1

Background Error Modeling for WRF-Var using the NMC method

3 Hongli Wang* and Xiang-Yu Huang and Juanzhen Sun

4 National Center for Atmospheric Research, Boulder, CO, USA

6 Dongmei Xu

7 Nanjing University of Information Science and Technology, Nanjing, China

9 Shuiyong Fan and Jiqin Zhong

10 Beijing Meteorological Bureau, Beijing, China

11

12 Man Zhang

13 University of Colorado, Boulder, CO

14

15

16

17 Submitted to JAMC

18 Aug 2013

19 Corresponding author:

20 Dr. Hongli Wang

21 hlwang@ucar.edu

22 MMM/NCAR

23 Now affiliated with CIRA/CSU.

1
24 ABSTRACT

25

26 Background error modeling plays a key role in a variational data assimilation system. The

27 NMC method has been widely used in variational data assimilation systems to model

28 climatological background error covariances. In this paper, the features of background

29 error modeling via the NMC method are investigated for the variational data assimilation

30 system of the Weather Research and Forecasting Model (WRF-Var). The background

31 error statistics are extracted from the short-term 3km resolution forecasts in June, July

32 and August 2012. It is found that (1) background error variances vary month to month

33 and also have a feature of diurnal variations in low-level atmosphere; (2) u- and v-wind

34 variances are underestimated and their auto-correlations length scales are overestimated

35 when the default control variable option in the WRF-Var is used.

36

37 Two additional control variable transforms (CVT) are proposed to form the background

38 error covariance matrix via the NMC method. One uses u-, v-wind, temperature, surface

39 pressure and pseudo relative humidity as control variables, and the other uses so called

40 alpha control variables to construct the background error covariance matrix. Single

41 observation assimilation experiments show that the two proposed methods give good

42 error variance modeling for u-, v-wind using the NMC method, and the alpha control

43 variable has the merits of incorporating geographically dependent covariance

44 information and producing multi-variate analysis. The preliminary results from data

45 assimilaton and forecast study for a real convective case shows that the use of the new

46 CVTs improve the wind and precipiation forecast up to 12 hours.

2
47

48 1. Introduction

49 The background error (BE) covariance matrix plays a key role in a variational data

50 assimilation system by weighting the importance of a priori state, by smoothing and

51 spreading information from observation points, and by imposing balance between the

52 model control variables (Daley 1991; Bannister 2008a,b). However, the estimation of the

53 BE statistics is not straightforward, since the truth is not known. Two methods are mainly

54 used in current data assimilation systems. The so-called NMC (named for the National

55 Meteorological Center, now called the National Centers for Environmental Prediction)

56 method (Parrish and Derber 1992) is one approach that is widely employed to estimate

57 the BE covariances. This method uses the differences between forecasts of different

58 lengths, but valid at the same time, to evaluate the short-range forecast errors. An

59 alternative method is to use an ensemble of short-term forecasts at a specific time to

60 evaluate the BE covariances (Houtekamer et al. 1996; Fisher 2003).

61 Various control variable transforms (CVTs) have been used in variational data

62 assimilation systems to model multivariate and univariate aspects of the BE covariances

63 approximately in a compact and efficient way (Bannister 2008a,b). Three kinds of BE

64 modeling for wind analysis are widely used in variational data assimilation. Vorticity and

65 (unbalanced) convergence/divergence are used in the data assimilation systems at

66 ECMWF (Courtier et al. 1996) and Meteo France (Fischer et al. 2005). Streamfunction

67 and unbalanced velocity potential are widely used as control variables in global data

68 assimilation systems, and some regional data assimilation systems (e.g. Ingleby 2001;

69 Barker et al. 2004, 2012; Zupanski 2005; Rawlins et al. 2007; Huang et al. 2009; Wang et

3
70 al. 2013). Whereas, velocities are employed as control variables in data assimilation

71 systems for mesoscale and convective scales (e.g. Zou et al. 1995; Sun and Crook 1997;

72 Gao et al. 1999; Zupanski et al. 2005; Kawabata et al. 2011). The velocity control

73 variables may be more suitable for mesoscale and convective scale data assimilation

74 since past theoretical analysis found that velocity control variables could combine the

75 background and observations for all scales (Xie et al. 2002; Xie and MacDonals 2011).

76 Assumptions are made to model BE covariances in an efficient and affordable way since

77 the BE matrix is of high dimensions. The present numerical weather prediction model

78 uses a large dimensional space, typically 107 dimensions or more, and so the BE matrix

79 has 1014 elements, which cannot be explicitly modeled. In practice, it is usually assumed

80 that the BE covariances are nearly homogeneous and isotropic. For example, the

81 variational data assimilation system of the Weather Research and Forecasting (WRF)

82 model (WRF-Var) assumes that the BE covariances are homogeneous and isotropic. It is

83 noted that both choice of CVT and assumptions made to model BE covariances as

84 mentioned above could have impact on extracting BE information from forecast

85 examples using either the NMC or the ensemble method.

86 The WRF-Var system has been extensively used in the research community and

87 operational centers (Barker et al. 2012; Huang et al. 2013). For examples, WRF-Var was

88 adopted in the Rapid Update Cycling Data assimilation and Forecasting System at

89 Beijing Meteorological Bureau (BJ-RUC; Chen et al. 2009), which has been run in

90 operation since June 2008. The WRF-Var system with radar data assimilation has shown

91 consistently positive impact on precipitation prediction (Wang et al. 2013a). The

92 operational application of the WRF-Var system at Taiwans Central Weather Bureau

4
 93 (CWB) has significantly reduced typhoon track forecast errors (Hisao et al. 2012).

94 The NMC method has been employed in the specification of BE statistics for WRF-Var.

95 WRF-Var is the basic component of the WRF models community data assimilation

96 system (WRFDA; Barker et al. 2012). Barker et al. (2004) suggested to apply empirical

97 tuning factors to the length scales calculated via the NMC method (ranging between 0.5

98 and 1). Previous studies (e.g. Xiao and Sun 2007; Sugimoto 2009; Li et al. 2012; Sun et

99 al. 2012; Wang et al. 2013a) showed that radar radial velocity data assimilation using

100 WRF-Var system with reduced lengthscales improved analyses and forecasts. These

101 provide the motivation to investigate the features of BE modeling via the NMC method to

102 further improve the performance of the WRF-Var system. Moreover, investigations on

103 BE modeling for WRF-Var will also benefit WRFDA.

104 In WRF-Var, stream function and velocity potential (CV option 5; CV5) are the default

105 option to produce wind analyses. In this paper, background error modeling for u-wind

106 and v-wind using CV5 are investigated, and two additional CVTs are proposed to form

107 the climatological background error covariance matrix via the NMC method. Up to the

108 authors knowledge, this is the first work using the two proposed CVTs to study the

109 climatological BE modeling in context of the WRF-Var system.

110 This paper is organized as follows. Section 2 provides a description of the NMC method

111 and features of BE statistics over Beijing region. Two new CVTs are introduced to use

112 climatological background errors with the NMC method in section 3. Single observation

113 data assimilation experiments using several CVTs are presented in section 4. The impacts

114 of the new CVTs on real data assimilation and forecast for a convective case occurred in

5
115 Beijing is presented in section 5. A summary and discussion is given in the final section.

116 2. Background error modeling

117 2.1 The NMC method

118 A common method to model the BE covariance matrix is to take the difference between

119 pairs of forecasts of different lead times but each valid at the same time (Parrish and

120 Derber 1992). Forecast differences are usually calculated over a reasonably long period

121 of time (e.g. a month). This makes the NMC method suitable for climatological forecast

122 error statistics. In WRF-Var, the background error covariance matrix may be considered

123 the following expression

124 B (x 24 x12 )(x 24 x12 )T (1)

125 where x 24 and x12 are 24 h and 12 h forecasts respectively valid at the same time. The

126 overbar denotes an average over time and/or space. The two forecasts can be written in

127 terms of truth and their errors

128 x 24 = x truth + 24 + b 24 (2a)

129 x12 = x truth + 12 + b12 (2b)

130 Here, x truth is the true atmospheric state at the valid time. 24 and 12 are the random

131 errors, and b 24 and b12 are the biases in each forecast. Assuming there is no bias or the

132 bias is constant in time, b 24 = b12 , the forecast difference is

133 x diff = x 24 x12 = 24 12 (3)

134 The BE covariance matrix is written as

135 B (x diff )(x diff )T

6
136 = ( 24 12 )( 24 12 ) T

137 = ( 24 )( 24 ) T + (12 )(12 ) T ( 24 )(12 ) T (12 )( 24 ) T (4)

138 It is seen that the BE modeling using Eq.1 including three parts: 24 h BE, 12 h BE and
139 their correlations. In a real data assimilation and forecast system, analyses are usually
140 update every 6/3 hours. It indicates that the background error covariance generated by
141 forecast differences with longer lead times needs to be tuned in real applications.
142
143 2.2 Error variance estimation using the BJ-RUC operational forecasts
144 Before discussing the BE modeling, it is crucial to check the features of error variance in
145 raw dataset estimates using Eq. 4. The BE modeling should reproduce those features as
146 much as possible. Operational forecasts produced by the BJ_RUC system (Chen et al.
147 2009) during the period of 1 June-31 Aug 2012 are used to calculate short-term BE
148 statistics in this paper. The statistics are over the inner domain (Fig.1) with 3 km grid
149 spacing. The forecast differences between 24 h and 12 h forecasts valid at the same times
150 are employed to model the BE statistics. Hereafter, the forecast differences between
151 forecasts with different lead times but valid at the same times are names as the NMC
152 ensembles in this paper.
153
154 Horizontal distributions of standard deviation of the error variances at three model levels
155 are shown in Fig. 1. It is obvious that even over the inside small domain (about
156 1300x1600 km2) the variance distributions show geographically dependent structures.
157 These features are significant for wind at about the 30th model level, for temperature at
158 about 11th model level. It is noted that the WRFDA uses a domain averaged error
159 variance, thus does not account for these location dependent features. As shown in the
160 next section, an effective yet efficient scheme is poposed for the location dependent error
161 variance modeling for WRFDA. In general, the value of the standard deviation of the
162 wind error is about 3 m s-1 at the 11th and 25th model levels. The value increases to be
163 about 5 m s-1 in the extratropical jet (33rd) level. For temperature, the value is about 1 K
164 in most domain. For relative humidity, the values of the error are between 10-20%. The

7
165 vertical structures of the standard deviation of the error variances are depicted in Fig. 2. It
166 is seen that the error of wind and temperature (Fig. 2a, 2b and 2c) exhibit two maximum
167 regions, one below the 11th model level (approximate 850 hPa), and the other is around
168 30th model level (approximate 250 hPa). Fig.1 and 2 also reveal that the errors for u, v
169 and T in the north region of the model grid are larger than in the south region below the
170 10 th model level. This is related the surface observation density in BJ-RUC as will be
171 discussed in section 5.
172
173 The monthly variations of background errors are found. Figure 3 depicts the BE variances

174 for u , v , T , and RH . It is seen that the BE variances for wind and temperature in June

175 and July are relatively larger than in August. The diurnal variations of forecast error near

176 surface are clearly shown in Fig. 4. The error variances for wind and temperature in the

177 low atmosphere in the evening 1200 UTC (local time 20Z) are larger than those in the

178 morning 0000 UTC (local time 08Z). The above results indicate that even with the

179 climatological BE statistics the time dependent variances can be achieved. Surface

180 observations are important data sources for a regional rapid update cycle data

181 assimilation system. The above results indicate that BE covariances accounting for the

182 diurnal variation and geographically dependent error variance may benefit the surface

183 data assimilation.

184
185 2.2 B modeling in WRF-Var

186 The BE covariance estimated by Eq.4 is not directly used in a variational data

187 assimilation system. The CVTS and some assumptions are made to model BE covariance

188 in an efficient and affordable way. In the WRF-Var system, a control variable transform

189 x = Uv is used to model background errors. v represents control variable vector and

190 x stands for analysis increment vector. The U transform maps control variables from

8
191 control space to analysis space. The CVT x = Uv is implemented through a series of

192 operations x = U p U v U h v (Barker et al. 2004). The default control variables (CV option

193 5; CV5) in WRF-Var includes the streamfunction , the unbalanced part of velocity

194 potential u , the unbalanced part of temperature Tu , the unbalanced part of surface

195 pressure Psu , and pseudorelative humidity RH . The term unbalance refers to the

196 residual from the balance with the streamfunction.

197 The operators U p , U v and U h are described in Barker et al. (2004) in details. The

198 horizontal transform U h is to model the auto-correlation of a control variable using

199 recursive filters. The horizontal correlations are assumed to be homogeneous (i.e. not

200 dependent on geographic position) and isotropic for each control variable. The vertical

201 transform U v is performed via an empirical orthogonal function (EOF) decomposition of

202 the vertical component of BE on model levels. The variances and vertical correlations of

203 each control variable are modeled in this stage. In the default CV5, the time- and domain-

204 averaged vertical component of the BE is used indicating that the variances and vertical

205 correlations are constant on each model level and do not depend on geographic positions.

206 In addition, the horizontal correlation for each control variable is not model in the

207 physical space but in the EOF spaces. The physical variable transform U p involves

208 balance transform and conversion of control variables to analysis variable increments.

209 The statistical balance transform is applied in this stage, defined by

9
 I 0 0 0 0

C , I 0 0 0 u
210 T = C 0 I 0 0 Tu (5)
T ,
Ps C Ps , 0 0 I 0 Psu
RH
0 0 0 0 I RH

211 where I is the identity matrix, and C , , CT , and C Ps , stand for statistical regression

212 matrixes between , T , Ps and . The analysis variables of u-wind ( u ), v-wind ( v ),

213 and specific humidity ( q ) can be obtained by a transform as follows.

u Cu , Cu , 0 0 0


v Cv , Cv , 0 0 0

214 T = 0 0 I 0 T
0 (6)

Ps 0 0 0 I 0
Ps
q 0 0 0 0 Cq , rh RH

215 Cu , , C u , , C v , , C v , and C q, rh map variables , , T , Ps and RH to analysis

216 variables u , v , T , Ps and q . Noted that temperature and pressure are required to

217 obtain C q, rh .

218

219 2.3 Features of B variance and correlation modeling

220 It is seen that the CVT and several assumptions are taken to model the BE matrix

221 approximately in a compact and efficient way. Practically CVT (CV5) is used to model

222 background errors for WRF-Var. A natural question to ask is: what information is

223 lost/filtered out by the CVT and the above assumptions? We compare the error standard

224 deviation in the NMC ensembles to that modeled by WRF-Var using CV5 to answer the

225 question.

10
226

227 The standard deviation in the three-month forecasts directly estimated by Eq. (1) without

228 CVT is named as STD_NMC. The standard deviation derived from B that is generated

229 by WRF-Var gen_be utility is named as STD_CV5. This is achieved through sampling

230 the B matrix in control variable space and then computing statistics in analysis space

231 after CVT (Andersson et al. 2000). 200 samples are taken in this paper. Single

232 observation experiment is another alternative method to show BE variance in observation

233 space, which will be used in section 3.

234

235 The vertical profiles of STD_NMC and STD_CV5 are plotted in Fig. 5. It is seen that

236 STDs of all the variables are underestimated and especially for u , v . In the BJ_RUC

237 system for radar data assimilation, the lengthscales are tuned to be half of the original

238 ones. With tuned B (CV5), the u , v variances (STD_CV5_L05 in Fig. 5) are

239 comparable to STD_NMC. This indicates the CVT in WRF-Var may contribute to the

240 significant underestimates of the background error STD of winds.

241

242 Horizontal correlations for u (black curves) and v (red curves) in the NMC ensembles,

243 modeled by CV5 at 25th model level (approximate 500 hPa surface) along east-west

244 direction and north-south direction are shown in Fig. 6. The correlations using the NMC

245 ensembles are directly calculated in u and v spaces so that they can be used as time- and

246 domain averaged statistics reference for CV5. It is seen that CV5 slightly overestimates

247 the correlation length for u along east-west direction and significantly overestimates the

248 correlation length along north-south direction. A possible reason is that with the

11
249 climatological BE statistics one gets an average correlation model that is not the optimal

250 one for any particular situation. Other possible reasons for the overestimation of

251 correlation length for u and v are that the horizontal correlation modeling is carried out on

252 EOF (decomposed from time-averaged vertical covariance) spaces and only one Gaussian

253 fitting is used to model the pattern of horizontal correlations. There is a noticeable

254 negative correlation for u in south-north direction, and for v in weat-east direction. This

255 feature is modeled by CV5, however CV5 produces a larger negative correlation.

256

257 3. New CVTs for B modeling

258 In this paper, two CVTs are proposed to account for background errors statistics in the

259 NMC ensembles. A natural choice is to use u and v as control variables for wind

260 analyses since u and v have been used as control variables in data assimilation systems

261 for mesoscale and convective-scale forecasts (e.g. Zou et al. 1995; Sun and Crook 1997;

262 Gao et al. 1999; Zupanski et al. 2005; Kawabata et al. 2011).

263 3.1 New CVT: CV7

264 The new formulation, which uses u , v , T , Ps and RH s (pseudo relative humidity) as

265 control variables, is developed in WRF-Var for mesoscale and convective-scale data

266 assimilation. The new control variable transform can be written as

267 x = U p 2Uv U h v 2 (7)

268 We followed WRF-Var procedure to use recursive filter and EOFs to model horizontal

269 and vertical correlations respectively which are implemented through U v U h . The

270 homogeneous and isotropic filters, which are used for CV5, are applied to each control

271 variable. In addition, the time- and domain-averaged vertical component of BE is used

12
272 indicating that the BE statistics do not depend on geographic position. It is noted that

273 U p 2 in Eq.7 only involves conversion analysis increments in relative humidity to specific

274 humidity as shown in Eq.6. No statistical balance transform (Eq.5) is applied in this

275 transform.

276

277 The variance and correlation modeling for u and v using CV7 are investigated. The STD

278 derived from the BE matrix using CV7 is shown in Fig. 5 (STD_CV7). It is shown that

279 the use of u and v , which are WRF model prognostic variables, as control variables

280 gives a good STD modeling. However, the horizontal correlation lengths are

281 underestimated (Fig. 6). The possible reason for the underestimation of correlation length

282 for u and v are that the horizontal correlation modeling is carried out on EOF

283 (decomposed from time-averaged vertical covariance) spaces and only one Gaussian

284 fitting is used to model the pattern of horizontal correlations. It is noted that to better

285 model the anisotropic error correlation for u and v anisotropic filters need to be developed

286 in the WRF-Var.

287

288 3.2 New CVT: alpha control variable

289 Another possible approach that directly uses the NMC ensembles to form the BE

290 covariance matrix is described here. The analysis increment is expressed in a subspace

291 expanded by the NMC ensembles,

K
292 x' = ( a k x kd ) (8)
k=1

13
293 where K is the total number, and the vector x kd ( k = 1, K ) is the kth unbiased the NMC

294 ensemble normalized by K1/2.

295 (
xdk = xdiff
k )
x / K (9)

296 In practice, the time-averaged bias x is removed from the forecast differences. The

297 vector a k stands for the augmented control variables for the kth forecast difference. a k

298 will be called alpha control variable hereafter in this paper. The symbol denotes the

299 Schur product of the vectors a k and x kd . Let X' = (x1d , x 2d ,..., x kd ,..., x dK ) , it is obvious that

300 (X' X' )! S is the B covariance matrix defined in Eq.1 but with covariance localization
T

301 defined by S = a k (a k )T .

302

303 The transform (Eq.8) has been developed in the WRF hybrid ensemble-3DVar data

304 assimilation system. The readers are referred to Wang et al. (2008) for details. In their

305 scheme, an ensemble of forecast perturbations was used to incorporate flow dependent

306 error covariance of the day. We adopt this idea but use the so called NMC ensembles

307 instead of real ensemble perturbations to form the BE covariance matrix.

308

309 Localization schemes are required to mitigate the sample noise. In the WRF hybrid

310 ensemble-3DVar system, both horizontal and vertical localization can be applied.

311 Specifically, the horizontal and vertical correlation localizations are implemented through

312 recursive filters (Wang et al. 2008) and vertical correlation matrix respectively. A general

313 formulation to form the vertical covariance matrix can be written as

14
 d (l1 , l2 ) 2
314 Cov (l1 , l2 ) = exp( )
D(l1 ) 2 (10)

315 where Cov(l1, l2 ) represents the correlation between model levels l1 and l2 . d is the

316 distance in a specified coordinate between model level l1 and l2 and D stands for the

317 level-dependent vertical localization radius.

318

319 The default vertical correlation matrix in WRF-Var is defined in model level space,

l1
320 specifically, d (l1, l2 ) = l2 l1, D(l1 ) = 10 , N is the total number of model levels.
N

(l2 l1 ) 2
321 Cov(l1 , l2 ) = exp( ) (11)
l1 2
(10 )
N

322 It is seen that the level-dependent localization radius D(l1 ) only depends on the number

323 index of model level indicating that an observation at model level with large number

324 index will be widely spread in vertical direction. This may reduce the impact of

325 observations that are located in low model levels.

326

327 In addition to the above formulation (Eq.11), a specific application of Eq.10 in height

328 coordinate is also tested

( Z (l2 ) Z (l1 ))2

329 Cov (l1 , l2 ) = exp( ) (12)
(Z (l1 ))2

330 Z is domain-averaged height at a model level. For simplicity, a constant vertical

331 localization radius Z (l1 ) is used as done by Li et al. (2012). The above two vertical

332 localization specifications (Fig. 7) will be tested in next section.

15
333

334 In summary, we proposed two CVTs to incorporate climatological B with the NMC

335 method in WRF-Var. The first one uses u , v , T , Ps and RH s as control variables,

336 which is named as CV7 in WRF-Var. The other "alpha" control variable approach can

337 provide geographic location dependent BE covariance modeling, which will be clearly

338 shown in the single observation data assimilation experiments in next section.

339

340 4. Single observation data assimilation experiments

341 The B matrix weights the background state and spreads out observation information in

342 horizontal and vertical directions in space. Increments from single observation

343 experiments can be used to estimate BE variance and demonstrate how the BE covariance

344 spreads the observation information spatially, which give a graphic representation of the

345 BE structure function (Huang et al. 2009; Gustafsson et al. 2012). The increment x a at

346 the single observation location can be expressed in a scaled form

347 x a = ( d ) /( + e) (13)

348 where the scale d is innovation, e is the observation error variance, is the BE

349 variance in observation space. Given d , e , and x a are known, the BE variance can

350 be easily derived from Eq. (13).

e x a
351 = (14)
d x a

352

353 4.1 Experimental design

16
354 To better understand the differences in BE representations in the three formulations,

355 single observation data assimilation experiments are conducted to show the BE

356 covariance structures ( B ). First the experiments (Table. 1) assimilating a single u

357 observation are presented. The innovation is 1.0 m s-1, and the observation error is 1.0 m

358 s-1. The variance scaling factor for each control variable is 1.0 which is the default value

359 in WRF-Var. The single observation is set either at location (271,212,25) or at location

360 (364,150,25) in model grid. The two locations are named as P1 and P2 hereafter.

361

362 The experiments U-CV5-P1 and U-CV5-P2-L05 are designed to see impact of CV5 BE

363 and BE tuning on analysis increments. The length scaling factors with values of 0.5 are

364 adopted from BJ_RUC for radar data assimilation (Wang et al. 2013a). Compared to the

365 above two experiments, two experiments U-CV7-P1 and U-CV7-P2 are carried out to

366 show impact of CV7. CV7 only produces univariate analyses. An incremental diabatic

367 digital filtering initialization (DFI) scheme (Lynch and Huang 1994) is used to spread

368 observation information in space and develop multi-variate analysis increments. An

369 experiment named U-CV7-P2-DFI (Table. 1) is carried out in which analysis increments

370 are from experiment U-CV7-P2. The model carries out a 30 minutes backward adiabatic

371 integration and a 60 minutes forward diabatic integration with starting time at 0000 UTC

372 July 21 2012.

373

374 Four experiments with using the alpha control variable are conducted. The first two

375 experiments are without vertical localization and the other two with vertical localization

376 with the purpose to check the influence of the vertical localization scheme. In WRF-Var,

17
377 the vertical localization is not taken in the default option. Along with U-alpha-L300-P1,

378 U-alpha-L300-P2 is used to investigate how much geographical position dependend BE

379 can be achieved.

380

381 4.2 Analysis increments

382 In this subsection, first the horizontal and vertical structures of u analysis increments

383 will be examined. Then the multivariate features of analysis increments in selected

384 experiments will be described.

385

386 4.2.1 Structures of u increments

387 The horizontal the u increments on the 25th model level (approximate 500 hPa pressure

388 level) are shown in Fig. 8. The vertical south-north section of the u increments across

389 the single observation location are shown in Fig. 9. The maximum values of the u

390 analysis increments in U-CV5-P1 and U-CV7-P1 are 0.66 (Fig. 8a and Fig. 9a) and 0.89

391 m s-1 (Fig. 8c and Fig. 9c) respectively. The corresponding BE standard deviations

392 derived using Eq.14 are 1.39 and 2.84 m s-1, respectively, indicating that CV5

393 underestimates wind variance more than CV7. This result is consistent with the BE

394 standard deviation estimations presented in section 2.2 (Fig. 5a). It is noted that the four

395 experiments using the alpha control variable produces almost the same value of the

396 maximum analysis increments (Fig. 8e,f,g,h and Fig.9e,f,g,h) as that of U-CV7-P1 (Fig.

397 8c and Fig. 9c). The two new formulations provide a consistent BE variance modeling.

398

18
399 Comparing the horizontal spread of observation information in Fig. 8a to Fig. 8c or Fig.

400 9a to Fig. 9c, it is found that CV5 produces a widest spread of observation information

401 that is consistent with the correlation in Fig. 6. By reducing lengthscale by a half, the

402 maximum value of the u increment in U-CV5-P2-L05 (Fig. 8b and Fig. 9b) is increased

403 to 0.88 m s-1 which is almost same as those in U-CV7-P2 (Fig. 8c and Fig. 9c). The

404 results confirm that reducing the lengthscale increases the wind variance in Fig. 5a,b.

405 Increasing the gradients of streamfucntion and velocity potential by reducing length

406 scales leads to the incensement of the u variance. This may partially explain why radar

407 data assimilation using WRF-Var system with reduced lengthscales improved analyses

408 and forecasts (Xiao and Sun 2007; Sugimoto 2009; Sun et al. 2012; Wang et al. 2013b).

409

410 In CV7, B does not depend on horizontal locations as clearly shown in Fig. 8c,d and Fig.

411 9c,d. However, experiments using the alpha control variable show location dependent

412 error structures (Fig. 8e,f,g,h and Fig. 9e,f,g,h). The structures of horizontal increments in

413 the experiment using CV5 (Fig. 8a,b and Fig. 9a,b) much assemble those in the

414 experiments using the alpha control variables with large horizontal localization length

415 scales (Fig. 8e,f and Fig. 9e,f).

416

417 The vertical localization scheme (Eq.11) works well to keep observation information

418 around the observation location as shown in Fig. 9g and h. Comparing to Fig. 9e and f,

419 the negative u increment in low-level atmosphere is not seen. At the observation 25th

420 level, the information can be spread about 5 km in vertical direction (Fig. 7a).

421

19
422 The horizontal and vertical increments in the experiment U-CV7-P2 are shown in Fig. 10.

423 It is seen that the u analysis increment is not isotropic but stretched in west-east

424 direction (Fig. 10a). In addition, the maximum wind increment is reduced to 0.76

425 comparing to 0.89 before DFI (Fig. 9d). The extended west-east structure is derived by

426 DFI balancing in addition to the wind advection in backward and forward model

427 integration. The similar results are found by using a back and forth nudging algorithm

428 (Boilley and Mahfouf 2012). Comparing to Fig. 9d, Fig. 10b shows that the vertical

429 increments are spread in west-east direction. Small negative increments under about the

430 10th model are derived by DFI as well. The structures of positive increments shown in Fig.

431 10 much resemble those produced by CV5 (Fig. 8b and Fig. 9b) indicating that statistical

432 correlation presented in CV5 implicitly includes the model dynamical information.

433

434 4.2.2 Multivariate analyses by assimilating a single u observation

435 The features of the multivariate analysis increments are analyzed. Figure 11 shows v

436 increments by assimilating a single u observation. The results from the experiments U-

437 CV5-P2, U-CV7-P2-DFI, U-alpha-P2-L300-VL1, and U-alpha-P2-L100-VL1 are plotted.

438 It is interesting that CV7-P2-DFI produces a similar pattern for v to that in U-CV5-P2.

439 However, the increment amplitude is much smaller. Comparing experiments U-alpha-P2-

440 L300-VL1 (Fig. 11c) and U-alpha-P2-L100-VL1 (Fig. 11d), it is obvious that the smaller

441 horizontal localization length scale limits the increment to a local area around the single

442 observation. Figure 12 shows T increments by assimilating a single u observation. It is

443 obvious that the amplitudes of increments are small indicating that the correlations

444 between temperature and u-wind are weak at the model level. Large sensitivity to the

20
445 horizontal localization scales (Fig. 12b,c) is found when the alpha control variables are

446 used.

447

448 In general, the two proposed methods give good variance modeling in the NMC

449 ensembles. Compared to CV7, the use of the NMC ensembles through the alpha control

450 variables has the benefits of incorporating the geographically dependent covariance

451 information and producing a multi-variate analysis. We do see some noises in the single

452 observation experiments using the alpha control variable. It is expected that the use of

453 data from dense observing network such as radar, satellite and surface stations will

454 alleviate the sampling noises. In addition, the use of an real ensemble of forecasts valid at

455 the time can also reduce the samples noises. For CV7, DFI can help the analysis to

456 produce mulit-variate analysis.

457

458 4.3 Sensitivity to vertical localization schemes

459 As shown in Fig. 7a, the default vertical localization scheme (Eq.11) has small vertical

460 localization radius in low-level atmosphere. Experiments with the vertical localization

461 scheme (Eq.12) are conducted to show analysis sensitivity to vertical correlation matrix

462 specifications. In the experiments using the vertical localization matrix (Eq.12), the

463 localization radius is 3 km, which has been used for radar radial velocity assimilation (Li

464 et al. 2012). The three experiments are listed in Table. 2.

465

466 The vertical west-east section of the u increments for the single u observation

467 experiments U-alpha-L100-P2-VL2, U-alpha-L100-P3-VL1 and U-alpha-L100-P3-VL2

21
468 are shown in Fig. 13. Comparing the experiments U-alpha-L100-P1-VL1 and U-alpha-

469 L100-P1-VL2 (Fig. 9h and Fig. 13a), the vertical correlation schemes produce almost the

470 same analysis increment structures. But for an observation at 15th model, the observation

471 spread in the vertical direction is very different (Fig. 13b and c). Compared to the default

472 formulation Eq.11, the Eq.12 might be more efficient to spread observation vertically in

473 the low-level part of the atmosphere.

474

475 5. Real data assimilation and forecasting experiments

476 In this section, the impact of the new CVTs on data assimilation and forecasts of a

477 convective case occurred on 21 July 2012 in Beijing is studied. First, the synoptic

478 background of the case will be briefly described, and then experimental setup and results

479 will be presented.

480 5.1 The Bejing 7.21 extreme rainfall event

481 A warm season convective case that caused extreme precipitation with hourly rainfall

482 rates exceeding 85 mm occurred in Beijing on 21 July 2012 is studied. The 850 hPa

483 large-scale circulations at 0000 UTC are shown in Fig. 15a. The heavy rain took place in

484 a favorite synoptic environment that southwesterly monsoonal flow with high moisture in

485 the planetary boundary layer are over the Beijing region. It seen that a northeast-

486 southwest-oriented trough axis and a well-defined wind-shift line associated with a low-

487 leve vortex moving northward into southern Beijing. The south flow supplies moisture

488 sources for this extreme precipitation event.

489 5.2 Experimental setup

22
490 The BJ-RUC operating at Beijing Meteorology Bureau ()is employed. BJ-RUC employs

491 a one-way, two-domain nested grid. The horizontal grid resolutions of the outer and inner

492 domains are 9 and 3 km, respectively. First, a long (e.g. 30-h) forecast for the outer

493 domain is made at 1800 UTC 20 July 2012. The initial and boundary fields are

494 interpolated from National Centers for Environmental Prediction (NCEP) Global

495 Forecast System (GFS) 1 1 analyses and forecasts for this run, which then provides

496 the boundary conditions for the inner domain.

497 Four numerical experiments are conducted to examine the impacts of the background

498 error covariance modeling on analyses and forecasts. The control experiment

499 (CONTROL) is the 18-h 3km forecast initiated from GFS analysis at 1800 UTC 20 July

500 2012 at the inner domain. Its 6-h forecast provides the background in the data

501 assimilation experiments. The three data assimilation experiments using CVTS of CV5,

502 CV7 and alpha are named as 3DVAR-CV5-L05, 3DVAR-CV7, and 3DVAR-alpha,

503 respectively. In 3DVAR-CV5-L05, operational tuning factors with values of 0.5 are used

504 for each control variable. The forecasts in experiment using CV5 without tuning are not

505 discussed in this paper since they are less accurate than 3DVAR-CV5-L05. We do not

506 run cycles at later times (e.g. at 0300 UTC) because the precipitation began on 0200 UTC

507 21 July 2012. A 12-h forecast is made for each analysis. The forecasts after that time

508 have less value to forecasters. All the operational data including radar radial winds are

509 assimilated. The distributions of the observations assimilated are shown in Fig. 14. The

510 operational observations include radiosonde, sysnop, automated weather system surface

511 observations, and Global Position System precipitable water, aircraft data and radar radial

512 velocity.

23
513 5.3 Results

514 5.3.1. Analysis increments

515 We first examine the analysis increments to show the influence of the three background

516 error modeling implementations on data assimilation. The wind and relative humidity

517 increments at the 11th model level (aproximate 850 hPa) are shown in Fig. 15b,c,d.

518 Though the maximum locations of analysis increments for the wind and relative humidity

519 in the three experiments are similar, the patterns of the increments are quite different. The

520 maximum (minimum) values in 3DVAR-alpha are larger (smaller) then 3DVAR-CV5

521 and 3DVAR-CV7. This is because CV5 and CV7 produce an overall slightly error

522 variances modeling as shown in Fig. 5. The location dependent error modeling benefits

523 the analysis in 3DVAR-alpha by producing the increments with large amplitudes and

524 well-defined structures in the north model region. The reason is that in general the error

525 variances in the north region are larger than in south region at the model level (Fig. 1 and

526 Fig. 2). It is also seen that the structures of the relative analysis increments are quite

527 different to 3DVAR-CV5 and 3DVAR-CV7 by using location dependent error

528 covariance.

529 5.3.2 Verifications

530 The analyses and forecasts are verified against radar radial velocity observations. Figure

531 16 shows the root mean square erorr (RMSE) of wind forecasts against radar radial

532 velocity observations at anlaysis time and 12 h forecast time. As expected that the

533 RMSEs in three data assimilation experiments are smaller than the CONTROL

534 experiment. In the three data assimilation experiments, 3DVAR-alpha bestly fit the

24
535 observations, then 3DVAR-CV7 and 3DVAR-CV5. This result is well consistand with

536 the error variance modeling by the three CVTs (Fig. 5a and 5b). The two data

537 assimilation experiments 3DVAR-CV7 and 3DVAR-alpha using the new CVTs improve

538 radar radial velocity forecasts up to 12 hours over the control run (Fig. 16b).

539 The forecast skill of hourly precipation are examined by the neighborhood-based

540 fractions skill score (FSS; Roberts and Lean 2008). If a forecast is prefect, the value of

541 FSS is 1.0. An hourly radar quantitative precipitation estimate (QPE) that is produced

542 operationally at BMB is taken as the observation for verifying the model forecasts. Fig.

543 17 shows FSS for thresholds of 1 mm-1 and (b) 10 mm h-1. Overall, the experiment

544 3DVAR-alpha produces a best FSS among the three data assimilatoin experiments. The

545 experiment 3DVAR-CV7 produces a slightly better skill than the experiment 3DVAR-

546 CV5. Fig. 18 shows the 12 hourly and 1 hourly accumulated radar QPE at 1200 UTC 21

547 July 2012. It is seen that the sprious precipitation in the northwest to Beijing in the

548 CONTROL is not corrected in the threre data assimilation experiments. the precipitation

549 in the north region of Beijing is presented in the experiment 3DVAR-CV7 and 3DVAR-

550 alpha, and the later gives the best 12 hourly accumulated precipitation forecasts (Fig.

551 18a). For the hourly precipitation at 1200 UTC, the control run has a southwest

552 displacement bias, which is noticably reduced in the experiments 3DVAR-CV5 and

553 3DVAR-alpha. In summary, the preliminary results shows that the use of the new CVTs

554 improve the wind and precipiation forecast up to 12 hours.

555 6. Summary and discussion

556 In this paper, the features of background error modeling via the NMC method are

25
557 investigated in details for the WRF-Var system. The aim of this work is to further

558 improve the performance of the WRF-Var system through the best use of the

559 climatological background error covariance estimation. The short-term regional 3km

560 resolution forecasts in June, July and August 2012 from BJ-RUC are used to extract

561 background error statistics. The two new CVTs are proposed and described to incorporate

562 climatological B via the NMC method in WRF-Var. Up to the authors knowledge, this

563 is the first work using the two proposed CVTs to study climatological BE modeling in

564 context of the WRF-Var system. The features of several CVTs are investigated in detail

565 using three months of high-resolution operational forecasts.

566 The main work is summarized as follows.

567 The location dependent BE variances of various variables vary from month to
568 month.

569 The diurnal variation of BE in low-level part of atmosphere is also found.

570 WRF-Var CV5 BE modeling underestimates wind error variance but

571 overestimates wind error length scale.

572 Two CVTs that are proposed to incorporate climatological B with the NMC

573 method in WRF-Var are investigated. The first one uses u , v , T , Ps and RH s

574 as control variables. Another approach is to employ "alpha" control variables to

575 incorporate location dependent error covariance. The detailed technical

576 descriptions are provided for the two new formulations.

577 The two proposed methods give good variance modeling in the NMC ensembles. The use

578 of the NMC forecast differences through the alpha control variable has the benefits of

26
579 incorporating geographically dependent covariance information and producing multi-

580 variate analysis. However, analysis increments are sensitive to horizontal and vertical

581 localization radii using the alpha control variable. The preliminary results from data

582 assimilaton and forecast study for a real convective case shows that the use of the new

583 CVTs improve the wind and precipiation forecast up to 12 hours.

584 Though the new proposed CV7 only produces the univariate analyses, the multi-variate

585 analyses are achieved by use of digital filter or initialization technique in this paper. The

586 developments will benefit other components such as 4DVar, hybrid Var-ensemble data

587 assimilation in the WRF community data assimilation system (Barker et al. 2012; Huang

588 et al. 2013).

589 In WRF-Var, the climatological statistical correlations between relative humidity and

590 other control variables can be taken into account with CV option 6 (CV6) (Chen et al.

591 2013). These climatological statistical correlations between relative humidity and other

592 variables can be achieved by the use of the alpha control variable as well. Moreover, CV7

593 can also be used to model BE covariance of the day using an ensemble of short-term

594 forecasts.

595 The monthly and diurnal variations of variances can be considered in climatological BE

596 modeling in WRF-Var. In reality, the BE covariances may be substantially flow

597 dependent. The current BE statistics using the NMC method may not be optimal to

598 provide the BE covariance of the day for mesoscale and convective scale data

599 assimilation. A super ensemble including the NMC forecast differences and short-term

600 ensemble forecasts could be used to blend climatological error covariances and flow-

27
601 dependent error covariances of the day in a hybrid system. The investigation of the

602 impact of the new CVTs on numerical weather prediction with real radar data

603 assimilation is underway. The results will be reported in a following paper.

604

605 Acknowledgments. This work was supported by a project between UCAR and IUM/BMB.

606 NCAR is sponsored by the National Science Foundation. Any opinions, findings, and

607 conclusions or recommendations expressed in this publication are those of the authors

608 and do not necessarily reflect the views of the National Science Foundation.

28
 References

609 Andersson, E., M. Fisher, R. Munro and A. McNally, 2000: Diagnosis of background

610 errors for radiances and other observable quantities in a variational data assimilation

611 scheme, and the explanation of a case of poor convergence. Quart. J. Roy. Meteor. Soc.

612 126, 14551472.

613 Bannister, R. N., 2008a: A review of forecast error covariance statistics in atmospheric

614 variational data assimilation. I: Characteristics and measurements of forecast error

615 covariances. Quart. J. Roy. Meteor. Soc., 134, 19511970.

616 Bannister, R. N., 2008b: A review of forecast error covariance statistics in atmospheric

617 variational data assimilation. II: Modelling the forecast error covariance statistics.

618 Quart. J. Roy. Meteor. Soc., 134, 19711996.

619 Barker, D. M., W. Huang, Y.-R. Guo, A. J. Bourgeois, and Q. N. Xiao, 2004: A three-

620 dimensional variational data assimilation system for MM5: Implementation and initial

621 results. Mon. Wea. Rev., 132, 897914.

622 Barker, D. M., 2005: Southern High-Latitude Ensemble Data Assimilation in the

623 Antarctic Mesoscale Prediction System. Mon. Wea. Rev., 133, 34313449.

624 Barker, D. M., and Coauthors, 2012: The Weather Research and Forecasting (WRF)

625 Models Community Variational/Ensemble Data Assimilation System: WRFDA. Bull.

626 Amer. Meteor. Soc., 93, 831843.

627 Boilley, Alexandre, and Jean-Franois, Mahfouf, 2012: Assimilation of low-level wind in

29
628 a high-resolution mesoscale model using the back and forth nudging algorithm. Tellus

629 A 2012, 64, 18697, http://dx.doi.org/10.3402/tellusa.v64i0.18697

630 Chen, M., S. Fan, J. Zhong, X.-Y. Huang, Y.-R. Guo, W. Wang, Y. Wang, and B. Kuo,

631 2009: A WRF-based rapid updating cycling forecast system of BMB and its

632 performance during the summer and Olympic Games 2008. 10th WRF Users

633 Workshop, June 23-26, 2009, Boulder, CO. [Available online at

634 http://www.mmm.ucar.edu/wrf/users/workshops/WS2009/abstracts/P3B-37.pdf.]

635 Chen, Y., S.R.H. Rizvi, X.-Y. Huang, J. Min, and X. Zhang, 2013: Balance

636 characteristics of multivariate background error covariance and their impact on

637 analyses and forecasts in Tropics and Arctic. Meteor. Atmos. Phys., 121, 79-98.

638 Courtier P., E. Andersson, W. Heckley, J. Pailleux, D. Vasiljevic, A. Hollingsworth, M.

639 Fisher, and F. Rabier, 1998: The ECMWF implementation of three-dimensional

640 variational assimilation (3D-Var). I: Formulation. Q. J. R. Meteorol. Soc. 124: 1783

641 1807.

642 Daley, R. 1991: Atmospheric data analysis. Cambridge University Press, pp. 460.

643 Fisher M., 2003: Background error covariance modelling. P. 45 64 in ECMWF Seminar

644 on Recent developments in data assimilation for atmosphere and ocean, 8 12

645 September 2003. ECMWF: Reading UK.

646 Fischer C., T. Montmerle, L. Berre, L. Auger, S.E. tefnescu, 2005: An overview of the

647 variational assimilation in the ALADIN/France numerical weather prediction system.

648 Q. J. R. Meteorol. Soc. 131: 34773492.

30
649 Gao, J., M. Xue, A. Shapiro, and K. K. Droegemeier, 1999: A variational analysis for the

650 retrieval of three-dimensional mesoscale wind fields from two Doppler radars, Mon.

651 Wea. Rev., 127, 2128-2142.

652 Houtekamer PL, L. Lefaivre, J. Derome, H. Ritchie, and H.L. Mitchell, 1996: A system

653 simulation approach to ensemble prediction. Mon. Weather Rev. 124: 12251242.

654 Gustafsson, N., X.-Y. Huang, X. Yang, K.S. Mongensen, M. Lindskog, O. Vignes, T.

655 Wilhelmsson, and S. Thorsteinsson, 2012: Four-dimensional variational data

656 assimilation for a limited area model. Tellus A, 64, 14985, DOI:

657 10.3402/tellusa.v64i0.14985.

658 Hsiao, L-F, D-S Chen, Y-H Kuo, Y-R Guo, T-C Yeh, J-S Hong, C-T Fong, and C-S Lee,

659 2012: Application of WRF 3DVAR to Operational Typhoon Prediction in Taiwan:

660 Impact of Outer Loop and Partial Cycling Approaches. Wea. Forecasting, 27, 1249

661 1263.

662 Huang, X.-Y., and coauthors, 2009: Four-Dimensional Variational Data Assimilation for

663 WRF: Formulation and Preliminary Results. Mon. Wea. Rev., 137, 299-314.

664 Huang, X.-Y., and coauthors, 2013: The 2013 WRFDA overview. 14th WRF Users'

665 Workshop, Boulder, Colorado, 24-28 June 2013. Extended Abstract 1.2.

666 [http://www.mmm.ucar.edu/wrf/users/workshops/WS2013/WorkshopPapers.php]

667 Ingleby, N., 2001: The statistical structure of forecast errors and its representation in The

668 Met. Office global 3-D variational data assimilation scheme. Quart. J. Roy. Meteor.

669 Soc., 127, 209231.

31
670 Kawabata, T., T. Kuroda, H. Seko, and K. Saito, 2011: A cloud-resolving 4DVAR

671 assimilation experiment for a local heavy rainfall event in the Tokyo metropolitan area.

672 Mon. Wea. Rev., 139, 19111931.

673 Li, Y., X. Wang, and M. Xue, 2012: Assimilation of Radar Radial Velocity Data with the

674 WRF Hybrid Ensemble3DVAR System for the Prediction of Hurricane Ike (2008).

675 Mon. Wea. Rev., 140, 35073524.

676 Lynch, P. and X.-Y. Huang, 1994: Diabatic initialization using recursive

677 filters. Tellus 46A, 583-597.

678 Parrish, D., and J. Derber, 1992: The National Meteorological Centers spectral statistical

679 interpolation analysis system. Mon. Wea. Rev., 120, 17471763.

680 Rawlins, F., and coauthors, 2007: The Met Office global four-dimensional variational

681 data assimilation scheme. Quart. J. Roy. Meteor. Soc., 133, 347362.

682 Roberts, N. M., and H. W. Lean, 2008: Scale-selective verification of rainfall

683 accumulations from high-resolution forecasts of convective events. Mon. Wea. Rev.,

684 136, 7897.

685 Skamarock, W. C., and Coauthors, 2008: A description of the advanced research WRF

686 version 3. NCAR Tech. Note TN-475+STR, 113 pp.

687 Sugimoto, S., N. A. Crook, J. Sun, Q. Xiao, and D. M. Barker, 2009: An Examination of

688 WRF 3DVAR Radar Data Assimilation on Its Capability in Retrieving Unobserved

689 Variables and Forecasting Precipitation through Observing System Simulation

32
690 Experiments. Mon. Wea. Rev., 137, 40114029.

691 Sun, J., and N. A. Crook, 1997: Dynamical and Microphysical Retrieval from Doppler

692 Radar Observations Using a Cloud Model and Its Adjoint. Part I: Model Development

693 and Simulated Data Experiments. J. Atmos. Sci., 54, 1642-1661.

694 Sun, J., S. B. Trier, Q. Xiao, M. L. Weisman, H. Wang, Z. Ying, M. Xu, and Y. Zhang,

695 2012: Sensitivity of 012-h Warm-Season Precipitation Forecasts over the Central

696 United States to Model Initialization. Wea. Forecasting, 27, 832855.

697 Xiao, Q., and J. Sun, 2007: Multiple-Radar Data Assimilation and Short-Range

698 Quantitative Precipitation Forecasting of a Squall Line Observed during IHOP_2002.

699 Mon. Wea. Rev., 135, 33813404.

700 Wang, H., J. Sun, S. Fan, and X.-Y. Huang, 2013a: Indirect Assimilation of Radar

701 Reflectivity with WRF 3D-Var and Its Impact on Prediction of Four Summertime

702 Convective Events. J. Appl. Meteor. Climatol., 52, 889902.

703 Wang, H., J. Sun, X. Zhang, X.-Y. Huang, and T. Aulign, 2013b: Radar Data

704 Assimilation with WRF 4D-Var: Part I. System Development and Preliminary Testing.

705 Mon. Wea. Rev. 141, 22242244.

706 Wang, X., D. M. Barker, C. Snyder, and T. M. Hamill, 2008: A Hybrid ETKF3DVAR

707 Data Assimilation Scheme for the WRF Model. Part I: Observing System Simulation

708 Experiment. Mon. Wea. Rev., 136, 51165131.

709 Xie, Y., C. Lu, and G. L. Browning, 2002: Impact of formulation of cost function and

710 constraints on three-dimensional variational data assimilation. Mon. Wea. Rev., 130,

711 24332447.

33
712 Xie Y. and A. E. MacDonals, 2011: Selection of momentum variables for a three-

713 dimensional variational analysis. Pure and Applied Geophysics, 169, 335351.

714 Zou, X., Y.-H. Kuo, and Y.-R. Guo, 1995: Assimilation of atmospheric radio refractivity

715 using a nonhydrostatic adjoint model. Mon. Wea. Rev., 123, 2229-2249.

716 Zupanski M, D. Zupanski, T. Vukicevic, K. Eis, and T. Vonder Haar. 2005: CIRA/CSU

717 four dimensional variational data assimilation system. Mon. Weather Rev. 133: 829

718 843.

34
719 Table. 1 List of single observation data assimilation experiments

Experiment name CV Location Horizontal Vertical

localization scaling localization
U-CV5-P1 5 P1
U-CV5-P2-L05 5 P1
U-CV7-P1 7 P1
U-CV7-P2 7 P2
U-CV7-P2_DFI 7 P2 DFI is used
U-alpha-L300-P1 alpha P1 300 No
U-alpha-L300-P2 alpha P2 300 No
U-alpha-L100-P1-VL1 alpha P1 100 VL1
U-alpha-L100-P2-VL1 alpha P2 100 VL1
720 * In experiment U-CV5-P2-L05, Length scaling factors for and u are 0.5.

721

722 Table. 2 List of sensitivity experiments

Experiment name CV Location Horizontal Vertical

localization scaling localization
U-alpha-L100-P2-VL2 alpha P2 100 VL2
U-alpha-L100-P3-VL1 alpha P3 100 VL1
U-alpha-L100-P3-VL2 alpha P3 100 VL2
723 * P3 stands for single observation location at (364,150,15) in model grid.

35
724 11th model level 25th model level 33th model level

u
u

725

v
v

726

727

RH

728

729 Fig.1. Horizontal distribution of geographically dependent standard deviation of the error
730 variances for u (1st row), v (2nd row), T (3rd row) and RH (4th row) at the 11 th model
731 level (left column), the 25 th (middle column) and the 33 th model level (right column).

36
732

733
734 Fig. 2. Standard deviation of the error variances as functions of the model level and
735 distance (km) in south-north direction for (a) u, (b) v, (c) T and (d) RH as functions of
736 distance in south-north direction and the number of model levels.

37
737

738
739 Fig. 3. Profiles of forecast error in terms of STD for (a) u, (b) v, (c) T, (d) RH estimated

740 by the NMC method for June, July and August 2012.

38
741
742 Fig. 4. Profiles of forecast error in terms of STD for (a) u, (b) v, (c) T, (d) RH estimated

743 by NMC method at 0000 UTC (black curve) and 1200 UTC (red curve).

39
744
745 Fig. 5. The vertical profiles of STD_NMC, STD_CV5, STD_CV5_L05 and STD_CV7.

746

40
747
748 Fig. 6. Horizontal correlations for u (black curves) and v (red curves) in the NMC
749 ensembles (diamond mark), by BE modeling using CV5 (square mark) and CV7 (circle
750 mark) at 25th model level (approximate 500 hPa surface) along east-west direction (upper
751 panel) and north-south direction (low panel).

41
752

753
754 Fig. 7. The correlation matrixes for the alpha control variable localization using (a) Eq.10,

755 and (b) Eq.11.

42
756
757 Fig. 8. The structures of the u increments on the 25th model level for the single u

758 observation experiments. (a) U-CV5-P1, (b) U-CV5-L05-P2, (c) U-CV7-P1, (d) U-CV7-

759 P2, (e) U-alpha-L300-P1, (f) U-alpha-L300-P02, (g) U-alpha-L100-P1-VL1, and (h) U-

760 alpha-L100-P2-VL1.

43
761
762 Fig. 9. The vertical structures of the u increments as functions of the model level and the

763 distance in the west-east direction in the single u observation experiments. (a) U-CV5-

764 P1, (b) U-CV5-L05-P2, (c) U-CV7-P1, (d) U-CV7-P2, (e) U-alpha-L300-P1, (f) U-alpha-

765 L300-P02, (g) U-alpha-L100-P1-VL1, and (h) U-alpha-L100-P2-VL1.

44
766
767 Fig. 10. The u increments in the experiment U-CV7-P2-DFI. (a) the structures of at the

768 25th model level, and (b) the vertical structures of the u increments as functions of the

769 model level and the distance in the west-east direction.

45
770

771
772 Fig. 11. The strucgtures of the v increments at the 25th model level in the single u data

773 assimilation experiments (a) U-CV5-P2-L05, (b) U-CV7-P2-DFI, (c) U-alpha-L300-

774 P2, and (d) U-alpha-L100-P2.

46
775
776 Fig. 12. The strucgtures of the T increments at the 25th model level in the four single u

777 data assimilation experiments (a) U-CV5-P2, (b) U-alpha-L300-P2, (c) U-alpha-L100-

778 P2-VL1.

779

47
780
781 Fig. 13. The vertical structures of the u increments as functions of model level and

782 distance in the west-east direction in the single u observation experiments (a) U-

783 alpha-L100-P2-VL2, (b) U-alpha-L100-P3-VL1, (c) U-alpha-L100-P3-VL2.

784
785

48
786
787
788 Fig. 14 The distributions of the observations assimilated in the three data assimilation
789 experiments.
790

791

792
793 Fig. 15. (a) Background, and the analysis increments for the experiment, (b) 3DVAR-
794 CV5-L05, (c) 3DVAR-CV5, and (d) 3DVAR-alpha. Relative humidity (%) is shown in
795 shaded contours.
796

49
797

798

799
800 Fig. 16. Root mean square erorr of wind forecasts against radar radial velocity
801 observations at (a) 0000 UTC and (b) 1200 UTC 21 July 2012.
802

50
803
804
805 Fig. 17. FSS of the four experiments for thresholds of (a) 1 mm-1 and (b) 10 mm h-1.
806
807

51
808

809

810

811

812

813

52
814 Fig. 18. (a) 12 hour accumlated precipitation and (b) 1 hour accumlated precipitation at

815 1200 UTC 21 July 2012.

53