Experimental results on rainfall estimation in complex terrain with a mobile X-band polarimetric...

Atmospheric Research 94 (2009) 579–595

Contents lists available at ScienceDirect

Atmospheric Research

j ourna l homepage: www.e lsev ie r.com/ locate /atmos

Experimental results on rainfall estimation in complex terrain with a mobileX-band polarimetric weather radar

Marios N. Anagnostou a,⁎, John Kalogiros b,Emmanouil N. Anagnostou a,c, Anastasios Papadopoulos a

a Department of Inland Waters, Hellenic Center for Marine Research (HCMR), Greeceb Institute for Environmental Research and Sustainable Development, National Observatory of Athens, Greecec Department of Civil and Environmental Engineering, University of Connecticut, Storrs, Connecticut, United States

a r t i c l e i n f o

⁎ Corresponding author. Hellenic Centre for MarineInland Waters, 46.7 km Athens-Sounion Ave., P.O.Anavissos Attikis, Greece.

E-mail address: [email protected] (M.N. A

0169-8095/$ – see front matter © 2009 Elsevier B.V.doi:10.1016/j.atmosres.2009.07.009

a b s t r a c t

Article history:Received 23 March 2009Received in revised form 24 June 2009Accepted 18 July 2009

In this work the capability of reliable rainfall measurements with small weather radars incomplex terrain for flood forecasting purposes is examined. Rain measurements were carriedout during winter–spring 2007 with a mobile X-band dual-polarization radar in thenorthwestern mountainous part of the island of Crete in Greece. In this area a 2D-videodisdrometer and a network of raingauges was installed for radar calibration and evaluation ofrainfall measurements, respectively. The complex terrain of the experimental site maysignificantly reduce the performance of rain measurements due to ground clutter and partial ortotal beam blockage. A beam blockage algorithm using high resolution terrain data was appliedin order to find areas with significant terrain effects and estimate correction of the radarmeasurements. Rain attenuation correction was based on modern sophisticated algorithmsusing differential phase measurements. The accuracy of rainfall estimation from standard orpolarimetric algorithms at plan position indicator (PPI) scans was examined for high-temporalresolution (1 min) rainfall rates and accumulated rainfall values for winter and spring time rainevents. For high elevation measurements, which were required in order to avoid terrain effectsin large areas of interest, a correction for the vertical-profile-of-reflectivity (VPR) was alsoapplied to the radar data. An average VPR model used in the corresponding correction ofreflectivity was constructed based on range–height indicator (RHI) scans. It was concluded thatquantitative high resolution X-band radar observations of rainfall in complex terrain is possibleafter careful application of all the above processing steps.

© 2009 Elsevier B.V. All rights reserved.

Keywords:RainfallX-bandDual-polarizationVertical profile of reflectivityComplex terrain

1. Introduction

X-band (λ~3 cm) polarimetric weather radar systemsprovide a cost effective solution for high-resolution precipita-tion observations due to their small size and low demand intransmitter power. The introduction of polarimetric approachesin X-band systems can now account for the attenuation anddifferential attenuation effects (Testud et al., 2000; Matrosovet al., 2002, 2005; Anagnostou et al., 2006a,b, 2007) which used

Research—Institute ofBOX 712, GR-19013,

nagnostou).

All rights reserved.

to be a significant drawback for the use ofX-band frequencies inquantitative precipitation estimation (QPE). Such estimatesfrom X-band dual polarization radar systems can be used to fillin spatial coverage gaps of the large low-frequency (C-/S-band)operational weather radars and to provide high-resolutionrainfall observation over small scale mountainous basins andurban areas. Measurements provided by dual-polarizationobservations in addition to the horizontal polarization reflec-tivity (ZH) provided by standard single polarization systems arethe differential reflectivity, (ZDR), i.e. the ratio of horizontal tovertical polarization reflectivity, the differential propagationphase shift (ΦDP) and copolar correlation coefficient (ρhv).

In addition, X-band weather radars are more suitable forsmall range flood studies in complex terrain. However, the

mailto:[email protected]

http://dx.doi.org/10.1016/j.atmosres.2009.07.009

http://www.sciencedirect.com/science/journal/01698095

580 M.N. Anagnostou et al. / Atmospheric Research 94 (2009) 579–595

deployment of operational weather radars in mountainousterrain introduces issues as a result of the partial or completebeam blockage of the ray due to the ground clutter at lowradar scan elevations. For QPE in complex terrain, radar rainmeasurements are taken at higher elevation scans, which arerelated to rapid increase of radar beam height with distance,in order to avoid either ground clutter or blockage from theterrain. This observational geometry in combination with lowheight of the freezing level causes significant errors in rainfallestimation because the radar resolution volumes are probablylocated in the melting layer or snow regions (Kitchen andJackson, 1993) even if it rains on the ground. The beaminterception through precipitation melting regions is oftenassociated with an enhancement of the reflectivity. Theprimary cause of this reflectivity enhancement (bright band,BB) is the rapid increase in the dielectric constant ofhydrometeors at the top of the melting layer followed by anincrease of the fall velocities of melting snowflakes towardthe end of the melting process (Fabry and Zawadzki, 1995).

In this paper, we present an evaluation of availableprocessing methods for X-band polarimetric radar overcomplex terrain. The difference from previous studies isthat we evaluate the combination of all the processing steps(beam blockage, rain attenuation, rainfall algorithm, verticalprofile of reflectivity-VPR) for an X-band weather radar inan integrated fashion for a difficult case due to terrain limita-tions. In Section 2 we discuss a standard and two advancedX-band polarimetric algorithms for rainfall estimation, whilein Section 3 we discuss the study area and data acquired withthe X-band mobile radar (XPOL) of the National Observatoryof Athens during two major rain events, a stratiform eventduring winter (04/01/2007) and a more convective eventduring spring time (22/03/2007). In Section 4we qualitativelycompare the rainfall algorithms for the lowest elevatedgauges, close to the coast. In Section 5we identify the problemof estimating rain in high elevation, we introduce the alreadyavailable VPR correction algorithm and we show the resultsfor the highland gauges. Finally, Section 6 contains the con-clusions of this work.

2. The XPOL algorithm

2.1. Polarimetric radar variables

As stated in the introduction the polarimetric radar param-eters used in quantitative rain estimation are the horizontalpolarization reflectivity, ZH (mm6m−3), vertical polarizationreflectivity (ZV, mm6m−3) and differential reflectivity, ZDR(dB) and the specific differential propagation phase shift, KDP(° km−1). These variables depend on the raindrop size dis-tribution, DSD, and the drop scattering amplitudes as follows(Brandes et al., 2004):

ZH;V =4λ4

π4 jKw j2 ∫Dmin

Dmax j fHH;VVðDÞ j2NðDÞdD ð1Þ

ZDR = 10logZHZV

� �ð2Þ

KDP =180λπ

∫Dmin

DmaxRe½ fHHð0;DÞ−fVV ð0;DÞ�NðDÞdD ð3Þ

where D (mm) is the raindrop sphere equivalent volumediameter and Dmin and Dmax are the diameters of smallest andlargest drops in the distribution. The fHH,VV(D) and fHH,VV(0,D)are the backscattering and the forward scattering amplitudes ofa drop at horizontal and vertical polarization, respectively,Kw isthe dielectric factor of water, λ (cm) is the radar wavelength,and N(D) (mm−1m−3) is the count of raindrop of size D perunit volume. The fHH,VV(D) and fHH,VV(0,D) parameters dependon the assumed raindrop shape-size relationship as discussedin subsequent section. The horizontal polarization reflectivity,ZmH (mm6m−3), and differential reflectivity, ZmDR (dB), mea-sured by the radar at range gate “r”, are related to the cor-responding intrinsic (non-attenuated) radar parameters (ZHand ZDR) as follows:

ZmHðrÞ = ZHðrÞ × 10−0:2∫r0AHðsÞds ð4Þ

ZmDRðrÞ = ZDRðrÞ × 10−0:2∫r0ADP ðsÞds: ð5Þ

AH and ADP (dB km−1) are the co-polar specific anddifferential rain-path attenuation, respectively. Quantitativeuse of X-band radar data in estimation of precipitation para-meters (rainfall rates,water content, raindrop sizedistribution)requires knowledge of the non-attenuated radar parameters(i.e., ZH and ZDR), which is a subject discussed next.

2.2. Rain attenuation correction

The fundamental aspect that brought X-band back to theinterest of hydrometeorologists for rainfall estimation is thatthe horizontal versus vertical polarization differential phaseshiftΦDPmeasurement can beused as a constrainingparameterfor the effective estimation of specific copolar, AH, anddifferential, ADP, attenuation profiles (e.g., Testud et al., 2000;Vulpiani et al., 2005; Matrosov et al., 2002; Anagnostou et al.,2006a,b; Park et al., 2005a). As shown by Anagnostou et al.,(2006a,b), this aspect minimizes the uncertainty due to rainpath attenuation at X-band due to the fact that ΦDP is not apower related parameter (provided that backscattering signalsare above the minimum detectable level) and it is almostlinearly relatedwith the range integrated co-polar attenuation,expressed in dB. Once the AH range profile is estimated bymeans of a rain path attenuation ΦDP-constrained technique,ADP can then be retrieved directly fromAH given that AH and ADPare almost linearly related (i.e., AH ∝ ADP).

The attenuation correctionmethod used in this study is ZPHI(Testud et al., 2000), which is based on ΦDP measurements,combined with a ΦDP–ZDR constraint self consistent method(Bringi et al., 2001). This method was applied to each ray inseparate rain cells defined by correlation between polarizationchannels higher than 0.8. In addition, each rain cell along a radarraywas further checked for sub-cells in order to avoid thewrongdetection of possibly merged neighboring cells using only theρHV correlation limit. This would affect the attenuationcorrection method because rain cells may have different rainmicrophysical parameters and, thus, different parameters ofsignal attenuation. The sub-cellswere detected from significantpeaks ofKDP, whichwere identified as a factor of 2 change ofKDPbetween a local maximum (peak) and adjacent minima.Neighboring sub-cells with too short (3 km length limit) or too

581M.N. Anagnostou et al. / Atmospheric Research 94 (2009) 579–595

smallΦDP change (6° limit)weremerged in order to be possibleto apply the attenuation correction method to them too.

In the case of ZDR constraint the difference of ZDR from thebeginning of the rain cell was used in the fitting procedureinstead of the actual value of ZDR (Bringi et al., 2001; Park et al.,2005a; Anagnostou et al., 2006a,b) to make the method fullyindependent of radar calibration offsets. Also, in order to avoidpossible backscattering phase shift effects (otherwise called “δ”effect, see Hubbert and Bringi 1995) within rain cells we used aleast median error (outliers rejection) instead of least squareerror criterion in the fitting of theoretical to measured ΦDP inthe modified ZPHI algorithm. Alternatively, we could use afiltered version of ΦDP according Hubbert and Bringi (1995)iterative algorithm, but this method smoothed too much theΦDP measurements. Finally, KDP was estimated from thegradient of a 3 km in length Savitzky-Golay linear (first orderpolynomial) filter ofΦDPmeasurements. Measurements ofΦDP

with significant error (noisy outliers) were first removed bycomparison with a running median filter. With a 3° randomerror ofΦDP measurements, which was estimated from 3 (km)length high pass filter, and 150 (meters) long measurementgates the resulting KDP error is about 0.4 (° km−1). In addition,KDP is inversely proportional to radar wavelength and KDPvalues are higher in X-band radars, which means that therelative KDP error is smaller, compared to S or C-band radars.

Fig. 1 (a–c) shows an example of PPI plots of XPOLobserved (raw) horizontal polarization reflectivity ZH(Fig. 1a), the corrected for attenuation (Fig. 1b) ZH and thedifferential phase shift ΦDP in Fig. 1c. In the southwest regionof the domain (in sectors 210°–240° and 330°–360°) we notesignificant differential phase shift and consequently, en-hanced specific attenuation with an average attenuationcoefficient of 0.24 (dB/°) In X-band these phase changescorrespond to 6 and 4 (dB) attenuation. In addition theattenuation coefficient is variable and it is estimated in everyrain cell according to the modified ZPHI algorithm describedabove. This is the reason of the appearance of ray lines withincreased reflectivity in the attenuation corrected ZH. Thedifference between raw and attenuation corrected ZH in Fig. 1is justified in X-band radars even thoughΦDP changes are notbig in some sectors (Anagnostou et al., 2006a,b).

Fig. 1. Attenuation correction of PPI scan based

2.3. Rainfall estimation algorithms

Rainfall is a stochastic process that varies both in space andtime thus making its accurate estimation at small scale anextremely difficult task. Since radar does not measure rainfalldirectly but rather we relate the measured variables (e.g.,horizontal polarization reflectivity, ZH) with rainfall properties,accurate estimation of rainfall requires amodel thatwould bestdescribe those physical properties of rain. For decades, rainfallestimateswere derived froma single radar parameter— i.e., theequivalent radar reflectivity factor (Ze) (e.g., Atlas and Ulbrich1990; Joss andWaldvogel 1990). In order to better understandthe characteristics of rainfall, research in the advances of radartechnology has turned to dual-polarization systems whichprovide additional information on the physical parameters ofrainfall. In the present study, we discuss three radar-rainfallestimation techniques developed for (XPOL) X-band dual-polarization radar measurements.

The algorithms are based on power–law relationshipsdetermined on the basis of least squares fitting to rain rate,R, and radar parameters ZH, ZDR and KDP computed from2D-video (2DVD) disdrometer raindrop spectra measure-ments using a 1-min averaging window. These relationshipswere obtained on the basis of T-matrix scattering simula-tions (Barber and Yeh, 1975; Ishimari 1991) at a 10 °C airtemperature, which included a “normalized” Gamma dis-tribution (Bringi and Chandrasekar, 2001), an assumedlinear drop axis ratio model (Pruppacher and Beard, 1970;Matrosov et al., 2002)with a Gaussianmodel with zeromeanand 5° standard deviation for droplets canting angle (Bringiet al., 2003) and a broad range of DSD parameters accordingto the disdrometermeasurements collected in the urban areaof Athens, Greece in year 2006 (the number of good rainevents with rain accumulationmore than 10 mmwas 13 andtheir average time duration of continuous rain was one day)due to insufficient number of current DSD observations(technical failure of 2DVD after of 23rd of December) fromCrete. The dielectric constant of the water was evaluated foran average atmospheric temperature (10 °C). Fig. 2 (a–c)shows a frequency plot of the three governing parameters(D0, NW and μ) that characterize the normalized Gamma

on ΦDP in rain below the melting layer.

Fig. 2. Frequency plots of (a) D0 (mm), (b) log10NW (NW in mm−1m−3) and (c) μ (unitless), parameters of the “normalized” Gamma distribution determinedbased on 2DVD measurements from Athens and Crete in 2006.


distribution model. DSD parameters and rainfall estimationfrom radar data have been compared to the results ofdisdrometer placed at a distance of 10 km from the radar(Kalogiros et al., 2006) in Athens area during 2006 forstratiform and convective events and they have shown quitegood agreement. Parameterizations of DSD parametersbased on polarimetric radar data (ZH, ZDR and KDP) havebeen estimated from disdrometer data (for the definition ofthe population of DSD parameters) and T-matrix calcula-tions. Even though these intermediate parameterizations ofDSD parameters are used in the rainfall estimators they arefunctions of polarimetric radar data. In this paper we don'texamine the accuracy of DSD parameters estimations but thereliability of the final relationships for rainfall estimation.Below we discuss the rainfall algorithms derived in thisstudy.

The named here as the standard or classic algorithm, is theone that uses the horizontal polarization reflectivity (ZH) asinput and it is devised as follows:

RSTD = 3:36 × 10−2Z0:58Hmm ð6Þ

where,RSTDunits are inmmh−1 andZHmmunits are inmm6m−3.The other two rainfall algorithms, which are called herepolarimetric, are:

RKDP= 16:32K0:79

DP ð7Þ

where,RKDPunits are inmmh−1 and theKDPunits are in ° km−1,

and

RP = 1:3 × 10−3NW ðZHmm=NW Þ0:58 ð8Þ

where, theRpunits are inmmh−1, theZHmmunits aremm6m−3

and the Nw (mm−1m−3) is the “intercept” parameter of the“normalized” Gamma type drop size distribution model.

2.4. Microphysical retrieval algorithm

As shown by the integral Eqs. (1)–(3), information on theDSD, as well as hydrometeor phase (liquid, solid, mixed) andshape are needed to relate polarimetric radar measurements

to precipitation and other radar parameters. Past investiga-tions based on models and observations have shown that theshape of raindrops can be well approximated by oblatespheroids (e.g., Pruppacher and Beard 1970; Beard andChuang 1987; Bringi et al., 1998). The spheroid minor-to-major axis ratio (r) can be approximately related to theequivolumetric spherical diameter (D, in mm). A raindropshape–size relationship used here is a linear model between rand D originally presented by Pruppacher and Beard (1970):

r = 1:03−βD ð9Þ

A point to note about this relationship is that β may varyand that this variability can be determined on the basis ofpolarimetric radar parameters (Matrosov et al., 2002; Gorgucciet al., 2000, 2001; Bringi et al., 2002).

The “intercept” parameterNW (mm−1m−3) of the normal-ized Gamma distribution of DSD assuming the linear axis ratiorelation (Eq. 9) is estimated by the following equation (Parket al., 2005a, Kalogiros et al., 2006):

log10NW = 2:48Z0:099Hmm ξð−0:067β−1:15Þ

DR ð10Þ

where β is the effective slope (mm−1) of the dependence ofrain droplets axes ratio on their diameter and it is given by:

β = 5:69KDP

ZH

� �0:41ðξDR−1Þ0:58 ð11Þ

where ZHmm units are in mm6m−3 and the ξDR is 10(ZDR/10) inlinear units.

3. Study area and data

As part of European research project HYDRATE (Hydro-meteorological data resources and technologies for effectiveflash flood forecasting), the National Observatory of Athensdual-polarization XPOL radar, a 2D-video (2DVD) disdrom-eter and a dense network of raingauges (a pair in every site)were deployed in the northwest part of Crete island, near thecity of Chania in Greece, during the winter–spring period of2007. Fig. 3 shows the position of the radar, the 2DVD (site F0)

Fig. 3. Location of the NOA's dual-polarization X-band radar (XPOL), the 2DVD disdrometer site (F0), and the six dual-gauge validation sites (F1–F6).


and the validation gauge sites (F1–F6), as well as the localtopography of the area. Each gauge site (F1–F6) consists of apair of tipping bucket type (with 0.2 mm resolution) rain-gauges. The XPOL radar was placed 4 km southwest of the cityof Chania and 7 km east of the municipal airport of Chania,Souda.

Fig. 4. Radar beam occlusion power percentages (%) due to local topogr

The radarwas operated onlywhen therewas forecast for rainevent in the nearby experimental area from the beginning ofJanuary till the endofMarch2008. The radarwasoperatedfirst ina range height indicator (RHI) mode taking four RHI measure-ments from 0° to 30° elevations and then in a planar positionindicator (PPI)mode takingmeasurements in thewesternpart of

aphic effect for antenna elevation at (a) 1°, and (b) 3° elevations.

Table 1Raingauge and 2DVD position relative to XPOL radar and percentage (%) of beam blockage for each radar elevation due to local topography.

Site# Distance fromXPOL (km)

Azimuth fromXPOL (o)

Height ASL(m)

Beam blockage1ο (%)




F0 1.65 310.1 25 0 0 0 0F1 25.10 215.5 1041 99.7 75.4 3.6 0F2 19.02 218.2 490 95.8 26.7 0.1 0F3 17.20 258.3 100 0.1 0 0 0F4 37.08 232.8 435 99.1 37.1 0.1 0F5 25.12 254.1 102 1.2 0 0 0F6 21.21 269.7 30 0 0 0 0


the island within a 180° sector scan, and for 1°, 2°, 3° and 4°elevation sweeps. Antenna rotation rate was 10 (°/s) and thetime period for a full volume scanwas about 3minutes. The twoupper elevation sweeps are selected to avoid ground clutter andbeamblockage in the south–southwest section of the sector. Thepercentage (%) of beam blockage due to ground clutter wastheoretically calculated for each antenna elevation and ispresented in Fig. 4 for the experimental area andbeamelevationsat 1° and 3°. In the calculations we used high resolution (90 m)terrain data, a three dimensional model of radar beam, beamrefraction due to the atmosphere and the effect of Earthcurvature. In Table 1 we summarize the % of beam blockage forall beamelevations at each raingauge location. Aswenotice fromthe table radar data at beam elevations below 2° can be used forquantitative rainfall estimation and comparison with validationgauges mainly in the lowland experimental area (heights lessthan 100 m above sea level). However, the more mountainousground validation sites (i.e., F1, F2 and F4) can only be comparedagainst the 3° and 4° radar elevation measurements.

Raindrop size distribution data for the different rainfallevents were available for this study on the basis of measure-ments from the 2DVD disdrometer. The location of the 2DVDas shown in Fig. 2 is about 1.5 km north from the XPOL andclose to the sea level. The DSD spectra included 1-min averagedrop counts of equivalent-volume diameter (D, mm). XPOLparameters such as horizontal reflectivity (ZH), differentialreflectivity (ZDR) and specific differential phase shift (KDP)

Fig. 5. Scatter plot of (a) horizontal (ZH) and (b) vertical (ZV) polarization reflectivit2DVD spectra observations.

were determined from the observed spectra and calculatedscattering amplitudes using the T-matrixmethod as describedin the previous section. The XPOL datawere compared againstthe coincident simulated ZH and ZV (dBZ) reflectivity valuesderived from the 2DVD measured spectra. The bias error isestimated as the standard error of themean (and not standarddeviation) difference between parameters (ZH, ZV, and ZDR)estimated by the disdrometer and the radar (Fig. 5). Theapparent deviation from linearity in the ZH and ZV radar todisdrometer scatter plots at high dBZ values is simply due tothe limited number of data at these values during the timeperiod that the disdrometer was operational. If more datapoints were available probably they would be scattered moresymmetrically around the calibration line. The estimated biaserrors for ZH and ZDR (about 200 data points from instant radarvalues and one minute averaging disdrometer values) are 0.3and 0.02dB respectively, which are quite low (actual biaserrors may be twice as these statistical average values). Notethat the bias error of ZDR is smaller than ZH bias because thebias errors of ZH and ZV (ZDR=ZH−ZV) are correlated.

4. Comparison of XPOL-derived rain rates with in-situ data

After the XPOL measurements quality control (i.e., cluttercorrection, beam occlusion, and adjustment of ZH and ZVparameters) and having corrected them for the specific (AH)and differential (ADP) rain-path attenuation, rainfall estimates

y measured by XPOL versus coincident values theoretical derived from 3-min


for the two selected events were then calculated on the basisof the earlier presented relationships (Eqs. 6–8). In order toavoid brightband effects (which will be discussed in the nextSection) and therefore achieve better quantitative compar-isons of radar rainfall estimation with the coincident rain-gauge observations, we selected validation sites (F3 and F6)with less than 5 (%) beam occlusion and used radar estimatesfrom the lowest (1°) antenna elevation. For the gauge/radarcomparison, rainfall was averaged in 1-min time intervals.Evaluation is performed based on visual and statisticalcomparison methods. Visual include time series used toshow the co-variation of the three technique estimates incomparison to the corresponding rainfall rates from thevalidation sites. Statistical methods include bias, correlationcoefficient and relative rootmean square error (rRMSE) of theestimated from radar (hereafter named estimated) versus the

Fig. 6. Time series comparison of the first rain event (04/01/2007) of XPOL rainfastandard— red line) compared to rainfall measured at the two lowland validation sitlegend, the reader is referred to the web version of this article.)

raingauge measured (hereafter named reference) rainfall.Bias is defined as the ratio of mean between each estimateand the reference:

Bias =∑estimated∑reference

ð12Þ

while the relative root mean square error is defined as:

rRMSE =

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1N∑ðreference−estimatedÞ2

q1N∑reference

ð13Þ

Figs. 6 and 7 show the time series of the three radarrainfall estimations (RSTD, RKDP and Rp) coincident withraingauge measurements from the two lowland validation

ll retrievals (standard or polarimetric — black or polarimetric bias adjustedes (bold blue line). (For interpretation of the references to color in this figure

Fig. 7. Similar time series plot to Fig. 6, but for the second rain event (22/03/2007).


sites (F3 and F6) for the two storm events (04/01/08 and 22/03/08). In both figures, the upper panel shows the time seriesof raingauge rainfall measurements (bold blue line) againstthe standard (black line) and the polarimetric bias adjustedstandard rainfall estimates (red line). The middle and lowerpanels show the polarimetric RKDP and Rp rainfall estimates(black line), respectively.

The accuracy of reflectivity calibration and random errorsin radar parameters as mentioned above are quite low. Usingthese errors it is trivial to estimate an error of rainfall retrievalusing algorithm Eq. (8). However, this is not the actual errorin rainfall estimation. There are also errors due to theapproximation of actual rainfall introduced from the param-eterization. This expected to be a higher than the error inreflectivity calibration. Rainfall algorithms like this are basedon regression techniques applied to scattering simulationsand measured (like in our case using the video disdrometer

data) or estimated populations of DSD parameters. Anyalgorithm like RSTD which uses reflectivity information (ZHor ZDR) alone for rainfall estimation, cannot avoid uncertain-ties in the radar retrieval of rainfall due to DSD variabilitywith site and rainfall type (stratiform, convective). In casethat no disdrometer is available or we want to use gauges forevaluation of algorithm performance instead of calibratingRSTD an alternative method to estimate the parameters (i.e.calibrate) of RSTD rainfall function is to use RKDP as a reference(Gorgucci et al., 1992), because KDP is immune to calibrationbiases. In our case the RSTD algorithm was based on Athensduring the season 2005–2006 disdrometer data because thecurrent Crete disdrometer data was not sufficient for suchpurposes due to an early system failure. The mean ratio ofRSTD and RKDP is determined for the 1st elevation (1°), forRKDP N0.1 mm h−1, and for beam occlusion less that 5 (%).Then the mean ratio is averaged for the entire event. This

Table 2aRadar–gauge rainfall comparison statistics for the January 4, 2007 rain event.

F3/F6 Correlation Bias rRMSE

RSTD 0.92/0.86 0.72/0.79 0.69/0.65RKDP 0.88/0.90 1.01/1.02 1.16/0.64RP 0.89/0.92 1.00/0.96 1.78/0.46

Table 2bRadar–gauge rainfall comparison statistics for the March 22, 2007 rain event.

F3/F6 Correlation Bias rRMSE

RSTD 0.88/0.86 0.52/0.59 0.99/1.02RKDP

0.84/0.86 0.65/0.69 0.86/0.89RP 0.85/0.84 0.79/0.78 0.79/0.89


“polarimetric” bias adjustment is then applied to the standardrainfall estimation algorithm for the entire PPI every timestep.

Table 2 (a–b) summarizes the bulk error statistics (corre-lation, bias, and rRMSE) for the two storm events. According tothe results of a previous more extensive study (Kalogiros et al.,2006)we expect the algorithms to performgood on a statisticalbasis but probably not in a specific event, which means that achange of bias or relative root mean square error between thetwo events in the comparison with the gauges is justified.Unfortunately, we did not have more events from Cretecomplex terrain site in order to estimate the performance ofthe algorithms on a longer period. The statistics are computedfor rain rates above 0.1 (mmh−1). A first observation is that allthree algorithms exhibit adequate correlations (r2=0.92 and0.86 for RSTD, r2=0.88 and 0.90 for RKDP and r2=0.89 and 0.92for Rp, respectively) with the rainfall measurements from the F3and F6 in situ sites. Overall, the two polarimetric (Rp and RKDP )algorithms exhibit better performance compared to thestandard rainfall algorithm. Specifically, in the January 4thstratiform like rain event from the comparison with F6 site therRMSE (Table 2a) for the RSTD the value is 0.65, while for of Rp is0.46 and the value for the RKDP is 0.64. However, for the sameevent and for the F3 site the rRMSE of RSTD is 0.69 compared tothe Rp that is 1.78 and RKDP is 1.16, showing that the abovestatement is not always true.We also notice an overestimationof the standard algorithm in the peak of the event while theother two polarimetric algorithms perform better.

For the 22nd of March storm event, which is of convectivelike type, both ground sites show a peak around minute 600of that date. The standard rainfall algorithm underestimatesthe overall event compared to gauges from site F3, while thealgorithm performs slightly better in the comparison withgauges from site F6 (see Table 2b), even though, this site isfurther away in range from the radar (~21 km). Bothpolarimetric algorithms performed better compared to thestandard algorithm at both sites. Specifically, the bias of thepolarimetric algorithm Rp is close to 0.8 for both sites whilethe standard algorithm is 0.52/0.59 for sites F3/F6, respective-ly. In Fig. 7b and for the standard algorithm at minute 650–750 we notice an underestimation compared to rainfallmeasured by gauges at site F3 and in the second peak of thesame event during the 700–750 minute, for the comparisonwith the gauge from site F6. Moreover, after minute 1200 andat the first peak (minute ~600) of the rain event, both

polarimetric algorithms show good agreement with the rainobserved from the gauges at both sites.

Finally, by comparing the bulk statistics of the standardrainfall algorithm with and without the polarimetric biasadjustment, we note that the adjustment improves the biasby almost 25% (0.72/0.79 improves to 1.01/0.98 for theJanuary event and for the sites F3/F6, respectively). For theMarch event the improvement is slightly lower compared tothe January event (bias of 0.52/0.59 improves to 0.59/0.65 forthe sites F3/F6, respectively). Also, the relative rRMSE of thestandard algorithmwith the adjustment for the January eventincreases from 0.69/0.65 to 1.14/1.00, while for the Marchevent it reduces from 0.99/1.02 to 0.92/0.97 for sites F3 andF6, respectively.

Fig. 8 (a–d) shows rainfall accumulation maps (in mm) ofthe 4th of January storm event (~11 hrs) from the standard(RSTD) and polarimetric (Rp) algorithms at the first twoelevations (1° and 2°). The circle (positioned in the directiony=0 km and in the x direction between −40 and −20 km)in Fig. 8 (a–b) shows the significant overestimation in theconvective part of the event (30 km west of the radar) of theRSTD compared to theRp Fig. 8 (c–d).Weclearly notice a BB zonewith high rainfall values at 2° elevation. On the other hand, thepolarimetric algorithm in Fig. 8 (c–d) at 1° appears moreuniform compared to the RSTD in the convective part of theevent. This is more apparent at the second elevation (2°). TheBB due to the enhancement effect in horizontal and differentialreflectivity is also present in the polarimetric estimationalgorithm. Similar results we noticed for the 22nd of Marchstorm event. This enhancement effect in reflectivity factor,which is mostly noticed in stratiform type storms and at higherantenna elevations, need to be corrected to remove any rainfallestimation overestimations (within the melting layer) orunderestimations (above the melting layer and within thesnowregion). Thenext chapter explains theVPReffect andhowthe polarimetric information can be used to map the BB andcorrect the reflectivity to improve rainfall estimation.

5. VPR correction method and performance

Conditions of low freezing levels is typical when the radarsare deployed in complex terrain and for low or moderateprecipitation events (i.e., stratiform rain type) associated withweak vertical motion and the occurrence of melting layer. InFigs. 9 (a–b) and 10a we see sample RHI and PPI plots where abrightband,which can be found in themelting-layer height or inthe zero isotherm, is apparent. It is called bright band, because ofthe enhancement effect at the reflectivity factor (Fabry andZawadzki, 1995). Therefore, it is necessary to develop an ap-proach to identify the rain, melting-layer, and snow regions inthe PPI and quantify the surface rainfall for each range gatefalling in those regions due to quick rise of radar beam withdistance at high elevation angles. One of the approaches sug-gested to account for these effects is to use a priori informationon the vertical profile of reflectivity (VPR; e.g., Andrieu andCreutin 1995; Bellon et al., 2005). Under this approach thereflectivity measured aloft is related in a mean sense to anexpected reflectivity at the ground. This ground-level reflectivityestimate is thenused in rainfall estimation. This studyusedaVPRapproach tailored to X-band polarimetric radar measurementsby Matrosov et al. (2007). The main feature of this approach is

Fig. 9. RHI plot of (a) horizontal reflectivity (ZH) and (b) copolar correlation coefficient (ρHV).

Fig. 8.Accumulated rainfall map estimated from the standard (a and b)RSTD and polarimetric (c and d)Rp algorithm for about 11 hrs at one and 2° elevation of 04/01/2007.


Fig. 10. Sample PPI plot for 4° elevation on 10:03 UTC March 22, 2007 of reflectivity (a) uncorrected for VPR and (b) corrected for VPR. The blue line identifies thelower and upper boundary of the identified BB.


that polarimetric observations in mixed phase hydrometeors(i.e., in the melting layer) and snow are used to define themelting layer boundaries instead of utilizing an average a prioriVPR and melting-layer heights based on climatological or otherauxiliary data. Reflectivity-based estimators are subsequentlyused for QPE once reflectivity values reduced to ground level areestimated based on an average VPR and beam-specific informa-tion on the altitudes of the melting layer boundaries. At shorterranges and up to the melting layer base rain-path attenuation isestimated according to Section 2.2. At longer ranges the radarresolutionvolumesare partially or completelyfilledwithmixed-phase hydrometeors from the melting region or snow particlesabove the freezing level and the reflectivity profile is character-izedbya constant slopeof reduction (Kitchen and Jackson, 1993;Fabry and Zawadzki, 1995).

The RHI scans performedwith the XPOL radarwere used toconstruct the idealized average model (Fig. 11) of the mean

Fig. 11. Shape parameters of ideal vertical profile of reflectivity (VPR) fitt

vertical reflectivity-profile (Matrosov et al., 2007) and relatethe BB reflectivity enhancement in the melting layer to therain rate below it. The number of RHI scanswith clear BB in thetwo rain events which were used in the VPR analysis in thispaper was 307 in the first event (4/1/07) and 268 in thesecond event (21/3/07). Themodel parameters are: the brightband (BB) reflectivity enhancement of themean profile that isclose to triangle shape, ΔZ, the mean reflectivity in the rainlayer that is assumed to be constant (Kitchen 1997; Fabry andZawadzki, 1995), ZH, the difference between reflectivities inthe rain layer and in the snow region just above the freezinglevel generally that is about 1.5dB and it varies relatively little,ΔZ1, the mean reflectivity gradient in the snow region whichis linear decreasing with almost constant slope, tan(β), andthe base and top heights of melting layer htop and hbase,respectively. The estimation of the parameters of the averageVPRprofilewas based on the average profile fromall RHI scans

ed to the average profile of ZH from XPOL measured RHIs of Fig. 9a.


and not individual profiles like the one in Fig. 11. The slope βwas estimated from the layer above htop (found by applyingcriteria similar withMatrosov et al., 2007, see Section 5.1) andbelow the first significant minimum of the average profile.This slopewas assumed to be the slope of the ideal VPR profilei.e. the one with no increase of reflectivity above the meltinglayer due to the natural spatial (horizontal or vertical)variability of reflectivity in individual profiles.

Due to beam broadening effect and attenuation in rain themaximum BB enhancement ΔZ decreases as the distanceincrease from the radar (notice the smoothing of BB withdistance from the radar in Figs. 9 and 10). Based on theobserved XPOL RHIs the following relationship was derivedfor the current data set:

ΔZ = −0:13r + 5:48 ð14Þ

where r units are in kilometers and ΔZ units are in dB. Fromthe analysis of RHIs the rest of the VPR model parameterswere also estimated. The average difference ΔZ1 value was1.17dB and the mean value of the reflectivity gradient in thesnow region above the freezing level (i.e., tan(β)) was about10.1 (dB km−1) with a standard deviation of 1.2 (dB km−1). Asmall difference of 64 (meters) in altitude between reflectiv-ity peak and ρHV correlation minimum (the latter lower thanthe first) was also found. The values of the above parametersare different than the ones presented by Matrosov et al.(2007), which implies that they cannot be considered asconstants but must be estimated from RHI data for eachsystem and probably climatic region or rain event.

5.1. Bright band (BB) identification

Before applying any algorithm that corrects for the verticalreflectivity variability, a clear identification of the exactlocation of the melting layer within the radar resolutionvolume needs to be established. In a stratiform type event thereflectivity by itself is enhanced in the melting layer region.However, positioning of BB along a ray is challenging due toits horizontal variability which is often significant andcomplicates the determination of melting layer boundaries.The co-polar correlation coefficient ρHV is a very usefulpolarimetric parameter that can be used for identifying theupper and lower boundaries of BB because allows for a robustdiscrimination among the regions of rain, melting layer andsnow region as it was showed by Matrosov et al. (2007). Themagnitude of this coefficient is related to the variety ofhydrometeor shapes present in the radar resolution volume.For the XPOL radar, the ρHV in rain is generally greater than0.96–0.98 while in snow regions the values of this coefficientare greater than 0.86–0.95. The melting layer boundaries areidentified as the altitudes where ρHV first decreases at valuesless than 0.96 (base) or first increasing at values above 0.95(top) with increasing altitude. Also, the minimum value ofρHV in the melting layer should be below 0.93.

5.2. Application of the VPR correction approach

The VPR correction approach is designed to estimate re-flectivities in the underlying rain layer when the radar beam

overshoots this layer and is completely or partially filled withmixed-phase hydrometeors and snow. Although it is knownthat the use of such polarimetric parameters as ZDR and KDP inrain can improve QPE (e.g., Anagnostou et al., 2004; Park et al.,2005b; Matrosov et al., 2005) there are no practicalapproaches to date to relate rain values of these polarimetricparameters to the corresponding values observed in meltingand snow regions aloft. Therefore, this study further focuseson the verification of rainfall estimates obtained with thestandard algorithm Eq. (6) with and without using VPRadjustments. It is underlined that although polarimetric QPEestimators are not utilized here, the radar polarimetricmeasurements were used to identify the BB and correctingthe reflectivity data for attenuation in rain. In addition, wenote that the VPR correction is applied to PPI scans but not RHIscans because we are interested on the correction of PPIrainfall maps. RHI correction is of no meaning because in RHIscans the exact VPR profile is measured.

Fig. 10a shows a sample PPI of ZH at 4° elevationwhere theblue lines indicate the upper and lower boundary of BBestimated by the ρHV limits mentioned above. As soon as theranges of themelting layer (i.e., the reflectivity enhancement)are established for each radar ray, the mean VPR model isapplied to all the ranges where the radar resolution volume iseither in themelting layer or in the snow region. According tothe mean VPR model described in the beginning of thissection, measured reflectivities on a beam to beam basis arerecalculated to rain values that are expected at the ground.Rain regions identified previously using ρHV have beencorrected for rain-path attenuation. The corrected PPI(shown in Fig. 10b) even though exhibits regions ofreflectivity enhancement there are not areas of vastly differentreflectivities values. Fig. 12 (a–d) shows accumulation mapsusing the standard Z–R relation at 2° elevation with andwithout accounting VPR adjustment. The current VPR correc-tion is a statistical correction (average VPR is used) based onthe full dataset. Thus, the cumulated rainfall test is anappropriate method for evaluation of the performance of theVPR correctionwhichwas also used byMatrosov et al. (2007).It is apparent that the convective part (core) of the cell(continuous black line circle positioned in the y-directionbetween −20 and 0 km and in the x-direction between −10and −30 km) becomes “part” of the bright band (dashedblack line circle positioned in the y-direction between −20and 0 km and in the x-direction between−25 and−20 km).On the other hand, Fig. 12b shows a visual improvement as theenhanced reflectivity zone is more dispersed and, thus, lesspronounced relative to those of Fig. 12a. The core part of theconvective cell is still present, but it is more spread out.Consistent results we notice in Fig. 12 (c–d) presenting the 3°elevation. We observe that the BB at the higher elevationsangles (i.e., 3° in Fig. 12c) is more thick compared to the BB )(positioned in the y-direction between the−20 and 0 km andin the x-direction between 0 and−20 km) at lower elevationangle (i.e., 2° in Fig. 12a). Fig. 12d shows now a morehomogeneous plot (continuous black line circle positioned inthe y-direction between−20 and 0 km and in the x-directionbetween 0 and −20 km). However rainfall spatial variabilitycould change this and probably gave this ‘opposite’ result inthe example shown here. Secondly, we observe a significantincrease of the estimated rain further away from the BB zone.

Fig. 12. Rain accumulation maps (in mm) of the standard algorithm (RSTD) of precipitation event on January 4, 2007 at 2° elevation (a) without correcting and (b)correcting for VPR, subplots (c) and (d) same as (a) and (b) but for 3° elevation.


Similar increase in rain is observed at 2° elevation (Fig. 12b).Comparison with rain estimates at 1° elevation (Fig. 8a),where no BB effects are present, show that the currently usedmodel profile method for VPR correction does not removecompletely the BB effect and tends to overestimate rainbeyond the melting layer due to the assumption of constantdecrease of reflectivity with altitude (i.e. distance in PPIs) inthe snow region.

For a more quantitative evaluation of the method a timeseries accumulated comparison of radar rain estimates, withand without VPR adjustment, at the two lowland sites areshown in Fig. 13 (a–b). We compare the polarimetric biasadjusted standard radar rainfall algorithm for 1° (orange line)and 3° (red line) elevation with the rain measured from thegauges. The gray line is the 3° estimate but corrected for VPR.From both figures we note an improvement (estimates come

closer to in situ point measurements) of the radar rainestimation from the higher elevation beam (3°) when weapply VPR correction. More precisely, in Fig. 13a, we observethe rainfall estimation with VPR correction (gray line) at 3°elevation reduced to values closer to rainfall estimation at 1°elevation. The F3 (Modi) site is close to radar (~17 km) and at3° elevation, which is about 900 m above the ground, theobserved reflectivity is within the manifestation effect thatthe brightband creates (according to the theoretical model ofFig. 11, the BB starts from ~700–1500 m). This explains errors(overestimation) introduced to the rainfall estimation algo-rithm. However, in general terms the VPR corrected rainfallalgorithm reaches the rainfall estimation at 1° elevation (11%correction). On the other hand, in Fig. 13b we observe anincrease in rainfall estimation from VPR correction. Theexplanation for the increase is that site F6 (Maleme) is further

Fig. 13. Rain accumulation plots on January 4, 2007, of raingauge, polarimetric bias adjusted RSTD, with and without VPR correction for the 1° and 3° elevation, overthe lowland sites (a) F3 (Modi) and (b) F6 (Maleme).


away in range compared to F3. The range from the radar isabout 21 km where a 3° elevation beam is at about 1100 mabove the ground. Taking into account beam broad andsmoothing effects the radar beam should be slightly withinthe snow region. This means that the radar is expected tounderestimate surface rainfall as confirmed through compar-isons with rain measured from the raingauge at site F6 andthe VPR correction should increase the radar estimates. Thecorrection due to VPR in this case is in the order of 19%.

Applying the same radar rainfall algorithmand comparing itagainst the two highland sites F1 (Omalos at 1000m above sealevel) and F2 (Lakkoi 500 m above sea level) we observe thefollowing results shown in Fig. 14 (a–b). In both figures weshow the radar rainfall estimates for the 3° (orange and darkgray lines) and 4° (red and light gray lines) elevations. The grayand dark gray lines are radar rainfall estimates corrected forVPR. Overall, we note a not so good agreement between radar

rainfall estimation and raingauge observations in the highlandgauge sites. However, there is a notable improvement in therainfall estimation (particularly at site F2) from VPR correctionapplied on radarmeasurements from both 3° and 4° elevations.It is shown that the radar rain estimation at 4° elevation withthe VPR correction is getting much closer to the raingaugeobservations, while the estimates from the 3° are improving aswell. The F1 site is at altitude of 1000 m and the radar ray at 3°and 4° elevation over the range of that particular site iscalculated to be about 310 and 750 m above the ground.Similarly, for site F2 that is at height of 500 m the radar ray isabout 490 and 830 m above the ground. Therefore, due to thesmall sampling area of the raingaugeswe are not expecting thegauge at the ground tomeasure the total volume that the radarobserves (volume instead of point measurement) at eachdifferent height; hence, these significant differences betweenthe radar rainfall estimations and the observed rain from the

Fig. 14. Rain accumulation plots on January 4, 2007, of raingauge, polarimetric bias adjusted RSTD, with and without VPR correction for the 3° and 4° elevation, overthe two highland sites (a) F1 (Omalos) and (b) F2 (Lakkoi).


gauges are probably expected. This is more pronounced in thisspecific case where the rain was produced by a squall line(Mesoscale system) that was moving fast. We note that a syn-chronization of radar and gauges time series was applied usingan approximate fall velocity of water droplets and maximumcorrelation of corresponding rainfall time series.

Finally, in Fig. 15 results for F4 site (Kantanos at 435 mabove sea level and 37 km away from the radar) are shown.On January 4, 2007 the total rainfall at this site as well as at F5site was too small. Thus, we show here results from the rainevent onMarch 22, 2007. On that day the bright band was toohigh and the radar ray passing over F5 site was located belowthe bright band for most of the time even for the 4° elevation(see Fig. 10). Thus, the VPR correction for the radar rainfallestimation at site F5 was too small and we don't show it here.We note, however, that the comparison with gauge data

showed that on that rain event there is high underestimationof rainfall from the standard rainfall algorithm (see the resultsfrom sites F3 and F6 in Section 4 and Table 2b). This was thecase for all the sites expect the clearly mountainous sites F1and F4 and is attributed to different rain microphysics overthe mountains area. Fig. 15 show that for the 3° elevation(1550 m above sea level over F4) there is a minor VPRcorrection, which brings the radar estimate a little furtheraway from gauge measurements. This result indicates thatthe VPR method with a fixed predefined profile of reflectivityas the one used here may sometimes give even wrong sign ofcorrection. The VPR correction for 4° elevation (2200 m abovesea level over F4) is in the correct direction and brings rainfallestimation very close to the gauge measurements, which is aquite satisfactorily result considering the very long distanceof the site from the radar.

Fig. 15. Rain accumulation plots on March 22, 2007 of raingauge, polarimetric bias adjusted RSTD, with and without VPR correction for the 3° and 4° elevation, overthe highland site F4 (Kantanos).


6. Summary and conclusion

In this paper an integrated application of available proces-sing methods for high resolution X-band polarimetric radarmeasurements of rainfall in complex terrainwas presented. Therequired (or at least the most crucial) processing steps arecalibration of radar reflectivity, beam blockage calculation, rainattenuation correction, application of rainfall estimation algo-rithms, and correction for the vertical profile of reflectivity(VPR). For accurate calibration of radar reflectivity as well as todevelop polarimetric rainfall algorithms we used high qualitymeasurements from a video disdrometer located at site close tothe radar with no intermediate obstacles. Theoretical T-matrixscattering calculations were used to obtain radar products (ZH,ZDR, and KDP) from disdrometer data. For beam blockageestimation high resolution terrain information (90 m resolutionor less) and a three dimensionalmodel of radar beam is needed.This information is significant to exclude from further proces-sing highly occluded areas or to correct reflectivity from areaswith minor occlusion (e.g. 3dB reduction of ZH). For correctionfor rain attenuation a modified ZPHI method was used. Thismethod uses the polarimetric radar products (mainly ΦDP) inorder to calculate attenuation parameters by data fitting in raincells defined by a threshold onρHV andKDP. The choice of rainfallalgorithm is also a crucial step for accurate rainfall estimation.Polarimetric algorithms (i.e. formulae combining polarimetricproducts with radar reflectivity) are more accurate and can beapplied under different rain types and in different geographicalareas because less assumptions are made in their development.We haddeveloped and tested polarimetric algorithmsusing thedisdrometer data in past campaigns in the area of Athens. Thefinal and quite important processing step is VPR correctionwhich is a required step in complex terrain measurements dueto the requirement for high elevation angles in order to avoid

beam blockage as much as possible. This correction aims at theremoval of bright band (melting layer) effect on reflectivityvalues during stratiform type rain events. In this paper we usedan already available method, which uses a statistical (average)profile of reflectivity obtained from the full data set. Thismethod can only be applied to standard rainfall algorithms,which use only radar reflectivity, but not to modern polarimet-ric algorithms and is further discussed below.

Two raineventswere studied in thiswork, amore convectivetype storm during the spring and a stratiform type rain eventduring the winter. Comparison of radar rain estimates withgauges showed good bulk statistics in low altitudes (below100 m) with correlation higher than 0.8, while at higheraltitudes correlation dropped below 0.6. For the convectiveevent the standard rainfall algorithmunderestimates theoverallevent compared to gauges at low altitude sites. Polarimetricalgorithms performed better with low bias compared to thestandard algorithm. However, for the stratiform like rain event,even though the standard algorithm showed again underesti-mation of rainfall it gave lower relative root mean square errorcompared to the two polarimetric algorithms. This implies thatin less intensive events (like the stratiform rain event) thepolarimetric algorithms might show larger relative randomerror compared to the standard algorithm. In general and inagreement with previous more extensive data set polarimetricalgorithms seem to perform better than standard Z–R algo-rithms. In addition, we included a method to “calibrate” thestandard algorithm using a polarimetric algorithm like RKDP asa reference. The standard Z–R algorithm cannot avoid uncer-tainties in the estimation of rainfall due to DSD variability withsite and rainfall type (stratiform, convective). By comparing thebulk statistics of the standard rainfall algorithm with andwithout this “polarimetric” bias adjustment, we found animprovement of bias up to 25% for the stratiform event.


The final significant correction in complex terrain rainmeasurements by weather radar is related to the requirementfor higher antenna elevations for the PPI scans to avoid terrainblockage. High elevation scans are in addition to others errorsassociated with errors introduced from the enhancement inreflectivity due to bright band (melting layer) effects duringstratiform rain events. A recently published (Matrosov et al.,2007) mean VPR model and melting layer identification bypolarimetric measurements of copolar correlation coefficientbetween horizontal and vertically polarized radar signal wasapplied to correct those errors. Comparisons of XPOL rainfallestimates (rain accumulation maps and comparison withgauges) with and without VPR correction indicated that thecorrection for BBand snoweffect in and above themelting layerimproves rain estimation and provides more consistent rainfallmaps. However, the efficiency of this algorithm is probablygood in a statistical sense due to the usage of a statistical meanVPR profile “for a particular radar and experimental setting” andfor “polarimetric NEXRADapplications” asMatrosov et al. (2007)pointed out in their conclusions but not good enough on a PPIper PPI basis for high resolution (temporal and spatial) studieswith X-band radars. Also, even though polarimetric radars usepolarimetric algorithms for rainfall estimation this is notpossible with the current algorithm which is strictly limitedto classic Z–R algorithms. A further advanced VPR correctionscheme which would be more suitable for the needs of highresolution X-band radars will be the subject of future work.

Acknowledgements

Thiswork is part of theHYDRATEproject (STREP category—GOGE-2006-0370024) funded by EC 6th Framework Program.We would like to thank here the Polytechnic of Crete,department of Civil and Environmental Engineering and theprofessor J. Tsanis for providing the deployment position andthe technical help during the deployment of the radar. Wewould like also to thank the municipals of Kantanou (F4) andVoukoulies (F5), the military camp of Maleme (F6) and thepeople of Modi (F3), Lakkoi (F2) and Omalou (F1) for allowingthe deployment of the in situ sites. Last but not least, the authorwould like to give special thanks to theGIS group of HCMR, ViviDrakopoulou and Hara Kyriakidou, for the art work of the mapin Fig. 3.

References

Anagnostou, E.N., Anagnostou, M.N., Krajewski, W.F., Kruger, A., Miriovsky, B.J.,2004. High-resolution rainfall estimation from X-band polarimetric radarmeasurements. J. Hydrometeorol. 5, 110–128.

Anagnostou, E.N., Grecu, M., Anagnostou, M.N., 2006a. X-band polarimetricradar rainfall measurements in Keys area Microphysics project. J. Atmos.Sci. 63, 187–203.

Anagnostou, M.N., Anagnostou, E.N., Vivekananda, J., 2006b. Correction forRain-Path Specific andDifferential Attenuation of X-bandDual-PolarizationObservations. IEEE Transaction of Geosciences and Remote Sensing 44,2470–2480.

Anagnostou, M.N., Anagnostou, E.N., Vivekanandan, J., 2007. Comparison ofRaindrop Size Distribution Estimates fromX-Band and S-Band PolarimetricObservations. IEEE Geosciences and Remote Sensing Letters 4, 601–605.

Andrieu, H., Creutin, J.D., 1995. Identification of Vertical Profiles of RadarReflectivity for Hydrological Applications Using an Inverse Method. Part I:Formulation. Journal of Applied Meteorology 34, 225–239.

Atlas, D., Ulbrich, C.W, 1990. Early foundations of themeasurement of rainfallby radar. Radar in Meteorology. American Meteorological Society, Boston,USA, pp. 86–97.

Barber, P., Yeh, C., 1975. Scattering of electromagnetic waves by arbitrarilyshaped dielectric bodies,. Appl. Opt. 14, 2864–2872.

Beard, K.V., Chuang, C., 1987. A new model for the equilibrium shape ofraindrops. J. Atmos. Sci. 44, 1509–1524.

Bellon, A., GyuWon, Lee, Zawadzki, I., 2005. Error Statistics of VPR Correctionsin Stratiform Precipitation. J. Appl. Meteorol. 44, 998–1015.

Brandes, E.A., Zhang, G., Vivekanandan, J., 2004. Drop size distribution retrievalwith polarimetric radar:Model andApplication. J Appl.Meteor. 43, 461–475.

Bringi, V.N., Chandrasekar, V., Xiao, R., 1998. Raindrop axis ratios and sizedistributions in Florida rainshafts: An assessment of multiparameterradar algorithm. IEEE Trans. Geosci. Remote Sens. 36, 703–715.

Bringi, V.N., Chandrasekar, V., 2001. Polarimetric Doppler Weather Radar.Cambridge University Press, Cambridge, UK. 662pp.

Bringi, V.N., Keenan, T.D., Chandrasekar, V., 2001. Correcting C-band radarreflectivity and differential reflectivity data for rain attenuation: A self-consistent method with constraints. IEEE Trans. Geosci. Remote Sens. 39,1906–1915.

Bringi, V.N., Huang, G.J., Chandrasekar, V., Gorgucci, E., 2002. Amethodology forestimating the parameters of a gamma raindrop size distribution modelfrom polarimetric radar data: Application to a squall-line event from theTRMM/Brazil campaign. J. Atmos. Oceanic Technol. 19, 633–655.

Bringi, V.N., Chandrasekar, V., Hubbert, J., Gorgucci, E., Randeu, W.L.,Schoenhuber, M., 2003. Raindrop Size Distribution in Different ClimaticRegimes from Disdrometer and Dual-Polarized Radar Analysis. J. Atmos.Sci. 60, 354–365.

Fabry, F., Zawadzki, I., 1995. Long-term observations of the melting layer ofprecipitation and their interpretation. J. Atmos. Sci. 52, 838–851.

Gorgucci, E., Scarchilli, G., Chandrasekar, V., 1992. Calibration of radars usingpolarimetric techniques. IEEE Trans. Geosci. Remote Sens. 30, 853–858.

Gorgucci, E., Scarchilli, G., Chandrasekar, V., Bringi, V.N., 2000. Measurementof Mean Raindrop Shape from Polarimetric Radar Observations. J. Atmos.Sci. 57, 3406–3413.

Gorgucci, E., Scarchilli, G., Chandrasekar, V., Bringi, V.N., 2001. Rainfall Estimationfrom Polarimetric Radar Measurements: Composite Algorithms Immune toVariability in Raindrop Shape–Size Relation. J. Atmos. Ocean. Tech. 18,1773–1786.

Hubbert, J.V., Bringi, V.N., 1995. An iterative filtering technique for the analysis ofcoplanardifferential phaseanddual-frequency radarmeasurements. J. Atmos.Oceanic Technol. 12, 643–648.

Ishimari, A., 1991. ElectromagneticWave Propagation, Radiation, and Scattering.Prentice Hall. 637pp.

Joss, J., Waldvogel, A., 1990. Precipitation measurements and hydrology. In:Atlas, D. (Ed.), Radar in Meteor. Amer. Meteor. Soc, Boston, pp. 577–606.

Kalogiros, J, Anagnostou, M, Anagnostou, E, 2006. Rainfall retrieval frompolarimetric X-band radar measurements. Proceedings of 4th EuropeanConference on Radar in Meteorology and Hydrology (ERAD), Barcelona,Spain, pp. 145–148. 18–22 September.

Kitchen, H., Jackson, P.M., 1993. Weather radar performance at long range —simulated and observed. J. Appl. Meteor. 32, 975–985.

Kitchen, H., 1997. Towards improved radar estimates of surface precipitationat long range. Quart. J. Roy. Meteor. Soc. 123, 145–163.

Matrosov, S.Y., Clark, A., Martner, B.E., Tokay, A., 2002. X-band polarimetricradar measurements of rainfall. J. Appl. Meteor. 41, 941–952.

Matrosov, S.Y., Kingsmill, D.E., Martner, B.E., Ralph, F.M., 2005. The Utility ofX-Band Polarimetric Radar for Quantitative Estimates of RainfallParameters. J. Hydrometeorol 6, 248–262.

Matrosov, S.Y., Clark, KurtA., Kingsmill, DavidE., 2007. “A Polarimetric RadarApproach to Identify Rain. Melting-Layer, and Snow Regions for ApplyingCorrections to Vertical Profiles of Reflectivity”. Journal of AppliedMeteorology and Climatology. 46, 154–166.

Park, S.-G., Maki, M., Iwanami, K., Bringi, V.N., Chandrasekar, V., 2005a.“Correction of Radar Reflectivity and Differential Reflectivity for RainAttenuation at X Band. Part I: Theoretical and Empirical Basis”. J. Atmos.Ocean. Tech. 22, 1621–1632.

Park, S.-G., Maki, M., Iwanami, K., Bringi, V.N., Chandrasekar, V., 2005b.Correction of radar reflectivity and differential reflectivity for rainattenuation at X band. Part II: Evaluation and application. J. Atmos.Oceanic Technol. 22, 1633–1655.

Pruppacher, H.R., Beard, K.V., 1970. A wind tunnel investigation of theinternal circulation and shape of water drops falling at terminal velocityin air. Quart. J. Roy. Meteor. Soc. 96, 247–256.

Testud, J., Le Bouar, E., Obligis, E., Ali-Mehenni, M., 2000. The rain profilingalgorithm applied to polarimetric weather radar. J. Atmos. Ocean. Technol.17, 332–356.

Vulpiani, G., Marzano, F.S., Chandrasekar, V., Lim, S., 2005. Constrained iterativetechnique with embedded neural network for dual-polarization radarcorrection of rain path attenuation. IEEE T. Geosci. Remote. 43, 2305–2314.

Experimental results on rainfall estimation in complex terrain with a mobile X-band polarimetric...

Documents

Transcript of Experimental results on rainfall estimation in complex terrain with a mobile X-band polarimetric...