Large-Scale Hydrological Modeling for Calculating Water Stress

Since NSE and RSR are strongly affected by outliers, the trimmed means(22) (by ... The cross-correlation between the two data sets (i.e., water availa...
4 downloads 3 Views 1MB Size
Subscriber access provided by UNIV OF NEBRASKA - LINCOLN

Article

Large-scale hydrological modeling for calculating water stress indices: Implications of improved spatio-temporal resolution, surface-groundwater differentiation and uncertainty characterization Laura Scherer, Aranya Venkatesh, Ramkumar Karuppiah, and Stefan Pfister Environ. Sci. Technol., Just Accepted Manuscript • DOI: 10.1021/acs.est.5b00429 • Publication Date (Web): 30 Mar 2015 Downloaded from http://pubs.acs.org on April 4, 2015

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a free service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are accessible to all readers and citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

Environmental Science & Technology is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 33

Environmental Science & Technology

1

Large-scale hydrological modeling for calculating

2

water stress indices: Implications of improved

3

spatio-temporal resolution, surface-groundwater

4

differentiation and uncertainty characterization

5

Laura Scherer1*, Aranya Venkatesh2, Ramkumar Karuppiah2, Stephan Pfister1

6

1

ETH Zurich, Institute of Environmental Engineering, 8093 Zurich, Switzerland

7

2

ExxonMobil Research and Engineering Company, Annandale, NJ, 08801, United States

8

*Phone: +41-44-632-31-72. E-mail: [email protected].

9 10

ABSTRACT

11

Physical water scarcities can be described by water stress indices. These are often determined

12

at an annual scale and a watershed level; however, such scales mask seasonal fluctuations and

13

spatial heterogeneity within a watershed. In order to account for this level of detail, first and

14

foremost, water availability estimates must be improved and refined. State-of-the-art global

15

hydrological models such as WaterGAP and UNH/GRDC have previously been unable to

16

reliably reflect water availability at the subbasin scale. In this study, the Soil and Water

17

Assessment Tool (SWAT) was tested as an alternative to global models, using the case study

18

of the Mississippi watershed. While SWAT clearly outperformed the global models at the 1 ACS Paragon Plus Environment

Environmental Science & Technology

19

scale of a large watershed, it was judged to be unsuitable for global scale simulations due to

20

the high calibration efforts required. The results obtained in this study show that global

21

assessments miss out on key aspects related to upstream/downstream relations and monthly

22

fluctuations, which are important both for the characterization of water scarcity in the

23

Mississippi watershed and for water footprints. Especially in arid regions, where scarcity is

24

high, these models provide unsatisfying results.

25 26

TOC ART

27 28 29

INTRODUCTION

30

Water scarcity is a global issue affecting human and ecosystem health in many regions across

31

the world.1 Globalized markets mean that the distances between production and consumption

32

are large, making reliable information on the supply chain of products difficult to obtain.2 To

33

account for impacts related to water consumption, water stress indices (WSI) have recently

34

been integrated in water footprinting3 and life cycle assessment (LCA), a methodology to

35

quantify the cradle-to-grave environmental impacts of a product or service.4 Multiple methods

36

have been developed to assess the impacts of freshwater use,5 all of which utilize existing

37

water availability data from hydrological models at global scale. Several of these models were 2 ACS Paragon Plus Environment

Page 2 of 33

Page 3 of 33

Environmental Science & Technology

38

compared by Haddeland et al.6 who identify huge differences among the models, especially in

39

arid regions. This highlights the difficulty of hydrological modeling at a large scale, which, in

40

turn, demands a proper evaluation of model performance and uncertainty. These hydrological

41

models often lack transparency of algorithms and parameterizations after calibration.

42

Furthermore, model performance is only assessed in a simplified manner and uncertainty is

43

not communicated.

44

Several concepts have been proposed to improve the assessment of WSI. Loubet et al.7 split

45

watersheds into smaller units and analyzed WSI per subbasin, while Boulay et al.8

46

differentiated surface and groundwater in the stress assessment. A clear research gap is

47

identified, given that differentiation of water sources becomes more relevant at higher spatial

48

and temporal resolution and that the combined effect of these aspects on WSI is still missing.

49

While Boulay et al.9 tested the effect of using different global models to characterize the

50

uncertainty in WSI, they did not analyze the uncertainty in the underlying models themselves.

51

This study attempts to utilize large-scale hydrological modeling, using the Soil and Water

52

Assessment Tool (SWAT)10 to obtain improved estimates of WSI. This approach is an

53

alternative to using outputs from existing global hydrological models, while simultaneously

54

refining the spatial resolution from watershed to subbasin level. While SWAT was originally

55

developed for watershed simulations, it has already been used for larger scale simulations,

56

such as in Africa,11 Iran12 and the Danube watershed.13 This study also aims to estimate WSI

57

that differentiate between surface and groundwater use and availability, which has often been

58

ignored in previous LCA studies.

59

In order to analyze the model performance of global models in comparison to regional SWAT

60

results and to investigate additional insights afforded by higher spatial resolution in the

61

calculation of WSI, the tool was tested on the Mississippi watershed, which drains more than 3 ACS Paragon Plus Environment

Environmental Science & Technology

62

3,000,000 km2. Model performance and uncertainty were each assessed using a series of

63

methods. The resulting water availability was compared to outputs from the global models

64

WaterGAP6,14 and UNH/GRDC.15 The more detailed model set up in this work, providing

65

monthly results on subbasin level for ground and surface water, was used to estimate

66

alternate, improved WSIs. In addition, the modeling uncertainty was quantified and

67

propagated to these (new) WSI. Last but not least, scenario analysis was used to estimate

68

potential future water stresses following increased irrigation demand.

69 70

MATERIALS AND METHODS

71

Water balance modeling. Water balance simulations of the Mississippi watershed were

72

performed using SWAT version 2009.16 SWAT is a physically-based hydrological model that

73

simulates the water cycle within a watershed, using inputs such as temperature, precipitation,

74

and land use, along with extensive calibration and validation. The entire Mississippi

75

watershed was delineated into 451 subbasins and was run over a period of 30 years from 1971

76

to 2000, where the first five years were used as a warm-up period, i.e. the output was

77

disregarded during post-processing. Further details are presented in the Supporting

78

Information (SI).

79

A few modifications were made to the existing SWAT model, in order to improve the

80

formulation and results. Firstly, to better simulate potential evapotranspiration, the Fortran

81

source code was adapted by modifying the Priestley-Taylor method that can be used to

82

estimate this parameter within SWAT. The original coefficient of 1.28 in the Priestley-Taylor

83

equation, representing both humid and arid climates, was replaced by a conditional statement

84

using different coefficients for humid (1.26) and arid (1.74) climates, where the latter was

85

defined by a condition of relative humidity < 60%.17 4 ACS Paragon Plus Environment

Page 4 of 33

Page 5 of 33

Environmental Science & Technology

86

Secondly, when using ArcSWAT18 as a pre-processing interface to provide input data to

87

SWAT, one weather station of each climate variable is typically assigned to each subbasin. In

88

this study, instead of using the weather station closest to the subbasin’s centroid, all available

89

weather information (time series of grid cells) within one subbasin was averaged. In addition,

90

since the original SWAT model restricts the number of weather stations for solar radiation

91

and relative humidity to 300 while the watershed was discretized into 451 subbasins, the

92

source code was modified to circumvent this restriction.

93

Thirdly, outflow from a reservoir was simulated based on a modified target release scheme.

94

The SWAT model assumes that all reservoirs are used for flood control and distinguishes

95

between flood and non-flood season. Since not all reservoirs conform to this assumption, the

96

scheme was modified to represent multi-purpose reservoirs according to Wu and Chen.19

97

The input data used in setting up the model are described in the SI section S2.

98

Calibration and validation. The model was calibrated against monthly river discharge at 131

99

gauges (Figure S1) within the watershed. Monthly time series at these gauges were split

100

equally into two parts and the more recent data were used for calibration, while the earlier

101

data were used for validation. The parameters selected for calibration are described in S3.

102

River discharge was calibrated manually and for each subbasin individually. The model

103

performance was evaluated based on visual inspection and four efficiency criteria: i) Nash-

104

Sutcliffe efficiency (NSE), ii) percent bias (PBIAS), and iii) root mean square error (RMSE)

105

normalized to standard deviation (RSR) as recommended by Moriasi et al.,20 as well as iv)

106

bR2 as suggested by Krause et al.21 All these criteria compare simulations to observations, and

107

provide goodness-of-fit metrics that emphasize different aspects of the hydrograph, such as

108

systematic or conditional bias. A further discussion on the relevance of these metrics is

5 ACS Paragon Plus Environment

Environmental Science & Technology

109

presented in the SI (section S3). Performance ratings based on these metrics are provided in

110

Table S2.

111

Overall model performance for the entire Mississippi watershed was determined by averaging

112

subbasin performance, measured by the four efficiency criteria, using three different

113

weighting criteria: (i) Equal weights (e.g., the NSE for the entire watershed is the simple

114

average of NSEs for the individual subbasins), (ii) weighted by catchment area and length of

115

time series of all observations for individual subbasins (using the square root of time series

116

length to lower its influence compared to the catchment area), or (iii) weighted by the

117

observed river discharge in individual subbasins. Since NSE and RSR are strongly affected by

118

outliers, the trimmed means22 (by removing the 5% highest and lowest values) or geometric

119

means23 were additionally calculated.

120

For comparison with manual calibration, automatic calibration was carried out using the

121

SWAT Calibration and Uncertainty Programs (SWAT-CUP) version 201224.

122

In addition, the observed evapotranspiration (ET)25 was compared to estimated

123

evapotranspiration. By evaluating a component of the hydrological cycle besides river

124

discharge, a deeper insight into the quality of the water balance is obtained. This is especially

125

interesting with regards to groundwater recharge, which cannot be directly validated by runoff

126

gauges. The overall model performance for ET was either averaged using equal weights or

127

weighted by the subbasin area.

128

Model uncertainty. In one of the algorithms implemented in SWAT-CUP’s automated

129

calibration procedure, SUFI-2, simulated river discharge is characterized by a prediction

130

uncertainty band. Overall model uncertainty is assessed by two metrics (i) the p-factor, which

131

refers to the percentage of observed data that is bracketed by the model predicted uncertainty

132

band, and (ii) the mean difference between the upper and lower limits of that band (r-factor) 6 ACS Paragon Plus Environment

Page 6 of 33

Page 7 of 33

Environmental Science & Technology

133

normalized by the standard deviation of the observed variable.26,27 Ideally, all measurements

134

would be bracketed by the band (i.e., p-factor of 100%) and its width (i.e., r-factor) would

135

approach zero. The prediction uncertainty band estimated by SUFI-2 is a result of the

136

algorithm used in automated calibration, which is not estimated in the case of manual

137

calibration. Therefore, in this study, uncertainty ranges were developed to include 75% or

138

95% of the observed data (i.e., p-factor of 75% or 95%). Measurement uncertainty was

139

considered by assuming a relative error of either 10%28,29 or 25%30 for the observed river

140

discharge data. Two cases for calculating r-factors were considered. Either i) it was inspected

141

if the two bands overlapped at all (simple overlap; SO), or ii) by how much the bands

142

overlapped on average (percentage overlap; PO).

143

Model comparison and spatial association analysis. The SWAT model outputs of river

144

discharge and groundwater recharge were spatially compared to the outputs from the

145

WaterGAP model14 as used in the WATCH project31 and from UNH/GRDC.15 WaterGAP

146

covers the entire simulation period while the UNH/GRDC model is only available for the

147

years 1986 to 1995. Both models have a resolution of 0.5°. The grid cells were disaggregated

148

to a finer resolution keeping the original cell values (i.e. without interpolation), and

149

subsequently aggregated to subbasins as defined in the SWAT model. A third global

150

modeling framework, Aqueduct32 was also evaluated. However, a proper comparison was

151

impeded by the aggregation of the available data to annual averages of a different time period

152

(1950 – 2008) at the level of subbasins larger than those delineated in this study (178

153

compared to 451 for the Mississippi). Only the 44 discharge gauges that were at most 100 km

154

from an outlet of a subbasin were included in the analysis.

155

The cross-correlation between the two datasets (i.e., water availability from two models) was

156

evaluated using the bivariate spatial association measure (BSA) developed by Lee,33 which

157

integrates Pearson’s correlation coefficient and Moran’s index for spatial autocorrelation. 7 ACS Paragon Plus Environment

Environmental Science & Technology

158

Values of BSA closer to ±1 typically indicate a strong positive correlation between two

159

spatial datasets. Furthermore, the level of agreement (LA) between water availabilities from

160

two models was also determined. Since exact matches of water availabilities between models

161

would be unrealistic LA refers to the percentage of observations that lie within the

162

intersection of two intervals, each representing 25% relative error in water availability from

163

the two models. If the bands overlap completely, 100% LA is expected.

164

The goodness of the SWAT model (evaluated by the efficiency criteria) was spatially

165

correlated to various parameters that might explain the weaknesses of the model: the aridity

166

index (AI) as the ratio of mean annual precipitation to mean annual potential

167

evapotranspiration34, the average slope of the main river channel within each subbasin as

168

retrieved from the SWAT model, the average number of days with snowfall per year (SD), i.e.

169

precipitation days with below zero temperatures, and the fragmentation of the river network

170

caused by upstream reservoirs (RF). For the RF metric, all 52 reservoirs within the

171

Mississippi watershed available in the Global Lakes and Wetlands Database (GLWD)35 were

172

considered.

173

Water stress indices. Water stress indices (WSI) were calculated according to Pfister et al.36

174

and Pfister and Bayer37 for 24 different cases (see also SI section S6) taking into account (i)

175

total water, surface water or groundwater resources, (ii) annual or monthly resolution, (iii)

176

water withdrawal or water consumption and (iv) upstream level or subbasin level.

177

At the upstream level, the upstream water use was added to the local water use as both lessen

178

the local water availability.7 In contrast, at the subbasin level only local water use was

179

considered, and only the surface runoff generated in the subbasin was taken into account

180

instead of total river discharge.

8 ACS Paragon Plus Environment

Page 8 of 33

Page 9 of 33

Environmental Science & Technology

181

Surface and groundwater availability uncertainties were estimated and propagated to the

182

resulting WSI: for surface water, absolute uncertainty was set to the value for which 75% (p-

183

factor) of the measurement band (assuming 10% measurement uncertainty) overlapped with

184

the simulation band (PO). For groundwater recharge, no direct measurements were available

185

and the uncertainty was assessed by averaging the relative difference between model results

186

and measurement for two proxies for groundwater recharge that can be retrieved from the

187

river hydrograph: the 90% flow duration value (Q90) which is exceeded 90% of the time, and

188

the ratio Q90/Q50.38 Uncertainties of ungauged subbasins were set equal to the uncertainties

189

of the next downstream gauge. Finally, the uncertainties of water availability were propagated

190

to the WSI according to the Gaussian law of error propagation (S6).39

191

Scenario analysis. While WSI are calculated based on past data, potential future changes in

192

climate conditions and water demand can impact WSI.40 The impacts of potential changes in

193

precipitation, temperature and irrigation demand in 2040 to 2069 on WSI were examined as

194

part of a scenario analysis. Data on climate change projections were retrieved from an

195

ensemble average of 23 general circulation models at a 0.5° resolution41 and downscaled to

196

0.125° for the US,42 covering most of the Mississippi watershed. The IPCC emission scenario

197

A1B42 was considered and data were provided as monthly averages. Changes in irrigation

198

demand were obtained from Pfister et al.40 as monthly averages for the 2050s for four

199

scenarios.

200 201

RESULTS

202

Calibration and validation. The calibrated SWAT model performed well throughout the

203

watershed according to PBIAS, suggesting low systematic bias. On the other hand, NSE, RSR

204

and bR2 exhibited a clear East-West spatial pattern (Figure 1). Overall Mississippi watershed 9 ACS Paragon Plus Environment

Environmental Science & Technology

Page 10 of 33

205

model performance strongly depended on the weighting procedure. PBIAS and bR2 yielded

206

satisfactory results for calibrated river discharge in all cases, whereas NSE and RSR only

207

produced satisfactory results when weighted by the mean observed discharge (Table 1). NSE

208

and RSR were greatly influenced by outliers, so trimming 5% at each side or calculating

209

geometric means improved the results.

210

Model performance was lower for the validated time series, as expected. Satisfactory results

211

could only be produced with regard to PBIAS and bR2 (Table S3). The performance after

212

autocalibration was significantly lower than after manual calibration (Table S4 - Table S5)

213

and model results from autocalibration were therefore disregarded from further analyses.

214 215

Figure 1. Model performance of calibrated river discharge using four different criteria (green

216

indicates good performance and red indicates poor performance). 10 ACS Paragon Plus Environment

Page 11 of 33

Environmental Science & Technology

217 218

Table 1. Mississippi model performance of calibrated river discharge for different criteria and

219

weighting of subbasins (satisfactory values in bold).

NSE NSE (trimmed) PBIAS RSR RSR (geometric) bR2

No weight -0.91 0.14 8.43 0.99 0.84 0.51

Area / Length 0.15 0.30 4.29 0.79 0.71 0.64

Discharge 0.59 0.59 3.77 0.59 0.57 0.73

220 221

Actual evapotranspiration was not calibrated. Nevertheless, the model yielded satisfactory

222

results using all efficiency criteria, with the exception of the untrimmed NSE where outliers

223

had a strong influence (Table S6).

224

Model uncertainty. Like model performance, the overall uncertainty depended on the

225

weighting procedure. r-factors could only yield satisfactory values (< 1)26,27 when 75% of the

226

data were bracketed by the uncertainty band, instead of 95%. Lower r-factors were obtained

227

when measurement uncertainty was considered as a binary decision based on whether there

228

was an intersection or not of the uncertainty bands of simulated and measured river discharges

229

(SO). The higher the measurement uncertainty, the more likely the two bands intersect which

230

leads to an apparently lower overall uncertainty. On the other hand, r-factors increased with a

231

higher measurement uncertainty when taking into account the percent overlap (PO) of the

232

bands (Table 2, Figure S2).

233 234

11 ACS Paragon Plus Environment

Environmental Science & Technology

Page 12 of 33

235

Table 2. Uncertainty of surface water availability combining different procedures and for

236

different weighting of subbasins (satisfactory values in bold).

p-factor (%) 75 75 75 75 75 95 95 95

Measurement uncertainty (%) 0 10 10 25 25 0 10 10

r-factor Strategy SO SO PO SO PO SO SO PO

No weight

Area / Length

Discharge

0.93 0.77 0.93 0.56 0.97 2.01 1.81 2.03

1.17 0.94 1.18 0.60 1.24 2.00 1.75 2.03

1.10 0.84 1.12 0.48 1.20 1.85 1.57 1.87

237 238

Model comparison and spatial association analysis. The SWAT model performed

239

considerably better than the global models WaterGAP and UNH/GRDC (Table S7). The

240

global models did not produce satisfactory values for any of the efficiency criteria.

241

Nevertheless, WaterGAP yielded excellent results for the furthest downstream gauge close to

242

the watershed outlet and outperformed the SWAT model in that case. The reason for this

243

improved performance lies in the fact that WaterGAP matches the runoff results to

244

measurement stations.14 Likewise, Aqueduct yielded a low PBIAS at the gauge closest to the

245

outlet, but an unsatisfactory result when considering internal gauges (S5). The same test could

246

not be carried out for the UNH/GRDC model because observation and simulation periods did

247

not overlap for the gauge closest to the outlet.

248

Spatial patterns agreed relatively well between the models with correlations, more precisely

249

bivariate spatial associations (BSA), larger than 0.4 (Table S8). The high BSA for

250

groundwater availability was due to the more homogeneous spatial distribution of recharge

251

which is favored by the measure.33 The level of agreement (LA) between models was

252

observed to be low, indicated by values of 40% and lower (Table S8). Groundwater

12 ACS Paragon Plus Environment

Page 13 of 33

Environmental Science & Technology

253

availabilities agreed better in arid regions whereas no trend was notable for surface water

254

availabilities.

255

The poor performance of the SWAT model was mainly observed in regions of high aridity

256

(with low aridity indices; Table S9). The aridity index and slope of the main channel in

257

subbasins both depicted an East-West gradient (Figure S3) similar to the model efficiency

258

criteria in Figure 1. However, the relationship between model efficiency and slopes was less

259

evident when determining the partial BSA, in a manner analogous to the partial correlation

260

coefficient,43 and thereby removing the influence of aridity. The partial BSA between model

261

efficiency criterion bR2 and slope, for instance, only amounted to -0.11. In contrast, the partial

262

BSA between bR2 and AI when controlling for slopes still amounted to 0.66. Snowfall (or

263

SD) and dams (or RF) hardly influenced the model performance (Table S9).

264

Water stress indices. The mean WSI for the Mississippi watershed ranged from 0.3 to 0.8

265

(Table 3), depending on the source of water, the temporal and spatial scales as well as the

266

type of water use (consumption or withdrawal). The ranges for individual gauges were even

267

broader (Table S11). Water stress was higher in the arid West than in the humid East (Figure

268

2, Figure S4 - Figure S9). The annual averages of monthly WSI mostly exceeded the overall

269

annual average. Likewise, WSI at the upstream level generally surpassed WSI at the subbasin

270

level. No clear tendency of WSI could be observed between surface and groundwater or

271

withdrawal and consumption. Annual WSI related to water consumption at the subbasin level

272

were quite high.

273

Uncertainties (relative errors) of annual WSI were mostly lower than the monthly WSI.

274

Likewise, relative uncertainties in WSI for total water resources were lower than for surface

275

and groundwater resources alone, since it was considered unlikely that availabilities of both

13 ACS Paragon Plus Environment

Environmental Science & Technology

Page 14 of 33

276

were simultaneously over- or underestimated to the full extent. This follows from the law of

277

error propagation which is explained in more detail in the SI (section S6).

278

Table 3. Consumption weighted water stress indices and their uncertainties.

Withdrawal Consumption

Upstream Subbasin Upstream Subbasin

Annual WSI (Uncertainty in %) Total Surface Groundwater water water 0.48 (21) 0.49 (31) 0.47 (15) 0.40 (20) 0.45 (28) 0.29 (20) 0.51 (13) 0.43 (15) 0.56 (21) 0.80 (10) 0.76 (16) 0.72 (12)

Average monthly WSI (Uncertainty in %) Total Surface Groundwater water water 0.70 (15) 0.72 (19) 0.62 (26) 0.62 (19) 0.69 (22) 0.45 (31) 0.71 (15) 0.61 (20) 0.71 (14) 0.64 (18) 0.57 (25) 0.60 (17)

279

280 281

Figure 2. Distributed average water stress indices of the Mississippi watershed considering

282

water consumption and upstream level (a: annual total water resources, b: monthly total water

283

resources, c: monthly surface water resources, d: monthly groundwater resources). 14 ACS Paragon Plus Environment

Page 15 of 33

Environmental Science & Technology

284 285

While the severity of water stress differed most when comparing annual to monthly WSI,

286

spatial patterns differed most when comparing surface water and groundwater WSI (Table

287

S12).

288

Scenario analysis. Between 2040 and 2069, precipitation in the Mississippi watershed was

289

projected to increase by ~2%. On the other hand, temperature was projected to rise by more

290

than 1°C, leading to enhanced evapotranspiration. As a result, total water availability was

291

simulated to increase by about 5% whereas groundwater availability would slightly decrease.

292

Water consumption was assumed to increase due to an increase in agricultural activities as

293

assessed by four scenarios. This increase ranged from being negligible in the expansion

294

scenarios (< 1%) to high in case of intensification of current agricultural areas (> 10%).

295

Altogether, the scenarios led to a slightly lower water stress (Table S13) compared to current

296

averages. This trend was more pronounced at the monthly time scale. Only groundwater stress

297

may increase when considered at an annual time scale. Differences between the scenarios

298

were insignificant.

299 300

DISCUSSION

301

Calibration. Sound calibration and validation combines multiple criteria since different

302

criteria focus on different aspects of a hydrograph and no single one performs ideally.20,21,44,45

303

In this study, model performance was judged based on NSE, PBIAS and RSR as

304

recommended by Moriasi et al.20 and bR2 as suggested by Krause et al.21 (see also S3). The

305

SWAT model of the Mississippi watershed mostly performed well in the Eastern humid

306

region but poorly in the Western arid region, which is explained in part by the spatial 15 ACS Paragon Plus Environment

Environmental Science & Technology

Page 16 of 33

307

association between aridity index and model performance (Table S9). The overall model

308

efficiency depended on the weighting procedure of the subbasins. It could be argued that it is

309

more relevant to simulate high discharges well as they contribute more to the total water

310

availability. When weighting by discharge, the SWAT model performed satisfactorily (NSE)

311

to very well (PBIAS, Table 1). However, weighting by discharge is subjective. Area and time

312

series length would be more objective weighting parameters from a data perspective, but it

313

might be less relevant from a water resource perspective. In that case, the model performed

314

very well regarding PBIAS, satisfactorily regarding bR2, but poorly regarding RSR and the

315

most comprehensive measure, NSE. This highlights that model evaluation using a single

316

criterion can be misleading and supports the recommendation of combining multiple criteria,

317

as mentioned above. Considering the large watershed and high climatic diversity, the model

318

performance is considered satisfactory.

319

PBIAS was the criterion which indicated best model performance. However, several

320

parameters influence “losses” in the system and enable the modeler to match the water

321

balance.46 Such losses are recharge to deep aquifers, which do not contribute baseflow to

322

rivers, capillary rise, which is better described as evapotranspiration from shallow aquifers,

323

and soil evaporation. Although care was taken that the parameters maintained physical

324

meaning as much as possible (Table S1), controlling all these parameters in conjunction still

325

resulted in highly influencing the water balance and therefore might result in overfitting the

326

model. This is also reflected in the lower performance during the validation period.

327

Calibration of the SWAT model given the large number of unknown parameters is a

328

mathematically ill-posed problem, as available observations do not constrain the model

329

sufficiently to find a unique solution. Instead, different parameter sets can fit the observed

330

discharge records in a similarly satisfactory way, and this creates an equifinality problem.47

331

This stresses that a good discharge representation does not necessarily mean that the entire 16 ACS Paragon Plus Environment

Page 17 of 33

Environmental Science & Technology

332

water balance including groundwater recharge is represented well. It is therefore important to

333

also validate against evapotranspiration as an alternate water balance component. Equifinality

334

is higher, the more parameterized a model is. Physical-based, distributed models such as

335

SWAT are particularly prone to be overparameterized. In order to get results closer to

336

physical reality than mathematical convenience, Kirchner48 and Beven47 advocate reducing

337

model complexity and minimizing the number of calibration parameters. This is especially

338

relevant for scenario analyses where conditions significantly deviate from the previously

339

experienced and calibrated conditions.

340

While manual calibration is time-consuming and tedious, it allows combining multiple

341

numerical performance criteria with visual inspection. On the other hand, it also involves

342

subjectivity.49 In contrast, automation tremendously speeds up calibration and makes it more

343

objective. The autocalibration procedure tested in this study was based on a single criterion

344

objective function. As discussed above, a single criterion is inadequate for evaluation or

345

optimization as there can be substantial trade-offs among performance criteria.49 Yet the

346

model even performed worse in terms of bR2 which was selected as objective function. This

347

might be explained by the fact that the Mississippi watershed is huge and encompasses high

348

spatial variability which could not be captured by simply autocalibrating based on

349

parameterizing soil texture and land use classes similarly for the entire region. Although the

350

manual calibration clearly outperformed autocalibration, it is infeasible on a global scale with

351

limited efforts. Hence the automated procedure has to be improved. Automated multicriteria

352

optimization could be more promising. It would provide a set of Pareto optimal solutions

353

from which a specific solution could be singled out based on further analyses and priorities

354

according to model application.22,49

355

Existing global hydrological models were mostly calibrated against gauges at watershed

356

outlets, with a lesser emphasis on internal spatial heterogeneity. Therefore, WaterGAP 17 ACS Paragon Plus Environment

Environmental Science & Technology

Page 18 of 33

357

performed very well at the gauge closest to the outlet, but performed poorly when considering

358

internal gauges (Table S8). The same trend was observed for Aqueduct. Therefore, state-of-

359

the-art global models should be used for applications at subbasin scale with caution. In

360

contrast to global models, internal gauges were used to calibrate the SWAT model (131 in

361

total). As a result, the SWAT model outperformed the global models at the subbasin level

362

(Table S8) and is expected be a more suitable basis for estimating more robust, spatially-

363

explicit WSI.

364

An additional problem that exists for all hydrological models is man-managed reservoirs,

365

which confound the calibration processes. Reservoirs are operated in many different ways and

366

it is challenging to gather information on individual operation rules for large scales such as

367

the Mississippi watershed, continents or the entire world.50,51 In our study, a generic operation

368

equation was used, but parameters were fitted separately for each reservoir.

369

Uncertainty. Hydrological models inherit uncertainty from multiple sources: errors in forcing

370

data and boundary conditions, sub-optimal parameterizations, flaws in model structure and

371

errors in recorded model output.28 The difficulty lies in the fact that these uncertainties

372

interact and cannot easily be disaggregated from residual time series.52 There is currently no

373

consensus on a particular uncertainty estimation method. The method chosen in this study was

374

a modified version of the algorithm in SUFI-2.26,27 Estimates of measurement uncertainty

375

were based on literature, and vary greatly between 3% and 42% depending on measurement

376

strategy and river channel conditions.29 Moreover, it should be kept in mind that some

377

strategies of including measurement uncertainty may lead to a poor model appearing

378

satisfactory,53 as can be the case when uncertainty is considered as a binary decision on

379

intersection of the simulation and observation uncertainty bands. Such strategies should

380

therefore be avoided. Although the SWAT model generally exhibited uncertainties exceeding

381

a satisfactory level (Table 2), they might still be within acceptable limits when propagated to 18 ACS Paragon Plus Environment

Page 19 of 33

Environmental Science & Technology

382

WSI (Table 3). In terms of LCA or water footprinting, it is essential to provide results with

383

uncertainty. Complete uncertainty analyses were beyond the scope of this study. However, it

384

should be noted that uncertainties in water use and function choice for the WSI would

385

increase uncertainty.54

386

The large differences among global hydrological models as found by Haddeland et al.6 were

387

confirmed by this study. Although the spatial patterns agreed relatively well between the

388

models with BSA larger than 0.4, the LA was unsatisfactory (Table S8). Thus, the choice of

389

water availability model entails a major source of uncertainty for assessing water scarcity.

390

One possibility to deal with that would be to consider multiple models as suggested by

391

Haddeland et al.6

392

Water stress. WSI varied greatly depending on the type of water resource, the temporal

393

resolution, the spatial level and the water use type considered (Table 3). The severity of water

394

stress differed most when comparing annual to monthly WSI (Table S12); hence, annual WSI

395

as typically applied can be misleading. However, temporal resolution should be seen in

396

conjunction with travel time. While water flows for about four months from the source of the

397

Missouri to the outlet of the Mississippi watershed, travel time is much lower in individual

398

subbasins. Therefore, an increase in spatial resolution should be accompanied by an increase

399

in temporal resolution and vice-versa. Spatial patterns varied most when comparing surface

400

water and groundwater WSI (Table S12, Figure 2). Considering also their disparities in

401

anthropogenic and ecological significance, it is essential to assess water stress separately.

402

Water stress was already differentiated in a study by Boulay et al.,8 which, however, lacks the

403

spatial and temporal details of this work.

404

Loubet et al.7 were the first to calculate WSI at the subbasin level and they considered

405

upstream water availability and use. In case of transboundary water resources, water 19 ACS Paragon Plus Environment

Environmental Science & Technology

Page 20 of 33

406

availability and use within a subbasin might be more appropriate to ensure independence from

407

upstream users. The same coefficients were used to calculate WSI at both spatial levels which

408

might be improper and needs further analysis. Currently, most WSI used in LCA incorporate

409

water withdrawal instead of water consumption (e.g., ref 37). This can be justified by the fact

410

that water released back to the environment might be of degraded quality and that the point of

411

release might be located in a different spatial unit than the point of extraction. Additionally, it

412

is argued that water consumption estimates are generally derived from withdrawals and might

413

be less accurate. On the other hand, if the spatial units are not too small, it can be assumed

414

that the points of extraction and release are situated in the same spatial unit and the water

415

released is unlikely to be of such a low quality that it is unusable for any user. The water

416

consumption seems more relevant when addressing water scarcity issues.7 All the above

417

highlights that further development of WSI is required and that a consensus needs to be

418

reached.

419

The indices calculated by Pfister et al.36 mostly underestimated water stress compared to those

420

in this study (Figure S12), which could be explained by the overestimation of water

421

availability in WaterGAP (Table S7). Aqueduct, on the other hand, comprises simplistic water

422

risk indicators such as WTA based on annual averages. The underlying model GLDAS is a

423

land surface model which focuses on vertical fluxes rather than hydrological processes and

424

whose evaluation against river discharge is limited (only 66 gauges at the global scale and R2

425

as a performance measure).55 It is likely less suitable to analyse WSI at a higher resolution

426

(spatially and temporally) than models such as those implemented in this study.

427

Future scenarios. Current analysis for the Mississippi watershed as a whole shows that the

428

dominant East-West trend requires high water consumption reductions in the West in order to

429

avoid high stress levels. Based on the analyzed future scenarios, this trend is expected to

430

decrease slightly. 20 ACS Paragon Plus Environment

Page 21 of 33

Environmental Science & Technology

431

The calibrated SWAT model was also used for analyzing irrigation demand scenarios.

432

Changes in evapotranspiration were considered in the hydrological modeling, but not in the

433

scenario generation of irrigation demand.40 Increased evapotranspiration partly offsets the

434

effect of increased precipitation on water availability and thereby amplifies irrigation demand

435

and consequently aggravates water stress. Also, the model was evaluated by a split sample

436

test where the observation records were divided into two equivalent time periods. Since both

437

time periods used for calibration and validation usually represent similar conditions, such a

438

test is not very insightful for extrapolations into the future.48 Furthermore, model

439

overparameterization can affect projections of hydrological impacts in scenario analyses.51

440

Nevertheless, the scenarios foresee relatively low increase in water demand and climate

441

effects appear less severe compared to current anthropogenic stress levels, but this in itself

442

varies depending on the region of the world.56

443

Practical implications. Global water stress has typically been analyzed at the country or

444

watershed level. These scales mask the spatial variability, most notably in large countries or

445

watersheds.57 Global hydrological models such as WaterGAP, UNH/GRDC and GLDAS

446

(underlying Aqueduct) are currently challenged to reliably reflect water availability at the

447

subbasin scale. Such models could improve their transparency by moving beyond model

448

evaluation based on single criteria, and more robust uncertainty analyses. Although the

449

SWAT model of the Mississippi watershed clearly outperformed the global models, it is

450

unlikely to substitute them on a global scale. First, manual calibration as carried out in this

451

study would be infeasible for global coverage in the short-term and could at most be applied

452

to hotspots of severe water scarcity. Second, the SWAT model with its complexity is prone to

453

be overparameterized. This increases equifinality and reduces the reliability, especially of

454

scenario analyses. Therefore, the modeling algorithms of the global models need to be

455

improved and more observational data such as internal gauges have to be incorporated in the 21 ACS Paragon Plus Environment

Environmental Science & Technology

Page 22 of 33

456

calibration process. Until such improvements take place it is recommended that (i) global

457

models be evaluated based on multiple performance criteria, (ii) the uncertainty of these

458

models be assessed, (iii) the weaknesses of the models be related to other spatial factors such

459

as aridity index in order to set priorities as to where to improve the models, and (iv) an

460

ensemble average be derived from multiple hydrological models. While improved

461

representations of water availability are a prerequisite for spatially and temporally higher

462

resolved WSI, LCA practitioners should focus on developing indices and impact assessment

463

methods that distinguish surface and groundwater. In the meantime, any global water scarcity

464

indicator should be considered carefully, taking into account the high uncertainties and spatial

465

differences presented in this paper. This is especially relevant for global analyses of supply

466

chains as done by water footprint or LCA studies.

467 468

ASSOCIATED CONTENT

469

Supporting Information

470

Input data, further method descriptions, full results, and maps of water stress indices as well

471

as SHP files of water stress indices are available free of charge via the Internet at

472

http://pubs.acs.org.

473 474

AUTHOR INFORMATION

475

Corresponding Author

476

*Phone: +41-44-632-31-72. E-mail: [email protected].

477 22 ACS Paragon Plus Environment

Page 23 of 33

Environmental Science & Technology

478

ACKNOWLEDGEMENTS

479

The authors thank Karim Abbaspour for his support in using ArcSWAT and SWAT-CUP,

480

Stefanie Hellweg and Adam Usadi for their helpful comments and Catherine Raptis for proof-

481

reading the manuscript. This work was funded by ExxonMobil Research and Engineering,

482

Corporate Strategic Research.

483 484 485 486

REFERENCES

487

(1) UNEP. Measuring water use in a green economy. A report of the working group on

488

water efficiency to the International Resource Panel; United Nations Environment

489

Programme, 2012.

490

(2) Hubacek, K.; Feng, K.; Minx, J. C.; Pfister, S.; Zhou, N. Teleconnecting Consumption

491

to Environmental Impacts at Multiple Spatial Scales. J. Ind. Ecol. 2014, 18 (1), 7–9; DOI

492

10.1111/jiec.12082.

493 494

(3) ISO 14046:2014. Water footprint requirements and guidelines working group TC 207/SC5/WG8. http://www.iso.org/.

495

(4) Hellweg, S.; Milà i Canals, Llorenç. Emerging approaches, challenges and

496

opportunities in life cycle assessment. Science 2014, 344 (6188), 1109–1113; DOI

497

10.1126/science.1248361.

498

(5) Kounina, A.; Margni, M.; Bayart, J.-B.; Boulay, A.-M.; Berger, M.; Bulle, C.;

499

Frischknecht, R.; Koehler, A.; Milà i Canals, Llorenç; Motoshita, M.; Núñez, M.; Peters, G.;

500

Pfister, S.; Ridoutt, B.; Zelm, R.; Verones, F.; Humbert, S. Review of methods addressing

23 ACS Paragon Plus Environment

Environmental Science & Technology

Page 24 of 33

501

freshwater use in life cycle inventory and impact assessment. Int. J. Life Cycle Ass. 2013, 18

502

(3), 707-721; DOI 10.1007/s11367-012-0519-3.

503

(6) Haddeland, I.; Clark, D. B.; Franssen, W.; Ludwig, F.; Voß, F.; Arnell, N. W.;

504

Bertrand, N.; Best, M.; Folwell, S.; Gerten, D.; Gomes, S.; Gosling, S. N.; Hagemann, S.;

505

Hanasaki, N.; Harding, R.; Heinke, J.; Kabat, P.; Koirala, S.; Oki, T.; Polcher, J.; Stacke, T.;

506

Viterbo, P.; Weedon, G. P.; Yeh, P. Multimodel Estimate of the Global Terrestrial Water

507

Balance: Setup and First Results. J. Hydrometeorol. 2011, 12 (5), 869–884; DOI

508

10.1175/2011JHM1324.1.

509

(7) Loubet, P.; Roux, P.; Núñez, M.; Belaud, G.; Bellon-Maurel, V. Assessing Water

510

Deprivation at the Sub-river Basin Scale in LCA Integrating Downstream Cascade Effects.

511

Environ. Sci. Technol. 2013, 47 (24), 14242–14249; DOI 10.1021/es403056x.

512

(8) Boulay, A.-M.; Bulle, C.; Bayart, J.-B.; Deschênes, L.; Margni, M. Regional

513

Characterization of Freshwater Use in LCA: Modeling Direct Impacts on Human Health.

514

Environ. Sci. Technol. 2011, 45 (20), 8948–8957; DOI 10.1021/es1030883.

515

(9) Boulay, A.-M.; Motoshita, M.; Pfister, S.; Bulle, C.; Muñoz, I.; Franceschini, H.;

516

Margni, M. Analysis of water use impact assessment methods (part A): evaluation of

517

modeling choices based on a quantitative comparison of scarcity and human health

518

indicators. Int. J. Life Cycle Ass. 2015, 20 (1), 139-160; DOI 10.1007/s11367-014-0814-2.

519

(10) Arnold, J. G.; Srinivasan, R.; Muttiah, R. S.; Williams, J. R. Large area hydrologic

520

modeling and assessment part I: Model development. J. Am. Water Resour. As. 1998, 34 (1),

521

73–89; DOI 10.1111/j.1752-1688.1998.tb05961.x.

522

(11) Schuol, J.; Abbaspour, K. C.; Yang, H.; Srinivasan, R.; Zehnder, Alexander J. B.

523

Modeling blue and green water availability in Africa. Water Resour. Res. 2008, 44 (7),

524

W07406; DOI 10.1029/2007WR006609.

24 ACS Paragon Plus Environment

Page 25 of 33

Environmental Science & Technology

525

(12) Faramarzi, M.; Abbaspour, K. C.; Schulin, R.; Yang, H. Modelling blue and green

526

water resources availability in Iran. Hydrol. Process. 2009, 23 (3), 486–501; DOI

527

10.1002/hyp.7160.

528

(13) Pagliero, L.; Bouraoui, F.; Willems, P.; Diels, J. Large-Scale Hydrological

529

Simulations Using the Soil Water Assessment Tool, Protocol Development, and Application

530

in the Danube Basin. J. Environ. Qual. 2014, 43 (1), 145–154; DOI 10.2134/jeq2011.0359.

531

(14) Döll, P.; Kaspar, F.; Lehner, B. A global hydrological model for deriving water

532

availability indicators: model tuning and validation. J. Hydrol. 2003, 270 (1–2), 105–134;

533

DOI 10.1016/S0022-1694(02)00283-4.

534

(15) Fekete, B. M.; Vörösmarty, C. J.; Grabs, W. High-resolution fields of global runoff

535

combining observed river discharge and simulated water balances. Global Biogeochem. Cy.

536

2002, 16 (3), 15-1; DOI 10.1029/1999GB001254.

537

(16) Neitsch, S. L.; Arnold, J. G.; Kiniry, J. R.; Williams, J. R. Soil and Water Assessment

538

Tool. Theoretical documentation version 2009. Texas Water Resources Institute Technical

539

Report No. 406; Texas Water Resources Institute, 2011.

540 541

(17) Shuttleworth, W. J. Evaporation. In Handbook of hydrology; Maidment, D. R., Ed.; McGraw-Hill: New York, 1993; pp 4.1–4.53.

542

(18) Olivera, F.; Valenzuela, M.; Srinivasan, R.; Choi, J.; Cho, H.; Koka, S.; Agrawal, A.

543

ArcGIS-SWAT: A geodata model and GIS interface for SWAT. J. Am. Water Resour. As.

544

2006, 42 (2), 295–309; DOI 10.1111/j.1752-1688.2006.tb03839.x.

545 546

(19) Wu, Y.; Chen, J. An Operation-Based Scheme for a Multiyear and Multipurpose Reservoir to Enhance Macroscale Hydrologic Models. J. Hydrometeorol. 2012, 13 (1).

547

(20) Moriasi, D. N.; Arnold, J. G.; van Liew, M. W.; Bingner, R. L.; Harmel, R. D.; Veith,

548

T. L. Model evaluation guidelines for systematic quantification of accuracy in watershed

549

simulations. T. ASABE 2007, 50 (3), 885–900.

25 ACS Paragon Plus Environment

Environmental Science & Technology

Page 26 of 33

550

(21) Krause, P.; Boyle, D. P.; Bäse, F. Comparison of different efficiency criteria for

551

hydrological model assessment. Adv. Geosci. 2005, 5, 89–97; DOI 10.5194/adgeo-5-89-

552

2005.

553 554 555

(22) Ruppert, D. Trimming and Winsorization. Encyclopedia of Statistical Sciences; John Wiley & Sons, Inc, 2004. (23) Limpert, E.; Stahel, W. A.; Abbt, M. Log-normal Distributions across the Sciences:

556

Keys

557

3568(2001)051[0341:LNDATS]2.0.CO;2.

558 559

and

Clues.

BioScience

2001,

51

(5),

341–352;

DOI

10.1641/0006-

(24) Abbaspour, K. C. SWAT-CUP 2012: SWAT Calibration and Uncertainty Programs. A user manual, 2013.

560

(25) Zhang, K.; Kimball, J. S.; Nemani, R. R.; Running, S. W. A continuous satellite-

561

derived global record of land surface evapotranspiration from 1983 to 2006. Water Resour.

562

Res. 2010, 46 (9), W09522; DOI 10.1029/2009WR008800.

563

(26) Abbaspour, K. C.; Johnson, C. A.; van Genuchten, M. Th. Estimating Uncertain Flow

564

and Transport Parameters Using a Sequential Uncertainty Fitting Procedure. Vadose Zone J.

565

2004, 3 (4), 1340–1352; DOI 10.2113/3.4.1340.

566

(27) Abbaspour, K. C.; Yang, J.; Maximov, I.; Siber, R.; Bogner, K.; Mieleitner, J.;

567

Zobrist, J.; Srinivasan, R. Modelling hydrology and water quality in the pre-alpine/alpine

568

Thur

569

10.1016/j.jhydrol.2006.09.014.

watershed

using

SWAT.

J.

Hydrol.

2007,

333

(2–4),

413–430;

DOI

570

(28) Butts, M. B.; Payne, J. T.; Kristensen, M.; Madsen, H. An evaluation of the impact of

571

model structure on hydrological modelling uncertainty for streamflow simulation. J. Hydrol.

572

2004, 298 (1–4), 242–266; DOI 10.1016/j.jhydrol.2004.03.042.

573

(29) Harmel, R. D.; Cooper, R. J.; Slade, R. M.; Haney, R. L.; Arnold, J. G. Cumulative

574

uncertainty in measured streamflow and water quality data for small watersheds. T. ASABE

575

2006, 49 (3), 689. 26 ACS Paragon Plus Environment

Page 27 of 33

576 577

Environmental Science & Technology

(30) Di Baldassarre, G.; Montanari, A. Uncertainty in river discharge observations: a quantitative analysis. Hydrol. Earth Syst. Sc. 2009, 13 (6).

578

(31) Centre for Ecology and Hydrology, Wageningen UR. WATCH 20th century model

579

output datasets. ftp://ftp.iiasa.ac.at/WorkBlock6/WaterGap/WFD_1963_2001/ (accessed

580

April 11, 2014).

581 582 583

(32) Reig, P.; Shiao, T.; Gassert, F. Aqueduct Water Risk Framework. Working Paper; World Resources Institute: Washington, DC, 2013. (33) Lee, S.-I. Developing a bivariate spatial association measure: An integration of

584

Pearson's

r

and

Moran's

585

10.1007/s101090100064.

I.

J.

Geogr.

Syst.

2001,

3

(4),

369-385;

DOI

586

(34) UNEP. World atlas of desertification; Arnold: London, 1992.

587

(35) Lehner, B.; Döll, P. Development and validation of a global database of lakes,

588

reservoirs

and

wetlands.

589

10.1016/j.jhydrol.2004.03.028.

J.

Hydrol.

2004,

296

(1–4),

1–22;

DOI

590

(36) Pfister, S.; Koehler, A.; Hellweg, S. Assessing the Environmental Impacts of

591

Freshwater Consumption in LCA. Environ. Sci. Technol. 2009, 43 (11), 4098–4104; DOI

592

10.1021/es802423e.

593

(37) Pfister, S.; Bayer, P. Monthly water stress: spatially and temporally explicit

594

consumptive water footprint of global crop production. J. Clean. Prod. 2013 (0); DOI

595

10.1016/j.jclepro.2013.11.031.

596

(38) Nathan, R. J.; McMahon, T. A. Evaluation of automated techniques for base flow and

597

recession

598

10.1029/WR026i007p01465.

599 600

analyses.

Water

Resour.

Res.

1990,

26

(7),

1465–1473;

DOI

(39) Drosg, M. Basics on uncertainties. Dealing with Uncertainties; Springer, 2009; pp 17– 48.

27 ACS Paragon Plus Environment

Environmental Science & Technology

Page 28 of 33

601

(40) Pfister, S.; Bayer, P.; Koehler, A.; Hellweg, S. Projected water consumption in future

602

global agriculture: Scenarios and related impacts. Sci. Total Environ. 2011, 409 (20), 4206–

603

4216; DOI 10.1016/j.scitotenv.2011.07.019.

604

(41) Meehl, G. A.; Covey, C.; Taylor, K. E.; Delworth, T.; Stouffer, R. J.; Latif, M.;

605

McAvaney, B.; Mitchell, John F. B. THE WCRP CMIP3 Multimodel Dataset: A New Era in

606

Climate Change Research. B. Am. Meteorol. Soc. 2007, 88 (9), 1383–1394; DOI

607

10.1175/BAMS-88-9-1383.

608

(42) Maurer, E. P.; Brekke, L.; Pruitt, T.; Duffy, P. B. Fine-resolution climate projections

609

enhance regional climate change impact studies. Eos T. Am. Geophys. Un. 2007, 88 (47),

610

504; DOI 10.1029/2007EO470006.

611 612

(43) Kleinbaum, D.; Kupper, L.; Nizam, A.; Rosenberg, E. Applied regression analysis and other multivariable methods; Cengage Learning: Boston, 2013.

613

(44) Legates, D. R.; McCabe, G. J. Evaluating the use of “goodness-of-fit” measures in

614

hydrologic and hydroclimatic model validation. Water Resour. Res. 1999, 35 (1), 233–241;

615

DOI 10.1029/1998WR900018.

616

(45) Boyle, D. P.; Gupta, H. V.; Sorooshian, S. Toward improved calibration of hydrologic

617

models: Combining the strengths of manual and automatic methods. Water Resour. Res.

618

2000, 36 (12), 3663–3674; DOI 10.1029/2000WR900207.

619

(46) van Griensven, A.; Meixner, T.; Grunwald, S.; Bishop, T.; Diluzio, M.; Srinivasan, R.

620

A global sensitivity analysis tool for the parameters of multi-variable catchment models. J.

621

Hydrol. 2006, 324 (1–4), 10–23; DOI 10.1016/j.jhydrol.2005.09.008.

622

(47) Beven, K. A manifesto for the equifinality thesis. The model parameter estimation

623

experiment

MOPEX

MOPEX

624

10.1016/j.jhydrol.2005.07.007.

workshop

2006,

320

28 ACS Paragon Plus Environment

(1–2),

18–36;

DOI

Page 29 of 33

Environmental Science & Technology

625

(48) Kirchner, J. W. Getting the right answers for the right reasons: Linking measurements,

626

analyses, and models to advance the science of hydrology. Water Resour. Res. 2006, 42 (3),

627

W03S04; DOI 10.1029/2005WR004362.

628

(49) Madsen, H. Automatic calibration of a conceptual rainfall–runoff model using

629

multiple objectives. J. Hydrol. 2000, 235 (3–4), 276–288; DOI 10.1016/S0022-

630

1694(00)00279-1.

631

(50) Meigh, J. R.; McKenzie, A. A.; Sene, K. J. A Grid-Based Approach to Water Scarcity

632

Estimates for Eastern and Southern Africa. Water Resour. Manag. 1999, 13 (2), 85-115;

633

DOI 10.1023/A:1008025703712.

634 635 636 637

(51) Hanasaki, N.; Kanae, S.; Oki, T. A reservoir operation scheme for global river routing models. J. Hydrol. 2006, 327 (1–2), 22–41; DOI 10.1016/j.jhydrol.2005.11.011. (52) Beven, K. On doing better hydrological science. Hydrol. Process. 2008, 22 (17), 3549–3553; DOI 10.1002/hyp.7108.

638

(53) Harmel, R. D.; Smith, P. K. Consideration of measurement uncertainty in the

639

evaluation of goodness-of-fit in hydrologic and water quality modeling. J. Hydrol. 2007, 337

640

(3–4), 326–336; DOI 10.1016/j.jhydrol.2007.01.043.

641

(54) Pfister, S.; Hellweg, S. Uncertainty in LCIA of water impacts on human health.

642

Surface

water

use



human

643

http://www.ifu.ethz.ch/ESD/downloads/reports/index_EN, 2011.

health

impacts;

644

(55) Zaitchik, B. F.; Rodell, M.; Olivera, F. Evaluation of the Global Land Data

645

Assimilation System using global river discharge data and a source-to-sink routing scheme.

646

Water Resour. Res. 2010, 46 (6), W06507; DOI 10.1029/2009WR007811.

647

(56) Gerten, D.; Heinke, J.; Hoff, H.; Biemans, H.; Fader, M.; Waha, K. Global Water

648

Availability and Requirements for Future Food Production. J. Hydrometeorol. 2011, 12 (5),

649

885–899; DOI 10.1175/2011JHM1328.1.

29 ACS Paragon Plus Environment

Environmental Science & Technology

Page 30 of 33

650

(57) Jeswani, H. K.; Azapagic, A. Water footprint: methodologies and a case study for

651

assessing the impacts of water use. J. Clean. Prod. 2011, 19 (12), 1288–1299; DOI

652

10.1016/j.jclepro.2011.04.003.

30 ACS Paragon Plus Environment

Page 31 of 33

Environmental Science & Technology

TOC Art 84x47mm (220 x 220 DPI)

ACS Paragon Plus Environment

Environmental Science & Technology

Model performance of calibrated river discharge using four different criteria (green colors indicate good performance and red colors indicate poor performance). 152x127mm (220 x 220 DPI)

ACS Paragon Plus Environment

Page 32 of 33

Page 33 of 33

Environmental Science & Technology

Distributed average water stress indices of the Mississippi watershed considering water consumption and upstream level (a: annual total water resources, b: monthly total water resources, c: monthly surface water resources, d: monthly groundwater resources). 152x114mm (220 x 220 DPI)

ACS Paragon Plus Environment