Note that inthe September OSE the bias of all parametersis smaller in comparison with the June OSE.The reason is an overall good skill of the modelfor this September period, which is further con-firmed by the precipitation verification.

In the September OSE, the improvement of the2 m temperature bias and of the dew point tem-perature bias is 7% and 13%, respectively. Thesurface verification of the June OSE presents apositive impact of the GPS data on the modelperformance. The dew point temperature bias isimproved by about 0.1 K, or about 14%, and the2 m dry bulb temperature is improved by about25%. This is a positive signal in favor of GPSobservations. A minor improvement in the PMSLis also achieved in the GPS analysis in June 2002OSE.

5.3 Precipitation verification

To provide an overview of the GPS impact on themodel humidity field, the IWV differences (GPSminus reference) are plotted in Fig. 6 for theSeptember 10, 2001, June 20 and 23, 2002. TheIWV difference plots have been selected as theyare associated with precipitation episodes dis-cussed later in this section. In general duringthe September 2001 OSE (Fig. 6a) a moderateimpact is observed when assimilating GPS. Theaverage difference at analysis time is in the rangeof � 20% for an average IWV amount of20 kg=m2. However, it is noteworthy that themodification of IWV field is only for limitedregions and completely vanishes in the þ12 hforecast (not shown). Most of IWV modificationsare over northern Italy and the Gulf of Genoa(see Guerova et al, 2004 for a full analysis ofthis period). During June 2002 period (Fig. 6b)the GPS assimilation resulted in significant mod-ifications of the aLMo IWV field over the entiremodel domain, with an average impact of � 30%at analysis time for an average IWV amount of32 kg=m2. Moreover, both the 00 and 12 UTCforecasts exhibit substantial IWV differences, inorder of � 20%, up to the end of the forecast

Table 2. Near-surface verification of September, Januaryand June OSEs with the SYNOP data. Bias and std ofpressure reduced to mean sea level (PMSL), dry bulb tem-perature (T 2M) and dew point temperature (TD 2M) at 2 mare listed for the reference and GPS analyses

Sept.ref

Sept.GPS

Juneref

JuneGPS

PMSL bias [Pa] �4.69 �5.49 34.25 33.92PMSL std [Pa] 70.64 70.31 105.1 105.8TD 2M bias [K] �0.68 �0.60 �0.82 �0.72TD 2M std [K] 2.29 2.26 2.73 2.69T 2M bias [K] 0.15 0.14 0.27 0.21T 2M std [K] 2.29 2.26 2.55 2.52


(þ30 h). Note that the plots presented in Fig. 6are 00 UTC analysis, which means that the radio-sondes humidity profiles have also been availableat this time; in this respect the GPS contributionin June OSE is substantial. This is possibly aconsequence of the GPS network density andhigh temporal availability, which allows betterrepresentation of water vapor structure duringperiods of active weather.

To further investigate the GPS impact on themodel skills, the precipitation data have beenexamined. The radar composite coveringGermany, France, Belgium and Netherlands has

been available through M. Tomassini (DWD).Additionally, the Swiss radar composite and thesurface precipitation from the ANETZ networkare used. ANETZ is the Swiss automatic surfacestations network. To get better evaluation of tim-ing and structure of precipitation patterns, the sixhour instead of 24 hour accumulated precipita-tion is used.

On September 10, 2001, the weather in centralEurope was driven by cold polar air advectionoccurring between the anticyclone located westof the British Isles and an occluding cyclone overthe Baltic Sea (Fig. 7d). In this situation a weak

Fig. 6. IWV difference GPS minus reference experiment at:(a) 00 UTC analysis time on 10 September 2001, (b) 00 UTCanalysis time on 20 June 2002, and (c) 18 UTC analysis timeon 23 June 2002


impact on the precipitation field over Switzerlandis seen in the GPS forecast between 00 and þ6 h.The six hour accumulated precipitation from thereference and GPS forecast can be seen in Fig. 7aand 7b. Comparing the plots two precipitation pat-tern can be distinguished; one west of Switzerlandat the upwind slope of the Jura mountain and asecond much more intense precipitation patternat the upwind side of the Swiss Alps. Betweenthe Jura mountain and the Alps the Swiss Plateauis located. The Swiss Plateau is the precipitationfree region as reported in the surface observa-tions in Fig. 7c. The surface observations fromANETZ are chosen due to lack of radar data. Incomparison aLMo predicts light precipitation up

to 1 mm=6 h in the Swiss Plateau. The intenseprecipitation pattern at the windward slope ofthe Alps is overestimated in the forecast. Theobservations show moderate precipitation inten-sity between 2 to 7.2 mm=6 h over central andeastern Switzerland, while the model predicts10 mm=6 h or more. The observations give aslight advantage for the GPS forecast in termsof reduction of precipitation intensity and bettercorrespondence for precipitation free region. Inthis respect the GPS data seem to have a potentialto drive the model towards more realistic fore-cast, but they are not able to compensate for themodel deficiencies probably responsible for thismassive precipitation overestimation. This case

Fig. 7. Accumulated precipitation on 10 September 2001 00 UTC to 06 UTC over Switzerland from: (a) reference forecast,(b) GPS forecast, and (c) surface observations from ANETZ, (d) surface pressure map for Europe at 12 UTC on September 9


is one of the few cases during the SeptemberOSE where some differences in the precipitationpatterns could be recognized over the entireaLMo domain. In addition the daily precipitationstatistic for Switzerland does not show significantdifferences between the forecasts with and with-out GPS. This weak impact could be related tothe lack of upwind GPS sites in this experiment.French stations are only assimilated in the June2002 experiment.

During the June 2002 OSE period, intense pre-cipitation events (above 20 mm=6 h) are reportedon the 20–21 and 23–24 June. The predominantair flow directions during this period are sum-marized in Table 3 for Switzerland and Europe.The circulation in this period was driven byanticyclone crossing to the north of the Alpsand moving eastward. In this weather regimetwo interesting cases can be singled out: one witha positive impact of GPS data on the model fore-cast and one with negative impact on the analy-sis. The first case is the 00 h to þ06 h forecast on20 June 2002 presented in Fig. 8. Comparing thereference precipitation forecast with the GPStwo distinct intense precipitation pattern (darkblue to yellow color) can be distinguished inFig. 8a, one south of Switzerland and one tothe northwest. In the GPS forecast (Fig. 8b) thereare two intense precipitation regions to the north-west, one over the Jura mountain and one furtherto the west already seen in the reference forecast.In comparison the Swiss radar data (Fig. 8c)shows intense precipitation over the Jura moun-tain. This proves that the precipitation patternseen in the GPS forecast is a real one. From

the difference plot (Fig. 6b), one can see thatthe IWV over Switzerland and northwest ofSwitzerland has been increased up to 8 kg=m2

as a result of assimilation of GPS data (the in-tense orange patch). It is noteworthy that thecloud water content in the Jura region (notshown) was increased between 1–4 g=kg in theGPS run at the altitude of 4 km (650 hPa). Inaddition the vertical velocity at 700 hPa (notshown) presents air ascent over the Jura moun-tain region in the GPS forecast. Thus it can beconcluded that in this case the GPS data provideimportant information at the right time and placeand improve the forecast. It is necessary to drawattention to the intense precipitation pattern tothe south of Switzerland, i.e. on the south slopeof the Alps. There both forecasts tend to over-estimate the intensity and to displace the locationto the east compared to the observations.

The second case, this time with a negativeimpact on the aLMo analysis, has been seen inthe night of 23–24 June in the hours between 18and 06 UTC. The 6 h precipitation sum 18 to 00UTC from the reference and GPS analyses areplotted in Fig. 9a and 9b for the model domaincentered at Switzerland but covering the Alpsand the radar data are in Fig. 9c. Comparingthe model analyses two distinct precipitation pat-terns can be recognized; one to the northwest ofSwitzerland, associated with the cold front pas-sage (Fig. 9d), and a second one to the south. Forboth precipitation structures GPS data assimila-tion resulted in an increase of the intensity of pre-cipitation. There is also a tendency to locate theprecipitation in a smaller region compared to

Table 3. Air flow directions and weather systems in the period 18–24 June 2002 for Switzerland and Europe

Date Switzerland Europe

18 June south flow cyclone over Iceland, anticyclone over Europe19 June south flow cyclone and anticyclone moving to the east20 June south to southwest flow cyclone west of British Isles,

anticyclone over eastern Europe21 June west flow cyclone over British Isles, anticyclone withdraw

to the south of the Alps22 June west flow cyclone occlusion23 June west to southwest flow occluding cyclone over Baltic sea,

anticyclone moves to the north coverssouthern (Spain, Italy) and southeast Europe

24 June west flow cyclogenesis west of Spain


the reference analysis. In particular, the GPSassimilation resulted in a substantial increaseof large-scale precipitation amount south ofSwitzerland with a maximum well above200 mm=6h, tripling the reference value. Theradar data are certainly in a better agreementwith the reference than with the GPS experiment.The explanation of this case could be traced tothe too fast advection of humidity structurestowards Switzerland in aLMo, by the predomi-nant south-westerly air flow, provoking a con-tinuous feeding of the model with water vaporthrough assimilation of the Italian GPS siteTorino. A detailed study of the IWV amount on23 and 24 June 2002 indicates a peak value of

55 kg=m2 in northern Italy and 45 kg=m2 inSwitzerland (Fig. 4a), an extremely high watervapor content rarely observed at these latitudes.The specific humidity profiles from Torino (notshown) indicate moisture substantially increasedat the altitude up to 2 to 2.5 km in the model GPSanalysis. The IWV difference plot (Fig. 6c)shows that in the Torino region the GPS assim-ilation resulted in an increase of total water vaporup to 13 kg=m2, i.e., above 20%.

A summary of the findings from the qualitativeintercomparison between aLMo precipitation andthe radar data gives an impact limited to the earlyforecast hours (up to þ6 h), which is consistentwith the findings reported by Gutman et al

Fig. 8. Six hour accumulated precipitation 00 to 06 UTC on 20 June 2002 from: (a) reference forecast, (b) GPS forecast, and(c) radar observation, (d) surface pressure map at 12 UTC on 19 June 2002


(2004). It is noteworthy that the period analysedhere is rather short thus the results on precipita-tion improvement has to be taken with caution.At present at Meteoswiss one month assimilationof GPS data (June 2002) is performed, which ispossibly bringing a better insight on the IWVimpact on precipitation.

6. Conclusions

With the development of high-resolution meso-scale models rose the need of reliable watervapor information with high temporal and spatialresolution. One such source of information arethe ground-based networks of Global PositioningSystem (GPS) receivers. The focus of the

European COST Action 716 was to evaluate theapplicability of GPS in operational NumericalWeather Prediction (NWP) through data vali-dation and assimilation experiments. The GPSimpact on the operational NWP system ofMeteoSwiss, the aLpine Model (aLMo), isreported in this manuscript.

The GPS data from more than 200 Europeansites have been provided within the Near-RealTime (NRT) demonstration project of COST716. The quality of NRT data have been vali-dated with the Post Processed (PP) data fromSwitzerland. It can be concluded that the twodata sets have similar quality with few cases ofsmoothing observed in the NRT data duringactive weather events. The IWV bias of the

Fig. 9. Six hour accumulated precipitation 18 to 00 UTC on 23 June 2002 from: (a) reference analysis, (b) GPS analysis, and(c) radar observation, (d) surface pressure map at 12 UTC on 23 June 2002


NRT data compared with the PP data is less than0.5 kg=m2. The data from three COST ProcessingCenters, namely LPT, GFZ and GOP, are deliv-ered within 1 h 45 min observation window inmore than 90% of the cases. The three centersprovide data for 95% of the GPS sites availablein aLMo domain.

Through assimilation experiments with aLMoit was found that the nudging scheme is well ableto correct the IWV deficiencies observed in thereference model. A stronger forcing with ashorter time scale could even improve this be-havior in presence of fast meteorological events.The reconstruction of humidity profile is one ofthe weakest elements of the process; combiningGPS with other humidity information or usingSlant Path GPS observations could be envisagedto improve this weakness. Comparing radiosondewith GPS, a dry radiosonde bias has been foundover Northern Italy. This is consistent with thestudies of Ohtani and Naito (2000) and Haaseet al (2002) and seems to be a deficiency in theradiosonde observation during day time.

Three Observing System Experiments have beenconducted, assimilating GPS data in the aLMo: 5days in September 2001 and January 2002 and 7days in June 2002. However, due to inconsistentusage of humidity correction scheme the resultsfrom winter OSE are inconclusive. The strongestGPS impact on aLMo IWV field was obtained inJune 2002 OSE, and is visible up to the end of theforecast (þ30 h). For this period the aLMo refer-ence analysis exhibits a dry bias over Switzerland,which is well corrected by assimilating GPS; 2 mtemperature and dew point temperature analysishave also been improved over the whole domain.However, the impact on the precipitation analysisand forecast is mixed. A missing structure isrecovered in the precipitation forecast on June20, 2002. A negative impact on precipitation ana-lysis reported on June 23 is possibly due to modelweakness in a special weather situation overnorthern Italy. The impact on precipitation fore-cast is limited to the first 6 hours and to intenseprecipitation events. In our preliminary September2001 OSE (Guerova et al, 2004) one case of posi-tive impact on the cloudiness is reported and theprecipitation daily cycle of the assimilation cyclewas improved.

The use of GPS in NWP has a foreseeablelong-term future due to the importance of humid-

ity information for high resolution mesoscalemodels and the good spatial coverage and tem-poral availability of these data. The future poten-tial of GPS will be further extended with the newEuropean project – GALILEO (the EuropeanSatellite Navigation System) in operationalservice from 2008. The GPS derived water va-por gradients could help improving the spatialspreading of humidity in active weather regimesand in regions with strong water vapor inhomo-geneity. One further improvement could be theretrieval of GPS humidity profiles (tomography),with the limiting factor being the accurate SlantPath estimates. The operational use of GPS inNWP models depends also on future data avail-ability; GPS networks belong mainly to thegeodetic community and are not incorporated inthe WMO Meteorological Observing System.

Acknowledgement

This work is supported by the Swiss Federal Office of Edu-cation and Science under grant no. C99.0046. We acknow-ledge F. Schubiger, M. Arpagaus, P. Kaufmann and A. Rossafrom MeteoSwiss for providing verification data. We aregrateful to the Swiss Centre for Scientific Computing forresources and support; observing system experiments requireindeed a lot of computing power!

References

Bettems JM (2002) EUCOS impact study using the limited-area non-hydrostatic NWP model in operational use atMeteoSwiss. SMA-MeteoSchweiz Pub. 62. (Availablefrom SMA-MeteoSchweiz, Kraehbuehlstrasse 58, 8044,Zurich, Switzerland)

Bevis M, Businger M, Herring TA, Rocken C, Anthes RA,Ware RH (1992) GPS Meteorology: Remote sensing ofatmospheric water vapor using the Global PositioningSystem. J Geophys Res 97: 15787–15801

Brockmann E, Guerova G, Troller M (2002) Swiss Activitiesin combining GPS with Meteorology. Mitteilung desBundesamtes f€uur Kartographie und Geod€aasie (TorresJA, Hornik H, eds). EUREF Publ. Vol. 23, Frankfurt,Germany, pp 95–99

COST-716 (2001) Format Specification for COST-716Processed GPS Data, Prepared by: D. Offiler, Met Office,Version 1.0f (http:==www.oso.chalmers.se=geo=cost716.html – WG2 – Documents)

COST-716 (2002) GPS Meteorology – USERREQUIREMENTS, Prepared by S. J. M. Barlag and S.De Haan, KNMI, Version 2.2, (http:==www.oso.chalmers.se=geo=cost716.html – WG3 – Documents)

Cucurull L, Navascues B, Ruffini G, Elosegui P, Ruis A,Vila J (2000) The use of GPS to validate NWP systems:


The HIRLAM model. J Atmos Oceanic Technol 17:773–787

Davies H (1976) A lateral boundary formulation for multi-level prediction models. Quart J Roy Meteor Soc 102:405–418

De Pondeca M, Zou X (2001) A case study of the variationalassimilation of GPS zenith delay observations into amesoscale model. J Appl Meteor 40: 1559–1576

Doms G (2002) The LM cloud ice scheme. COSMO News-letter No. 2 (Doms G, Schaettler U, eds), pp 128–136.(Available from Deutscher Wetterdienst, P.O. Box100465, 63004 Offenbach, Germany)

Doms G, Schaettler U (2001) COSMO Newsletter No 1, 114pp. (Available from Deutscher Wetterdienst, P.O. Box100465, 63004 Offenbach, Germany)

Doms G, Schaettler U, Montani A (2004) Model systemoverview: Data assimilation. COSMO Newsletter No. 4(Doms G, Schaettler U, Montani A, eds), pp 16–21.(Available from Deutscher Wetterdienst, P.O. Box100465, 63004 Offenbach, Germany)

Elgered G (2001) An overview of COST Action 716:Exploitation of ground-based GPS for climate and nu-merical weather prediction applications. Phys Chem Earth(A) 26: 6–8, 399–404

Falvey M, Beaven J (2002) The impact of GPS precipitablewater assimilation on mesoscale model retrievals of oro-graphic rainfall during SALPEX’96. Mon Wea Rev 130:2874–2888

Guerova G, Brockmann E, Quiby J, Schubiger F, Matzler Ch(2003) Validation of NWP mesoscale models with SwissGPS Network AGNES. J Appl Meteor 42: 1, 141–150

Guerova G, Bettems JM, Brockmann E, Matzler Ch (2004)Assimilation of the GPS-derived Integrated Water Vapor(IWV) in the MeteoSwiss Numerical Weather Predictionmodel – a first experiment. Phys Chem Earth 29: 2–3,177–186

Guerova G, Brockmann E, Schubiger F, Morland J,Matzler Ch (2005) An integrated assessment of mea-sured and modeled IWV in Switzerland for the period2001–2003. J Appl Meteor (accepted)

Guo YR, Kuo YH, Dudhia J, Parsons DB, Rocken C (2000)Four-dimensional data assimilation of heterogeneousmesoscale observations for a strong convective case.Mon Wea Rev 128: 619–643

Gutman SI, Holub K, Sahm SR, Stewart JQ, Smith TL,Benjamin SG, Schwarr BE (2004) Rapid retrieval andassimilation of ground based GPS-Met observations at the

NOAA Forecast Systems Laboratory: Impact on weatherforecast. Jap Meteor Soc 82: 1B, 351–360

Haase J, Ge M, Vedel H, Calais E (2002) Accuracy andvariability of GPS Tropospheric Delay Measurementsof water vapor in the Western Mediterranean. Bull AmMeteor Soc (submitted)

Kopken C (2001) Validation of integrated water vapor fromnumerical models using ground-based GPS, SSM=I, andwater vapor radiometer measurements. J Appl Meteor 40:1105–1117

Kuo YH, Zou X, Guo YR (1996) Variational assimilationof precipitable water using a non-hydrostatic mesoscaleadjoint model. Part I: Moisture retrieval and sensitivityexperiments. Mon Wea Rev 124: 122–147

Nakamura H, Koizumi K, Mannoji N, Seko H (2004) Dataassimilation of GPS precipitable water vapor to the JMAmesoscale numerical weather prediction model and itsimpact on rainfall forecast. Jap Meteor Soc J 82: 1B,441–452

Ohtani R, Naito I (2000) Comparisons of GPS-derivedprecipitable water vapors with radiosonde observationsin Japan. J Geophys Res 105: 26917–26929

Schraff Ch (1997) Mesoscale data assimilation and predic-tion of low stratus in the Alpine region. Meteorol AtmosPhys 64: 21–50

Smith T, Benjamin S, Schwartz B, Gutman S (2000) UsingGPS-IPW in a 4-D data assimilation system. Earth PlanetsSpace 52: 921–926

Tomassini M, Gendt G, Dick G, Ramatschi M, Schraff Ch(2002) Monitoring of integrated water vapor fromground-based GPS observations and their assimilationin a limited area model. Phys Chem Earth 27: 341–346

Van der Marel H, Brockmann E, de Haan S, Dousa J,Johansson J, Gendt G, Kristiansen O, Offiler D,Pacione R, Rius A, Vespe F (2003) COST-716 NearReal-Time demonstration project. Jap Meteor Soc J(submitted)

Vedel H, Huang XY, Haase J, Ge M, Calais E (2003) Impactof GPS zenith tropospheric delay data on the precipitationforecasts in Mediterranean France and Spain. GeophysRes Lett (submitted)

Corresponding author’s address: G. Guerova, SwissFederal Institute of Technology (EPFL), EPFL-ENAC-LMCA, Lausanne, 1015, Switzerland (E-mail: [email protected])

164 G. Guerova et al: Assimilation of COST 716 Near-Real Time GPS data

Date post:	22-Nov-2023
Category:	Documents
Upload:	independent
View:	0 times
Download:	0 times