Radar adjusted data versus modelled precipitation: a case study over Cyprus

Transcription

1 Advances in Geosciences, 7, 85 90, 2006 SRef-I: /adgeo/ European Geosciences Union 2006 Author(s). This work is licensed under a Creative Commons License. Advances in Geosciences Radar adjusted data versus modelled precipitation: a case study over Cyprus M. Casaioli 1, S. Mariani 1,2, C. Accadia 1,*, M. Gabella 3, S. Michaelides 4, A. Speranza 5, N. Tartaglione 6 1 Agenzia per la Protezione dell Ambiente e per i Servizi Tecnici (APAT), Rome, Italy 2 epartment Mamatics, University Ferrara, Ferrara, Italy 3 epartment Electronics, Politecnico di Torino, Turin, Italy 4 Meteorological Service, Nicosia, Cyprus 5 epartment Mamatics Informatics, University Camerino, Camerino, Italy 6 epartment Physics, University Camerino, Camerino, Italy * present organization: EUMETSAT, armstadt, Germany Received: 30 October 2005 Revised: 28 ecember 2005 Accepted: 29 ecember 2005 Published: 30 January 2006 Abstract. In framework European VOLTAIRE project (Fifth Framework Programme), simulations relatively heavy precipitation events, which occurred over isl Cyprus, by means numerical atmospheric models were performed. One aims project was indeed comparison modelled rainfall fields with multi-sensor observations. Thus, for 5 March 2003 event, 24-h accumulated precipitation BOlogna Limited Area Model (BO- LAM) forecast was compared with available observations reconstructed from ground-based radar data estimated by rain gauge data. Since radar data may be affected by errors depending on distance from radar, se data could be rangeadjusted by using or sensors. In this case, Precipitation Radar aboard Tropical Rainfall Measuring Mission (TRMM) satellite was used to adjust ground-based radar data with a two-parameter scheme. Thus, in this work, two observational fields were employed: rain gauge gridded observational obtained by merging range-adjusted radar rain gauge fields. In order to verify modelled precipitation, both nonparametric skill scores contiguous rain area (CRA) were applied. Skill score results show some differences when using two observational fields. CRA results are instead quite in agreement, showing that in general a 0.27 eastward shift optimizes forecast with respect to two observational analyses. This result is also supported by a subjective inspection shifted forecast field, whose gross features agree with pattern more than non-shifted forecast one. However, some open questions, especially regarding effect or range adjustment techniques, remain open need to be addressed in future works. Correspondence to: M. Casaioli (marco.casaioli@apat.it) 1 Introduction Meteorological observations, derived from different instruments, can be compared with model outputs in order to verify models mselves quantify ir skill since, focusing on relationship between observations forecasts, it is possible to underline model s ability to forecast correctly a meteorological event. Forecast verification can also provide insight into way atmospheric processes are modelled. Precipitation is ten object verification studies: point measurements over l are compared with modelled precipitation while over sea it is not possible, usually, to perform same kind comparison. Neverless, satellite or ground-based remote sensing can help evaluating rainfall over sea. Performing this kind operation is one purposes European VOLTAIRE project (Fifth Framework Programme; After selecting several relevant wear events over isl Cyprus, precipitation measured at rain gauges (hereafter RG) retrieved by both ground-based radar (hereafter GR) Precipitation Radar on board Tropical Rainfall Measuring Mission (TRMM) was collected. These case studies were also modelled using hydrostatic BOlogna Limited Area Model (BOLAM). In this work, we focus on relatively heavy precipitation event occurred on 5 March This was due to a cyclonic circulation slowly moving from western to eastern Mediterranean Sea (Tartaglione et al., 2005). Since GR data can be prone to distance-dependent errors, some range adjustment may be helpful. In this study, TRMM precipitation radar data were used to adjust GR data, employing a (physically easy to be interpreted) two-coefficient scheme. The forecast field was compared with both a RG

2 In order to verify precipitation forecast by BOLAM over Cyprus, both RG GR data, which are managed by Meteorological Service Cyprus, were used. The RG 86 M. Casaioli et al.: Fig. Radar 2. The adjusted 24-h network GR location are shown in Fig. 1. data observational versus modelled precipitation precipitation contours (mm) 2 Casaioli et al.: Radar adjusted data versus modelled precipitation contiguous rain area (CRA; Ebert McBride, 2000) were applied for verifying modelled precipitation. The paper is organized as follows: in section 2, observational data set radar range adjustment scheme are discussed; atmospheric model BOLAM is presented in section 3; a brief discussion applied verification methodologies is proposed in section 4; results conclusions are, finally, reported in section UTC 5 March to 0600 UTC 6 March 2003, obtained app Barnes scheme. isl, we decided to merge GR data with RG gri The GR raw data can be corrected /or adjusted, s some factors can alter beam response. GR measures from a lateral direction. In present case, distanc tween echo radar varies from 10 to 120 km. changes by a factor over 100, since volume incre. 2 Ground-radar rain gauge data cause this large ratio distances, scattering vo In order to verify precipitation forecast by BOLAM over with square distance. The scattering volum Fig. Fig. 1. Cyprus, 1. Geographical Geographical both distribution RG GRover over data, which isl isl arecyprus managed Cyprus by 147 rain Meteorological gauge stations, Service marked as white Cyprus, circles, were used. location The RG TRMM has a similar size in all locations. This advan 147 rain gauge stations, marked as white circles, location network ground-based radar GR location (34.98 ground-based radar (34.98 N, are32.73 shown E, N, in1310 Fig. m1. above mean Fig The TRMM 24-h observational suggestsprecipitation its use for contours estimating (mm) frominfluence o sea level), marked as a black circle. Orography is E, indicated 1310 minabove colourmean06: UTC 5 March to 06: UTC 6 March 2003, obtained applying GR sampling volume (Gabella et al., 2005). For both ra sea level), scale; visible marked in as figure a black is circle. TroodosOrography mountain range. is indicated in colour Barnes scheme. average reflectivity Z in same 10-km circular ri scale; visible in figure is Troodos mountain range. computed. Let us consider < GR > 2π < TRMM isl, we decided to merge GR data with RG gridded gridded an observed precipitation field obtained values average reflectivity, averaged in azimuth For selected event, precipitation recorded at. The Cyprus GR can provide information also on precipitationthe over both by merging RG data with rainfall retrieved from GR raw GR waters data surrounding can TRMM. be corrected Because isl /or adjusted, Cyprus. ir size, How- some factors ables proximity show can alter similar Troodos beambehaviour, response. high massif GR except measures in for south- rain decreased since se two RG range-adjusted network during GR data. 24 hours, from 0600 UTC 5 Marchever, to 0600 BothUTC non-parametric 6 Marchskill 2003 scores was (Hanssen considered. Kuipers, Sinceeastern from asitivity part lateral direction. GRTripylos with In distance. present peak incase, Averaging northwestern distanceover part between Fig. large RG 1965; givesschaefer, only point 1990; Wilks, measurements, 1995; Stephenson, a two-pass 2000) Barnes(see 1) echo produced rings two reduces radar obscured varies deviations from (no-data) 10 caused toradar 120 km. sectors. bybe- causeinan this attempt large ratio to obtain distances, a better observational scattering volume rain cells scheme contiguous (Barnes, rain 1964, area1973; Koch (CRA; etebert al., 1983) McBride, was usedthus, intensity. in order 2000) towere obtain applied a gridded for verifying 0.09 observational modelled precipitation. over : changes Cyprus by a factor its surrounding over 100, since sea, even volume considering increases that Thus, following factor: first observational The paper is organized astollows: be compared in Sect. 2, with observational Fig. 1. data Geographical set distribution radar range over adjustment isl scheme Cyprus are isl, TRMMwe hasdecided a similar much with precipitation square occurred distance. The souastern scattering volume part BOLAM was obtained by using only RG data (see Fig. 2). However, F = < GR tosize merge > in all GR data locations. with This RG advantage gridded 147 rain gauge stations, marked as white circles, location 2π discussed; atmospheric model BOLAM is presented in. TRMM suggests its use for estimating influence in this Sect. case, ground-based 3; a briefuse discussion radar only (34.98 RG N, applied for forecast E, 1310 verification verification m above mean methodologies is proposed for at least in Sect. two4; reasons. maygr The sampling GR raw< volume data TRMM can (Gabella > corrected 2π et al., 2005). /orfor adjusted, both radars, since sea level), marked as a black circle. Orography is indicated in colour be insufficient scale; visible in figure is Troodos results mountain conclusions range. are, some average factorsreflectivity can alter Z beam in response. same 10-km GRcircular measures ringrain is is statistically described by using a weighted regression First, finally, RGreported are only in Sect. available 5. on Cyprus western part from computed. a lateral Let direction. us consider In < GR present > 2π case, < TRMM distance> be- values echo average radar reflectivity, varies from averaged 10 toin120 azimuth, km. Be- for 2π tween log(f ) log(), where is distance betw coastal grid points (where observed precipitation is spreadtween For selected event, precipitation recorded at cause both GR this echo large TRMM. ratiobecause GR. distances, The ir best size, relationship, scattering se two volume variables show by asimilar factorbehaviour, over 100, except sincefor volume decreased increases sen- among by RG Barnes network procedure) during 24may hours, suffer froman0600 underestimation UTC 5 Marchwith found by Gabella et al. (2005), seems to be followin respect 2 toground-based to0600 internal UTC 6radar March grid points rain was gauge Second, considered. data precipitation Since changes RG gives only point measurements, a two-pass Barnes with sitivity square GR with distance. distance. Averaging The scattering over ( volume large area ) comparison is limited by presence surrounding sea. Inscheme order to(barnes, verify precipitation 1964, 1973; Koch forecast et al., by 1983) BOLAM wasover used TRMM rings has areduces similar size deviations in all caused locations. by rain This cells advantage high F The Cyprus GR can provide information also on precipitation over waters surrounding isl Cyprus. How- ( ) db 10 log(f )=a 0 + a 1 log = Cyprus, in order both to obtain RGa gridded GR0.09 data, observational which are managed : by intensity. TRMM suggests its use for estimating influence g first Meteorological observationalservice tocyprus, be compared were used. with The BOLAM GR Thus, sampling following volume (Gabella factor: et al., 2005). For both radars, RG ever, network wasproximity obtained by GR using location only Troodos are RGshown data high (see in massif Fig. Fig. 1. 2). in However, average south-easternfor part selected vent, Tripylos precipitation peak inrecorded northwestern at RG part 2π = < GR reflectivity > 2π Z in = same km 10.1 circular logring is, in this case, use only RG for forecast verification may computed. < TRMM Let us> consider 40 <GR> (1) 2π <TRMM> 2π be insufficient for at least two reasons. values average reflectivity, averaged in azimuth, for (see network Fig. 1) during produced 24 h, from two 06:00 obscured UTC 5(no-data) March to 06:00 radar UTC First, RG are only available on Cyprus western partsectors. both is statistically GR where TRMM. adescribed 0 Because aby 1 are using air weighted regression size, se regression coefficients two vari- show coefficient similar behaviour, for normalization except for distance. decreased sen- between log(f ) log(), where is distance between g Thus, 6 March in an attempt 2003, wastoconsidered. obtain a better Since RG observational gives only point coastal grid points (where observed precipitation is spread ables measurements, a two-pass Barnes scheme (Barnes, 1964, echo GR. The best relationship, among ones over Cyprus by Barnes its procedure) surrounding may suffer sea, aneven underestimation considering withthat sitivity It GRiswith possible, distance. thisaveraging way, to over use twolarge different area kinds 1973; Koch et al., 1983) was used in order to obtain a gridded found by Gabella et al. (2005), seems to be following: much respect 0.09 precipitation to internal observational occurred grid points. : Second, firstsouastern precipitation rings reduces deviations caused by rain cells high observational part data: original ( range-adjusted ) ones. The (ori comparison is limited by presence surrounding sea. intensity. to be compared with BOLAM was obtained by using only F The Cyprus GR can provide information also on precipitation over waters surrounding isl Cyprus. How- ( ) db 10 log(f )=a 0 + a 1 log = Thus, following factor: RG data (see Fig. 2). However, in this case, use only g RG for forecast verification may be insufficient for at least ever, proximity Troodos high massif in souastern = log, (2) two reasons. 40 part Tripylos peak in northwestern part F = < GR > 2π (1) (see First, Fig. RG1) are produced only available two obscured on Cyprus (no-data) western radar part sectors. where < a TRMM 0 a > 1 are 2π regression coefficients g is a coastal Thus, grid in anpoints attempt (where to obtain observed a better precipitation observational is spread coefficient for normalization distance. byover Barnes Cyprusprocedure) its surrounding may suffersea, an underestimation even consideringwith that is statistically It possible, described this way, bytousing use two a weighted differentregression kinds GR between data: log(f original ) log(), range-adjusted where is ones. distance The (original between respect muchtoprecipitation internaloccurred grid points. Second, souastern precipitation part comparison is limited by presence surrounding sea. echo GR. The best relationship, among ones

3 Casaioli et al.: Radar adjusted data versus modelled precipitation M. Casaioli et al.: Radar adjusted data versus modelled precipitation 87 found by Gabella et al. (2005), seems to be following: ( ) F db 10 log(f ) = a 0 + a 1 log = g = log ( ), (2) 40 where a 0 a 1 are regression coefficients g is a coefficient for normalization distance. It is possible, this way, to use two different kinds GR data: original range-adjusted ones. The (original adjusted) GR data were accumulated in aforementioned time interval, n, in order to make m available on a latitude-longitude grid (with grid spacing 0.09 ); remapping procedure (Baldwin, 2000; Accadia et al., 2003) was also applied. However, in order to choose best representative GR field, it was decided to compare precipitation retrieved from raw range-adjusted GR data with rain gauge ones. Fig. 3 shows results this comparison. A near zero correlation is found when original precipitation data is used in comparison with rain gauges (Fig. 3a), whereas a significantly (taking also into account its confidence interval) higher correlation is obtained by using precipitation reconstructed by range-adjusted GR data (Fig. 3b). In this latter case better agreement between se two fields appears more evident. Therefore, we shall hereafter consider as GR data range-adjusted ones. Then, observational (OBSPREC; Fig. 4) is improved by merging RG gridded field GR rangeadjusted data using following formula for each verification grid point: OBSPREC = (GR GR + RG RG) ; (3) (GR + RG) this way, when re is no contribution from RG (over sea, for instance), OBSPREC is given only by GR; same is true for RG, on grid points where GR is shielded by mountains, for instance. When estimations by both instruments were available, y were respectively weighted by ir values. (a) RG gridded field vs. non-adjusted GR field 3 Forecast precipitation The model used in this verification case study is BO- (b) RG gridded field vs. range-adjusted GR LAM, a hydrostatic primitive-equation model (Malguzzi field Tartaglione, 1999; Buzzi Foschini, 2000). The stard 6 h, 0.5 resolution, 60-model-level ECMWF analyses were first horizontally interpolated onto 30-km model domain Fig. 3. Precipitation scatterplots 24-h rain gauge gridded field Fig. 3. covering entire Mediterranean region. The outputs this against Precipitation 24-h radar scatterplots field over 73 grid points 24-h where rain bothgauge data were gridded model were used as initial boundary conditions (one-way field against available. Correlation 24-h radar (CORR.), fields over its confidence 73 grid interval points (Fisher, where both nesting) for a domain with a finer grid (10 km) encompassing only eastern Mediterranean region. val (Fisher, (MSE), 1925), root mean bias square (BIAS), error (RMSE) stardareerror indicated. (ST. A ERR.), linear re-meagression fit (solid line) a 95% confidence ellipse are also shown. data were 1925), available. bias (BIAS), Correlation stard error (CORR.), (ST. ERR.), mean its confidence square errorinter- Modelled precipitation was n 24-h accumulated (see square error (MSE), root mean square error (RMSE) are indicated. Fig. 5) for same time interval described in previous A linear regression fit (solid line) a 95% confidence ellipse are section. These data, originally on BOLAM native grid, also shown. were remapped (Accadia et al., 2003) on same grid adjusted) GR data were accumulated in aforemen-

4 for instance), OBSPREC is given only by GR; same is true for RG, on grid points where GR is shielded by mountains, for instance. When estimations by both instru- were available, y were respectively weighted 88ments didby M. not Casaioli exceed et al.: 39Radar mm adjusted (24h) data 1. versus modelled precipitation ir values. confidence ellipse (see Fig. 6) are related to higher values observed precipitation, mainly due to GR data; RG 4 Casaioli et al.: Radar adjusted data versus modelled precipitation gridded ere both ce inter-.), mean ndicated. llipse are oremenvailable 9 );., 2003) tive GR etrieved rain Fig The 24-h observational precipitation contours (mm) from were 0600 remapped (Accadia 0600 et al., 62003) 06:00 UTC 5 March to 06:00 UTC on 6 March 2003, 2003, same grid obtained according according observed to to Eq. Eq. 3. (3). precipitation field, which was used as verification grid. 3 Forecast precipitation The model used in this verification case study is BO- LAM, a hydrostatic primitive-equation model (Malguzzi Tartaglione, 1999; Buzzi Foschini, 2000). The stard 6h, 0.5 resolution, 60-model-level ECMWF analyses were first horizontally interpolated onto 30-km model domain covering entire Mediterranean region. The outputs this model were used as initial boundary conditions (one-way nesting) for a domain with a finer grid (10 km) encompassing only eastern Mediterranean region. Modelled precipitation was 24-h accumulated (see Fig. 5) for same time interval described in previous section. These data, originally on BOLAM native grid, Fig The 24-h BOLAM forecast field contours (mm) from 06: UTC 5 March to :00 UTC6 6 March Visual comparison between forecast observational observed (Fig. precipitation 6) displays field, a fair which agreement, was used although as verification correlation is still low. Almost all pairs outside 95% grid. confidence Visual comparison ellipse (seebetween Fig. 6) are forecast related to observational higher values (Fig. observed 6) displays precipitation, a fair agreement, mainly duealthough to GR data; RG cor- relation did not exceed is still 39 low. mm Almost (24h) 1 all. pairs outside 95% confidence ellipse (see Fig. 6) are related to higher values observed precipitation, mainly due to GR data; RG did not exceed 39 mm (24 h) 1. Fig. 6. As in Fig. 3, but for 24-h forecast against 24-h obtallied upon 2 2 contingency tables, summarize in a categorical way possible combinations forecast observed events above or below selected precipitation thresholds. The set thresholds, depending on event magnitude, does not include values greater than 40 mm (24h) 1. However, single case-study verification only by means se scores may lead to unstable results, due to limits statistical sample. Thus, CRA (Ebert McBride, Fig. 6. As2000), in Fig. 3, allowing but for to assess 24-h forecast spatial field against structure 24-h observational forecast error, field over was 532 alsogrid performed. points. This object-oriented method is simply based on pattern matching two contiguous areas, defined as observed forecast precipitation areas, delimited by a chosen isohyet, which was set to 0.5 mm (24h) 1. In a complex rain field (such as one inset question), thresholds, it is preferable depending to match on event rainfall magnitude, patterns instead not include precipitation values greater maxima, than 40thus mm (24correlation h) 1. max- does imization was employed as CRA pattern-matching criterion. However, The CRA single case-study is performed verification by shifting only by forecast means precipitation se scores field may both lead latitudinally to unstable results, longitudinally. due to limits However, statistical in order sample. to eliminate Thus, matches CRA that are (Ebert not statistically McBride, significant, 2000), allowing a threshold to assess minimum spatial correlation structure value between forecast shifted error, forecast was alsoobservations performed. was Thisimposed. object-oriented This value method depends is simply on based effective on pattern number matching independent two contiguousused areas, in defined comparison, as observed which is a function forecast precipita- num- samplebetionareas, grid points delimited where by a chosen isohyet, performed which(this was can set to change 0.5 mm from (24shift h) 1 to. shift) In a complex autocorrelation rain field (such as both one observed in question), forecast it is preferable fields. to The match F test, with rainfall a 95% patterns confi-indencstead level, was precipitation n applied maxima, to assess thus statistical correlation signifi- maxcancimization each wasmatch employed (see Pansky as CRA pattern-matching Brier, 1958; Xie criterion. The 1995). CRA is performed by shifting forecast Arkin, precipitation field both latitudinally longitudinally. Fig. 6. As in Fig. 3, but for 24-h forecast against 24-h observational field over 532 grid points. 4 Methodology In order to give a quantitative assessment model s skill in predicting event, non-parametric skill scores, such as ETS, BIA, HK, ORSS, PO FAR, were initially used. These scores (for details see Hanssen Kuipers, 1965; Schaefer, 1990; Wilks, 1995; Stephenson, 2000), which are 4 Methodology In order to give a quantitative assessment model s skill in predicting event, non-parametric skill scores, such as ETS, BIA, HK, ORSS, PO FAR, were initially used. These scores (for details see Hanssen Kuipers, 1965; Schaefer, 1990; Wilks, 1995; Stephenson, 2000), which are tallied upon 2 2 contingency tables, summarize in a categorical way possible combinations forecast observed events above or below selected precipitation thresholds. The 5 In Results order to eliminate conclusions matches that are not statistically significant, a threshold minimum correlation value between It shifted is now forecast possible to match observations obtained was results imposed. into This a coherent value depends picture. on Model effective skill scores number calculated independent with respect samples to RG used gridded in comparison, (i.e., which only over is a function isl Cyprus) number are fairly good (Table 1), whereas when including GR rangeadjusted data (i.e., also over sea) quite poor scores are grid points where is performed (this can change from shift to shift) autocorrelation both observed forecast fields. The F test, with a 95% confidence obtained (Table 2). level, was n applied to assess statistical significance Table each 1. match Skill(see scorepansky results comparison Brier, 1958; betweenxie BOLAM Arkin, forecast 1995). rain gauge for 5 March 2003 event. Skill score values less or equal zero are not indicated. Threshold BIA ETS PO FAR HK ORSS

5 Table 2. Skill score results comparison between BOLAM sults show that shifting 0.27 eastward, BOLAM forecast forecast observational field (rain gauges + radar) for gives a better match with observational (correlation increases from 0.36 to 0.43). M. 5Casaioli March 2003 et al.: event. Radar adjusted data versus modelled precipitation 89 Table Threshold 1. Skill scorebia results ETS comparison PO between FAR HKBOLAM ORSS forecast mm (24h) 1 rain gauge for 5 March 2003 event. Skill score values less or equal zero are not indicated Threshold 15.0 BIA ETS PO FAR HK ORSS mm (24 h) fields, 40.0 obtaining good skill scores when area is large ( more than one type observation is included) is very unlikely; whereas, it is easy to obtain good skill scores on smaller arear (in particular when only one type observation Table is included). 2. Skill score Thisresults could comparison also indicate between a major BOLAM impact forecast observational observational error (e.g., field (rain GR range-effect, gauges + radar) RGfor representativity 2003 event. error) on skill scores when using combined 5 March observational data set. Although we do not present results inthreshold this paper, itbia is worth ETS mentioning PO that FAR worst HK skill ORSS scores were mm (24 obtained h) 1 by using in observational composite (i.e., RG GR) original GR data (rar0.262 than those 0.230obtained by using 10.0 range-adjusted GR data) This result0.306 also appears in Fig (a), where precipitation scatterplot original GR20.0 data versus RG0.131 data exibits worst0.288 indicators 0.531(cor- relation 40.0 MSE) This confirms 0.162that radar range adjustment is a key procedure for having reliable GR data. Besides, BO- LAM tends to overforecast precipitation almost everywhere (see BIA score in Tables % confidence ellipse in Fig. 6), especially over sea. This could indicate that forecast rainfall qualitatively changes from l to 5 Results conclusions sea, but this may be proven only on basis a long-term It is now possible to match obtained results into a coherent picture. Model skill scores calculated with respect to statistical verification BOLAM over Cyprus. Furrmore, model skill scores may be penalized by RG gridded (i.e., only over isl Cyprus) forecast oversmoothness are fairly good (Table 1), 1. In fact, many details that are visible in over sea are not present in forecast whereas when including GR range-adjusted field (see in data Fig. (i.e., 4 also rainfall over peaks sea) located quite south, poor scores southwest are obtained north-east (Table 2). isl, respectively, compare There with Fig. are many 5). However, reasons such whichdiscrepancies could account ( forir sucheffect differences on skill inscores) skill scores. cannotfor be example, unequivocally whenattributed comparing to forecastobtaining error, since good skill observational scores when error (namely, area is large ( radar two fields, more range-effect) than one type may sensibly observation affect is estimation included) is maximum very unlikely; rainfall whereas, peaks. it is easy to obtain good skill scores on smaller For arear what (inconcerns particular when CRAonly results, one type whenusing observation RG is included). gridded This result for verification, could also indicate a betteraagreement major impact (using correlation observational maximization) error (e.g., between GR range-effect, observations RG representativity fieldserror) is obtained on by skill shifting scores when forecast using field 0.27 combined E forecast observational 0.09 N (seedata Fig. set. 7(a) Although against Fig. we2do not Tartaglione present t results 2005). in this This paper, result it is worth quite similar mentioning to that one obtained worstby skill us- al., scores ing were RG-GR obtained composite by usingfor verification. observational In fact, composite CRA re- (i.e., RG + GR) original GR data (rar than those obtained by using 1 Numerical models oversmooth precipitation fields with respect range-adjusted GR data). This result also appears in to reality (Chéruy et al., 2004) Fig. 3a, where precipitation scatterplot original GR data versus RG data exibits worst indicators (correlation MSE). This confirms that radar range adjustment is a key procedure for having reliable GR data. Besides, BO- LAM tends to overforecast precipitation almost everywhere (see BIA score in Tables % confidence ellipse in Fig. 6), especially over sea. This could indi- (a) Forecast shifted 0.27 E 0.09 N with respect to Barnes rain gauge-based (b) Forecast shifted 0.27 E 0 N with respect to RG- GR composite Fig. 7. The 24-h BOLAM forecast field contours (mm) (mm) from from 06: UTC 5 Marchtoto :00 UTC6 6 March 2003, as as result result CRA. An a priori estimate impact possible observational that error forecast on se rainfall resultsqualitatively is not possible, changes thus we from propose l to cate sea, a subjective but this may checkbeproven only on (Fig. 4) basis against a long-term original shifted verification forecast BOLAM (Figs. 5 over 7(b)). Cyprus. The idea is to statistical identify Furrmore, qualitative model features skill scores may be penalized illustrated by in Fig. 4 which are less prone to possible observational error, forecast oversmoothness 1. In fact, many details that are visible in over sea are not present in forecast to check m against both non-shifted shifted model fields. For instance, it is quite improbable that radar field (see in Fig. 4 rainfall peaks located south, southwest north-east isl, respectively, compare errors introduce a macroscopic pattern shifting. The gross structure rainy area on isl s southwestern with Fig. side 5). However, non-rainy such discrepancies area on ( northwestern ir effect on sideskill is better scores) matched cannotby be unequivocally shifted forecasts attributed than by to forecast original error, one. since Over observational sea, checkerror is difficult (namely, since a few radar range-effect) may sensibly affect estimation maximum rainfall peaks. For what concerns CRA results, when using RG gridded for verification, a better agreement (using 1 Numerical models oversmooth precipitation fields with respect to reality (Chéruy et al., 2004).

6 90 M. Casaioli et al.: Radar adjusted data versus modelled precipitation correlation maximization) between observations forecast fields is obtained by shifting forecast field 0.27 E 0.09 N (see Fig. 7a against Fig. 2 Tartaglione et al., 2005). This result is quite similar to one obtained by using RG-GR composite for verification. In fact, CRA results show that shifting 0.27 eastward, BOLAM forecast gives a better match with observational (correlation increases from 0.36 to 0.43). An a priori estimate impact possible observational error on se results is not possible, thus we propose a subjective check (Fig. 4) against original shifted forecast (Figs. 5 7b). The idea is to identify qualitative features illustrated in Fig. 4 which are less prone to possible observational error, to check m against both non-shifted shifted model fields. For instance, it is quite improbable that radar errors introduce a macroscopic pattern shifting. The gross structure rainy area on isl s southwestern side non-rainy area on northwestern side is better matched by shifted forecast than by original one. Over sea, check is difficult since a few observed peaks, probably due to convective activity, are absent in BOLAM forecast ( hydrostatic model did not forecast m at all!). However, position main rainy area, located to northwest isl, is better centred in shifted forecast than in original one. The proposed results are a starting point for future works. The effect forecast oversmoothness on skill scores could be tested. Besides, effect different possible range adjustment techniques on observational could be also studied, in order to estimate reduce observational error to assess physical reliability CRA results. Overall, results point out how much caution is needed when dealing with rainfall observations as ground truth. When satellite, rain gauge radar data are simultaneously available over a large area, a reliable data base suitable for precipitation forecast verification can be built. Orwise, model verification task can be quite difficult, since it cannot be easy to distinguish effect on verification results forecast error from one related to observational error. Moreover, even a rough estimate observational error magnitude cannot be available. In such a condition, exemplified in here presented case study, both forecast should be regarded as an error-affected representation an unknown reality. Acknowledgements. This research was supported by EU Framework 5th Programme Part A: Environment Sustainable evelopment Contract EVK CT (VOLTAIRE project). Authors are grateful to ECMWF for initial boundary conditions for BOLAM. Comments suggestions received during VOLTAIRE workshops were very useful. References Accadia, C., Mariani, S., Casaioli, M., Lavagnini, A., Speranza, A.: Sensitivity precipitation forecast skill scores to bilinear interpolation a simple nearest-neighbor average method on high-resolution verification grids, Wea. Forecast., 18, , Baldwin, M.: Quantitative Precipitation Forecast Verification ocumentation, Tech. rep., NCEP/EMC, mmb/ylin/pcpverif/scores/docs/mbdoc/pptmethod.html, Barnes, S. L.: A technique for maximizing details in numerical wear map, J. Appl. Meteorol., 3, , Barnes, S. L.: Mesoscale objective using weighted timeseries observations, NOAA, National Severe Storm Laboratory, Norman, OK 73069, tech. memo. ERL NSSL-62, Buzzi, A. Foschini, L.: Mesoscale meteorological features associated with heavy precipitation in sourn Alpine region, Meteorol. Atmos. Phys., 72, , Chéruy, F., Speranza, A., Sutera, A., Tartaglione, N.: Surface winds in Euro-Mediterranean area: The real resolution numerical grids, Ann. Geophys., 22, , Ebert, E. E. McBride, J. L.: Verification precipitation in wear systems: determination systematic errors, J. Hydrol., 239, , Fisher, R. A.: Statistical Methods for Research Workers, Oliver Boyd, Edinburgh (UK), Gabella, M., Joss, J., Perona, G., Michaelides, S.: Range adjustment for Ground based Radar, derived with space-borne TRMM Precipitation Radar, IEEE Transactions on Geoscience Remote Sensing, 44, 1, , doi: /tgrs , Hanssen, A. W. Kuipers, W. J. A.: On relationship between frequency rain various meteorological parameters, Meded. Verh., 81, 2 15, Koch, S. E., desjardins, M., Kocin, P. J.: An interactive Barnes objective map scheme for use with satellite conventional data, J. Clim. Appl. Meteor., 22, , Malguzzi, P. Tartaglione, N.: An economical second order advection scheme for numerical wear prediction, Quart. J. Roy. Meteorol. Soc., 125, , Pansky, H. A. Brier, G. W.: Some applications statistics to meteorology, Pennsylvania State University, University Park, Schaefer, J. T.: The critical success index as an indicator warning skill, Wea. Forecast., 5, , Stephenson,. B.: Use Odds Ratio for diagnosing forecast skill, Wea. Forecast., 15, , Tartaglione, N., Mariani, S., Accadia, C., Speranza, A., Casaioli, M.: Comparison raingauge observations with modeled precipitation over Cyprus using contiguous rain area, Atmos. Chem. Phys., 5, , Wilks,. S.: Statistical methods in atmospheric sciences: An introduction, Academic Press, San iego (USA), Xie, P. Arkin, P. A.: An intercomparison gauge observations satellite estimates monthly precipitation, J. Appl. Meteorol., 34, , Edited by: V. Kotroni K. Lagouvardos Reviewed by: anonymous referee