Journal of Water and Climate

0 downloads 0 Views 7MB Size Report
stationarity in precipitation pdf in future climate conditions. Similarly, Haberlandt et al. ...... conducive (or more un-conducive) to rainfall formation (Lutgens and Tarbuck 2013;. 395. Yarnal 1993). ..... Earth System Science Portals. 650. 651. 652.
Journal of Water and Climate Evaluation of climate change impacts on extreme rainfall events characteristics using a synoptic weather typing based daily precipitation downscaling model --Manuscript Draft-Manuscript Number:

JWC-D-16-00107R1

Full Title:

Evaluation of climate change impacts on extreme rainfall events characteristics using a synoptic weather typing based daily precipitation downscaling model

Article Type:

Research Paper

Corresponding Author:

sara nazif IRAN (ISLAMIC REPUBLIC OF)

Corresponding Author Secondary Information: Corresponding Author's Institution: Corresponding Author's Secondary Institution: First Author:

Hamed Tavakolifar, PhD candidate

First Author Secondary Information: Order of Authors:

Hamed Tavakolifar, PhD candidate Ebrahim Shahghasemi sara nazif

Order of Authors Secondary Information: Abstract:

Climate change affects all phenomena in the hydrologic cycle, especially extreme events. GCMs are used for investigate its impacts. Due to their low resolutions, downscaling methods are developed. In this study, a novel downscaling method called synoptic statistical downscaling model is proposed for daily rainfall downscaling with an emphasis on extreme events characteristics preservation. In this method, first, synoptic weather typing process is done to detect the circulation types using the SOM method. For each CT, two regression models are developed for rainfall occurrence and amount downscaling. To improve the models performance in preservation of the extreme rainfall characteristics the weighted mean square error method is used for model's development. The proposed model is applied to a region located in central Iran. The results show that the developed model can downscale all percentiles of rainfall events with an acceptable performance and there is no assumption about the similarity of future rainfall data with the historical observations. The outputs of CCSM4 for two representative concentration pathways of RCP4.5 and RCP8.5 are used to investigate the climate change impacts in the study region. The results show 40% and 30% increase in the number of extreme rainfall events under RCP4.5 and RCP8.5, respectively.

Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation

Manuscript

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

Click here to download Manuscript first revision .docx

1

Evaluation of climate change impacts on extreme rainfall events characteristics

2

using a synoptic weather typing based daily precipitation downscaling model

3 4

Hamed Tavakolifar

5

Ph.D. candidate, School of Civil Engineering, College of Engineering, University of

6

Tehran, Tehran, Iran, Email: [email protected]

7 8

Ebrahim Shahghasemi

9

Assistant Professor, School of Civil Engineering, College of Engineering, University of

10

Tehran, Tehran, Iran, Email: [email protected]

11 12

Sara Nazif

13

Assistant Professor, School of Civil Engineering, College of Engineering, University of

14

Tehran, Tehran, Iran, Email: [email protected]

15 16

Abstract

17

Climate change has impacted all phenomena in the hydrologic cycle, especially extreme

18

events. GCMs (general circulation models) are used for investigate climate change

19

impacts but due to their low resolutions, downscaling methods are developed to provide

20

data with resolutions high enough for regional studies from GCMs outputs. The

21

performance of downscaling methods is commonly acceptable in preserving average

22

characteristics, but they do not preserve the extreme events characteristics especially

23

amount and its distribution. In this study, a novel downscaling method called synoptic

24

statistical downscaling model (SSDM) is proposed for daily precipitation downscaling

1

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

25

with an emphasis on extreme events characteristics preservation. In the proposed method,

26

first, synoptic weather typing process is done to detect the circulation types (CTs) using

27

the Self-organizing Maps (SOM) method based on the 500-hPa geopotential heights data.

28

For each CT, two regression models are developed for rainfall occurrence and amount

29

downscaling. To improve the models performance in preservation of the extreme rainfall

30

characteristics beside the mean events statistics, the weighted mean square error (WMSE)

31

method is used for model’s development. The proposed model is applied to a region

32

located in central Iran. The results show that the developed model can downscale all

33

percentiles of precipitation events with an acceptable performance and there is no

34

assumption about the similarity of future rainfall data with the historical observations.

35

The outputs of CCSM4 GCM for two representative concentration pathways (RCP) of

36

RCP4.5 and RCP 8.5 are used to investigate the climate change impacts in the study

37

region. The results show 40% and 30% increase in the number of extreme rainfall events

38

under RCP4.5 and RCP 8.5, respectively.

39 40

Keywords: Downscaling, Synoptic weather typing, SOM, Extreme events, Circulation

41

type, Synoptic statistical downscaling model

42 43

1

Introduction

44

The investigation of local rainfall events characteristics, especially extreme ones, is a key

45

issue in hydrologic research in different fields such as urban drainage system management

46

and flood control strategies. Due to climate change impacts, the intensity and frequency

47

of extreme rainfall events are increasing. Based on the 5th report of Intergovernmental

48

Panel on Climate Change (IPCC, 2014), the frequency and intensity of heavy

2

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

49

precipitation events have increased in the Northern Hemisphere. General circulation

50

models (GCMs) are the primary available tools for investigating the response of the

51

hydro-climate system to climate change impacts (IPCC 2014).

52

In Figure 1, the spatial and temporal scale of various weather phenomena, the time and

53

space domains of available GCMs outputs and the time and space domains required in

54

some hydrological investigations are compared. As shown in this figure, the coarse

55

gridding of GCMs outputs cannot capture all types of mesoscale phenomena which have

56

important influences on temporal and spatial patterns of local rainfall in small regions

57

which are considered in flood risk or urban drainage management. For example, in terms

58

of spatial scale, storms in urban basins (usually less than 500 km2) are the mesoscale

59

phenomenon (Willems et al. 2012) and in terms of temporal scale, daily rainfall is the

60

largest time scale applicable in urban flood management. Because of the spatial and

61

temporal scale mismatch between the outputs of GCMs and the scale at which

62

hydrological investigations about climate change impacts are commonly carried out,

63

downscaling methods are applied (Wilby et al. 1998).

64

Two main approaches in downscaling are dynamical (numerical) downscaling and

65

statistical (empirical) downscaling (Benestad et al. 2008). In dynamical downscaling,

66

limited-area regional climate models (RCM) are applied and the GCMs outputs are used

67

as their initial conditions (Feser et al., 2011). In the statistical downscaling (SD) methods,

68

statistical or empirical models are used to relate the GCMs outputs and local rainfall

69

(Wilby and Dawson 2013). Different methods, techniques and approaches are proposed

70

in the SD context including stochastic weather generators (WG) (Baigorria and Jones

71

2010; Richardson 1981; Semenov 2008; Semenov and Barrow 1997; Wilks 2010),

72

transfer function (TF) approaches (Chau and Wu 2010; Chen et al. 2010; Dibike et al.

3

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

73

2008; Hessami et al. 2008; Karamouz et al. 2013b; a; Najafi et al. 2011; Pasini 2009;

74

Tavakol-Davani et al. 2013; Tomassetti et al. 2009; Wilby et al. 2002; Willems and Vrac

75

2011) and weather typing (WT) approaches (Arnbjerg-Nielsen 2008; Cheng et al. 2010,

76

2011; D’onofrio et al. 2010; Hewitson and Crane 2006; Vrac and Naveau 2007).

77

Previous studies have commonly focused on reproducing the mean behavior of the daily

78

rainfall. However, some studies devoted to downscale daily precipitation with an

79

emphasis on the extreme events (Benestad 2009; Cheng et al. 2011; Hundecha and

80

Bárdossy 2008). It is difficult to preserve the characteristics of extreme daily rainfalls

81

during the downscaling process due to local phenomena that affect extreme daily rainfall

82

amounts, the considerable number of dry days and a non-Gaussian statistical distribution

83

on wet days (Benestad 2009). In many downscaling approaches, there is a tendency to

84

underestimate the extreme precipitation events (Arnbjerg-Nielsen 2008; D’onofrio et al.

85

2010; Fealy and Sweeney 2007).

86

To overcome this shortage, some methods are proposed such as bias correction and

87

change factor methods. Sarr et al. (2015) compared the performance of the delta-change

88

method with a quantile–quantile transformation (QQ) method in downscaling the amount

89

of daily extreme rainfall events with 5 to 100 years return period in 6 stations in Senegal.

90

The results showed that these methods in some cases over estimate and in some stations

91

underestimate the extreme rainfall events amount. A comparison between different types

92

of these methods can be found in Sunyer et al. (2015), Tryhorn and DeGaetano (2011)

93

and Wang et al. (2015). According to their reviews, considering merely a limited number

94

of extreme events in the bias correction method and inability to reproduce timing

95

characteristics (chronology) of daily rainfall such as the length of dry/wet spells are some

96

limitations of these approaches.

4

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

97

Another approach to overcome the limitations of ordinary SD models for downscaling

98

daily precipitation is combining WT methods with SD models. The main idea in

99

developing this approach is that the precipitation occurrence in different periods of time

100

with a particular synoptic scale weather pattern follows a similar scheme (Willems et al.

101

2012).

102

Cheng et al. (2010) applied a synoptic WT using principal component analysis (PCA) and

103

the logit and nonlinear regression procedures as the within-weather-type daily rainfall

104

simulation models. Although, the results show that the proposed method has a good

105

potential to simulate daily rainfall occurrence and amount, it does not display a stable

106

performance in all ranges of rainfall amounts. Hertig and Jacobeit (2013) also used the

107

PCA (Principal Component Analysis) on 700-hPa geopotential heights to determine

108

circulation patterns (CP) in the synoptic scale and the Poisson regression model for daily

109

precipitation simulation. Their model overestimated the precipitation below the 90%

110

percentile and underestimated precipitations above this threshold.

111

Using k-means clustering technique to determine synoptic CPs, D’onofrio et al. (2010)

112

simulated daily rainfalls according to daily precipitation probability density function (pdf)

113

in each CP. Their approach was fairly straight and simple, but the results were

114

overestimated for extreme values and wet/dry sequences. In addition, they assumed

115

stationarity in precipitation pdf in future climate conditions. Similarly, Haberlandt et al.

116

(2015) developed a stochastic rainfall downscaling model of which the parameters were

117

conditioned based on CPs. They investigated the non-stationarity of the relationship

118

between rainfall and CPs and concluded that the modification of the parameters of the

119

stochastic rainfall downscaling model for its future application is necessary.

5

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

120

Despite the improvements in SD precipitation downscaling methods in previous studies,

121

preserving all statistical characteristics of the observed local daily rainfall by the

122

downscaling model is still a challenge. The previous studies cannot simulate the daily

123

precipitation pdf for all ranges of rainfall amount without the assumption of the constant

124

precipitation pdf for current and future climate conditions.

125

As estimation of extreme rainfall events is a vital issue in urban areas in planning for

126

flood events and due to climate change impacts on characteristics of extreme rainfall

127

events. It is needed to consider climate change impacts in extreme rainfalls evaluation.

128

On the other side, the common methods used for evaluation of climate change impacts on

129

rainfall events, does not provide satisfactory results for extreme rainfall events

130

investigation. Therefore, the main aim of this study is to develop a downscaling model

131

which can preserve the distribution frequency or percentile values of the observed daily

132

precipitation for all ranges of rainfall with an emphasis on extreme events. The proposed

133

method that is called synoptic-statistic downscaling model (SSDM) is based on a

134

combination of synoptic WT of upper-level atmosphere conditions which affect the local

135

atmosphere states, and SD models for simulating daily rainfall in each weather type. The

136

performance of SD models is improved by applying some modifications in common

137

methods including skewness reduction and weighted least squares (WLS). These changes

138

have made SSDM capable of overcoming some of the reported challenges and

139

disadvantages in literature including reproducing extreme events frequencies, and

140

inherent non-Gaussian statistical distribution of daily rainfall. In the proposed method,

141

stability in relationships between predictors and local rainfall in each weather type as well

142

as the adequacy of recognized WTs for future events are assumed. There is no assumption

143

about the stationarity of the daily rainfall pdf in future conditions, and changes in the

6

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

144

frequency distribution of predictors (because of climate change) can result in daily rainfall

145

pdf changes. The proposed method is applied to an area located at the north of Tehran,

146

the capital of Iran, as the study area. The historical floods of this city have resulted in

147

considerable damages to people and properties in the past. Therefore, the investigation of

148

climate change impacts on precipitation regime of this region with an emphasis on

149

extreme events is of high importance.

150

In the next section, the proposed method in this study for rainfall downscaling, SSDM, is

151

described. Then, the case study and data used in the study are introduced. Next, SSDM’s

152

development for the study area is described and the results of its application are discussed.

153

Afterwards, climate change impacts on rainfall characteristics are investigated for the

154

study area by applying SSDM on outputs of a GCM. Finally, a summary and conclusion

155

are given.

156 157

2

Method

158

2.1

159

There are different types of precipitation (cyclone-related, frontal or convective). Most of

160

the precipitation types are synoptic scale phenomena. In the proposed SSDM approach,

161

the effects of synoptic scale phenomena on the mesoscale rainfall are considered using

162

the synoptic weather typing method. Regarding the multi-scale nature of the atmospheric

163

processes that affect the rainfall formation, different sets of climatic variables contribute

164

to the prediction of rainfall in different weather types (D’onofrio et al. 2010). In SSDM,

165

days are classified according to the circulation type (CT) which occurs in each day. SSDM

166

algorithm includes two main phases. Figure 2 shows these two phases and their main

SSDM Algorithm

7

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

167

steps and components as well as how they are related to each other. The detailed

168

description of SSDM algorithm is given in the following parts.

169

2.1.1 Phase I: Synoptic-Scale Classification Model

170

Synoptic climatology was defined by Yarnal (1993) as the study of the linkage between

171

the atmospheric circulation and an environmental response. A successful classification of

172

atmospheric conditions is an essential step in the investigation of this linkage. Therefore,

173

the first phase in SSDM algorithm is clustering the large-scale atmospheric CTs that

174

mainly influence the local surface weather conditions.

175



Self-Organizing Maps (SOM)

176

In this study, the SOM technique is used for synoptic based classification of CTs for their

177

successful application in climatological fields in the last decade (Hewitson and Crane

178

2002; Sheridan and Lee 2011). SOM was first presented in details by Kohonen (1995).

179

Inspired by the neural networks, the SOM methodology utilizes an array or lattice of a

180

low-dimensional (typically two-dimensional) matrix of neurons or nodes, which are

181

trained to produce a discretized representation or map of the input space. Each node

182

(considered a cluster) is defined by a reference vector of weighting coefficients, which

183

has the same dimension as the input data vector. At first, nodes are randomly distributed

184

in data space.

185

There are some standard SOM configurations used for its development including sheet,

186

cylinder and toroid. The structure used here for SOM development is a sheet shape

187

(matrix) of nodes with a rectangular lattice neighborhood structure. Since in the synoptic

188

scale studies the two dimensional variations of climatic variables is investigated, the sheet

189

shape is used similar to previous studies such as Hewitson and Crane (2002) and Sheridan

190

and Lee (2011). The cylinder and toroid map shapes of SOM are used when three-

8

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

191

dimensional analysis of data is considered. About the neighborhood condition in SOM

192

map development two cases of hexagonal and rectangular are available which both of

193

them are considered in this study and due to better performance of the rectangular

194

neighborhood it is adopted in this study. In addition to SOM strcuture, the appropriate

195

number of nodes in SOM structure should also be determined. Commonly, a subjective

196

method is used for this purpose (Sheridan and Lee 2011). In many cases, different

197

numbers of nodes are considered, and then by evaluation of the developed models’

198

performance, the best number of nodes that results in the best model performance, is

199

determined.

200

In training procedure, each input vector is presented to the SOM and the best matching

201

unit (BMU) is determined. BMU is the node that has the least Euclidean distance from

202

the input vector. Then the weighting coefficients of BMU and its neighboring nodes are

203

modified in a way that the distance from the input vector is minimized. The weighting

204

coefficients modifications are controlled by some factors such as learning rate.

205

Neighborhood adjustment is utilized to preserve the topological properties of the input

206

space. The training can be sequential (introducing input data one by one to the SOM) or

207

batch training (introducing all input data simultaneously to the SOM). The training

208

process has two stages of rough training (with large neighborhood radius and high

209

learning rate) and fine-tuning (with small radius and low learning rate). The termination

210

criteria for the training process can be the achievement of the minimum value of SOM

211

error measurement called average quantization error (Kohonen and Honkela 2007) or

212

reaching a predefined number of training epochs. A practical public domain software

213

package for implementing SOMs along with extensive references and documentations are

9

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

214

provided by Helsinki University of Technology’s Laboratory of Computer and

215

Information Science that is available at http://www.cis.hut.fi/projects/somtoolbox/.

216



Relating CTs and daily precipitation

217

In order to analyze the relationship between the daily precipitation and the driving large-

218

scale atmospheric CTs, a performance index (PI) (Zhang et al. 1997) beside the

219

interpretation of the composite maps (Yarnal et al. 2001) is utilized. Using this PI will

220

help to quantify this relationship and makes easier to determine CTs which result in wet

221

and dry days which is necessary in this study. PI will help to provide the same

222

interpretation for CTs in different cases which is a very important in reproducing the study

223

result. PI compares the mean daily precipitation for ith CT with the climatological daily

224

mean precipitation as follows: 𝑃𝐼(𝑖) =

𝑅𝑖 ⁄𝑛𝑖 𝑅 ⁄𝑛

(1)

225

where 𝑛𝑖 is the number of days included in ith CT, 𝑅𝑖 is the total amount of precipitation

226

in ith CT and 𝑅 is the total amount of precipitation received in the study region in the

227

considered period of 𝑛 days. PI values close to one show that the given CT leads to

228

precipitation amounts around the climatological mean. PI values close to zero mean that

229

the corresponding CT does not considerably contribute to precipitation, or the CT is un-

230

conducive to rainfall formation while PI values greater than one mean that the given CT

231

is conducive to considerable precipitation.

232

2.1.2 Phase II: Local Scale Statistical Model

233

At the second phase of the SSDM algorithm, two statistical models are developed for

234

each CT. The first model, called rainfall occurrence model (ROM), determines that a day

235

is wet or dry and the second model called rainfall amount model (RAM) estimates the

236

amount of precipitation for wet days. Both models are regression-based. A combination 10

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

237

of subjective and automatic methods is used for the selection of the independent variables

238

in both regression models. In order to improve the ability of regression models to

239

reproduce extreme rainfall events’ characteristics, transformation functions (TF) and

240

WLS are used respectively for reduction of daily precipitation skewness and estimation

241

models’ parameters.

242



Variable Selection

243

For developing a statistical rainfall simulation model, appropriate features of the

244

atmosphere are chosen as the predictors. There are numerous features of the atmosphere

245

that can be used for this purpose. The potential variables are initially screened based on

246

the physical process of rain formation and suggestions available in the literature for the

247

study region. This initial screening is followed by an automatic variable selection

248

procedure that determines the best set of variables to be used in the statistical models. In

249

this study, the stepwise linear regression is used for this purpose.

250

Stepwise regression develops a multivariate regression model in which the choice of

251

predictive variables is carried out by an automatic procedure. At each iteration, some

252

predictors that improve the explanatory power of the model are added to the subset of

253

predictive variables of the model. At the same time, variables whose elimination does not

254

considerably affect the model’s performance are removed from the subset of predictive

255

variables of the model. The null hypothesis states that the statistical relationship between

256

a considered variable and the rainfall is arbitrary. For a variable that is not currently

257

included in the model, if there is sufficient evidence to reject the null hypothesis, the

258

variable would be added to the model. Conversely, for a variable that is included in the

259

model, if there is insufficient evidence to reject the null hypothesis, the variable would be

260

removed from the model. The null hypothesis is checked through an F-statistic and by

11

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

261

computing the corresponding p-value. The p-value shows the statistical significance level

262

of the model and indirectly can be used to control the number of predictive variables in

263

the regression model. Smaller p-values result in higher level of statistical significance and

264

fewer variables in the model. In this study, the p-values for ROMs and RAMs are set to

265

0.001 and 0.05 respectively. The procedure of determination of predictive variables

266

through the stepwise regression is stopped when no more predictors can be justifiably

267

entered or removed from the stepwise model (Ryan 1997).

268



Skewness Reduction

269

In linear regression modeling, there is an assumption that variables follow the normal

270

distribution. Non-normally distributed variables (e.g. highly skewed or kurtotic variables)

271

can distort relationships and significance tests. In such cases, the simple linear regression

272

on the original data could provide misleading results, or may not be the best test available

273

(Draper and Smith 1998). Daily precipitation is a highly skewed variable, especially in

274

arid and semi-arid areas. Therefore, for skewness reduction of daily precipitation, in each

275

submodel (ROM and RAM), a TF is applied on daily rainfall. The considered TF for

276

ROM submodel is the Box-Cox transformation after checking different common transfer

277

functions. This TF is selected after checking different types of TFs. This TF is applied on

278

daily rainfall values as shown in Equation 2. TF of RAM is Box-Cox transformation

279

applied on the square root of daily precipitation as given in Equation 3 (Box and Cox

280

1964; Li 2005). Different power transformations can be considered in TF of RAM but in

281

this study, the best results were obtained using square root transformation.

282

a. TF of ROM submodel (𝜆) 𝑦𝑖

(𝑦𝑖 + 1)𝜆 − 1 = { 𝜆 ln(𝑦𝑖 + 1)

𝑖𝑓 𝜆 ≠ 0

(2)

𝑖𝑓 𝜆 = 0

12

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

283

b. TF of RAM submodel (𝜆) 𝑦𝑖

(𝑦𝑖 0.5 )𝜆 − 1 = { 𝜆 ln(𝑦𝑖 0.5 )

𝑖𝑓 𝜆 ≠ 0

(3)

𝑖𝑓 𝜆 = 0 (𝜆)

284

where 𝑦𝑖 is daily precipitation amount, 𝑦𝑖

285

is the Box-Cox transformation parameter. The TF of ROM is applied on all days but TF

286

of RAM is just applied on wet days.

287



is transformed daily precipitation data and 𝜆

Model Parameters Estimation

288

Ordinary least squares (OLS) is a standard method for estimating the unknown parameters

289

in a linear regression model, with the goal of minimizing the differences between the

290

observed and simulated values. The sum squared error (SSE) or mean squared error

291

(MSE) that are usually used as objective functions in OLS are calculated as follows (Ryan

292

1997): 𝑛

𝑆𝑆𝐸 = ∑(𝑦𝑖 − 𝑓(𝑥 ⃗⃗⃗𝑖 , 𝛽 ))2

(4)

𝑖=1 𝑛

1 𝑀𝑆𝐸 = ∑(𝑦𝑖 − 𝑓(𝑥 ⃗⃗⃗𝑖 , 𝛽 ))2 𝑛

(5)

𝑖=1

293

where 𝑦𝑖 is the observed value, 𝑓 is the developed linear regression model, 𝑥 is the

294

predictor variables vector, 𝛽 is the linear model parameters and 𝑛 is the number of

295

observations.

296

The observed daily precipitation histograms are commonly strongly skewed. The mean

297

and the median of the precipitation data are different because the data are not normally

298

distributed. As the majority of daily precipitation data are concentrated around the mean

299

value, the effect of extreme rainfall amounts is underestimated in the estimation of MSE.

300

As a result, the variance of the model errors is not independent of its value. Inconstant

13

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

301

error variance is often termed “heteroscedasticity” (Carroll and Ruppert 1988; Draper and

302

Smith 1998). But, one of the assumptions of the regression model is that the error variance

303

is the same everywhere. In heteroscedasticity case, OLS measures are inefficient because

304

they give equal weight to all observations regardless of the fact that those with large

305

residuals contain less information about the regression.

306

One approach to overcoming this issue is using the weighted least squares (WLS) for

307

estimating model parameters. In WLS, the weighted sum of square errors (WSSE) is

308

calculated as follows: 𝑛

𝑊𝑆𝑆𝐸 = ∑ 𝑤𝑖 (𝑦𝑖 − 𝑓(𝑥 ⃗⃗⃗𝑖 , 𝛽 ))2

(6)

𝑖=1

309

where 𝑤𝑖 is the weight of an observation. In the current study, weights are estimated based

310

on data variance in each percentile. The variance of daily rainfall amount data in each

311

percentile is calculated as follows: 𝑛𝑝

𝛿𝑝

2

1 = ∑(𝑦𝑗 − 𝑦̅)2 𝑛𝑝

(7)

𝑗=1

312

where 𝛿𝑝 2 is the variance of the observed precipitation in the pth percentile, 𝑛𝑝 is the

313

number of observations in the pth percentile and 𝑦̅ is the mean daily rainfall amount. Then,

314

𝑤𝑝 for each percentile is estimated as follows: 𝑤𝑝 =

𝛿𝑝 2

(8)

2 ∑100 𝑛=1 𝛿𝑛

315

where 𝑤𝑝 is the initial weight of the pth percentile. This weight is used for all rainfall

316

amounts between pth and (p+1)th percentiles. For more accuracy and flexibility of the

317

model in reproducing extreme rainfall events values, the initial weights are modified

318

during the model training process as follows:

14

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

319

a: for high percentiles (above 98 percent) and low percentiles (below 50 percent) the

320

initial weights are increased;

321

b: for middle percentiles (between 50 and 98 percent) the initial weights are decreased;

322

The amount of change in initial weights in each of the above-mentioned categories is

323

determined by trial and error to provide a better model performance. 

324 325

Regression Models

The mathematical formulations of ROM and RAM are as follows: 1 ⃗⃗⃗⃗𝑜 , ⃗⃗⃗⃗ 𝑅𝑂𝑀𝑘 (𝑥 𝛽𝑘 ) = { 0

−1 ⃗⃗⃗⃗𝑜 , ⃗⃗⃗⃗ 𝑖𝑓 (𝐹𝑘 (𝑥 𝛽 𝑘 ))(𝜆 ) ≥ 𝑅 𝑡 𝑘

(9)

𝑂𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

𝑘 𝑎 ⃗⃗⃗⃗𝑎 , ⃗⃗⃗⃗ 𝑅𝐴𝑀𝑘 (𝑥 𝛼 𝑘 ) = 𝛼0𝑘 + 𝛼1𝑘 𝑥1𝑎 + 𝛼2𝑘 𝑥2𝑎 + ⋯ + 𝛼𝑚 𝑥𝑚

(10)

326

where ⃗⃗⃗⃗ 𝑥 𝑜 and ⃗⃗⃗⃗ 𝑥 𝑎 are the vectors of the selected predictive variables for ROM and RAM

327

submodels respectively, resulting from stepwise regression procedure, ⃗⃗⃗⃗ 𝛽 𝑘 and ⃗⃗⃗⃗ 𝛼 𝑘 are the

328

coefficient vectors in linear models of ROM and RAM for the kth CT, and 𝑅 𝑡 𝑘 is the

329

rainfall event threshold that distinguishes between wet and dry days. If the output value

330

of the regression model developed for ROM model for a day is greater than or equal to

331

𝑅 𝑡 𝑘 , that day is considered a wet day and vice versa. The value of this threshold is a

332

parameter in the model that should be determined in a way to better fit the distribution of

333

the observed daily precipitation values. 𝜆−1 represents the reverse transformation

334

function. For each CT, ROM and RAM submodels (𝑅𝑂𝑀𝑘 and 𝑅𝑂𝑀𝑘 ) are developed.

335

The final statistical model (𝑅𝑘 ) is the product of ROM and RAM as follows: (𝜆−1 )

⃗⃗⃗⃗𝑎 , ⃗⃗⃗⃗ ⃗⃗⃗⃗𝑎 , ⃗⃗⃗⃗ 𝑅𝑘 (𝑥 𝑥 𝑜 ) = 𝑅𝐴𝑀 (𝑥 𝛼𝑘)

⃗⃗⃗⃗𝑜 , ⃗⃗⃗⃗ × 𝑅𝑂𝑀 (𝑥 𝛽𝑘 )

(11)

336

⃗⃗⃗⃗𝑎 , ⃗⃗⃗⃗ where 𝑅𝑘 (𝑥 𝑥 𝑜 ) is the transfer function which downscales the large scale (mesocale)

337

atmospheric variables to local precipitation.

338

3

Case study 15

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

339

The SSDM is applied for downscaling daily precipitation in an area located in the north

340

of Tehran. The location of the study area is shown in Figure 3. The data used for SSDM

341

development in the study area are described in the following.

342

3.1

343

Data 

Observed Daily Precipitation

344

In this study, the observed daily precipitation dataset provided by the APHRODITE

345

project for the Middle East is utilized (Yatagai et al. 2012). Its state-of-the-art gridded

346

daily precipitation datasets, with high-resolution grids over Asia, are based on

347

interpolations of data collected from the rain-gauge observation networks. The data

348

release for the APHRO_V1101_ME for the period of 1970 to 2006 (13514 days) is used.

349

The network grid of APHRODITE dataset, around Tehran’s urban area and the

350

considered grid in the north of Tehran are shown in Figure 3.

351

Days with equal or more than 1 mm precipitation are considered wet days. This threshold

352

is determined regarding the rainfall recording errors and by checking different values for

353

it. The long-term average of annual rainfall in the study area is about 383 mm and just

354

24% of days are rainy. The histogram of daily rainfall amounts is shown in Figure 4. It

355

can be seen that it is strongly skewed. The average daily rainfall amount is 4.4 mm.

356

According to the national urban drainage systems regulations, the drainage systems are

357

designed based on maximum daily rainfalls with 2 years return period. Therefore, in this

358

study the 2-year maximum daily rainfall amount that is determined based on annual

359

maximum daily rainfall time series, is used as the threshold of extreme rainfall events

360

which stands at 20 mm/day for the study area.

361



Large-Scale Atmosphere Variables

16

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

362

Large scale atmosphere variables on a 2.5° × 2.5° grid network over Iran and around

363

Tehran’s urban region from 1970 to 2006, obtained from the National Centers for

364

Environmental Prediction/National Center for Atmospheric Research (NCEP/NCAR)

365

reanalysis archive (Kalnay et al. 1996; Kistler et al. 2001), were utilized in this study as

366

the observed large-scale atmosphere variables to develop the SSDM model. Data grids in

367

synoptic domain (from 42.5° to 60° longitudes and from 25° to 40° latitudes including 70

368

grids) and mesoscale domain (from 50° to 52.5° longitudes and 35° to 37.5° latitudes

369

including 4 grids around the study area) were respectively utilized for synoptic scale

370

pattern classification and rainfall SD models development. This domain is determined

371

based on the previous studies in the study area such as Alijani (2002) and Raziei et al.

372

(2012 and 2013). These spatial domains are shown in Figure 3. NCEP Reanalysis data

373

provided by the NOAA/OAR/ESRL PSD, Boulder, Colorado, USA, are available at

374

http://www.esrl.noaa.gov/psd/.

375



GCM outputs and scenarios

376

Climate change scenarios simulation results of Community Climate System Model,

377

version 4 (CCSM4) GCM under 2 different representative concentration pathways

378

(RCPs) namely RCP4.5 and RCP8.5 in a near future period (from 2006 to 2050) were

379

utilized for evaluation of climate change impacts on daily precipitation in the study area

380

as an experimental analysis. CCSM4 developed at the National Center for Atmospheric

381

Research (NCAR), is one of the climate models included in both the Coupled Model

382

Intercomparison Project’s fifth phase (CMIP5) and the Intergovernmental Panel on

383

Climate Change (IPCC) Fifth Assessment Report (AR5) (Flato et al. 2013). The CCSM4

384

exchanges state information and fluxes between four main components of atmosphere,

385

land, ocean, and sea ice. The CCSM4 was introduced to the climate science community

17

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

386

in April 2010 (Gent et al. 2011). Historical experiment from 1970 to 2005 was also

387

utilized for the evaluation of CCSM4 model performance in representation of the

388

observed daily rainfall pattern in the study area. CCSM4 and NCEP reanalysis data are

389

used for the same grids.

390 391

3.2

392

3.2.1 Phase I) Synoptic Scale Pattern Classification

393

Developing SSDM Model



Synoptic Pressure Patterns

394

Upper-air pressure patterns provide insights into the atmospheric conditions that are more

395

conducive (or more un-conducive) to rainfall formation (Lutgens and Tarbuck 2013;

396

Yarnal 1993). Previous studies have shown that changes in moisture delivery to the

397

majority of Iran are domestically controlled by interrelated, synoptic-scale atmospheric

398

pressure patterns (Alijani 2002; Alijani and Harman 1985). Alijani and Harman (1985)

399

concluded that the uplift mechanisms account for more than 70 percent of rainfall events

400

in Iran based on a subjective analysis of daily surface and 500-hPa atmospheric pressure

401

patterns.

402

Raziei et al. (2013) investigated the relationship between daily large-scale atmospheric

403

CTs and wintertime daily precipitation over Iran during the period 1965–2000. In their

404

previous studies, they identified the dominant daily atmospheric CTs in the Middle East

405

and studied their relationship with the occurrence of meteorological dry/wet spells during

406

winter in western Iran (Raziei et al. 2012). They classified daily large-scale weather

407

conditions during the period 1965–2000 into 12 CTs by applying PCA to the 500-hPa

408

geopotential height fields. Results showed that just some limited number of identified

409

CTs affect the precipitation occurrence over the majority of the country while the

18

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

410

remaining CTs commonly provide regional or negligible contributions to precipitation.

411

Therefore, in this study, patterns of 500-hPa geopotential height were used to determine

412

the synoptic scale atmospheric conditions. For this purpose, the daily average 500-hPa

413

geopotential heights on the synoptic domain scale over Iran (as mentioned in section 3.1)

414

from 1970 to 2007 (13514 cases/days) were considered.

415



SOM Implementation

416

The software that is used in this study for SOM implementation is the public domain

417

software developed in MATLAB environment by Helsinki University of Technology’s

418

Laboratory of Computer and Information Science and is named "The SOM Toolbox"

419

(Vesanto et al. 2000). Here, the topology of SOMs is developed based on the sheet shaped

420

structures with a rectangular lattice neighborhood and the Euclidean distance is taken as

421

distance measure in training and clustering steps. Each case (the daily pattern of 500-hPa

422

geopotential height) is a vector of size 70 (number of grids in the study area). To eliminate

423

the effect of data variability in days with the same pressure pattern, the data for each grid

424

point were normalized as follows: 𝑔ℎ500𝑛𝑜𝑟𝑚 = 𝑖,𝑡

(𝑔ℎ500𝑖,𝑡 − 𝑔ℎ500𝑚𝑒𝑎𝑛 ) 𝑡 𝑣𝑎𝑟 𝑔ℎ500𝑡

(12)

425

where 𝑔ℎ500𝑛𝑜𝑟𝑚 and 𝑔ℎ500𝑖,𝑡 are the normalized and original values of 500-hPa 𝑖,𝑡

426

geopotential height in point i and day t, respectively, and 𝑔ℎ500𝑚𝑒𝑎𝑛 and 𝑔ℎ500𝑣𝑎𝑟 are 𝑡 𝑡

427

respectively the mean and variance values of 500-hPa geopotential heights of all grids in

428

day t.

429

In this study two dimensional SOMs with different sizes, varying from 3×3 to 6×6

430

(different numbers of nodes between 9 and 36) were considered. As mentioned before, PI

431

values far from one, show that the developed CTs better distinguish between wet and dry

19

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

432

days. Therefore, the stratification index (SI) is used to evaluate SOM models performance

433

based on the average distance of PIs from one. 𝑛

1 𝑆𝐼 = ∑(𝑃𝐼𝑖 − 1)2 𝑛

(13)

𝑖=1

434

where 𝑛 is the number of nodes in the SOM and 𝑃𝐼𝑖 is the PI of the ith node of the SOM.

435

A higher value of SI means more discrete stratification of daily precipitation. In Table 1,

436

the values of SI for developed SOMs with different sizes are given. According to these

437

results, SOMs with sizes of 4×4, 5×5, 5×6 and, 6×6 have higher SI values and therefore,

438

provide a better performance.

439

By increasing the number of SOM nodes, based on SI, the distinction of wet days from

440

dry days is improved, but choosing the SOM with a bigger size means that more statistical

441

models should be developed in the second phase of SSDM algorithm. Hence in this study

442

the 4×4 SOM size is chosen for classification of synoptic scale pressure patterns.

443



Detected CTs and their relationship with local daily precipitation

444

Figure 5 shows the 16 CTs which were determined in this study by applying SOM

445

method. Each CT map is drawn by averaging the 500-hPa geopotential height values of

446

all days that belong to that CT. To investigate the contribution of CTs to daily

447

precipitation in the study area, some statistical and subjective evaluations were done.

448

The frequency and the chance of rain occurrence in each CT are presented in Figure 6.

449

By moving from right to left in CTs array, a very strong shift from higher to lower

450

pressure pattern and a simultaneous increase in the chance of precipitation occurrence can

451

be seen. Although the conditions become wetter by moving up in CTs array, there is not

452

a clear behavior in pressure patterns. The CTs in the first column (from the right hand

453

side) are associated with high-pressure systems’ (an anticyclone) occurrence over Iran

20

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

454

with almost zero chance of precipitation. By receding high-pressure systems from the

455

study region or expansion of the mid-latitude westerlies over Iran, the CTs in column 2

456

and 3 are formed and the chance of precipitation is increased. The CTs in the left side of

457

CT array represent the low-pressure systems. These results indicate that days are

458

distributed fairly over the CTs, with relative minima in central part that is related to

459

transitional patterns from high-pressure to low-pressure systems.

460

In Table 2, the frequency of wet days, the proportion of total rainfall amount and

461

frequency of extreme events within 16 CTs are presented and in Table 3 the PI values

462

associated with each CT are given. Based on these results, about 75% of rainy days

463

happen just in 25% of the CTs (4 CTs) including CT(1,4), CT(2,4), CT(1,3) and CT(3,4).

464

These CTs are the most conducive situations to the rainfall formation in the study area.

465

About 85% of the total rainfall amount falls under these four CTs in the study area.

466

Additionally, extreme rainfall events are just observed in these CTs where about 80% of

467

extremes happen in CT (1,4). The PI values of these CTs are also higher than those of

468

other CTs. These results are in good agreement with Raziei et al. (2013) that concluded

469

that a number of identified CTs affect the precipitation occurrence over the majority of

470

Iran.

471

On the other hand, less than 4% of wet days and 2% of rainfall amount occur in 6 CTs

472

which are the CTs in the first column (from right) and the two last ones in the second

473

column in CTs array. Because the number of rainy days and rainfall amount in these CTs

474

are negligible, they can be considered un-conducive situations to rainfall formation in the

475

study area. Also, PI values of this group of CTs are smaller than those of other CTs. The

476

rest of CTs have PI values less than 1 which means that they lead to precipitation amounts

21

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

477

less than the climatological mean value. However, precipitation occasionally occurs in

478

the study area under these CTs.

479

Figure 7 shows the composite map of vorticity at 500-hPa level and relative humidity of

480

700-hPa level corresponding to each CT. In most rainfall-conducive CTs, the tendency to

481

expand the upper-level atmosphere layers that results in uplift motions in the study area

482

is detectable from positive vorticity that expands around the study area. Positive vorticity

483

and high relative humidity (above 60%) of mid-level atmosphere result in uplift motions

484

of wet air mass that can result in rainfall formation in the study area. On the other side, in

485

most rainfall-un-conducive CTs, where a very high-pressure zone is observed over the

486

study area, negative vorticity and low level of humidity in the study area result in dry

487

weather conditions. In other CTs, one of the main components of rain formation (humidity

488

or uplift motion) is not enough, but sometimes local instabilities can result in rainfall

489

occurrence in the study area. It can be concluded that the classification of CTs in the study

490

area could successfully relate the synoptic-scale atmospheric circulation and the local

491

precipitation.

492

3.2.2 Phase II) Local Scale Statistical Downscaling Models for Daily Precipitation

493

After detecting CTs, days were categorized based on CTs and then statistical models were

494

developed for each category to simulate daily rainfall. For each model, the first step is

495

determining the predictors. The set of predictors considered for both statistical models of

496

ROM and RAM development are listed in Table 4. These variables are used at a spatial

497

mesoscale domain (4 neighbor grids around the study area as shown in Figure 3) and at

498

the three levels of atmosphere including near surface, mid-level (700-hPa geopotential

499

height) and high-level (500-hPa geopotential height). Three kinds of variables are

500

considered in this set to represent different conditions of the rain occurrence and are used

22

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

501

in ROMs development. Dew point is the indicator of the moisture content of the air.

502

Vorticity is the indicator of the tendency of upward motion of the atmosphere, and dew

503

point depression is the indicator of the level of saturation. In RAMs development, the air

504

temperature, geopotential height, and relative humidity as well as the sea level pressure

505

are considered in addition to the variables included in the ROM.

506

Using the stepwise regression procedure, the predictor variables in ROMs and RAMs

507

were chosen. The TFs (Box-Cox and square root), were also applied on the observed

508

rainfall data for skewness reduction purpose. 80% of daily data were randomly selected

509

for calibration of the models and the remaining data were used for models’ validation.

510

Both ROMs and RAMs are linear models and WLS was used to determine RAMs

511

coefficients as described in section 2.2.2. The selected variables in each model are given

512

in Table 5. Even though some un-conducive CTs include a limited number of wet days,

513

developing SD models for them with appropriate performance is impossible because of

514

limited data. Therefore, all days corresponding to these CTs, are considered dry days and

515

are not included in Table 5.

516

3.3

517

The comparison between some statistical characteristics of the downscaled and observed

518

daily rainfall values for calibration and validation datasets as well as the total dataset are

519

presented in Table 6. The differences between mean and standard deviation of the

520

observed and simulated rainfall amounts are less than 5% and 13%, respectively, in both

521

calibration and validation datasets. Although the difference between skewness of the

522

observed and simulated rainfall amounts in calibration dataset is small (3.3%), in the

523

validation dataset it is considerable (about 40%). Based on this table, the differences

524

between modeled and simulated percentages of wet days, mean daily rainfall and standard

Daily rainfall simulation results

23

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

525

deviation of wet days' rainfall amount, considering both calibration and validation period,

526

are less than 10% which show a good agreement between the observed and modeled

527

rainfall amounts. However, it should be mentioned that in simulation of the maximum

528

rainfall value and skewness of wet days' rainfall amount in the validation period, the

529

model is not very successful.

530

For a better comparison of model performance in calibration and validation datasets, the

531

Q-Q plots of the modeled and observed daily rainfalls are presented for both datasets in

532

Figure 8. As can be seen, the model performance in calibration is very good (the all points

533

lie quite close to the line) but as it is expected, in the validation dataset, the model

534

performance is not as good as the calibration dataset. In the validation data set, rainfalls

535

above 8 mm are underestimated. The maximum difference between the modeled and

536

observed daily rainfall values in the validation data set is smaller than 13%.

537

To evaluate the model performance in extreme events representation, the time series of

538

the simulated and observed annual maximum daily rainfalls are compared as given in

539

Figure 9. The two-sample Kolmogorov-Smirnov (KS) test which is a nonparametric

540

hypothesis test, is used to evaluate the similarity between the empirical CDF (cumulative

541

distribution function) of these two time series. The results of this test confirm that both

542

time series have the same distribution at 95% significant level.

543

The observed and modeled distributions of wet days are also compared in Figure 10.

544

Based on this figure, the percentile plots and CDFs of the observed and simulated wet

545

days match with each other and this shows the successful performance of the model. This

546

also approved by Q-Q plots. The model’s performance in monthly time scale is also given

547

in Figure 11. Wilcoxon signed-rank test and Levene's test (Karamouz et al. 2011) are

548

used, respectively, to test the equality of the mean and variance of observed and simulated

24

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

549

monthly rainfall values as well as the number of wet days. Based on Wilcoxon signed-

550

rank test, mean monthly rainfall values of the simulated and observed series are equal at

551

95% significance level. However, the maximum differences between simulated and

552

observed mean monthly rainfall values are observed in the months of March and January.

553

The equality of the simulated and observed monthly mean of wet days is only rejected in

554

July. Based on Levene's test, the equality of the variance of the simulated monthly rainfall

555

with the observed values is approved at 95% significance level except in months of May,

556

July and August. The same result is obtained for monthly variance of the number of wet

557

days. This can be due to the limited number of wet days in these months.

558 559

4

560

The developed SSDM model for downscaling daily rainfall was used to evaluate the

561

probable climate change impacts on the rainfall characteristics in the near future horizon

562

(2006-2050) in the study area. For this purpose, the CCSM4 GCM model outputs for the

563

historical, RCP4.5 and RCP8.5 experiments were used.

564

The acceptability of the chosen GCM performance in the simulation of the synoptic and

565

mesoscale climate characteristics of the study area was evaluated through comparing the

566

simulations of CCSM4 in the historical period of 1970 to 2005 with NCEP records. This

567

evaluation was done in two aspects:

568

Climate change impacts evaluation in the study area



569 570 571

The differences between rainfall occurrence frequencies of CTs in CCSM4 results and NCEP records, especially for most rainy CTs.



Comparison of the Q-Q plots of the predictors developed based on the CCSM4 outputs and NCEP records in the study area.

25

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

572

The 500-hPa geopotential height patterns of CCSM4 simulations for the historical period

573

and NCEP records in the synoptic scale were classified by applying the developed SOM.

574

The frequencies of daily patterns in each CT are given in Table 7 for both NCEP and

575

CCSM4. The differences between frequencies of CTs in CCSM4 simulations and NCEP

576

records are very low, especially in rainy CTs. The Q-Q plots are also used to illustrate the

577

capability of CCSM4 in reproducing NCEP records of predictors in the study area. For

578

example, Figure 12 shows the Q-Q plots of the vorticity and dew point depression values

579

in mesoscale grids for both CCSM4 outputs and NCEP records. These results confirm the

580

ability of CCSM4 for reproducing predictors in the study area at an acceptable level of

581

accuracy. The percentiles of the downscaled daily rainfall series by applying developed

582

SSDM on CCSM4 outputs (for the historical experiment) and NCEP records are

583

compared in Figure 13. The agreement between them in all percentiles is remarkable. The

584

maximum difference is observed in the last percentile which is less than 1.5%.

585

By applying the developed SSDM model to CCSM4 outputs for RCP4.5 and RCP8.5

586

experiments in the future period (2006 to 2050), the possible climate change impacts were

587

investigated. The CTs occurrence frequencies are given in Table 8. The results show that

588

the CTs occurrence frequencies do not change considerably under climate change

589

impacts. There are small alterations in different cases, but there is no consistent scheme.

590

In Figure 14, the distribution of the simulated daily rainfall for historical, RCP4.5 and

591

RCP8.5 experiment results are compared. The results show that there is an upward trend

592

in high frequency daily rainfall amounts (above 80%) in both future climate pathways,

593

but in low frequency amounts, there is not any significant change. To better detect

594

changes in future daily rainfall characteristics in the study area, the Q-Q plots of the

595

simulated wet day amounts in historical and two future RCPs are illustrated in Figure 15.

26

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

596

Based on this figure, as a result of climate change, the intensity of extreme events is

597

increasing in both RCPs and this increase is slightly more in RCP4.5.

598

The characteristics of wet days under climate change impacts are compared in Table 8.

599

Based on this table, it seems that the percentages of wet days are decreasing under climate

600

change impacts, but the average rainfall values are increasing. Furthermore, the increase

601

in the variance of daily rainfall shows the increasing possibility of extreme rainfall values.

602

The number of extreme rainfall events in a year is also increasing under climate change

603

impacts.

604

5

605

In this study, SSDM is proposed for daily rainfall downscaling that can preserve the

606

statistical characteristics of the observed local daily rainfalls especially extreme rainfall

607

events that are used in different hydrological investigations such as flood management

608

issues. SSDM is an SD model conditioned on synoptic scale CTs. The model includes

609

two parts. The first part is clustering synoptic weather patterns into CTs and then

610

investigating the relationships between identified CTs and local daily rainfall events. The

611

SOM is trained to cluster synoptic patterns into CTs and then it is used to classify each

612

day's synoptic pattern to determined CTs. The second part of SSDM includes developing

613

occurrence and amount regression SD models in each CT to simulate local daily rainfall.

614

The developed model is based on two main assumptions. The first one is that the

615

regression equations developed between predictors and rainfall will be valid in the future.

616

The second assumption is that the considered CTs would cover all the future events.

617

Further investigations would be necessary for the evaluation of validity of these

618

assumptions under climate change impacts.

Summary and Conclusion

27

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

619

The model is applied to the northern part of Tehran as the study area. In the first step, 16

620

CTs were determined using SOM model developed based on 500-hPa geopotential height.

621

The investigation of the relationship between determined CTs and local daily rainfall

622

demonstrates that detected CTs can explain the main atmospheric features that result in

623

dry and wet conditions in the study area. In the second step, two models for rainfall

624

occurrence and amount simulation are developed. The models' parameters are estimated

625

using WLS and skewness reduction techniques to better cover the extreme events in the

626

model development. Applying these methods for estimating regression models’

627

coefficients has improved the ability of developed models in reproducing local rainfall

628

amounts. Beside this improvement, still the model performance in simulation of some

629

ranges of rainfall amount is not satisfactory. This could be because of local phenomena

630

that affect the rainfall variability and are not considered in the developed models. Further

631

investigations on comparison of the developed model in this study with the other models

632

and the methods that could overcome this methods weaknesses can be considered in

633

future studies.

634

Then experiment outputs of CCSM4 GCM model under AR5 climate change scenarios

635

(RCP4.5 and RCP8.5) were used to investigate climate change impacts in the study area.

636

The results show that under climate change impacts the number of extreme rainfall events

637

is increasing. It must be emphasized that since in this study only one GCM and one run

638

of climate change scenarios were used, the obtained results are experimental. More

639

investigations through considering different climate change scenarios and different

640

GCMs are needed for a more reliable judgment about climate change impacts on rainfall

641

regimes in the study area.

642

28

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

643

6

Acknowledgment

644

We acknowledge the World Climate Research Program's Working Group on Coupled

645

Modeling, which is responsible for CMIP, and we thank the National Center for

646

Atmospheric Research (NCAR) for producing their model outputs (CCSM4) and making

647

them available for us. For CMIP, the U.S. Department of Energy's Program for Climate

648

Model Diagnosis and Intercomparison provided coordinating support and led the

649

development of software infrastructure in partnership with the Global Organization for

650

Earth System Science Portals.

651 652 653

7

References

654

Alijani, B. (2002). “Variations of 500 hPa flow patterns over Iran and surrounding areas

655

and their relationship with the climate of Iran.” Theoretical and Applied

656

Climatology, 72(1–2), 41–54.

657 658 659

Alijani, B., and Harman, J. R. (1985). “Synoptic Climatology of Precipitation in Iran.” Annals of the Association of American Geographers, 75(3), 404–416. Arnbjerg-Nielsen, K. (2008). “Quantification of climate change impacts on extreme

660

precipitation used for design of sewer systems.” 11th International Conference on

661

Urban Drainage, Edinburgh, Scotland, UK, 1–10.

662

Baigorria, G. A., and Jones, J. W. (2010). “GiST: A Stochastic Model for Generating

663

Spatially and Temporally Correlated Daily Rainfall Data.” Journal of Climate,

664

23(22), 5990–6008.

665

Benestad, R. E. (2009). “Downscaling precipitation extremes.” Theoretical and Applied 29

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

666 667

Climatology, 100(1–2), 1–21. Benestad, R. E., Hanssen-Bauer, I., and Chen, D. (2008). Empirical-Statistical

668

Downscaling. Empirical-Statistical Downscaling, World Scientific Publishing

669

Company.

670 671 672 673 674 675 676

Box, G. E. P., and Cox, D. R. (1964). “An analysis of transformations.” Journal of the Royal Statistical Society. Series B, 26, 211–252. Carroll, R. J., and Ruppert, D. (1988). Transformation and Weighting in Regression. Springer-Science+Business Media, B.Y. Chau, K. W., and Wu, C. L. (2010). “A hybrid model coupled with singular spectrum analysis for daily rainfall prediction.” Journal of Hydroinformatics, 12(4), 458. Chen, S.-T., Yu, P.-S., and Tang, Y.-H. (2010). “Statistical downscaling of daily

677

precipitation using support vector machines and multivariate analysis.” Journal of

678

Hydrology, 385(1–4), 13–22.

679

Cheng, C. S., Li, G., Li, Q., and Auld, H. (2010). “A Synoptic Weather Typing

680

Approach to Simulate Daily Rainfall and Extremes in Ontario, Canada: Potential

681

for Climate Change Projections.” Journal of Applied Meteorology and

682

Climatology, 49(5), 845–866.

683

Cheng, C. S., Li, G., Li, Q., and Auld, H. (2011). “A Synoptic Weather-Typing

684

Approach to Project Future Daily Rainfall and Extremes at Local Scale in Ontario,

685

Canada.” Journal of Climate, 24(14), 3667–3685.

686 687

D’onofrio, A., Boulanger, J.-P., and Segura, E. C. (2010). “CHAC: a weather pattern classification system for regional climate downscaling of daily precipitation.”

30

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

688 689

Climatic Change, 98(3–4), 405–427. Dibike, Y. B., Gachon, P., St-Hilaire, A., Ouarda, T. B. M. J., and Nguyen, V. T.-V. V.

690

(2008). “Uncertainty analysis of statistically downscaled temperature and

691

precipitation regimes in Northern Canada.” Theoretical and Applied Climatology,

692

OURANOS Consortium on Regional Climatology and Adaptation to Climate

693

Change, Montreal, QC, Canada, 91(1–4), 149–170.

694 695 696

Draper, N. R., and Smith, H. (1998). Applied Regression Analysis, 3rd Edition. Wiley Series in Probability and Statistics. Fealy, R., and Sweeney, J. (2007). “Statistical downscaling of precipitation for a

697

selection of sites in Ireland employing a generalised linear modelling approach.”

698

International Journal of Climatology, 27(15), 2083–2094.

699

Feser, F., Rockel, B., von Storch, H., Winterfeldt, J., and Zahn, M. (2011). “Regional

700

Climate Models Add Value to Global Model Data: A Review and Selected

701

Examples.” Bulletin of the American Meteorological Society, American

702

Meteorological Society, 92(9), 1181–1192.

703

Gent, P. R., Danabasoglu, G., Donner, L. J., Holland, M. M., Hunke, E. C., Jayne, S. R.,

704

Lawrence, D. M., Neale, R. B., Rasch, P. J., Vertenstein, M., Worley, P. H., Yang,

705

Z. L., and Zhang, M. (2011). “The community climate system model version 4.”

706

Journal of Climate, 24, 4973–4991.

707

Haberlandt, U., Belli, A., and Bárdossy, A. (2015). “Statistical downscaling of

708

precipitation using a stochastic rainfall model conditioned on circulation patterns -

709

an evaluation of assumptions.” International Journal of Climatology, John Wiley

710

and Sons Ltd, 35(3), 417–432.

31

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

711

Hertig, E., and Jacobeit, J. (2013). “A novel approach to statistical downscaling

712

considering nonstationarities: application to daily precipitation in the

713

Mediterranean area.” Journal of Geophysical Research: Atmospheres, 118(2),

714

520–533.

715

Hessami, M., Gachon, P., Ouarda, T. B. M. J. M. J., and St-Hilaire, A. (2008).

716

“Automated regression-based statistical downscaling tool.” Environmental

717

Modelling & Software, 23(6), 813–834.

718

Hewitson, B. C., and Crane, R. G. (2006). “Consensus between GCM climate change

719

projections with empirical downscaling: precipitation downscaling over South

720

Africa.” International Journal of Climatology, 26(10), 1315–1337.

721 722 723

Hewitson, B., and Crane, R. (2002). “Self-organizing maps: applications to synoptic climatology.” Climate Research, 22, 13–26. Hundecha, Y., and Bárdossy, A. (2008). “Statistical downscaling of extremes of daily

724

precipitation and temperature and construction of their future scenarios.”

725

International Journal of Climatology, 28(5), 589–610.

726 727

IPCC. (2014). Climate Change 2013 - The Physical Science Basis. (Intergovernmental Panel on Climate Change, ed.), Cambridge University Press, Cambridge.

728

Kalnay, E., Kanamitsu, M., Kistler, R., Collins, W., Deaven, D., Gandin, L., Iredell, M.,

729

Saha, S., White, G., Woollen, J., Zhu, Y., Leetmaa, A., Reynolds, R., Chelliah, M.,

730

Ebisuzaki, W., Higgins, W., Janowiak, J., Mo, K. C., Ropelewski, C., Wang, J.,

731

Jenne, R., and Joseph, D. (1996). “The NCEP/NCAR 40-Year Reanalysis Project.”

732

Bulletin of the American Meteorological Society, American Meteorological

733

Society, 77(3), 437–471.

32

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

734

Karamouz, M., Goharian, E., and Nazif, S. (2013a). “Reliability Assessment of the

735

Water Supply Systems under Uncertain Future Extreme Climate Conditions.”

736

Earth Interactions, 17(20), 1–27.

737

Karamouz, M., Nazif, S., and Zahmatkesh, Z. (2013b). “Self-Organizing Gaussian-

738

Based Downscaling of Climate Data for Simulation of Urban Drainage Systems.”

739

Journal of Irrigation and Drainage Engineering, 139(2), 98–112.

740

Kistler, R., Collins, W., Saha, S., White, G., Woollen, J., Kalnay, E., Chelliah, M.,

741

Ebisuzaki, W., Kanamitsu, M., Kousky, V., van den Dool, H., Jenne, R., and

742

Fiorino, M. (2001). “The NCEP–NCAR 50–Year Reanalysis: Monthly Means CD–

743

ROM and Documentation.” Bulletin of the American Meteorological Society,

744

American Meteorological Society, 82(2), 247–267.

745

Kohonen, T. (1995). “Self-Organizing Maps.” Springer-Verlag, Heidelberg.

746

Kohonen, T., and Honkela, T. (2007). “Kohonen network.” Scholarpedia, 2(1), 1568.

747

Li, F. (2005). “Box-Cox Transformations: An Overview (slide presentation).”

748 749 750

Department of Statistics, University of Connecticut. Lutgens, F. K., and Tarbuck, E. J. (2013). The Atmosphere, An Introduction to Meteorology. Pearson.

751

Mearns, L. (1995). “Analysis of daily variability of precipitation in a nested regional

752

climate model: comparison with observations and doubled CO2 results.” Global

753

and Planetary Change, 10(1–4), 55–78.

754 755

Najafi, M. R., Moradkhani, H., and Wherry, S. A. (2011). “Statistical Downscaling of Precipitation Using Machine Learning with Optimal Predictor Selection.” Journal

33

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

756 757

of Hydrologic Engineering, American Society of Civil Engineers, 16(8), 650–664. Pasini, A. (2009). “Neural Network Modeling in Climate Change Studies.” Artificial

758

Intelligence Methods in the Environmental Sciences, S. E. Haupt, A. Pasini, and C.

759

Marzban, eds., Springer Netherlands, Dordrecht, 424.

760

Raziei, T., Bordi, I., Santos, J. a., and Mofidi, A. (2013). “Atmospheric circulation types

761

and winter daily precipitation in Iran.” International Journal of Climatology, 33(9),

762

2232–2246.

763

Raziei, T., Mofidi, A., Santos, J. a., and Bordi, I. (2012). “Spatial patterns and regimes

764

of daily precipitation in Iran in relation to large-scale atmospheric circulation.”

765

International Journal of Climatology, 32(8), 1226–1237.

766 767

Richardson, C. W. (1981). “Stochastic simulation of daily precipitation, temperature, and solar radiation.” Water Resources Research, 17(1), 182–190.

768

Ryan, T. P. (1997). Modern Regression Methods. Wiley, New York.

769

Sarr, M. A., Seidou, O., Tramblay, Y., and El Adlouni, S. (2015). “Comparison of

770

downscaling methods for mean and extreme precipitation in Senegal.” Journal of

771

Hydrology: Regional Studies, 4, 369–385.

772 773 774

Semenov, M. A. (2008). “Simulation of extreme weather events by a stochastic weather generator.” Climate Research, 35(3), 203–212. Semenov, M. A., and Barrow, E. M. (1997). “Use of A Stochastic Weather Generator In

775

The Development Of Climate Change Scenarios.” Climatic Change, Kluwer

776

Academic Publishers, 35(4), 397–414.

777

Sheridan, S. C., and Lee, C. C. (2011). “The self-organizing map in synoptic

34

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

778

climatological research.” Progress in Physical Geography, 35(1), 109–119.

779

Sunyer, M. A., Hundecha, Y., Lawrence, D., Madsen, H., Willems, P., Martinkova, M.,

780

Vormoor, K., Bürger, G., Hanel, M., Kriaučiūnienė, J., Loukas, A., Osuch, M., and

781

Yücel, I. (2015). “Inter-comparison of statistical downscaling methods for

782

projection of extreme precipitation in Europe.” Hydrology and Earth System

783

Sciences, Copernicus GmbH, 19(4), 1827–1847.

784

Tavakol-Davani, H., Nasseri, M., and Zahraie, B. (2013). “Improved statistical

785

downscaling of daily precipitation using SDSM platform and data-mining

786

methods.” International Journal of Climatology, 33(11), 2561–2578.

787

Tomassetti, B., Verdecchia, M., and Giorgi, F. (2009). “NN5: A neural network based

788

approach for the downscaling of precipitation fields – Model description and

789

preliminary results.” Journal of Hydrology, 367(1–2), 14–26.

790

Tryhorn, L., and DeGaetano, A. (2011). “A comparison of techniques for downscaling

791

extreme precipitation over the Northeastern United States.” International Journal

792

of Climatology, 31(13), 1975–1989.

793

Vesanto, J., Himberg, J., Alhoniemi, E., and Parhankangas, J. (2000). “Self-organizing

794

map in Matlab: the SOM Toolbox.” Laboratory of Computer and Information

795

Science, Helsinki University of Technology, Helsinki, Finland.

796 797 798

Vrac, M., and Naveau, P. (2007). “Stochastic downscaling of precipitation: From dry events to heavy rainfalls.” Water Resources Research, 43(7), n/a--n/a. Wang, L., Ranasinghe, R., Maskey, S., van Gelder, P. H. A. J. M., and Vrijling, J. K.

799

(2015). “Comparison of empirical statistical methods for downscaling daily climate

800

projections from CMIP5 GCMs: a case study of the Huai River Basin, China.” 35

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

801

International Journal of Climatology, John Wiley and Sons Ltd, n/a-n/a.

802

Wilby, R. L., and Dawson, C. W. (2013). “The Statistical DownScaling Model: insights

803

from one decade of application.” International Journal of Climatology, 33(7),

804

1707–1719.

805

Wilby, R. L., Dawson, C. W., and Barrow, E. M. (2002). “sdsm — a decision support

806

tool for the assessment of regional climate change impacts.” Environmental

807

Modelling & Software, 17(2), 145–157.

808

Wilby, R. L., Wigley, T. M. L., Conway, D., Jones, P. D., Hewitson, B. C., Main, J.,

809

and Wilks, D. S. (1998). “Statistical downscaling of general circulation model

810

output: A comparison of methods.” Water Resources Research, 34(11), 2995–

811

3008.

812

Wilks, D. S. (2010). “Use of stochastic weathergenerators for precipitation

813

downscaling.” Wiley Interdisciplinary Reviews: Climate Change, John Wiley &

814

Sons, Inc., 1(6), 898–907.

815

Willems, P., Arnbjerg-Nielsen, K., Olsson, J., and Nguyen, V. T. V. (2012). “Climate

816

change impact assessment on urban rainfall extremes and urban drainage: Methods

817

and shortcomings.” Atmospheric Research, Elsevier B.V., 103, 106–118.

818

Willems, P., and Vrac, M. (2011). “Statistical precipitation downscaling for small-scale

819

hydrological impact investigations of climate change.” Journal of Hydrology,

820

402(3–4), 193–205.

821 822

Yarnal, B. (1993). Synoptic climatology in environmental analysis: a primer. Belhaven Press.

36

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

823

Yarnal, B., Comrie, A. C., Frakes, B., and Brown, D. P. (2001). “Developments and

824

prospects in synoptic climatology.” International Journal of Climatology, 21(15),

825

1923–1950.

826

Yatagai, A., Kamiguchi, K., Arakawa, O., Hamada, A., Yasutomi, N., and Kitoh, A.

827

(2012). “Aphrodite constructing a long-term daily gridded precipitation dataset for

828

Asia based on a dense network of rain gauges.” Bulletin of the American

829

Meteorological Society, 93, 1401–1415.

830

Zhang, X., Wang, X. L., and Corte-Real, J. (1997). “On the relationships between daily

831

circulation patterns and precipitation in Portugal.” Journal of Geophysical

832

Research, 102, 13495.

833

37

1 834 Table 1: The SI values for different SOM dimensions size 2 3 4 33 34 35 36 44 5SOM (Self-organizing Maps) size 6 7 Number of nodes 9 12 15 18 16 8 SI (stratification index) 0.493 0.476 0.467 0.471 0.527 9 835 10 11 12 836 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 38 63 64 65

45

46

55

56

66

20 24 25 30 36 0.521 0.513 0.555 0.546 0.583

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

837

Table 2: The distribution of daily rainfall characteristics in CTs Rain characteristics

838

CTs* (circulation types)

1

2

3

1 0.6% 3.0% 9.6% 2 0.6% 2.9% 5.0% Rainfall occurrence percentages 3 0.4% 0.8% 3.4% 4 0.5% 0.8% 2.3% 1 0.4% 1.9% 9.2% 2 0.4% 1.6% 3.6% Proportion to total rainfall amount 3 0.2% 0.4% 2.5% 4 0.2% 0.4% 1.4% 1 0.0% 0.0% 10.8% 2 0.0% 0.0% 0.0% Proportion to number of extreme events 3 0.0% 0.0% 0.0% 4 0.0% 0.0% 0.0% th th * The given value in the i row and the j column corresponds to CT (i,j)

839 840 841

39

4 41.5% 16.9% 8.1% 3.7% 54.2% 15.5% 6.0% 2.0% 78.4% 10.8% 0.0% 0.0%

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

842

843

Table 3: PI values of the developed CTs CTs* (circulation types) 1 2 3 4 1 0.08 0.42 1.31 3.16 2 0.13 0.40 0.75 1.56 3 0.07 0.19 0.60 0.81 4 0.05 0.15 0.32 0.38 th th * The given value in the i row and the j column corresponds to CT (i,j)

844 845

40

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

846

Table 4: The variable sets considered in development of statistical models of SSDM Submodel

Variable's name

ROM

Vor

°C

Dew

°C

Dep

s-1

Vor

°C

Dew

°C

Dep

Air Temperature

°C

Ta

Geopotential height

m

hgt

700hp, 500hp

Relative Humidity

%

Rhu

Surface, 700hp, 500hp

Sea Level Pressure

N/m2

Slp

Surface

Dew point Temperature Dew point depression Vorticity

RAM (Rainfall amount model)

Considered levels Surface, 700hp, 500hp Surface, 700hp, 500hp Surface, 700hp, 500hp Surface, 700hp, 500hp Surface, 700hp, 500hp Surface, 700hp, 500hp Surface, 700hp, 500hp

s-1

Vorticity (Rainfall occurrence model)

Unit Variable abbreviation

Dew point Temperature Dew point depression

847 848

41

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

849

Table 5: The selected variables (predictors) in the developed statistical models for each

850

CT

851

Predictors* ROM (Rainfall occurrence model) RAM (Rainfall amount model) (1,2) 700hp-Dep, 500hp-Dep 700hp-Rhu 700hp-Rhu, 500hp-Rhu, (1,3) 700hp-Dep, 500hp-Dep 700hp-Vor, 500hp-Vor 700hp-Dew, 500hp-Ta, 700hp-Dep, 500hp-Dep, 500hp-Rhu, (1,4) 700hp-Vor, 500hp-Vor 700hp-hgt, Surface-Slp, Surface -Ta 500hp-Rhu, (2,2) 700hp-Dep, 500hp-Dep 700hp-Vor, 700hp-hgt (2,3) 500hp-Dep 700hp-Rhu 700hp-Dew, 500hp-Dew, 700hp-Dep, 500hp700hp-Rhu, (2,4) Dep 500hp-Rhu, Surface-Ta (3,2) 500hp-Dep 700hp-Vor (3,3) 500hp-Dep 500hp-Rhu (3,4) 700hp-Dep, 500hp-Dep 700hp-Rhu (4,3) 500hp-Dep 500hp-Ta 500hp-Dep, (4,4) 700hp-Dep, 500hp-Dep 500hp-Rhu, 700hp-Rhu * The predictors' descriptions are given in Table 4. CT (circulation types)

852

42

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

853

Table 6: Comparing SSDM results and the observed rainfalls in calibration and validation

854

periods as well as total data. Statistical characteristics

Percent of wet days (%) Average wet days rainfall amount (mm) Standard deviation of wet days rainfall amount (mm) Skewness of wet days rainfall amount Maximum daily rainfall amount (mm) Wet days coefficient of correlation (R) Root Mean Squared Error (RMSE, mm) 855

Calibration Validation Observed Modeled Observed Modeled 23.6 21.5 24.6 21.3

Total Observed Modeled 23.8 21.5

4.3

4.5

4.7

4.6

4.4

4.5

4.1

3.7

4.6

4

4.2

4

3

3.1

2.6

3.7

2.9

3.2

47.1

46.3

32

45.8

47.1

46.3

0.44

0.44

0.44

2.45

2.73

2.51

856

43

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

857

Table 7: CTs occurrence distributions in observed (NCEP) and CCSM4 outputs for the

858

historical and future time periods

Source of considered data

859

CTs * (circulation types)

1

2

3

1 6.7% 5.1% 7.1% 2 3.9% 4.5% 4.9% NCEP (1970-2005) 3 4.0% 2.7% 4.5% 4 7.6% 3.4% 4.9% 1 5.7% 4.8% 6.3% 2 3.1% 3.6% 4.5% CCSM4-historical (1970-2005) 3 3.2% 2.8% 4.1% 4 8.5% 4.9% 5.5% 1 5.0% 5.1% 7.0% 2 3.4% 4.5% 4.7% CCSM4-RCP4.5 (2006-2050) 3 3.3% 3.7% 4.0% 4 8.7% 5.1% 6.0% 5 5.9% 5.0% 7.1% 6 3.6% 4.5% 5.0% CCSM4-RCP8.5 (2006-2050) 7 3.4% 3.6% 4.0% 8 7.9% 5.1% 5.1% * The given value in the ith row and the jth column corresponds to CT (i,j)

860

44

4 16.4% 9.9% 8.1% 6.4% 16.6% 9.6% 8.1% 8.7% 16.0% 8.6% 7.4% 7.6% 15.9% 8.5% 7.3% 8.1%

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

861

Table 8: Comparing the statistics of the daily precipitation in the study area for the

862

historical period and future under RCP4.5 and RCP8.5 simulations by CCSM4 Historical (1970-2005) Percent of wet days (%) 23.9 Average wet days rainfall values (mm) 4.2 Standard deviation of wet days rainfall 3.9 values (mm) Skewness of wet days rainfall values 3.6 Maximum daily rainfall value (mm) 53.3 Average number of extreme events in a 1.0 year Statistical characteristics

863

45

RCP4.5 (2006-2050) 21.7 4.5

RCP8.5 (2006-2050) 22.7 4.4

4.6

4.4

4 56.6

3.9 55.1

1.4

1.3

Figure 1

Figure 1- The spatial and temporal scale of weather phenomena, available GCMs outputs and hydrological impact evaluation (Adapted from Nese and Grenci (2011))

Figure 2

Start

Selecting synoptic scale variables

Configuring SOM structure (number of nodes, dimensions, …)

Training SOM for clustering the synoptic scale variables patterns

NCEP reanalyzing database Local precipitation observed data

Selecting mesoscale variables by applying stepwise regression for rainfall occurrence model (ROM) and rainfall amount model (RAM) in each CT

Identifying the Circulation Types (CTs)

Fitting the parameters of the linear regression model between selected variables (predictors) and transformed rainfall based on minimizing weighted root mean squared errors (RMSE) in each CT

Interpretation of the relationship between local precipitation and CTs (statistical and physical relations)

Fitted statistical models for each CT (ROMs & RAMs)

Trained SOM

Synoptic classification (SOM classifier) SOM(SVi) = CTk

Skewness reduction of rainfall values (applying Cox-Box transfer function)

Using related statistical models (ROMk & RAMk)

Ri = ROMk(MVi) × RAMk(MVi)

End

Synoptic scale variables pattern of day i (vector SVi)

Phase I: Synoptic Scale Classification Model

GCM outputs

Mesoscale variables vectors of day i (MVi)

Phase II: Local Scale Statistical Model

Figure 2- The algorithm of SSDM and its main phases and steps

Figure 3

Figure 3- Location of the study area in Iran (red line box), mesoscale domain around the study area (dashed line box, 4 points) and the positions of NCEP data grids in synoptic scale domain (70 points) (Left side) Tehran urban and upland areas, positions of APHRODITE data grids around Tehran and the considered area at the north part of Tehran (shaded box) (Right side)

Figure 4

Figure 4- The histogram of wet day precipitation amount in the study area

Figure 5

Figure 5- Circulation types (CTs) resulted from applying SOM method

Figure 6

Figure 6- The frequency of daily 500-hPa geopotential heights patterns in identified CTs and probability of rainfall occurrences in each CT

Figure 7

Figure 7- Composite maps of the mean daily vorticity at 500-hPa and mean daily relative humidity at 700-hPa

Figure 8

Figure 8- Q-Q plot of modeled and observed wet day rainfall values for calibration and validation datasets

Figure 9

Figure 9- Time series of modeled and observed annual maximum daily precipitation

Figure 10

Percentiles of Wet Days Rainfall Values

Empirical Cumulative Distribution Functionof Wet Days Rainfall Values

50.00

1

40.00

0.8

30.00

0.6

20.00

0.4

10.00

0.2

0

0.00 0

20

(a)

40

Modeled

Daily Precipitation (mm)

Quantile-

60

Percentiles

80

0

100

10

20

Observed

30

40

50

Daily Precipitation (mm)

(b)

Modeled CFD

Quantile Plot of Wet Days Rainfall Values

Observed CFD

Histogram of Wet Days Rainfall Values

50

0.35

F(x)

0.3

40 0.25 30

0.2 0.15

20 0.1 10

0.05 0

0

(c)

0

10

20

30

Quantile (Modeled , mm)

40

50

(d)

0

5

10

15

20

25

30

35

40

45

50

Daily Precipitation (mm)

Modeled

Observed

Figure 10- Statistical performance of SSDM in daily precipitation downscaling of north part of Tehran urban area from 1970 to 2006. (a): The percentile plots of wet day rainfall values of observed and modeled precipitation. (b): Observed and modeled empirical cumulative distribution function (CDF) of wet day rainfall values. (c): Q-Q plot of Observed and modeled precipitation values of wet day. (d): Histogram of observed and modeled precipitation values of wet day

Figure 11

Figure 11- Statistical performance of SSDM in monthly precipitation downscaling of north part of Tehran urban area from 1970 to 2006. (a) mean monthly rainfall. (b) monthly wet days. (c) monthly variance. (d) standard deviation of wet days

Figure 12

Figure 12- Q-Q plot of vorticity (top curves) and dew point depression (down curves) values in 500-hp level at two points of mesoscale domain from NCEP and CCSM4

Figure 13

Figure 13- Q-Q plot of SSDM modeled wet day rainfall values sing NCEP and CCSM4-historical predictors for the period of 1970-2005

Figure 14

Figure 14- Comparing the statistical characteristics of the downscaled daily precipitation of north part of Tehran by applying SSDM and CCSM4 outputs (in three experiments: historical (1970-2005), rcp45 and rcp85 (2006-2050)) as predictors and historical observed data. (a) and (b): The percentile plots of wet day rainfall values. (c) and (d): The empirical cumulative distribution function (CDF) of wet day rainfall values

Figure 15

Figure 15- Q-Q plot of wet day rainfall values: CCSM4-historical predictors (past climate condition) against CCSM4-rcp45 and CCSM4-rcp85 predictors (near future climate condition) for north part of Tehran