Four-Dimensional Variational Assimilation of Water Vapor Differential Absorption Lidar Data: The...

23
Four-Dimensional Variational Assimilation of Water Vapor Differential Absorption Lidar Data: The First Case Study within IHOP_2002 VOLKER WULFMEYER,HANS-STEFAN BAUER,MATTHIAS GRZESCHIK, AND ANDREAS BEHRENDT Institut für Physik und Meteorologie, Universität Hohenheim, Stuttgart, Germany FRANCOIS VANDENBERGHE Research Applications Program, National Center for Atmospheric Research, Boulder, Colorado EDWARD V. BROWELL,SYED ISMAIL, AND RICHARD A. FERRARE Atmospheric Sciences Division, NASA Langley Research Center, Hampton, Virginia (Manuscript received 20 October 2004, in final form 21 June 2005) ABSTRACT Four-dimensional variational assimilation of water vapor differential absorption lidar (DIAL) data has been applied for investigating their impact on the initial water field for mesoscale weather forecasting. A case that was observed during the International H 2 O Project (IHOP_2002) has been selected. During 24 May 2002, data from the NASA Lidar Atmospheric Sensing Experiment were available upstream of a convective system that formed later along the dryline and a cold front. Tools were developed for routinely assimilating water vapor DIAL data into the fifth-generation Pennsylvania State University–NCAR Me- soscale Model (MM5). The results demonstrate a large impact on the initial water vapor field. This is due to the high resolution and accuracy of DIAL data making the observation of the high spatial variability of humidity in the region of the dryline and of the cold front possible. The water vapor field is mainly adjusted by a modification of the atmospheric wind field changing the moisture transport. A positive impact of the improved initial fields on the spatial/temporal prediction of convective initiation is visible. The results demonstrate the high value of accurate, vertically resolved mesoscale water vapor observations and ad- vanced data assimilation systems for short-range weather forecasting. 1. Introduction Quantitative precipitation forecasting (QPF) is a challenging objective. For achieving an acceptable ac- curacy of the temporal/spatial distribution of the pre- dicted precipitation, a complicated chain of processes has to be modeled correctly. This includes the param- eterization of turbulent processes and subgrid-scale convection, cloud microphysics, and radiation. Also the knowledge of the initial fields, particularly the key vari- ables water vapor and wind, is essential. Because of a lack of observing systems, large data gaps in the initial fields still exist to date. And if datasets exist, suitable data assimilation systems are of- ten not available. Furthermore, deficiencies of the pa- rameterizations are significant (e.g., Doms et al. 2002; Zängl 2004a,b). Therefore it is not surprising that the skill of short-range and medium-range QPF is not suf- ficient to serve many requests of the large user com- munities. For instance, the current QPF skill does not permit extending the lead time for flash flood forecast- ing (Arnaud et al. 2002). An extensive analysis of global weather forecast models was performed by the Working Group of Nu- merical Experimentation (WGNE) (Ebert et al. 2003), which has been established by the Commission for At- mospheric Sciences (CAS) of the World Meteorologi- cal Organization (WMO). Typically, the QPF skill is 0.2 above persistence, even for rather moderate rain rates Corresponding author address: Volker Wulfmeyer, Institut für Physik und Meteorologie, Universität Hohenheim, Garbenstrasse 30, 70599 Stuttgart, Germany. E-mail: [email protected] JANUARY 2006 WULFMEYER ET AL. 209 © 2006 American Meteorological Society MWR3070

Transcript of Four-Dimensional Variational Assimilation of Water Vapor Differential Absorption Lidar Data: The...

Four-Dimensional Variational Assimilation of Water Vapor Differential AbsorptionLidar Data: The First Case Study within IHOP_2002

VOLKER WULFMEYER, HANS-STEFAN BAUER, MATTHIAS GRZESCHIK, AND ANDREAS BEHRENDT

Institut für Physik und Meteorologie, Universität Hohenheim, Stuttgart, Germany

FRANCOIS VANDENBERGHE

Research Applications Program, National Center for Atmospheric Research, Boulder, Colorado

EDWARD V. BROWELL, SYED ISMAIL, AND RICHARD A. FERRARE

Atmospheric Sciences Division, NASA Langley Research Center, Hampton, Virginia

(Manuscript received 20 October 2004, in final form 21 June 2005)

ABSTRACT

Four-dimensional variational assimilation of water vapor differential absorption lidar (DIAL) data hasbeen applied for investigating their impact on the initial water field for mesoscale weather forecasting. Acase that was observed during the International H2O Project (IHOP_2002) has been selected. During 24May 2002, data from the NASA Lidar Atmospheric Sensing Experiment were available upstream of aconvective system that formed later along the dryline and a cold front. Tools were developed for routinelyassimilating water vapor DIAL data into the fifth-generation Pennsylvania State University–NCAR Me-soscale Model (MM5). The results demonstrate a large impact on the initial water vapor field. This is dueto the high resolution and accuracy of DIAL data making the observation of the high spatial variability ofhumidity in the region of the dryline and of the cold front possible. The water vapor field is mainly adjustedby a modification of the atmospheric wind field changing the moisture transport. A positive impact of theimproved initial fields on the spatial/temporal prediction of convective initiation is visible. The resultsdemonstrate the high value of accurate, vertically resolved mesoscale water vapor observations and ad-vanced data assimilation systems for short-range weather forecasting.

1. Introduction

Quantitative precipitation forecasting (QPF) is achallenging objective. For achieving an acceptable ac-curacy of the temporal/spatial distribution of the pre-dicted precipitation, a complicated chain of processeshas to be modeled correctly. This includes the param-eterization of turbulent processes and subgrid-scaleconvection, cloud microphysics, and radiation. Also theknowledge of the initial fields, particularly the key vari-ables water vapor and wind, is essential.

Because of a lack of observing systems, large data

gaps in the initial fields still exist to date. And ifdatasets exist, suitable data assimilation systems are of-ten not available. Furthermore, deficiencies of the pa-rameterizations are significant (e.g., Doms et al. 2002;Zängl 2004a,b). Therefore it is not surprising that theskill of short-range and medium-range QPF is not suf-ficient to serve many requests of the large user com-munities. For instance, the current QPF skill does notpermit extending the lead time for flash flood forecast-ing (Arnaud et al. 2002).

An extensive analysis of global weather forecastmodels was performed by the Working Group of Nu-merical Experimentation (WGNE) (Ebert et al. 2003),which has been established by the Commission for At-mospheric Sciences (CAS) of the World Meteorologi-cal Organization (WMO). Typically, the QPF skill is 0.2above persistence, even for rather moderate rain rates

Corresponding author address: Volker Wulfmeyer, Institut fürPhysik und Meteorologie, Universität Hohenheim, Garbenstrasse30, 70599 Stuttgart, Germany.E-mail: [email protected]

JANUARY 2006 W U L F M E Y E R E T A L . 209

© 2006 American Meteorological Society

MWR3070

of 8 mm day�1. This low skill further decreases with anincreasing amount of precipitation. Comparable skillanalyses of mesoscale models are missing. Although itcan be expected that the short-term forecast skill ofnested high-resolution models will be better (Davis andCarr 2000; Mass et al. 2002), improvements are stillrequired to meet the user needs.

For advancing QPF, error sources have to be de-tected and quantified. However, it is very difficult toseparate the contributions of errors due to initial fieldsand parameterizations. This is only possible if accurateobservations with good coverage and resolution areavailable in the model domain. By means of assimila-tion of these data, more insight is gained in the sensi-tivity of the forecast model on initial fields.

Several research programs have been initiated forinvestigating and improving QPF. The ObservingSystem Research and Predictability Experiment(THORPEX), a Global Atmospheric Research Pro-gram, is a 10-yr project of the World Weather ResearchProgram (WWRP). Its primary goal is the extension ofskillful medium-range QPF to a duration of 14 days(www.wmo.int/thorpex). Tools are targeted satelliteand in situ observations that are assimilated in ad-vanced weather forecast systems.

For better process understanding, high-resolutionobservations of key atmospheric variables are essential.Because of the large effort in the operation of corre-sponding remote sensing systems, these observationsare concentrated in field experiments. Within the Me-soscale Alpine Program (MAP) (see www.map.meteoswiss.ch), a better understanding of the develop-ment of precipitation in a high-mountain region was akey component (Bougeault et al. 2001). The Interna-tional H2O Project (IHOP_2002) has been conductedfor investigating the role of the water vapor field in theinitiation of convection and its relation to precipitation.Further details are found online at www.atd.ucar.edu/dir_off/projects/2002/IHOP.html. IHOP_2002 was per-formed in a relatively flat terrain. Convergence of dif-ferent air masses leads to strong inhomogeneities in theinitial fields of atmospheric variables resulting in thepresence of a dryline. Data from IHOP_2002 are usedin this study.

Another experiment was the Convective Storms Ini-tiation Project (CSIP) (www.env.leeds.ac.uk/csip/),which was performed in summer 2005 in the southernUnited Kingdom. A primary goal was the understand-ing of the initiation of convection as well as the forma-tion of secondary cells in a maritime environment. An-other large QPF research program has been initiatedand started in Germany (www.meteo.uni-bonn.de/projekte/SPPMeteo/). An essential component of this

effort is the performance of a field campaign in summer2007 in a low-mountain region. This study, which iscalled the Convective and Orographically-induced Pre-cipitation Study (COPS), will provide a comprehensivedataset for testing hypotheses on the improvement ofQPF in regions with more complex orography (www.uni-hohenheim.de/spp-iop). Consequently, the experi-ments IHOP_2002, CSIP, and COPS can be consideredas a series of QPF campaigns in which the regions ofinterest are becoming more and more influenced byorography.

In all these campaigns, the assimilation of state-of-the-art water vapor measurements is essential for im-proving mesoscale initial fields and for performing pro-cess study. This work is a study concerning the impactof water vapor data assimilation on the quality of QPF.The tools developed in this study will also be appliedfor the preparation and performance of future cam-paigns.

In a number of field campaigns, highly accurate andhigh-resolution water vapor data were provided usingground-based and airborne lidar systems. However, weare aware of only two related studies where water va-por differential absorption lidar (DIAL) data were as-similated in a global atmospheric model (Kamineni etal. 2003, 2006). Even using the three-dimensional varia-tional data assimilation (3DVAR) technique, whichdoes not take full advantages of the high temporal reso-lution of the DIAL data, the results showed a positiveimpact on the forecast of hurricane tracks.

This paper investigates the impact of four-dimen-sional variational data assimilation (4DVAR) analysisof water vapor DIAL data on short-range mesoscaleforecasts. We are focusing on a case that has been ob-served during the IHOP_2002 campaign. On 24 May2002, initiation of convection occurred along an east-ward-moving dryline in western Texas and Oklahomasouth of a triple point that formed in western Oklaho-ma. In the preconvective environment, upstream of theregion of convection, the National Aeronautics andSpace Administration (NASA) Langley Research Cen-ter (LaRC) Lidar Atmospheric Sensing Experiment(LASE) system was operated on a DC-8 aircraft andperformed water vapor DIAL measurements. There-fore, this case was well suited for investigating the im-pact of the assimilation of high-resolution and accuratewater vapor profiles on model forecast quality. The fol-lowing science questions are addressed:

• How accurately are water vapor fields represented incurrent mesoscale models?

• Can these initial fields be improved using advancedwater vapor lidar sensing systems?

210 M O N T H L Y W E A T H E R R E V I E W VOLUME 134

• What is the impact of accurate remote sensing dataon predicted water vapor and precipitation fields?

• How long is the memory of improved water vapordata in the model forecast?

The paper is organized as follows. In section 2, ashort overview of IHOP_2002 is given. Meteorologicalanalyses of the selected case and the observations arepresented in section 3. In section 4, the performance ofLASE is introduced including an extended error analy-sis. The water vapor measurements are shown and ana-lyzed. In section 5, we describe the procedure for per-forming 4DVAR assimilation of LASE observations.In section 6, the results are presented and discussed.Finally, an overview of future activities is given.

2. IHOP_2002 overview

IHOP_2002 was a large field effort dedicated to theimprovement of QPF in the region of the dryline. TheIHOP_2002 domain surrounded Oklahoma, a criticalregion for severe thunderstorm development. The ex-periment took place from 13 May 2002 to 25 June 2002.The operations center was located at the National Oce-anic and Atmospheric Administration (NOAA)/SevereStorm Laboratory (NSSL) in Norman, Oklahoma.

The main focus was the observation of the precon-vective water vapor field and atmospheric dynamics. Ahuge combination of instrumentation was operated andconsisted of three airborne DIAL systems, airborneDoppler lidar, and radar, among others. Several air-crafts were also equipped with in situ turbulence sen-sors and dropsonde systems. A suite of ground-basedremote sensing and in situ systems was deployed atdifferent field sites. Further details on the instrumentsetup is found in Weckwerth et al. (2004), theIHOP_2002 Web site (see above), and the IHOP_2002field catalogue (www.joss.ucar.edu/ihop/catalog/index.html).

Four research groups were set up focusing on theatmospheric boundary layer (ABL), convective initia-tion (CI), QPF, and instrumentation (IN), respectively.The corresponding missions of each group were definedin the Operations Plan. The major part of the missionswas dedicated to CI and ABL research.

We investigated the whole IHOP_2002 dataset forfinding a case with the following properties:

• Significant CI• Strong precipitation event• High quality water vapor data available in the pre-

convective environment upstream of the convectiveevent

The latter condition ensured that improved informa-tion was advected to the region of interest where CIwas expected, if there was a positive impact of watervapor data assimilation to the initial fields. The optimalday for this study turned out to be 24 May 2002.

3. IHOP_2002 case on 24 May 2002

a. Meteorological conditions

The upper-level analysis showed a short-wave troughmoving west to east over the central Great Plains. Thiswas captured by the velocity field at 300-hPa height ofthe 1800 UTC European Centre for Medium-RangeWeather Forecasts (ECMWF) operational analysis (seeFig. 1, left panel). It was associated with a well-definednear-surface short-wave trough visible in the 850-hPatemperature field (Fig. 1, right panel). It was forecastedto move east-northeastward across the central highplains through 0100 UTC increasing the gradients ofgeopotential height, temperature, and deep-layer shearin the IHOP_2002 domain. The associated large-scaleascent also helped to maintain already steep midlevellapse rates of 8–9 K km�1. They were superimposed bysurface dewpoints of the order of 20°C in the moistsector. This behavior supported high values of convec-tive available potential energy (CAPE). The strong in-stability was indicated by low values of the lifted indexprovided by the Eta Model analysis (not shown).

A significant dryline developed during the day result-ing in a sharp gradient in the relative humidity overwestern Texas and eastern New Mexico. This was sup-ported by northward moisture transport from the Gulfof Mexico, which continued during the night by a broadlow-level jet. The situation became more complex dueto the presence of the cold front, which was pushingsouthward during the day. The dryline and the coldfront merged in a triple point, enhancing the probabil-ity of CI. These large-scale processes were fairly wellforecasted by different operational forecast systemsused for mission planning, for example, the Eta and theRapid Update Cycle (RUC) models.

A prediction of the locations of the cold front and thedryline together with probabilities for CI was per-formed by the Storm Prediction Center (SPC) (avail-able from the IHOP_2002 field catalogue; see www.joss.ucar.edu/ihop/catalog/). The region of high prob-ability of convection developed around 2000 UTC atthe southwestern corner of Oklahoma, moved slowlysoutheastward, and reached its maximum extentaround 2300 UTC. Together these were responsible fora high probability for the development of a mesoscale

JANUARY 2006 W U L F M E Y E R E T A L . 211

FIG

.1.E

CM

WF

oper

atio

nal

anal

ysis

ofth

e(l

eft)

hori

zont

alw

ind

velo

city

at30

0hP

ain

clud

ing

win

dve

ctor

san

d(r

ight

)85

0-hP

ate

mpe

ratu

refi

eld

incl

udin

gw

ind

vect

ors

at18

00U

TC

24M

ay20

02.

212 M O N T H L Y W E A T H E R R E V I E W VOLUME 134

Fig 1 live 4/C

convective system (MCS), which was predicted to formduring this night.

The development and penetration of the cold frontwas influenced by outflow boundaries of the previousday’s convective developments in northeastern Texasand Colorado. Around 1900 UTC, the cold front wasmoving slowly southward through the Texas panhandleand northwestern Oklahoma. High values of CAPEand low values of convective inhibition (CIN) werepresent over the Texas panhandle and southwesternOklahoma. According to the mesoscale discussion per-formed at SPC, convection should begin near the triplepoint at about 2000 UTC. The locations of the dryline,the cold front, and the triple point at 2125 UTC areindicated in Fig. 2 including the flight track of the DC-8aircraft carrying the LASE instrument.

The convection was initiated along a southwest–

northeast band from central Texas to the southwesterncorner of Oklahoma. This activity along and east of thedryline occurred in a deeply mixed area and extendedrapidly north-northeastward. This development waswell captured by the visible channel of the Geostation-ary Operational Environmental Satellite-8 (GOES-8)(available from the IHOP_2002 field catalogue). Theeast–west-oriented cold front continued to surge southand was undercutting the storm near Erick, Oklahoma.

In retrospect, the timing of the convection was wellpredicted by the operational regional models. How-ever, the location of the dryline and the triple pointwere predicted too far west. The CI did not take placeat the triple point but about 200 km southwest along thedryline. The intensity of convection was overestimated.Furthermore, details of convection, the location of out-flow boundaries, and the fine structure of the dryline

FIG. 2. GOES visible channel image at 2125 UTC on 24 May 2002. The blue asterisk shows the estimated location of the triple point.The locations of the dryline and the cold front are also indicated. The flight track of the DC-8 aircraft carrying the LASE instrumentis plotted too, separated into six measurement periods, P1–P6.

JANUARY 2006 W U L F M E Y E R E T A L . 213

Fig 2 live 4/C

were not well predicted (see also section 6). Furtherdiscussions of the meteorological conditions during thisday are found in Geerts et al. (2006), Martin and Xue(2006), Wakimoto et al. (2006), and Xue and Martin(2006a,b).

b. Observations

As the initial point for IHOP_2002 operations,Shamrock, Texas, was selected based on its proximityto the apparent western edge of the outflow over Okla-homa and the likelihood that this might be close to adryline–cold front intersection by 1900–2100 UTC.

A huge combination of instruments was operated. Adiscussion of the operation of other aircrafts can befound in the IHOP_2002 field catalogue and in Geertset al. (2006) and Wakimoto et al. (2006). In this work,we are focusing on the operation, the measurements,and the performance of LASE. The DC-8 took off at1717 UTC and landed at 2244 UTC. Figure 2 shows theflight pattern chosen for the DC-8 aircraft separatedinto six periods.

During period 1 (P1; hereafter periods are referredto P1, P2, etc.) from 1750 to 1822 UTC, LASE wasflying along 35.77°N from 97.7° to 102.2°W. Analysis ofGOES images showed that LASE started in the north-eastern part of the moist sector east of the dryline,crossed the triple point at about 100.5°W (1808 UTC),and overpassed a section of the cold air north of thecold front. Further details are discussed in Wakimoto etal. (2006). During P2, LASE performed a west–east legfrom 102.5° to 98.4°W along 35.53°N. LASE was mostlyflying south of the cold front and tipped the southerncorner of the cold front at 101.4°W. P3 was defined bya southwest track of LASE from 35.7°N, 98.4°W to34.2°N, 99.6°W. During P4, LASE turned northwestand performed a leg from 34.2°N, 99.6°W to 35.2°N,103.3°W. Afterward, LASE flew a leg from 103.3° to98.5°W along 35.3°N (P5). As during P2, the southerntip of the cold front was scanned at about 101.5°W.Finally, P6 was characterized by another leg from 98.6°to 102.8°W mainly along 35.8°N (see also Wakimoto etal. 2006). Consequently, the airplane flew three timesacross the dryline and three times over the region of thetriple point across the cold front.

The operation mode of LASE and the region cov-ered by the measurements were very well suited for adata assimilation study. The preconvective water vaporfield was mapped by LASE in great detail in the up-wind region of the expected CI area. It was expectedthat by means of data assimilation an improved initialwater vapor field was advected in the region of interest,potentially improving the forecast of the water vaporand precipitation fields. The major part of the LASE

data was unaffected by the presence of clouds. TheLASE measurements had the largest vertical and hori-zontal coverage of all DIAL systems so that the im-act by data assimilation was expected to be strongest.The water vapor number density measured with LASEcan nearly directly be assimilated, as the transforma-ion to the prognostic variables of mixing ratio or ofspecific humidity is simple (see section 5a). Anotherimportant point, which favored LASE for data assimi-lation, was the fact that its performance was extensivelycharacterized by validation campaigns with excellentresults with respect to bias and noise errors (Browell etal. 1997).

4. LASE measurements

a. DIAL methodology and error analysis

LASE is the first autonomous DIAL system (Browelland Ismail 1995) that has been operated on variousaircrafts (Browell et al. 1998). During IHOP_2002, itwas deployed on the NASA DC-8. The flight speed wastypically 230 m s�1 and the flight altitude was about 8km. The retrieved LASE water vapor profiles coveredthe entire troposphere from the ground to about 6.5 kmabove sea level (ASL). Further details about the LASEsystem are found in Moore et al. (1996).

The DIAL methodology was introduced in severalpublications (Browell et al. 1979; Ismail and Browell1989; Bösenberg 1998; Wulfmeyer and Bösenberg1998) so that only a short description of the measure-ment process is given here. For signal processing, thefollowing steps are performed:

1) Each single-shot backscattered lidar return signal ofthe frequency tuned to the water vapor line (onlinesignal) and the nonabsorbed signal (offline signal)are detected and stored with a vertical resolution of30 m.

2) The background is subtracted.3) Each signal is averaged over 1 min.4) The online and offline signals are divided.5) A running average over 330 m is taken on the ratio

to reduce noise.6) The logarithm of the smoothed ratio is taken.7) The derivative of this logarithm over 330 m is de-

termined along the entire profile.8) The water vapor number density profile NW is cal-

culated using the water vapor absorption cross-section profiles.

9) This water vapor profile is corrected for Rayleigh–Doppler broadening (see, e.g., Ansmann and Bösen-berg 1987).

214 M O N T H L Y W E A T H E R R E V I E W VOLUME 134

Further details of advanced DIAL processingschemes are found in Bauer et al. (2004) and Di Giro-lamo et al. (2004).

At this point water vapor number density profileswere available with a horizontal resolution of 14 kmand a vertical resolution of about 330 m. The horizontalerror weighting function of the data can be consideredunity for a distance of less than 14 km. At a distance ofmore than 14 km, the errors can be considered as un-correlated. In the vertical, the error weighting functionW is a triangle function with a leg size of 330 m (Wulfm-eyer et al. 2004).

Using this retrieval technique, the water vapor pro-file started at 330 m above the surface. After postpro-cessing, water vapor retrievals close to the ground werefilled in by utilizing on- and offline ground return sig-nals from the low gain channel (Browell et al. 1997).Averaged (210 m vertically and 600 m horizontally)atmospheric signals centered at 120 m above groundwere calculated and used in conjunction with groundreturn signals for obtaining average water vapor num-ber density at 60 m. Afterward, 14-km horizontal aver-ages were calculated. This resulted in a vertical cover-age of the water vapor profiles from 60 m above groundup to about 6.5 km. At this height full overlap betweenthe laser transmitter beam and the receiver wasachieved so that corresponding systematic errors couldbe neglected.

The power of the DIAL technology is not only itshigh spatial resolution but also the capability to specifysystem errors very accurately. This is particularly im-portant for data assimilation studies. In the followingdiscussions, the systematic errors are called bias. Thiserror cannot be removed by signal averaging. The otherpart is due to system noise. It is assumed that differentnoise sources (detector and amplifier noise, Poissonstatistics, etc.) are uncorrelated in each range bin. In thefollowing discussions, this kind of error is called noiseerror.

Systematic errors can be assessed using analytical er-ror propagation (Ismail and Browell 1989; Bösenberg1998; Wulfmeyer and Bösenberg 1998; Wulfmeyer andWalther 2001a,b), end-to-end simulators (Bauer et al.2004; Di Girolamo et al. 2004), and comparisons withother instruments. LASE has extensively been charac-terized within validation campaigns and the overall rmsaccuracy has been determined to be less than 6% overthe entire measurement range (Browell et al. 1996).

For data assimilation, it is important to separate thisoverall error into uncorrelated errors due to systemnoise and a bias due to remaining uncertainties in theretrieval. A major part of the systematic errors are dueto laser performance and limited knowledge of water

vapor spectroscopy. Previous systematic error analysesand comparisons suggest an overall systematic error ofLASE water vapor profiles of about 3%.

The signal-to-noise ratio (SNR) of the backscattersignal depends on the atmospheric backscatter proper-ties and the efficiency of the receiver system. An el-egant technique exists, which can be used to estimatenoise errors in each single water vapor profile. Thisrealistic and direct noise error analysis without relyingon any other information and instruments is anothersignificant advantage of the DIAL technique. The cor-responding analysis has been presented, for example, inWulfmeyer (1999) and Lenschow et al. (2000), so thatonly a short introduction is given here.

First of all, a high-resolution water vapor time seriesis calculated using a resolution of 1.4 km, which corre-sponded to an averaging time of 6 s. The autocovari-ance function of this time series is determined in eachheight. The difference between the zero lag and the firstlag is a good approximation of the noise variance at thishorizontal resolution. The noise contribution in the firstlag is nearly eliminated, as noise errors are uncorre-lated between each sample. Noise errors �n at a differ-ent horizontal resolution are estimated by taking thesquare root of the variance and by dividing by thesquare root of the horizontal resolution ratio betweenthe high-resolution time series and the horizontal reso-lution of interest (Wulfmeyer 1999). As in some cases,not at all heights in the 6-s profiles water vapor esti-mates were available (too large statistical errors); thestatistics were calculated from a 2-min running windowthereby filling in midprofile gaps. Within the 2-min win-dow an error estimate was reported if at least 3 of the20 water vapor values were valid. This produced someerror estimate values for a time and altitude wherethere was no water vapor value. The final step was toeliminate these by masking the error estimates to the1-min water vapor field.

Finally, the error covariance matrix R can be con-structed for each noise error profile �n, which reads

Rmk � �n,m2 Cmk . �1�

Here, �n,m is the noise variance at model level m, andCmk is the error correlation function, which is the au-tocorrelation function of W between model levels mand k. Hence, Cmk: �C(zm � zk) where zm and zk arethe heights of the model levels. Figure 3 presents anexample of R for mixing ratio m, which was calculatedusing Eq. (3) (see section 5a) and taking temperatureand pressure profiles from a fifth-generation Pennsyl-vania State University–National Center for Atmo-spheric Research Mesoscale Model (MM5) forecast atthe location of the DIAL measurement.

JANUARY 2006 W U L F M E Y E R E T A L . 215

b. Discussion of the measurements

Figure 4 presents the overall LASE water vapornumber density (upper panel) and noise error fields(lower panel). The white areas indicate clouds, whichcould not be penetrated by the LASE transmitter. Theblack line at ground indicates the height of the surfaceabove sea level. The water vapor and error profilesstart at 60 m above ground.

Figure 4 confirms the high-resolution and good ver-tical coverage of LASE. Using both absorption linesand frequency switching, over the entire domain watervapor profiles can be measured from near the groundup to 6.5 km. The black line, which occurs mainly at aheight of 5 km, indicates the region where the DIALretrieval is switched between different absorption linestrengths. The height level of switching decreased con-siderably in the region of the dryline, as the drying alsoaffected the lower troposphere up to a height of 5 km.

The lower panel of Fig. 4 demonstrates that the hori-zontal and vertical variability of noise errors should betaken into account in data assimilation efforts. The ver-tical variability of the noise is mainly due to changes inthe structure of humidity. Close to regions of frequencyswitching, the errors are largest, as here the SNRs ofboth online signals are minimal. Errors of up to 80%can be reached here but only in strongly limited areas.Otherwise, the errors are typically �20%, which is apromising performance for data assimilation.

We investigated LASE humidity measurements dur-

ing all periods P1–P6 in detail. Exemplarily, Fig. 5 (leftpanel) presents a zoom in the time–height cross sectionof LASE absolute humidity measurements during P2.The right panel of Fig. 5 shows the correspondingGOES image including the estimated locations of thecold front, the dryline, and the triple point. West of thedryline, the middle-tropospheric humidity (MTH) wasabout 1 g m�3 and increased to 3 g m�3 ahead of thedryline at 101.4°W. The structure of the humidity fieldreveals several moist filaments over the dryline indicat-ing horizontal mixing of air masses. The ABL depth inthe dry sector was about 1 km and the absolute humid-ity in the ABL about 8 g m�3. The dryline could beclearly detected at 100.5°W where the ABL humiditysharply increased to more than 12 g m�3 accompaniedby an increase of ABL depth to 2 km. For the remain-ing part of this period, ABL observations were shad-owed by clouds.

The measurements during the other periods con-firmed an ABL depth of about 1 km in the dry sectorwest of the dryline, which increased to about 2.3 kmeast of the dryline. A similar moisture jump was ob-served crossing the dryline during P4, while during P5,close to the time of initiation of convection, the dataindicated a larger moisture jump and a slightly largerABL depth in the moist sector. With increasing dis-tance from the dryline, the humidity decreased gradu-ally to 7 g m�3 at 103.3°W. The MTH was consistentlylarger in the moist sector east of the dryline. During P4,

FIG. 3. Vertical error correlation matrix R of LASE for mixing ratio m at 3000 m abovesea level (ASL). The units are in (g kg)�2. The matrix values between 0 and 100 corre-spond to a height from 1477 to 4477 m with a resolution of 30 m.

216 M O N T H L Y W E A T H E R R E V I E W VOLUME 134

FIG. 4. (top) Time–height cross section of water vapor number density measured with LASE. (bottom)Corresponding time–height cross section of relative noise error of LASE water vapor number density.

JANUARY 2006 W U L F M E Y E R E T A L . 217

Fig 4 live 4/C

P5, and P6, the MTH decrease west of the drylineshowed a 2D structure like a dry intrusion.

These results demonstrate the capability of airborneDIAL for high-resolution process studies. The data canbe used to detect the dryline and to investigate thechange of its moisture structure before CI takes place.The data indicate enhanced moisture convergence inthe region of the dryline and enhanced lifting of ABLair masses before convection was initiated. This obser-vation is in agreement with the mechanisms suggestedin Xue and Martin (2006b).

5. Four-dimensional variational assimilation ofDIAL data into MM5

MM5 has been chosen for data assimilation, as itprovides convenient tools for ingesting measurementsof different observation systems, which is a good start-ing point for lidar data assimilation. An overview ofMM5 and the available data assimilation schemes aregiven in Grell et al. (1995). A detailed introduction forthe MM5 4DVAR system can be found in Ruggiero etal. (2001).

We applied 4DVAR, as we expected the major im-pact of DIAL data assimilation on the quality of themodel forecasts. In 4DVAR, the cost function

J�x� � �x � xb�TB�1�x � xb�

� �i�0

n

yi � Hi�xi�TRi

�1yi � Hi�xi� �2�

is minimized in the data assimilation window. Here xand xb are the state vectors of the model and back-

ground field variables, respectively; B is the back-ground error covariance matrix; y are the observations;xi the corresponding model forecasts both valid at timesi; and Hi and Ri are the model forward operator and theobservation error covariance matrix, respectively.

4DVAR uses the whole information content of theLASE data, as the vertical and the temporal distribu-tions of the water vapor field are taken into account.This is particularly important in regions with high spa-tial/temporal variability in the water vapor field such asnear the dryline. 4DVAR considers in a reasonable waythe error characteristics of the observations, as in eachsingle profile the best estimate of the error covariancematrix can be used for the calculation of J [see Eq. (2)].Additionally, this continuous data assimilation tech-nique considers the physics of the atmospheric pro-cesses while minimizing the cost function. Other dataassimilation schemes are also available with the MM5release: 3DVAR (Barker et al. 2004) and Newtonianrelaxation known as four-dimensional data assimilation(FDDA) (Stauffer and Seaman 1990, 1994). Compari-sons between different assimilation schemes are ongo-ing but are not the subject of this paper.

In its current release, the MM5 4DVAR system onlyoffers diagonal background error covariance matricesB. This approximation, however, has proven to workwell for most studies conducted with the system (Zou etal. 1995; Xiao et al. 2000). This can probably be ex-plained by the ability of 4DVAR systems to generatephysically consistent structure functions during themodel integration. The specification of B at the initialtime seems therefore not crucial. For each control vari-

FIG. 5. (left) Zoom in of the time–height cross section of LASE absolute humidity measurements during P2 from 1822 to 1900 UTC.The estimated location of the dryline is also indicated. (right) Corresponding GOES-8 image including the flight track in red. The redcrosses mark a full flight interval of 5 min. The estimated locations of the cold front and the dryline are marked by blue and orangelines, respectively.

218 M O N T H L Y W E A T H E R R E V I E W VOLUME 134

Fig 5 live 4/C

able, the background error variances (diagonal ele-ments of the matrix) were specified by constructing thedifferences between the 3-h forecasted and the initialvalues at each grid point. At each vertical level, themaximal value of the differences is found and assignedto all grid points on that level. This creates a verticalprofile of forecast errors valid at all geographical loca-tions of the model. The forecast errors are then squaredto produce the diagonal elements of the backgrounderror covariance matrix.

There is also an option in the MM5 4DVAR systemto prescribe the background error variances from thetables published in Parrish and Derber (1992). Thesetables contain the observational errors used for the ra-dio-sounding assimilation in the National Centers forEnvironmental Prediction (NCEP) spectral statisticalinterpolation (SSI) global analysis system at the time ofthe publication. We experimented with both options:MM5 forecast differences and NCEP sounding obser-vational errors and did not find significant differences.The latter was used in our study. Note that assigningsounding observational error values to model back-ground errors does not imply that model forecasts havethe radio-sounding instrumental accuracy. Radio-sounding instrumental error accounts only for a smallpart of the observational error compared to the contri-bution of representativeness error. Rather, the num-bers published by NCEP shall be regarded as baselinesfor specifying observational or forecast errors.

The addition of a new capability in any variationaldata assimilation system requires the two followingcomponents: a physical description of the measurementprocess and a statistical description of the accuracy ofthis measurement. The former, known as the observa-tion or forward operator, is usually expressed as a com-puter code that takes the model prognostic variables asinput and produces the measurement as output whilethe latter is usually provided as the set of coefficientsthat form the observing system error covariance matrix.This matrix has been presented in section 4a. The de-velopment of the DIAL observation operator and itsimplementation in the code of the MM5 4DVAR sys-tem is described in the next section.

a. The LASE decoder

A theoretical optimal 4DVAR LASE data assimila-tion system shall assimilate water vapor number den-sity, which is the primary observable of the DIAL sens-ing technique. Such an optimal system is ultimately thegoal of this research work. There are, however, practi-cal considerations that are important to keep in mindwhen it comes to real-world applications. First, as it waspointed out in Joiner and Da Silva (1998), optimality is

not always the most efficient approach and even some-times not desirable. Second, the multiprocessor versionof the MM5 4DVAR software that we are using is rela-tively new (Ruggiero et al. 2001) and it takes some timeto get familiar with all the aspects of this complex soft-ware. Our first step, therefore, was to assimilate watervapor mixing ratio m derived from water vapor numberdensity NW from DIAL measurements, where m is theMM5 prognostic variable for moisture and can bereadily ingested into MM5 4DVAR with the existingcode of the public release.

The relation between m and NW is given by

m �NW

NL

MWRL

p

T� 1.608NW

, �3�

where NL is Loschmidt’s number, MW is the molecularweight of water vapor, and RL is the gas constant of dryair. The model pressure p and temperature T at thecorresponding height have to be used in order to per-form the inversion described by Eq. (3). In practice, pand T are extracted from a free forecast performed justbefore the assimilation run with the same initial andboundary conditions. In this approach, the model pres-sure and temperature used to derive the water vapormixing ratio are kept constant while they would vary inthe minimization progress if the number density wereassimilated. In principle, a direct assimilation of num-ber density observations may result in larger pressureand temperature analysis increments and, possibly, asmaller mixing ratio analysis increment, although this ishard to ascertain due to the complexity of the modelvariable interactions that take place in the physical pro-cesses of the adjoint model. In any case, we expect onlya minor impact on the results due to the very low crosssensitivity of mixing ratio on temperature and pressureincrements [see Eq. (3)].

The MM5 4DVAR software allows the user to inputobservational error data along with the observations.Instrumental error estimates from the analysis de-scribed in section 4a were used. So far, there has notbeen any attempt to quantify the representativeness er-ror related to the use of any instrument in conjunctionwith MM5. In absence of good estimates, we did notallow an error of less than 0.1 g kg�1 to take the rep-resentativeness error into account. This had also theadvantage of avoiding instabilities in the data assimila-tion system in regions with very low absolute LASEerrors, which otherwise would produce a strong weightin the cost function. The low representativeness errorcan be justified for two reasons: (i) as DIAL data pro-vide a cross section through model grid boxes, the

JANUARY 2006 W U L F M E Y E R E T A L . 219

DIAL representativeness error is expected to be lowerthan for any pointwise measurement technique such asradio soundings, and (ii) to reduce furthermore thiserror source, we nearly matched the model grid size tothe horizontal resolution of the DIAL data. A newtechnique for estimating representativeness error hasrecently been proposed (Frehlich and Sharman 2004),and we will investigate the application of this techniqueto this problem along with the development of the fullcapability to directly assimilate water vapor numberdensity data.

The experiments presented in section 6 were per-formed with minimal changes to the standard MM54DVAR code but most of the development effort hasbeen devoted to the preprocessing of the DIAL datafiles. All the preprocessing functionalities (derivationof the water vapor mixing ratio and its error, time andspace subsampling, quality control, reformatting, etc.)have been integrated into a software referred as theLASE decoder. This program writes out the processedDIAL data in the input formats of the three currentlyexisting MM5 data assimilation systems: 4DVAR,3DVAR, and FDDA. This will allow comparisons be-tween the three algorithms in the future.

b. Setup of data assimilation runs and free forecasts

The LASE decoder and MM5 4DVAR softwarewere installed on the NEC supercomputer of the Ger-man Climate Computing Center (DKRZ) in Hamburg.So far, the assimilation was performed on one CPUonly, since the Open MP parallelization does not workwell with the 4DVAR system yet. In the future it isplanned to increase the speed of the assimilation usingthe capabilities of the message passing interface (MPI)parallelization. Assimilation and forecast were per-formed on different grids and with different physicsoptions. Two different sets of parameterizations wereused for the minimization run and the free forecasts.Since for the minimization of the cost function an ad-joint of the parameterization is necessary, we were lim-ited to schemes for which adjoints are already available.Furthermore minimization and free forecasts were per-formed with different horizontal and vertical resolu-tions.

In the data assimilation run, a horizontal resolutionof 30 km, a vertical resolution of 20 hybrid sigma levels,and a time step of 60 s were chosen in order to performdifferent sensitivity studies in a limited amount of time.There were 40 � 40 horizontal grid points, which rep-resents a domain of about 1200 km � 1200 km. Becausethe LASE profiles are available every 6 s, a medianfiltering was applied to generate one DIAL profilealong the flight path every 120 s matched to the hori-

zontal resolution of the model. Consequently, a DIALprofile was assimilated every other model time step.The DIAL observation operator included a bilineareight-point interpolation to compute the model mixingratio profile at the DIAL locations. In the vertical, the330-m DIAL measurements were interpolated withineach model vertical layer, to create a single data pointat the middle of the hybrid sigma level. The initial con-ditions and boundary forcings were provided using theECMWF operational analysis interpolated to the hori-zontal and vertical resolutions used in the MM5.

The data assimilation window extended from 1800 to2100 UTC on 24 May, close to the time where CI wasobserved in the satellite data. No attempts were madeto change the length of the window. The model physicsincluded a moist stable precipitation scheme, Anthes–Kuo convection parameterization, the NCEP medium-range forecast model (MRF) planetary boundary layer(PBL) scheme (Hong and Pan 1996), and a simple ra-diative cooling scheme without consideration of the ef-fects of cloud.

The free forecast produced by performing 4DVARof DIAL measurements is called ASSIM; the run start-ing from the original initial conditions created from theECMWF analysis is called NOASSIM. In the free fore-casts the parameterizations were changed to the NASAGoddard microphysics scheme including graupel andhail (Chen and Lamb 1994), the Kain–Fritsch 2 cumu-lus convection scheme (Kain 2004), and the improvedrapid radiative transfer scheme (RRTM) (Mlawer et al.1997). The remaining setup of the model is the same asthe assimilation run.

The final goal was to perform a high-resolution fore-cast with 3.3-km horizontal resolution, 36 hybrid verti-cal levels (220 � 220 � 36), and with a 6-s time step.This high resolution was chosen to resolve small-scaleconvective events as well as possible. However, the dataassimilation run could only be performed with a reso-lution of 40 � 40 � 20 grid points. Direct interpolationof this moderate-resolution run to the desired resolu-tion could cause spurious numerical effects in the simu-lations. Therefore, a medium-resolution model run with10-km horizontal resolution and 36 vertical levels wasinterposed (85 � 85 � 36 grid points). The results ofthis intermediate simulation were used to create theinitial conditions for the simulations at the final reso-lution of 3.3 km. The initial condition for the inter-mediate and high-resolution simulations were createdwith the one-way nesting approach using the MM5NESTDOWN utility. It allows in its latest version ahorizontal and vertical reinterpolation of the initialfields. In all cases, the model runs were limited to theshort range and amounted to 12 h. Consequently, 9 h of

220 M O N T H L Y W E A T H E R R E V I E W VOLUME 134

free forecasts were available, as these were countedstarting with the end of the data assimilation window.

6. Results

a. Comparison of LASE data with initial fields

Without any assimilation efforts, LASE data can beused to validate water vapor fields in the large-scaleanalyses. Figure 6 (middle panel) shows LASE mixingratio measurements using Eq. (3) in combination withpressure and temperature profiles extracted from thefree MM5 forecast. The middle panel also demon-strates the successful operation of the LASE decoderproducing mixing ratio profiles at the correct grid boxinterpolated to the vertical grid applied in the modelused for the data assimilation run. The LASE data arenot completely independent from model results. How-ever, the cross sensitivity of mixing ratio to temperatureand pressure is low. The upper and lower panels sur-rounding the LASE data are the corresponding mixingratio data extracted from ECMWF and Eta analyses,respectively, at the corresponding grid box and time by

interpolating between two time steps. The interpolationin time did not change the results significantly, so thatthe main findings are evident from Fig. 6. Both modelsproduced a moisture gradient in the free troposphere,which did not agree with LASE results. This may bedue to a too low vertical resolution used in the models.The locations of the dryline were fairly well predicted.However, the ECMWF analysis was too dry by about2 g kg�1 throughout the ABL. This is particularly vis-ible in the dry sector west of the dryline. In this region,the ABL depth was strongly underestimated by about500 m. The Eta analysis produced higher humidityin the ABL in the moist sector. However, the analysiswas even dryer than ECMWF west of the drylineand deviates strongly from the LASE data. Therefore,we decided to use ECMWF data for model initializa-tion.

b. Impact of 4DVAR on initial fields

After the performance of the data assimilation run,we compared its impact on the water vapor, tempera-ture, and wind fields. First of all, we investigated the

FIG. 6. Comparison of LASE water vapor profiles produced with (middle) the LASE decoder with (top) ECMWF analysis and(bottom) Eta analysis interpolated to the same grid box and time step. The LASE observation periods P1–P6 are also indicated.

JANUARY 2006 W U L F M E Y E R E T A L . 221

Fig 6 live 4/C

minimization of the cost function and the correspond-ing adaptation of the water vapor field along the LASEflight track. The results are presented in Figs. 7. It isevident that the 4DVAR analysis clearly improved thewater vapor field. In ASSIM, the humidity increasedthroughout the ABL. In the dry sector the boundarylayer got deeper in much better agreement with reality.

Also, the moisture gradient in the middle troposphereis better represented in ASSIM. Before application of4DVAR, the differences between NOASSIM andLASE data were of the order of �4–5 g kg�1, whichwere reduced to less than �0.5 g kg�1 except a smallregion at the return point of LASE at the end of P4 (seeFig. 5).

FIG. 7. Comparison of (a) LASE water vapor profiles with (b) NOASSIM and (c) ASSIM model outputs along the flight tracks. Thedifference between (d) ASSIM and NOASSIM, (e) LASE and NOASSIM, and (f) LASE results and ASSIM water vapor fields are alsoshown. Note the different color bars for absolute mixing ratio and their differences.

222 M O N T H L Y W E A T H E R R E V I E W VOLUME 134

Fig 7 live 4/C

In Fig. 8, the impact of 4DVAR on the water vaporfield is presented in horizontal planes at four differentheight levels. The differences of mixing ratios betweenASSIM and NOASSIM demonstrate a strong modifi-cation of the water vapor field in the northwestern partof the model domain not only where LASE was flyingbut also south of this region. Close to the ground, in theTexas panhandle and in midwest Oklahoma an en-hancement of humidity by more than 3 g kg�1 was pro-duced, reducing the moisture gradient in the region ofthe dryline south of the triple point. In the other layers,an impact is visible extending farther south of the re-gion were LASE was flying. Close to the ABL top at1500 m, the moisture gradient decreased in the regionof the triple point but increased farther south along thedryline. A dryline of the order of 0.5 g kg�1 was foundin the middle troposphere, which affected the potentialof CI.

We also investigated the resulting differences in theinitial surface temperature and wind fields (not shown).The main difference in the temperature fields wasfound in the middle troposphere west of the drylinewhere lower temperatures were found in ASSIM. Thiswas accompanied by two low-level jets, which occurredsouth of the cold front west of the dryline as well as eastof the dryline in the midwest of Oklahoma. This causeda considerably stronger convergence south of the triplepoint at the southwestern corner of Oklahoma and in-creased the potential of CI in this region.

c. Impact on the time evolution of atmosphericfields

Until 0000 UTC 25 May, close to the surface, thehorizontal water vapor gradient was eroded along thedryline in both ASSIM and NOASSIM. The gradientwas reduced further because of developing convectionand corresponding outflow boundaries along thedryline. The latter tendency was also observed inWakimoto et al. (2006), as they found it difficult todetect the dryline in some locations after convectionwas initiated.

Figure 9 presents the differences of the water vaporfields between ASSIM and NOASSIM at four verticallevels at 0000 UTC. At all heights, the region influ-enced by the LASE data spread out and was clearlyvisible in the free forecasts up to 9 h (not shown). Thiswas due to a different development of CI in ASSIM andNOASSIM resulting in a strong redistribution of watervapor up to the middle troposphere.

We focus on the difference in the moisture and windfields at the time when CI took place. For example,Figs. 10 and 11 show the ASSIM and NOASSIM watervapor and wind fields, respectively, at four different

height levels. The moisture tongue at 36°N, 100°W, re-ducing the west-east moisture gradient in the region ofthe dryline and the triple point is enhanced in ASSIM(see Figs. 10 and 11, upper panels). This feature wasconnected to enhanced horizontal transport of watervapor around the low-level low, resulting in a strongerconvergence and moisture gradient in the ABL at thesouthwestern corner of Oklahoma. In ASSIM, the re-distribution of the wind field also resulted in a reducedconvergence in the region of the triple point.

This resulted in different developments of CI, whichare visible in the lower panels of Figs. 10 and 11. Usingwater vapor as tracer for vertical transport, we found inboth cases that convection was initiated at the south-western corner of Oklahoma. However, convection wasstronger and located more in the south in the ASSIMrun. More importantly, convection around the triplepoint was strongly suppressed in ASSIM in contrast toNOASSIM. We also studied these effects in detail usingplots of vertical velocity and convergence, but as theseshowed similar structures as in the upper-troposphericwater vapor field, the corresponding figures are notshown here.

Though CI took place at about the same time (1900UTC), convection evolved differently in ASSIM andNOASSIM. In NOASSIM, the regions merged whereconvection was initiated (see at 5500 m; Fig. 11) andformed a straight thin line from the southwestern cor-ner of Oklahoma to the region of the triple point. Asquall line with an angle of about 30° developed andmoved eastward through Oklahoma. In contrast, inASSIM, this line of convection was broken (see at5500 m; Fig. 10) and convection was enhanced along thedryline in the southwestern corner of Oklahoma. At thebeginning of the development, this region of CI spreadout northeastward but not as far as in NOASSIM. Thisis in good agreement with satellite observations.

Satellite and radar data showed that CI took placetoo early in ASSIM and NOASSIM by about 1 h. Con-vection was initiated very close to the modeled regionin ASSIM but did not take place in the region of thetriple point. Therefore, the suppression of CI at thetriple point in ASSIM was in much better agreementwith observations. This finding also suggests differencesin the precipitation fields.

d. Impact on the precipitation fields

We also investigated the influence of the assimilationof accurate and high-resolution water vapor data onQPF. Therefore, we calculated the precipitation accu-mulated within half an hour with (ASSIM) and without(NOASSIM) data assimilation. To investigate the ac-curacy of the temporal/spatial distribution of the fore-

JANUARY 2006 W U L F M E Y E R E T A L . 223

FIG

.8.H

oriz

onta

lcr

oss

sect

ions

ofth

edi

ffer

ence

sin

the

init

ial

fiel

dsof

mix

ing

rati

oin

gkg

�1

mod

eled

wit

h(A

SSIM

)an

dw

itho

utL

ASE

data

assi

mila

tion

(NO

ASS

IM)

inte

rpol

ated

to3.

3km

ofth

ehi

gh-r

esol

utio

nru

n.

224 M O N T H L Y W E A T H E R R E V I E W VOLUME 134

Fig 8 live 4/C

FIG

.9.S

ame

asF

ig.8

but

for

the

3-h

free

fore

cast

at00

00U

TC

25M

ay20

02.

JANUARY 2006 W U L F M E Y E R E T A L . 225

Fig 9 live 4/C

FIG

.10.

Hig

h-re

solu

tion

ASS

IMm

ixin

gra

tio

fiel

dsin

gkg

�1

wit

hw

ind

vect

ors

inm

s�1

atfo

urhe

ight

leve

lsat

2000

UT

Cw

here

init

iati

onof

conv

ecti

onoc

curr

ed.

226 M O N T H L Y W E A T H E R R E V I E W VOLUME 134

Fig 10 live 4/C

FIG

.11.

Sam

eas

Fig

.10

but

for

NO

ASS

IM.

JANUARY 2006 W U L F M E Y E R E T A L . 227

Fig 11 live 4/C

casted precipitation fields, the data are compared quali-tatively with the corresponding radar reflectivities mea-sured with the NEXRAD radar in Frederick,Oklahoma. We omitted a comparison with simulatedradar reflectivity using a model forward operator be-cause this does still not allow a quantitative comparisonof the results due to large uncertainties in model cloudand precipitation microphysics.

It has to be pointed out that because of the largechain of complicated processes involved, a perfectagreement between the forecasted precipitation andobservations cannot be expected. Furthermore, a posi-tive impact should only be visible in the regions af-fected by improved representations of initial fields,which were advected in the region of interest during themodel forecast time.

Deep convection developed around 2000 UTC andproduced precipitation at about 2038 UTC in a 20° linespreading out from the southeastern corner of Oklaho-ma to middle Oklahoma until 2300 UTC. Both ASSIMand NOASSIM deviate from observations insofar asthey show precipitation in the area of the triple point.However, as explained in section 6c, the developmentof CI and thus of precipitation is better modeled in theASSIM run, as convection south of the triple point issuppressed.

During the next hours in ASSIM, convection gotstronger in middle Oklahoma so that the precipitationpattern became similar in this region. However, at theborder between Texas and Oklahoma, the precipitationpattern remained different. The ASSIM precipitationpattern is in better agreement with the observation.Just like the radar image, it shows an almost perpen-dicular angle with respect to the borderline, which isclearly not the case for the NOASSIM pattern. Furthercomparisons between ASSIM and NOASSIM as well asbetween different data assimilation techniques will bethe subject of future research.

e. Summary and outlook

In this study, we performed the first 4DVAR assimi-lation of water vapor DIAL data. A case fromIHOP_2002 was selected where LASE water vapordata were available upstream of a region that showedCI. Consequently, a major impact on the onset anddevelopment of convection and of precipitation in ashort-range forecast could be expected.

A new tool, the LASE decoder, was developed forlidar data assimilation. The LASE decoder ingests bothmeasurements and error estimates from DIAL datafiles, automatically performs the vertical averaging andtime subsampling that best suit MM5 forecasting grids,and writes out DIAL observation and error files at the

MM5 4DVAR, 3DVAR, and FDDA input formats.This tool is very versatile and can be routinely appliedto other case studies.

A method for determining the 2D distribution of theDIAL noise errors was presented. These data wereused to construct the error covariance matrix includingthe vertical and horizontal weighting functions of theDIAL retrieval. This procedure takes advantage of theproperties of the DIAL technique of providing realisticnoise errors for each profile in real time without anyother sources of information. Another advantage of theDIAL technique is that bias correction was consideredunnecessary due to the high accuracy of the measure-ments. These properties make the DIAL method well-suited for future real-time data assimilation projects,for instance, within THORPEX Regional Campaignsor the upcoming COPS.

Our results also demonstrate that airborne water va-por DIAL is a powerful tool for model validation. Evenusing 4DVAR in large-scale analyses, significant defi-ciencies were still detected in the representation of thetropospheric water vapor field.

A significant impact of the DIAL data on the initialwater vapor field was found. The assimilation of LASEdata resulted in an erosion of the moisture gradientalong the dryline south of the triple point and in en-hanced southwesterly moisture transport southwest ofthe triple point. Moisture convergence was enhanced150 km south of the triple point along the dryline andinitiated CI at a location where it was indeed observed.The CI was reduced in the region of the triple point inbetter agreement with radar images leading to an im-proved modeling of the precipitation field. Up to 4 hafter start of the free forecast, in the region of the bor-der between Oklahoma and Texas, lidar data assimila-tion had a positive impact on the precipitation forecast.

In connection with this case, we are strongly moti-vated to perform further detailed validations of watervapor fields, using other DIAL data, and of precipita-tion fields, by comparison with gridded composites ofradar and rain gauge data. We plan to investigate andcompare the impact of different data assimilation tech-niques such as 3DVAR and FDDA. Studies will beperformed where the length of the data assimilationwindow and the resolution of the data assimilation runswill be varied.

We will extend our preprocessing capabilities to in-clude Doppler lidar wind, GPS, and dropsonde mea-surements, so that our study can take full advantage ofthe many diverse instruments deployed during theIHOP_2002 or future field campaigns.

Though this is a case study, the results indicate thatthere can be large errors in water vapor fields used for

228 M O N T H L Y W E A T H E R R E V I E W VOLUME 134

initializing mesoscale models. Particularly in regionswith inhomogeneous distributions of water vapor suchas close to the dryline, absolute errors in water vaporand misplacements of humidity gradients can result inlarge errors of QPF. We expect similar results in re-gions with significant orography. In the future, it is im-portant to perform more impact studies in order to ex-plore the reasons for the decay of the impact of watervapor data assimilation. Particularly interesting is thesimultaneous assimilation of wind and water vapor dataor a comparison of their impacts. Notwithstanding therestricted impact of lidar 4DVAR on precipitation fore-cast, our study demonstrates a clear positive impact onCI. Thus it confirms the basic scientific hypothesis ofIHOP_2002.

Improved initial fields based on data assimilation canbe used not only for better prediction of QPF but alsofor reanalyses, which are capable of identifying otherremaining errors, for instance, those caused by param-eterizations. Without improved initial fields, it will bevery difficult to separate corresponding errors. Theseefforts shall be intensified in upcoming field campaignssuch as COPS. They will also be coordinated with re-search programs like THORPEX and within the U.S.Weather Research Program (USWRP; Fritsch and Car-bone 2004). Consequently, the IHOP_2002 field cam-paign and its data can be considered as a unique start-ing point for improving QPF in the future.

Acknowledgments. We thank JOSS and all IHOP_2002scientists who contributed to the great IHOP_2002 fieldcatalogue. Much information about the mission and theweather situation was extracted from this source. Weappreciate the access to ECMWF analysis data. Wethank Tony Notari, George Insley, Jerry Williams, andNick Kepics for their support of LASE in IHOP_2002.We acknowledge the help of Susan Kooi concerning thecalculation of the LASE 2D noise error field. TheNASA Earth Science Research Division provided fi-nancial support for the deployment of LASE duringIHOP.

REFERENCES

Ansmann, A., and J. Bösenberg, 1987: Correction scheme forspectral broadening by Rayleigh scattering in differential ab-sorption lidar measurements of water vapor in the tropo-sphere. Appl. Opt., 26, 3026–3032.

Arnaud, P., C. Bouvier, L. Cisneros, and R. Dominguez, 2002:Influence of rainfall spatial variability on flood prediction. J.Hydrol., 260, 216–230.

Barker, D. M., W. Huang, Y.-R. Guo, and Q. Xiao, 2004: A three-dimensional variational (3DVAR) data assimilation systemfor use with MM5: Implementation and initial results. Mon.Wea. Rev., 132, 897–914.

Bauer, H., H.-S. Bauer, V. Wulfmeyer, M. Wirth, B. Mayer, G.Ehret, D. Summa, and P. Di Girolamo, 2004: End-to-endsimulation of the performance of WALES: Forward module.Proc. 22d Int. Laser Radar Conf., ESA SP-561, Matera, Italy,ESA, 1011–1014.

Bösenberg, J., 1998: Ground-based differential absorption lidarfor water-vapor profiling: Methodology. Appl. Opt., 37, 3845–3860.

Bougeault, P., and Coauthors, 2001: The MAP Special ObservingPeriod. Bull. Amer. Meteor. Soc., 82, 433–462.

Browell, E. V., and S. Ismail, 1995: First lidar measurements ofwater vapor and aerosols from a high-altitude aircraft. Proc.OSA Optical Remote Sensing of the Atmosphere, Salt LakeCity, UT, OSA, 212–214.

——, T. D. Wilkerson, and T. J. McIlrath, 1979: Water vapor dif-ferential absorption lidar development and evaluation. Appl.Opt., 18, 3474–3483.

——, and Coauthors, 1997: LASE validation experiment. Ad-vances in Atmospheric Remote Sensing with Lidar, A. Ans-mann et al., Eds., Spinger-Verlag, 289–295.

——, S. Ismail, and W. B. Grant, 1998: Differential absorptionlidar (DIAL) measurements from air and space. Appl. Phys.,67B, 399–410.

Chen, J.-P., and D. Lamb, 1994: Simulation of cloud microphysicaland chemical processes using a multicomponent framework.Part I: Description of the microphysical model. J. Atmos. Sci.,51, 2613–2630.

Davis, C., and F. Carr, 2000: Summary of the 1998 workshop onmesoscale model verification. Bull. Amer. Meteor. Soc., 81,809–819.

Di Girolamo, P., and Coauthors, 2004: Simulation of the perfor-mance of WALES based on an end-to-end model. Proc. 22dInt. Laser Radar Conf., ESA SP-561, Matera, Italy, ESA,957–960.

Doms, G., A. Gassmann, E. Heise, M. Raschendorfer, C. Schraff,and R. Schrodin, 2002: Parameterization issues in the non-hydrostatic NWP-model LM. Proc. Seminar on Key Issues inthe Parameterization of Subgrid Physical Processes, Reading,United Kingdom, ECMWF, 205–252.

Ebert, E. E., U. Damrath, W. Wergen, and M. E. Baldwin, 2003:The WGNE assessment of short-term quantitative precipita-tion forecast. Bull. Amer. Meteor. Soc., 84, 481–492.

Frehlich, R., and R. Sharman, 2004: Estimates of turbulence fromnumerical weather prediction model output with applicationsto turbulence diagnosis and data assimilation. Mon. Wea.Rev., 132, 2308–2324.

Fritsch, M., and R. Carbone, 2004: Improving quantitative pre-cipitation forecasts in the warm season: A USWRP researchand development strategy. Bull. Amer. Meteor. Soc., 85, 955–965.

Geerts, B., R. Damiani, and S. Haimov, 2006: Finescale verticalstructure of a cold front as revealed by an airborne Dopplerradar. Mon. Wea. Rev., 134, 251–271.

Grell, G. A., J. Dudhia, and D. R. Stauffer, 1995: A description ofthe fifth-generation Penn State/NCAR Mesoscale Model(MM5). NCAR Tech. Note NCAR/TN-398�STR, 122 pp.[Available from UCAR Communications, P.O. Box 3000,Boulder, CO, 80307.]

Hong, S.-Y., and H.-L. Pan, 1996: Nonlocal boundary layer ver-tical diffusion in a medium-range forecast model. Mon. Wea.Rev., 124, 2322–2339.

Ismail, S., and E. V. Browell, 1989: Airborne and spaceborne lidar

JANUARY 2006 W U L F M E Y E R E T A L . 229

measurements of water vapor profiles: A sensitivity analysis.Appl. Opt., 28, 3603–3614.

Joiner, J., and A. M. Da Silva, 1998: Efficient methods to assimi-late remotely sensed data based on information content.Quart. J. Roy. Meteor. Soc., 124A, 1669–1694.

Kain, J. S., 2004: The Kain–Fritsch convective parameterization:An update. J. Appl. Meteor., 43, 170–181.

Kamineni, R., T. N. Krishnamurti, R. A. Ferrare, S. Ismail, andE. V. Browell, 2003: Impact of high resolution water vaporcross-sectional data on hurricane forecasting. Geophys. Res.Lett., 30, 1234, doi:10.1029/2002GL016741.

——, ——, S. Pattnaik, E. V. Browell, S. Ismail, and R. A. Fer-rare, 2006: Impact of CAMEX-4 datasets for hurricane fore-casts using a global model. J. Atmos. Sci., 63, 151–174.

Lenschow, D., V. Wulfmeyer, and C. Senff, 2000: Measuring sec-ond- through fourth-order moments in noisy data. J. Atmos.Oceanic Technol., 17, 1330–1347.

Martin, W. J., and M. Xue, 2006: Sensitivity analysis of convectionof the 24 May 2002 IHOP case using very large ensembles.Mon. Wea. Rev., 134, 192–207.

Mass, C. F., D. Ovens, K. Westrick, and B. A. Colle, 2002: Doesincreasing horizontal resolution produce more skillful fore-casts? Bull. Amer. Meteor. Soc., 83, 407–430.

Mlawer, E. J., S. J. Taubman, P. D. Brown, M. J. Iacono, and S. A.Clough, 1997: Radiative transfer for inhomogeneous atmo-spheres: RRTM, a validated correlated k-model for the long-wave. J. Geophys. Res., 102, 16 663–16 682.

Moore, A. S., Jr., and Coauthors, 1996: Development of the LidarAtmospheric Sensing Experiment (LASE)—An advancedairborne DIAL instrument. Advances in Atmospheric Re-mote Sensing with Lidar, A. Ansmann et al., Eds., Spinger-Verlag, 281–288.

Parrish, D. F., and J. C. Derber, 1992: The National Meteorologi-cal Center’s spectral statistical–interpolation analysis system.Mon. Wea. Rev., 120, 1747–1763.

Ruggiero, F. H., G. D. Modica, T. Nehrkorn, M. Cerniglia, J.Michalakes, and X. Zou, 2001: Development of an MM5-based four dimensional variational analysis system for dis-tributed memory multiprocessor computers. Preprints,HPCMP 2001 User’s Group Conf., Biloxi, MS, U.S. NavalOceanographic Office.

Stauffer, D. R., and N. L. Seaman, 1990: Use of four-dimensionaldata assimilation in a limited-area mesoscale model. Part I:Experiments with synoptic-scale data. Mon. Wea. Rev., 118,1250–1277.

——, and ——, 1994: Multiscale four-dimensional data assimila-tion. J. Appl. Meteor., 33, 416–434.

Wakimoto, R. M., H. V. Murphey, E. V. Browell, and S. Ismail,2006: The “triple point” on 24 May 2002 during IHOP. PartI: Airborne Doppler and LASE analyses of the frontalboundaries and convection initiation. Mon. Wea. Rev., 134,231–250.

Weckwerth, T. M., and Coauthors, 2004: An overview of the In-ternational H2O Project (IHOP_2002) and some preliminaryhighlights. Bull. Amer. Meteor. Soc., 85, 253–277.

Wulfmeyer, V., 1999: Investigation of turbulent processes in thelower troposphere with water-vapor DIAL and radar–RASS.J. Atmos. Sci., 56, 1055–1076.

——, and J. Bösenberg, 1998: Ground-based differential absorp-tion lidar for water-vapor profiling: Assessment of accuracy,resolution, and meteorological applications. Appl. Opt., 37,3825–3844.

——, and C. Walther, 2001a: Future performance of ground-basedand airborne water vapor differential absorption lidar. I:Overview and theory. Appl. Opt., 40, 5304–5320.

——, and ——, 2001b: Future performance of ground-based andairborne water vapor differential absorption lidar. II: Simu-lations of the precision of a near-infrared, high-power system.Appl. Opt., 40, 5321–5336.

——, H.-S. Bauer, A. Behrendt, F. Vandenberghe, and E. V.Browell, 2004: Assimilation of DIAL data in mesoscale mod-els: An impact study during IHOP_2002. Proc. 22d Int. LaserRadar Conf., ESA SP-561, Matera, Italy, ESA, 639–642.

Xiao, Q., X. Zou, and B. Wang, 2000: Initialization and simulationof a landfalling hurricane using a variational bogus data as-similation scheme. Mon. Wea. Rev., 128, 2252–2269.

Xue, M., and W. J. Martin, 2006a: A high-resolution modelingstudy of the 24 May 2002 dryline case during IHOP. Part I:Numerical simulation and general evolution of the drylineand convection. Mon. Wea. Rev., 134, 149–171.

——, and ——, 2006b: A high-resolution modeling study of the 24May 2002 dryline case during IHOP. Part II: Horizontal con-vective rolls and convective initiation. Mon. Wea. Rev., 134,172–191.

Zängl, G., 2004a: The sensitivity of simulated orographic precipi-tation to model components other than cloud microphysics.Quart. J. Roy. Meteor. Soc., 130, 1857–1875.

——, 2004b: Numerical simulations of the 12–13 August 2002flooding event in eastern Germany. Quart. J. Roy. Meteor.Soc., 130, 1921–1940.

Zou, X., Y.-H. Kuo, and Y.-R. Guo, 1995: Assimilation of atmo-spheric radio refractivity using a nonhydrostatic adjointmodel. Mon. Wea. Rev., 123, 2229–2249.

230 M O N T H L Y W E A T H E R R E V I E W VOLUME 134