Kalra Et Al-2012-Water Resources Research

WATER RESOURCES RESEARCH, VOL. 48, W06527, doi:10.
1029/2011WR010667, 2012
Estimating annual precipitation for the Colorado River Basin

using oceanic-atmospheric oscillations
Ajay Kalra1,2 and Sajjad Ahmad1
Received 14 March 2011; revised 22 April 2012; accepted 14 May 2012; published 22 June 2012.
[1] Estimating long-lead time precipitation under the stress of increased climatic
variability is a challenging task in the field of hydrology. A modified Support Vector
Machine (SVM) based framework is proposed to estimate annual precipitation using
oceanic-atmospheric oscillations. Oceanic-atmospheric oscillations, consisting of Pacific
Decadal Oscillation (PDO), North Atlantic Oscillation (NAO), Atlantic Multidecadal
Oscillation (AMO), and El Niño–Southern Oscillation (ENSO) for a period of 1900–2008,
are used to generate annual precipitation estimates with a 1 year lead time. The SVM model
is applied to 17 climate divisions encompassing the Colorado River Basin in the western
United States. The overall results revealed that the annual precipitation in the Colorado
River Basin is significantly influenced by oceanic-atmospheric oscillations. The long-term
precipitation predictions for the Upper Colorado River Basin can be successfully obtained
using a combination of PDO, NAO, and AMO indices, whereas coupling AMO and ENSO
results in improved precipitation predictions for the Lower Colorado River Basin. The
results also show that the SVM model provides better precipitation estimates compared to
the Artificial Neural Network and Multivariate Linear Regression models. The annual
precipitation estimates obtained using the modified SVM modeling framework may assist
water managers in statistically understanding the hydrologic response in relation to large
scale climate patterns within the Colorado River Basin.
Citation: Kalra, A., and S. Ahmad (2012), Estimating annual precipitation for the Colorado River Basin using oceanic-atmospheric
oscillations, Water Resour. Res., 48, W06527, doi:10.1029/2011WR010667.
1. Introduction 2012] have been studied by many researchers. Considering

1.1. Background the impacts of floods and droughts on humans and environ-
ment and growing water demand in many parts of the
[2] Climatic fluctuations and increasing water demand in world, efforts have increased to study and predict regional
growing regions have captured the attention of scientific and global precipitation variation. Interannual, decadal, and
communities. This has led to the need for estimating re- multidecadal climatic signals, such as oceanic-atmospheric
gional and global precipitation of interannual and longer oscillations provide an exciting opportunity to estimate pre-
time scales [Karl and Knight, 1997; Hidalgo and Dracup, cipitation at longer lead-times. Oceanic-atmospheric oscil-
2003]. Although precipitation is predominantly episodic, lations, often termed as teleconnections, have been used by
variability in precipitation results in such catastrophic researchers across the world to study their relationship with
events as floods and drought. The impact of these cata- precipitation over land surfaces at interannual and longer
strophic events on water resources planning [Qaiser et al., time scales ; this relationship, in turn, controls such key
2011; Dawadi and Ahmad, 2012]; water management components as streamflow, soil moisture, and evaporation
[Ahmad and Simonovic, 2001, 2006], water infrastructure of the hydrological cycle [Diaz and Kiladis, 1992; Dracup
[Forsee and Ahmad, 2011; Ahmad and Simonovic, 2000, and Kahya, 1994; Rajagopalan and Lall, 1998; Diaz et al.,
2004], human settlements [Mosquera-Machado and Ahmad, 2001; Gutzler et al., 2002; Viles and Goudie, 2003;
2007; Simonovic and Ahmad, 2005], and the environment Hidalgo and Dracup, 2003; Xu et al., 2004; Hu and Feng,
[Venkatesan et al., 2011a, 2011b; Shrestha et al., 2011, 2001; Stephen et al., 2010; Puri et al., 2011]. The main
objective of this study is to estimate annual precipitation
1
using oceanic-atmospheric oscillations.
Department of Civil and Environmental Engineering, University of Ne-
vada, Las Vegas, Nevada, USA.
2
1.2. Oceanic Variability and Precipitation
Now at Division of Hydrologic Sciences, Desert Research Institute,
Las Vegas, Nevada, USA. [3] It is well established that the year-to-year variability of
precipitation is primarily associated with the sea surface tem-
Corresponding author: S. Ahmad, Department of Civil and Environ- perature (SST) anomalies in the tropical Pacific Ocean, known
mental Engineering, University of Nevada, Las Vegas, 4505 S. Maryland
Parkway, Las Vegas, NV 89154-4015, USA. ([email protected]) as the El Niño–Southern Oscillation (ENSO) [Ropelewski and
Halpert, 1986; Redmond and Koch, 1991; Piechota and
©2012. American Geophysical Union. All Rights Reserved. Dracup, 1996; Kane, 1999; Barlow et al., 2002]. Wang et al.
0043-1397/12/2011WR010667 [2000] observed positive anomalies of precipitation in the
W06527 1 of 24
W06527 KALRA AND AHMAD: PRECIPITATION VARIATION AND OCEANIC OSCILLATIONS W06527
central Pacific and eastern Asia during extreme phases of Network (ANN). Several researchers have used different
ENSO cycles. Lau and Wu [2001] found that ENSO accounts types of ANN algorithms to forecast precipitation [Raman
for 30% of climate-related variability in Asian summer mon- and Sunilkumar, 1995; Kuligowski and Barros, 1998;
soon rainfall. Hu and Feng [2001] revealed that warmer Tokar and Johnson, 1999; Hsu et al., 1995, 1997; Tokar
phases of ENSO are related to increased summer precipitation and Markus, 2000]. A detailed review of the ANN applica-
within the central United States. A pseudo El Niño phenom- tions in hydrology is available from the work of the ASCE
enon, known as the El Niño Modoki, has been significantly Task Committee [2000b].
linked to temperature and precipitation over many parts of the [7] Although ANNs are powerful mathematical struc-
world, including Japan, Korea, New Zealand, and even the tures, they have certain drawbacks that include getting
West Coast of the United States [Ashok et al., 2007; Weng trapped in local minima and subjectivity in selection of
et al., 2007]. model architecture [Suykens, 2001, Lin et al., 2009]. Due to
[4] In the southwestern United States, particularly in the these drawbacks, there is a need for a more sophisticated
Colorado River Basin (CRB), various types of teleconnec- AI-type data-driven model that is capable of efficiently rep-
tions are found to be related with the hydrologic variability resenting the multifaceted interaction between oceanic-
[Pulwarty and Melis, 2001]. Piechota and Dracup [1996] atmospheric oscillations and precipitation.
used Palmer Drought Severity Index (PDSI) to relate the [8] Another type of data-driven model becoming popular
major dry and wet spells with historic ENSO events in the in ANN-dominated fields is the Support Vector Machine
Lower Colorado River Basin (LCRB). Kahya and Dracup (SVM). SVM is based on the structural risk minimization
[1993] described 1941 and 1983 heavy rainfall events with (SRM) principle, which helps it to efficiently relate the rel-
ENSO phases in the southwestern U.S., including the CRB. ative input to the desired output and perform better on
Cayan et al. [1998] showed that there has been a change in unseen data [Vapnik, 1995, 1998]. In other words, SVM is
the pattern and amount of precipitation within the CRB. a statistical tool that approaches the problem of training
Merideth [2000] observed wet periods during the start and polynomial functions, radial basis functions, or neural net-
end of 20th century, whereas dry periods during the mid- work regression estimators in a way that is similar to neural
century in the Colorado Basin. Moreover, the National networks; however, at the same time, it uses a new
Oceanic and Atmospheric Administration (NOAA) Climate approach [Liong and Sivapragasam, 2002].
Prediction Center (CPC) routinely monitors the correlation [9] In comparing SVMs and ANNs, there are two major
between ENSO and southwest U.S. precipitation [California differences. First, ANN is constructed based on the princi-
Dep. of Water Resources, 2005]. ple of empirical risk minimization (ERM), whereas SVMs
[5] While many studies have related ENSO with precipi- are based on the SRM principle. The ERM principle only
tation, there are other modes of atmospheric oscillations, minimizes the total error (empirical risk), whereas the
for instance, Pacific Decadal Oscillation (PDO), Interdeca- SRM principle minimizes empirical risk and addresses the
dal Pacific Oscillation (IPO), North Atlantic Oscillation model complexity against its success at fitting the training
(NAO), Atlantic Multidecadal Oscillation (AMO), Arctic data set in order to avoid over fitting [Haykin, 2003]. This
Oscillation (AO), Indian Ocean Dipole (IOD), and Pacific results in a better capability for generalization. Second,
North American (PNA) indices. These also are important quadratic optimization problem is used to express the archi-
indicators of climate variability, and have been linked with tecture and weight of SVM. The optimization problem can
precipitation both individually and in conjunction with be solved using standard programming algorithms. Due to
ENSO [Bjerknes, 1966; Hurrell, 1995; Giannini et al., the set of weights used in SVM, useful information is
2001; Gershunov and Barnett, 1998; Higgins et al., 2000; revealed to the user in terms of model error and global opti-
Brito-Castillo et al., 2002; McCabe et al., 2004; Wang and mum is guaranteed. Contrary to this, the weight and archi-
Swail, 2001; Dickson et al., 2000; Qian et al., 2000; Ashok tecture of ANN is determine by trial and error procedure,
et al., 2001, 2003; Wedgbrow et al., 2002; Kiem and which is iterative and time consuming. For these reason,
Franks, 2004; Pui et al., 2011]. Compared to ENSO and SVM models have been applied successfully in hydrology
PDO, the AMO and NAO indexes exhibit considerable to forecast streamflow [Kalra et al., 2012; Lin et al., 2010;
long-term variability [Cancelliere et al., 2007; McCabe Kalra and Ahmad, 2009; Asefa et al., 2006; Yu and Liong,
et al., 2007]. 2007; Liong and Sivapragasam, 2002; Dibike et al., 2001],
precipitation [Lin et al., 2009; Tripathi et al., 2006], and
1.3. Artificial Intelligence Models soil moisture [Ahmad et al., 2010; Gill et al., 2006]. The
[6] From the past studies, it is evident that oceanic- majority of these applications have shown the superiority
atmospheric oscillations do influence precipitation. In fact, of SVM modeling over the traditional ANN modeling
there have been attempts to use oscillations as predictors to approach.
estimate precipitation. However, due to the complex inter- [10] However, even though SVMs are considered supe-
action between precipitation and oceanic oscillation, it is rior to ANNs and other regression methods, they still are
difficult to construct a physically based mathematical statistical data-driven models. A number of factors deter-
model [Lin et al., 2009]. An attractive alternative are mine the accuracy of an SVM [Boser et al., 1992; Guyon
the artificial intelligence (AI) models, also referred to as et al., 1993; Schölkopf et al., 1997; Smola et al., 1998]:
machine-learning or data-driven models. AI-type models (1) the choice of kernel, which is responsible for the data
are constructed by employing flexible and adaptive model transformation into the high-dimensional space in which
structures in an empirical format and can be used to deter- SVM performs regression ; (2) the determination of the
mine the relationship between inputs and outputs. By far, hyper-parameters used in the model, which can result in
the most popular AI-type of model is the Artificial Neural overfitting or underfitting and can affect the accuracy of the
2 of 24
model’s predictions. To overcome these issues, several [12] The use of climate oscillation information has the
algorithms and methods have been developed, and are potential to improve hydrologic forecasts within a basin
available in the literature [Smola and Schölkopf, 2004]. [Piechota and Dracup, 1996; Cayan et al., 1998; Coulibaly
et al., 2000; Kim et al., 2008; Kalra and Ahmad, 2009; Pui
et al., 2011]. This research is an attempt to perform a statisti-
1.4. Motivation for Current Research cal analysis, by incorporating oceanic-atmospheric oscilla-
[11] As evident from the documented literature, the qual- tions in a SVM model, to estimate annual precipitation with
itative relationship between oceanic-atmospheric oscilla- a 1 year lead time for the Colorado River Basin. Once the
tions and precipitation has been studied extensively annual precipitation with a 1 year lead time has been esti-
[Ropelewski and Halpert, 1986; Redmond and Koch, mated, such stochastic techniques as K Nearest Neighbor
1991; Hurrell, 1995; Piechota and Dracup, 1996; Cayan (KNN) can be used to temporally disaggregate the precipita-
et al., 1998; McCabe and Dettinger, 1999; Wang et al., tion into seasonal, or monthly rainfall, depending on the
2000; Giannini et al., 2001; McCabe et al., 2004]. How- need of the end user [Rajagopalan and Lall, 1998, 1999;
ever, little attention has been paid to statistically understand- Kalra and Ahmad, 2011]. Furthermore, annual precipitation
ing the coupled impact of oscillations on precipitation for the estimates are helpful in analyzing the sediment yield within
Colorado River Basin by using an AI-type data driven the basin, which varies as a function of annual precipitation
model. This current study uses the four oceanic-atmospheric [Wilson, 1973]. Although there have been attempts to
oscillations—PDO, NAO, AMO, and ENSO—to estimate increase the precipitation lead time, whatever modest skills a
annual precipitation with a 1 year lead time for 17 climate climatologist may have at predicting a 3–9 month lead time
divisions encompassing the Colorado River Basin. These are arises from ENSO and its effect [California Dep. of Water
the only four oceanic indices for which reconstructed data Resources, 2005; Kim et al., 2006, 2008].
have been developed by using tree ring information; data are [13] This current study aims to provide annual precipitation
available from the National Climate Data Network website totals with a 1 year lead time by using oceanic-atmospheric
http://www.ncdc.noaa.gov/paleo/recons.html [Carrier et al., indices in a data-driven model. It should be noted that when
2011]. These oscillations have been studied by numerous estimating annual cumulative precipitation, as done in the
researchers, using different time scales to show their interac- current study, the standard definition of ‘‘lead time,’’ i.e., lags
tion with the western U.S. hydroclimatology, particularly the between the predictors and the predictand, may not be appro-
Colorado River Basin [Piechota and Dracup, 1996; McCabe priate. The current study uses values for the previous year’s
et al., 2004; Webb et al., 2004; Kim et al., 2006, 2008; Ellis oceanic oscillations, and estimates annual precipitation totals
et al., 2010]. Hydroclimatic variability within the CRB, for the current year; this is referred to as a 1 year lead time.
using PDO and ENSO as climate indices, has been reported Thus, on 1 January of the current year, the cumulative total
extensively [Kahya and Dracup, 1993; Piechota and precipitation for the entire current year can be known using
Dracup, 1996; Mantua et al., 1997; Cayan et al., 1999; the proposed modeling approach.
Pulwarty and Melis, 2001]. PDO in combination with [14] Generally, in any SVM type regression problem, the
AO and PNA has been studied to improve the lead times time series of the data is split into specific training and test-
for streamflow and precipitation [Coulibaly et al., 2000; ing data sets. The model is trained on only one data set and
Coulibaly, 2006; Kim et al., 2006, 2008]; however, the tested on the other unseen data set. We propose a modified
coupled impact of PDO and NAO for the Upper Basin has framework of SVM modeling approach, in which a data
not been explored. Attempts have been made to study the point is held out and the model is trained on the remaining
relationship between AMO and the hydrologic conditions data and then tested on the held out data. This process is
within the CRB [Webb et al., 2004]. Contrary to this, the repeated for the entire length of the data by moving the
NAO primarily has been studied in relation to changes in model forward in time. This SVM approach has been previ-
mean sea level pressures (SLP) over the Arctic Ocean ously applied in evaluating contaminants in groundwater
[Walsh et al., 1996], trends in surface wave heights over the by Khalil et al. [2005] but has never been used in a climate
North Atlantic [Kushnir et al., 1997], and predicting storm related research. This is the first attempt to use this modi-
activity and shifts in storm tracks in the Atlantic Ocean fied SVM predictive framework in a hydrologic forecasting
[Hurrell, 1995]. Lesser attention has been given to the study incorporating large-scale climate patterns.
changes in precipitation in relation to NAO within the Colo- [15] Along with extending the forecast lead time by using
rado River Basin. Majority of the literature suggests that four oceanic-atmospheric oscillations, current research per-
tropical Pacific drives the hydrologic variability over CRB. forms a rigorous sensitivity analysis to statistically determine
However, there have been attempts to study the relationship the coupled impact and also the individual impact of each
between Atlantic SST and Colorado River Basin [Kim et al., oscillation mode with respect to annual precipitation. Addi-
2006, 2008; Ellis et al., 2010]. Although, it is well estab- tionally, the robustness of the SVM approach is verified
lished that tropical Pacific modes of climate, i.e., PDO and using a bootstrapped cross validation technique.
ENSO are related to the hydrologic conditions in the western [16] Furthermore, the SVM precipitation estimates are
U.S.; there are other Atlantic climate patterns that are also compared with a feed-forward, back-propagation ANN
believed to be statistically connected with the western U.S. model as well as a Multivariate Linear Regression (MLR)
hydroclimatology [Pulwarty and Melis, 2001]. The physical model. All models are evaluated using root-mean-square
reason of the relationship between Atlantic SST’s and west- error (RMSE), mean error (ME), RMSE-observation stand-
ern U.S. hydroclimatology is uncertain; it is still necessary ard deviation ratio (RSR), correlation coefficient (R), Nash
to investigate both Pacific and Atlantic climate modes Sutcliffe coefficient of efficiency (NSE), percent bias
[McCabe et al., 2007; Thomas, 2007; Kim et al., 2008]. (Pbias), and linear error in probability space (LEPS) skill
3 of 24
score (SK). It should be noted that an exhaustive compari- individually. It should be noted that portions of the San Juan
son between different methods for estimating precipitation, and Dirty Devil subbasins do not intersect with any of the
under different circumstances, is not the goal of this study. Upper Basin climate divisions, and therefore were not
Instead, the results of the SVM model were compared to included in the calculations. This accounts for approximately
ANN and MLR models to explore how well each model 5% of the flow. Therefore, the total flow percentage in Figure
was able to estimate precipitation within CRB by using 1c adds up to less than 100%. In the LCRB, majority of the
oceanic-atmospheric oscillations. average annual streamflow occurs in the winter and spring
[17] This paper is organized as follows. Section 2 seasons. The northwestern and central regions of LCRB gen-
presents the description of the study region. The precipita- erate approximately 70% flow, whereas approximately 35%
tion data for the climate divisions is described in section 3. flow is generated in the southern LCRB [Thomas, 2007].
The description of the SVM modeling approach for esti-
mating precipitation, and the performance measures used to
evaluate the model performance are given in section 4. Sec- 3. Data
tion 5 describes the statistical properties of oscillations and [20] The data sets used to estimate annual precipitation
annual precipitation. Section 6 provides a discussion of the are the oceanic-atmospheric modes of the Pacific and the
SVM modeling results. A comparison of the SVM precipi- Atlantic Oceans (Figure 2) as well as the precipitation time
tation estimates is presented with estimates obtained from series (Figure 3). A brief description of the precipitation
the ANN and MLR models. Section 7 summarizes and con- data is provided in this paper. The oscillation data used in
cludes the paper. the current analysis is similar to the authors’ previous
work; a detailed description can be found in the work of
Kalra and Ahmad [2009].
2. Study Region [21] The precipitation data used in this study is the aver-
[18] The Colorado River is the most regulated river in age monthly time series data for 17 climate divisions, cov-
the United States, and is governed by the ‘‘Law of the ering a period from 1901–2008. The monthly data set is
River’’ [Sax et al., 2000]. It encompasses seven states, and added to obtain the annual precipitation time series for
is a major source of water to the southwestern United States each of the climate division. This data is obtained from the
(Figure 1a). It provides industrial and municipal water to National Climate Data Center (NCDC) http://www.esrl.noaa.
nearly 25 million people by means of existing reservoirs, gov/psd/cgi-bin/data/timeseries/timeseries1.pl. The NCDC pre-
Lake Powell, and Lake Mead [Sax et al., 2000]. It provides pares the data over each climate division by taking an average
agriculture water for nearly 3 million acres, and produces of temperature and precipitation from stations within a divi-
11.5 billion kilowatt-hours of hydroelectric power. The sion, reported by the National Weather Service (NWS) Coop-
Colorado River Basin is composed of the Upper Basin erative Observer Program (COOP). The data set is corrected
(Wyoming, Colorado, Utah, and New Mexico) and the for time bias by adjusting for the variations in average monthly
Lower Basin (California, Nevada, and Arizona). The flow mean temperatures, as described in the work of Karl et al.
demand between the upper and the lower basin is estab- [1986]. The divisional averages are computed starting 1931 to
lished by the flow at Lee’s Ferry (depicted by a triangle in present, whereas each station value was computed directly
Figure 1a), which acts as the hydrologic divide. The major- within a state from 1895–1930 [Guttman and Quayle, 1996].
ity of the flow, nearly 90%, is generated in the Upper Basin Therefore the count and distribution of the stations within
from the spring-summer snowmelt [Prairie and Callejo, COOP have changed over time. This may be considered a li-
2005]. Based on the flow contribution, the Upper Basin is mitation in the data set, but the data corresponds well both spa-
subdivided further into eight subbasins (Figure 1b). tially and temporally to large-scale historic climate anomalies,
[19] The United States is divided into 344 climate divi- such as drought [Guttman and Quayle, 1996].
sions, based on the climatic boundaries (available at http:// [22] It should be noted that the CRB is composed of
www.esrl.noaa.gov/psd/data/usclimate/map.html). The Colo- highly varied elevations and climate regimes, and it is diffi-
rado River Basin encompasses 29 climate divisions. Out of cult to integrate all precipitation contributions into a single
these 29 climate divisions, 17 divisions have greater than 30% time series that is representative of the entire basin precipi-
of their area within the Colorado River Basin (Figure 1a). For tation. Therefore, the climate division data used in the anal-
the purpose of this study, the climate divisions have been ysis is helpful in representing the temporal and spatial
sorted according to different states, and have been numbered variation of precipitation within CRB. The annual spread
from 1–17. Table 1 show the nomenclature used to identify of the input data for each climate division is shown in verti-
each climate division within a particular state. Divisions 1–7 cal box plots in Figure 3. The horizontal line inside each
and 10–12 are within the Lower Basin, and Divisions 8–9 and box shows the median value. The box represents the 25th
13–17 encompass the Upper Basin. An area-weighted method and 75th percentile (interquartile range) values, and the
was employed in the Geographic Information System to com- whiskers extend from 5th to 95th percentile values. The dot
pute the flow contribution from each of the Upper Basin cli- inside the box shows the historic mean of the input data.
mate divisions. The climate divisions were merged with the [23] The box plots show that the annual precipitation
subbasins of the Upper Basin (Figure 1b) to compute the per- within the CRB exhibits a higher degree of variability, as
centage contribution of each subbasin within that respective indicated by wider box plots for the majority of the climate
division. This resulted in a maximum flow generated by Cli- divisions. Estimating this variability by using oceanic-atmos-
mate Division 8 (57%), followed by Climate Division 17 pheric oscillations is a challenging task. It is noteworthy that
(14%), and 16 (11%), as shown in Figure 1c. The remain- the division with maximum flow in the upper basin, Climate
ing divisions in the Upper Basin generate less than 6% flow, Division 8 (Figure 1b), does not correspond to the largest
4 of 24
Figure 1. Map showing the location of (a) the Colorado River Basin and the 17 climate divisions,
(b) percent flow contribution from UCRB to Colorado River, and (c) flow generated from each climate
division in the Upper Basin. The location of Lees Ferry is indicated by a triangle.
precipitation, as shown in Figure 3. The division with a lower Feng and Hu, 2007]. The precipitation in Climate Division
flow, Climate Division 14, experiences large precipitation 14 probably is due to lower elevation rainstorms, which are
amounts (Figure 3). The most probable cause for this differ- more susceptible to infiltration, evaporation, etc.; as a result,
ence is the effective precipitation, that is, precipitation that is this water does not end up contributing significantly to the
realized as runoff. Precipitation in Climate Division 8 mostly Colorado River Basin flow [Colle, 2004]. This emphasizes
is seen as upper elevation snowpack that gradually melts and the fact that there is not a linear correlation between precipi-
contributes significantly to streamflow [Hamlet et al., 2005; tation and streamflow within Colorado River Basin.
5 of 24
Table 1. List of Climate Divisions Used in the Study A more detailed discussion on the subject can be found in
the work of Vapnik [1995, 1998]. The concept of Support
Climate Division Name State Region
Vector Regression (SVR) is to nonlinearly map the input
1 NORTHWEST AZ Lower Basin data x into a higher dimensional feature space such that:
2 NORTHEAST AZ Lower Basin
3 NORTH CENTRAL AZ Lower Basin
f ¼ wðxi Þ þ b: (1)
4 EAST CENTRAL AZ Lower Basin
5 SOUTHWEST AZ Lower Basin
6 SOUTH CENTRAL AZ Lower Basin [26] In equation (1) (xi) is an input feature and both w
7 SOUTHEAST AZ Lower Basin and b are coefficients that are estimated by minimizing the
8 CO DRAINAGE BASIN CO Upper Basin regularized risk function,
9 NORTHWESTERN PLATEAU NM Upper Basin
10 SOUTHWESTERN MOUNTAINS NM Lower Basin
11 EXTREME SOUTHERN NV Lower Basin 1X N
1
12 DIXIE UT Lower Basin Rðf Þ ¼ C ð þ i Þ þ jjwjj2 (2)
13 SOUTH CENTRAL UT Upper Basin N i¼1 i 2
14 NORTHERN MOUNTAINS UT Upper Basin
15 UINTA BASIN UT Upper Basin 8
16 SOUTHEAST UT Upper Basin >
> X K X L
>
> y wj xji b " þ i
17 GREEN AND BEAR DRAINAGE WY Upper Basin >
>
i
>
> j¼1 i¼1
<
Subject to X K X L (3)
>
> wj xji þ b yi " þ i :
>
>
4. Method >
> j¼1 i¼1
>
>
:
[24] This section describes the modified SVM modeling i ; i 0
framework abstracted from Kalra [2012] and also provides a
description of the model evaluation performance measures. [27] In equations (2) and (3), C is the cost, K is the num-
ber of support vectors; and " is called the Vapnik’s insensi-
4.1. SVM Modeling tive loss function. The Vapnik’s "-insensitive loss functions
[25] This section provides a brief explanation of the under- act as a threshold in the sense that errors less than " are not
lying principles of SVM abstracted from Vapnik [1995]. considered. Additionally, i and i are the slack variables.
Figure 2. Time series plot of annual average oscillations used in the study.
6 of 24
Figure 3. Box plots depicting annual precipitation data from 1901–2008 for 17 climate divisions
encompassing the Colorado River Basin. The horizontal line inside the box shows the median value. The
box represents the 25th and 75th percentile values (an interquartile range), and the whiskers extend from
5th to 95th percentile values. The circular dot inside the box represents the long–term mean of annual
precipitation.
If errors are larger than ", the slack variables are used to process is repeated for all instances and stops when i ¼
determine the degree to which samples will be penalized. N þ 1. A brief description of the step-by-step algorithm is
Interested readers are referred to Kalra and Ahmad [2009] described below.
for the working mechanism as well as an example of the [31] Step 1. Let [Xi] represent the data matrix compris-
SVM modeling approach. ing of all observations used in the study of length N, where
[28] As mentioned in the work of Twarakavi et al. i is the featured instance and varies from 1:N.
[2006], a typical modeling framework of any AI model [32] Step 2. Partition matrix [Xi] into two submatrices,
consists of the following four steps: (1) preparation of spe- [Ai] and [Bi], such that [Ai] is of length N-i and [Bi] is of
cific training and testing data suitable for model, (2) train- length i.
ing the model using the specific trained data set, (3) testing [33] Step 3. Train the SVM model on [Ai] and test the
the trained model using the testing data set, and (4) cross- model on [Bi].
validating the model using the entire data set. Step 1 is [34] Step 4. Repeat steps 1–3 for all the featured
essential in every data-driven modeling application. Step 2 instances.
fits the model and step 3 is used to evaluate the model per- [35] Step 5. Evaluate model performance for all instan-
formance on unseen data. Step 4 is applied at the end to ces (pooled) of [Bi].
show the robustness of the model: it tests to see if different [36] Step 6. Apply steps 1–5 for other climate divisions.
training and testing datasets do not yield different results
that lead to different conclusions. To make the SVM mod- 4.1.2. Bootstrap Cross Validation
eling approach more robust and to improve the efficiency [37] To establish the robustness of the SVM approach, it
of the SVM approach as a better forecasting tool, the cur- is necessary to determine that different training and testing
rent research proposes a modified SVM modeling frame- samples does not yield different results. For this purpose a
work. The proposed modification tests the model for each cross validation approach is needed to determine the per-
sample in the data and is not limited to only single training formance of the predictive model. Several cross validation
and testing sample. Additionally, the robustness of the approaches, i.e., k-fold, leave out, bootstrapping etc. have
modified SVM prediction framework is verified using a been documented in available literature [Khalil et al.,
bootstrapping cross validation technique. 2005; Asefa et al., 2006; Twarakavi et al., 2006; Chowd-
hury and Sharma, 2009; Lin et al., 2009; Li et al., 2010].
4.1.1. Modified SVM Framework All of the cross validation approaches have their advan-
[29] The modified SVM approach framework is a special tages and disadvantages but help in assessing how the
case of the k-fold cross validation technique [Geisser, results of a statistical analysis will generalize on an inde-
1975; Stone, 1974] in which k ¼ 1. Each data point in the pendent data set. The current study used a bootstrap cross
time series is held out in turn; the model is trained on the validation technique to test the stability of the SVM model
remaining data set (N-k, where N is the number of observa- results. Bootstrap or bagging is a statistical procedure that
tions in the data set) and tested on the held-out data point, uses intensive resampling, with replacement to reduce the
i.e., k. This process is repeated for the entire length of data uncertainties in the data [Efron, 1979]. The bootstrap
(k ¼ N). method is used to generate different realizations of the
[30] Let i represent the current instance, which is also measured data that can be used to assess the mean and the
representative of the testing instance (i ¼ 1). For each variability of the estimates. In this study each SVM model
instance i, the training set will consist of [1, i) U (I, N]. is trained on a set of bootstrap samples and tested on the
This will train the model on all instances, except for entire measured data set. The resampling procedure is
instance i, and will test the model on instance i only. This repeated 100 times and results in 100 ensembles for each
7 of 24
measured value. The final estimated value is the mean of hydroclimatic variables [Raman and Sunilkumar, 1995;
these 100 ensembles. Kuligowski and Barros, 1998; Tokar and Johnson, 1999;
[38] Based on the modeling approach described, four Hsu et al., 1995; Ahmad and Simonovic, 2005; Melesse
SVM models are developed using annual oceanic-oscilla- et al., 2011]. A more detailed description on the theoretical
tions PDO, NAO, AMO, and ENSO for time step ‘‘t’’ in aspects of ANN is available in the work of the ASCE Task
order to estimate precipitation at ‘‘t þ 1’’ (t is in year) for Committee [2000a]. The other type of model developed is
17 climate divisions encompassing the CRB. Each climate the parametric Multivariate Linear Regression model,
division is considered independent, and separate SVM which consists of oceanic-oscillations as the predictors and
models are developed for each division. The SVM models annual precipitation as the predictand. The modeling
are developed by using the standard software package approach used to develop the corresponding ANN model
included in the Comprehensive R Achieve Network (avail- and MLR model is similar to the SVM approach.
able at http://www.r-project.org/). Model I is termed as the
‘‘base case,’’ and uses all four oceanic modes to estimate 4.2. Model Performance Evaluation
annual precipitation. To estimate precipitation, Models II [41] The current study uses seven performance meas-
and III use a combination of 3 and 2 oscillation modes, ures: RMSE, ME, RSR, R, NSE, Pbias, and LEPS SK.
respectively. Model IV uses a single oscillation mode to Lower RMSE and ME represent better model performance
estimate precipitation. The major reason for developing [Singh et al., 2005]. R determines the linear association
Models II–IV is to evaluate the role of individual and between the measured and predicted value. [Legates and
coupled oceanic-oscillations in estimating precipitation McCabe, 1999]. RSR standardizes RMSE by using the
within the basin. standard deviation of observations [Singh et al., 2004]; it is
[39] Based on equation (2), the performance of any SVR calculated as the ratio of RMSE and standard deviation of
formulation depends on the selection of hyper-parameters: observed data. Singh et al. [2004] published the guidelines
cost (C), insensitivity value ("), and the radial basis kernel for RSR based on lower RMSE values. An RMSE value of
width (). Previous studies have used the following three less than half the standard deviation of observation is con-
procedures to estimate hyper-parameters in any SVR for- sidered low. Therefore, Moriasi et al. [2007] used the rec-
mulation: (1) user expertise, (2) grid based search, and ommended value (per Singh et al. [2004]) of less than 0.5
(3) using the statistical properties of the training data in an RSR to categorize model performance as ‘‘very good,’’ and
analytical approach. In the current study, a grid-based suggested a less stringent rating of 10% points and 20%
search was adopted to compute the hyper-parameters. In points greater than 0.5 RSR to be ‘‘good’’ and ‘‘satisfac-
grid based search every possible combination within a fea- tory.’’ RSR as mentioned in the work of Moriasi et al.
sible hyper-parameters space is considered and prediction [2007] can be statistically expressed as
error is computed for each combination. The feasible pa-
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
rameter space for each hyper-parameter is constructed X n
using the minimum and maximum possible values that are ðYiobs Yiest Þ2
given a priori (0.001 < C < 1000, 0.001 < < 100, and RMSE i¼1
RSR ¼ ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi : (4)
0 < " < 1). An increment of 0.01 is selected that helps in STDEVobs X n
ðYiobs Y mean Þ2
ensuring the optimality and computational efficiency of the i¼1
grid based approach. The hyper-parameters, which result in
minimal mean square error, are selected as optimal values. [42] NSE is used to access predictive power of hydrolog-
It should be noted that the hyper-parameters for all years ical models [Nash and Sutcliffe, 1970]. NSE is computed as
are not exactly the same (as training data changes) but cor-
respond to a tight cluster. A number of previous studies sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X n
have used similar approach [Asefa et al., 2006; Gill et al., ðYiobs Yiest Þ2
2006; Kalra and Ahmad, 2009; Twarakavi et al., 2006, i¼1
2009; Ahmad et al., 2010]. Furthermore, the performance NSE ¼ 1 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi : (5)
X n
of SVM depends on the choice of kernel as being a kernel’s ðYiobs Y mean Þ2
parameterization problem. In this study, a radial basis ker- i¼1
nel is used in the SVM model; this has performed better
when compared with other kernels, such as linear, polyno- [43] NSE ranges from 1 to 1, where NSE of 1 corre-
mial, sigmoid or spline, as evident in the past studies sponds to a perfect match between the observed and pre-
[Schölkopf et al., 1997; Smola et al., 1998; Dibike et al., dicted values. As reported in the work of Moriasi et al.
2001; Yu and Liong, 2007]. [2007], Pbias measures the average tendency of the esti-
[40] Additionally, the performance of SVM model is mated data in comparison with the observed values and is
compared with the two other hydrologic time series model- computed as
ing approaches, i.e., ANN and MLR. A feed-forward back- X
n
propagation ANN model with one input layer, one hidden ðYiobs Yiest Þ 100
layer, and one output layer containing a single node was Pbias ¼ i¼1
: (6)
used in the current study. A tan sigmoid transfer function X
n
ðYiobs Þ
was used to transfer the input signal to the output through i¼1
the hidden layer neurons. A linear transfer function was
used at the output node. Furthermore, this ANN-type model [44] The optimal value of Pbias is 0. Underestimation
has been used in other modeling studies involving different and overestimation of prediction bias is given by positive
8 of 24
and negative values [Gupta et al., 1999]. In equations 4–6, [48] It should be noted that the variable evaluated in the
Y obs is the measured precipitation, Y est is the estimated current analysis is precipitation at the annual time step.
precipitation, and Y mean is the long-term mean of the meas- Moriasi et al. [2007] suggested performance measures for
ured data. evaluating model performance for estimating streamflow at
[45] The model-estimated precipitation is compared a monthly time step. Precipitation exhibits greater variabili-
against the ‘‘climatology,’’ using the LEPS score. Accord- ty than streamflow, so it becomes questionable to use
ing to Potts et al. [1996] LEPS score measures the distance monthly streamflow performance measures for evaluating
between the estimated value and the observed value in annual precipitation estimates. Having said this, the selec-
terms of their cumulative probability distributions. The tion of model evaluation guidelines depends on the scope
LEPS score is defined as: and magnitude of the research problem. Stricter guidelines
are needed for projects involving significant consequences,
S ¼ 3 ð1 jPf Po j þ P2f Pf þ P2o Po Þ 1: (7) such as congressional testimony or the development of new
laws [Moriasi et al., 2007]. Moderate performance ratings
[46] In equation (7), Pf and Po are the estimated and would suffice the purpose for explanatory research or tech-
observed cumulative probabilities. The LEPS score, which nology assessment where no litigation is involved (U.S.
is computed for all the data points, shows the good and bad EPA, 2002). Furthermore, U.S. Environmental Protection
forecast years. A higher skill score is obtained when accu- Agency [2002] has indicated that the measures used to
rately forecasting values farther from the mean. Therefore, accept, reject, or qualify model results should be documented
if the forecast is for a value near the mean, then it will have before evaluating the model. Also, the annual precipitation
a lower skill (bad forecast) and a correct forecast farther used in the analysis has reduced variability compared to
from the mean will have a higher skill (good forecast). The monthly precipitation values. Therefore, the stricter perform-
average skill SK based on the LEPS score is defined as: ance measures mentioned above are adequate to evaluate the
SVM, ANN, and MLR models.
X
S
SK ¼ X 100; (8) 5. Statistical Properties of Annual Precipitation
Sm
and Its Relation With Oscillation Modes
where summation S is the sum of all years of LEPS score. [49] First, the linear correlation coefficients were com-
If S is positive, then summation Sm is the sum of the best puted between the oscillation modes and lag 1 precipitation
possible forecast, with Pf ¼ Po. If S is negative, then sum- to evaluate the persistence over time of the different oscil-
mation Sm is the sum of worst forecast computed by Pf ¼ 1 lations and precipitation for the CRB. The correlation coef-
or 0. SK ranges from –100 to 100, where a SK of 0 repre- ficients generally are computed to examine the potential
sents the climatological score or equivalently, random data. predictors [Grantz et al., 2005; Singhrattna et al., 2005].
LEPS SK is considered ‘‘good’’ if it is greater than 10 Table 3 shows linear correlation coefficient of lag 1 between
[Potts et al., 1996]. To evaluate the accuracy of the predic- the oscillation modes and annual precipitation for 17 cli-
tions, i.e., typical forecast error at each climate division, mate divisions encompassing the CRB. Significant correla-
cumulative absolute forecasts errors are computed using tion for each of the combinations is highlighted in bold, and
the nonexceedence plots. the lowest and the highest correlation values for a particular
[47] Table 2, which is abstracted from Moriasi et al. set are also presented.
[2007], shows the ratings for performance measures for
RSR, NSE, R, and Pbias for estimating streamflow at a Table 3. Correlation Coefficient Between Oscillation Modes and
monthly time step. Based on Table 2, the model perform- Annual Precipitation for 17 Climate Divisions at 90% Significance
ance can be judged satisfactory if RSR 0.70 and NSE > Levela
0.5 [Moriasi et al., 2007]. Moriasi et al. [2007] recom- Climate Division Region PDO NAO AMO ENSO
mended a value of R > 60.5 for a model performance to
be considered satisfactory; however, the current study used 1 Lower Basin 0.15 0.19 0.12 0.06b
2 Lower Basin 0.24 0.18 20.33 0.16
a much greater verification metric than the recommended 3 Lower Basin 0.12 0.18 20.33 0.10
measures and considered the model performance as satis- 4 Lower Basin 0.08 0.19 20.32 0.11
factory when R was >0.7. 5 Lower Basin 0.23 0.16 20.24 0.13
6 Lower Basin 0.13 0.17 20.30 20.18
7 Lower Basin 0.18 0.18 20.34b 0.14
Table 2. Recommended Performance Measures at Monthly Time 8 Upper Basin 0.18 0.10 20.23 0.12
Stepsa 9 Upper Basin 0.28 0.11 20.24 0.13
10 Lower Basin 0.32b 0.08 20.19 0.12
Performance Rating RSR Rb NSE jPbiasj 11 Lower Basin 0.22 0.17 20.21 20.22
12 Lower Basin 0.29 0.05 0.14 20.29b
Very good 0.0–0.50 0.85–1.0 0.75–1.0 <10 13 Upper Basin 0.27 0.04 0.06b 20.24
Good 0.51–0.60 0.81–0.85 0.65–0.75 10 and >15 14 Upper Basin 0.07b 0.01b 0.07 20.18
Satisfactory 0.61–0.70 0.71–0.80 0.51–0.65 15 and >25 15 Upper Basin 0.20 0.02 0.15 20.16
Unsatisfactory >0.70 0.70 0.50 >25 16 Upper Basin 0.16 0.05 20.19 0.15
a
17 Upper Basin 0.15 0.22b 20.18 20.19
Performance measures for RSR, NSE, and Pbias are taken directly from 90% Significant 11 8 12 7
Moriasi et al. [2007].
b a
The performance measure for R is greater than the recommended The significant correlations are shown in bold.
b
measure. The lowest and highest correlation values for each subset.
9 of 24
[50] When correlating the PDO index with the 17 climate was analyzed by means of the SVM model; the results are
divisions for a 1 year lead time, the correlation values discussed in the ensuing section.
resulted in 11 climate divisions exceeding 90% signifi-
cance, with Climate Division 10 having the highest correla- 6. Results and Discussion
tion and Climate Division 14 having the lowest correlation. [52] First, 1 year lead time precipitation estimates for the
Five of these divisions are in the Upper Basin, and the 17 climate divisions encompassing the CRB, using the SVM
remaining six are in the Lower Basin. In the case of NAO, Model I (base case), are discussed. Second, the coupled and
8 climate divisions exceeded the 90% significance, with individual effects of oscillations on annual precipitation
Division 17 having the highest correlation and Division 14 (Models II–IV) are analyzed. Lastly, the annual precipita-
having the lowest correlation. Seven of these divisions are tion estimates obtained using the SVM model is compared
in the Lower Basin, and the remaining division is in the with the ANN and MLR model estimates.
Upper Basin. When relating AMO and ENSO, Climate
Divisions 12 and 7, respectively, exceeded the 90% signifi- 6.1. Model I
cance. PDO showed statistically significant correlation [53] In Model I, all four oscillations modes are used to
with the Upper Basin precipitation, whereas AMO showed estimate annual precipitation with a lead time of 1 year.
a comparable stronger correlation with the Lower Basin Figure 4 shows the scatter plots between the measured and
precipitation. On the other hand, NAO and ENSO did not estimated annual precipitation for the 17 climate divisions
show statistically significant correlation with the CRB pre- for the period of record (1901–2008). A good match is
cipitation (Table 3). Additionally, among the predictors, as obtained between the measured and estimated annual pre-
expected, only the combinations of PDO-ENSO (60.50) cipitation for the majority of the Upper Basin climate divi-
and NAO-AMO (60.41) had significant correlations at sions compared to the Lower Basin. This is evident by the
90% significance level. sample points following the 45 bisector line for majority
[51] However, to form the basis for a skillful forecast, of the Upper Basin Climate Divisions, indicating a good
coefficient values could not be considered alone. Correla- model fit. For Climate Divisions 1, 5, 6, 11, and 12 in the
tions analysis was a first step to verify the potential predic- Lower Basin, the model does fairly well at the low values;
tors for each climate division that showed significant however, a few of the high values are scattered away from
relationships with annual precipitation in the CRB. Further- the bisector, indicating that the model was not able to cap-
more, each climate index individually and in combination ture them satisfactorily. The correlation values are greater
Figure 4. Scatterplot between measured and SVM estimated precipitation for 17 climate divisions for
Model I. Dashed diagonal line is the 45 bisector.
10 of 24
than 0.70 for all the climate divisions. Although, satisfac- Similar results were noticed for RMSE and ME error statis-
tory correlation values are achieved, it is noticed that for tics, as evident in Table 4. A Pbias of approximately 62%
some climate divisions, estimated values are not different was achieved for all the climate divisions. A lower value of
for different years; however the observed values are differ- Pbias signifies that there is less bias in the estimated values
ent, evident from the values saturated along a horizontal and majority of the estimated values are similar to the
line. This is mostly witnessed in the Lower Basin climate measured values. Although, there are some values that are
divisions. overestimated and underestimated, still the values are
[54] Based on Table 2, Figure 5 shows the spatial map of within a tolerance level having lesser bias and may be con-
the 17 climate divisions, depicting the three performance sidered good. The LEPS SK value ranged between 53%
measures of RSR, R, and NSE. Based on RSR, the model and 64% (Table 4). A LEPS SK score greater than 10%
shows good precipitation estimates for the majority of the indicates a ‘‘good’’ model. Therefore, the precipitation esti-
climate divisions (except climate divisions 13 and 14) in mates obtained using the SVM model is much better than
the Upper Basin, and satisfactory estimates (except climate the climatology.
division 1) for the climate divisions in the Lower Basin. [56] Based on Model I result, the SVM model performs
[55] Based on Figure 1c, climate divisions generating satisfactorily in capturing the variability in annual precipita-
more than 90% of the Upper Basin flow had RSR error station for a 1 year lead time. The scatter plots (Figure 4) and
tistics in the range of 0.51–0.60 (good estimates per Table 2). the spatial maps of the performance measures (Figure 5)
All the divisions in the Lower Basin, except Division 1, had show that the model produced good precipitation estimates
RSR in the satisfactory range (0.61–0.70). The RSR measure for the Upper Basin divisions generating two-thirds of the
indicated that the model performed unsatisfactorily in esti- flow; satisfactory estimates were obtained for the Lower
mating precipitation for Climate Division 1 (RSR > 0.7). Basin divisions, except for Climate Division 1.
The correlation statistics R agreed with the results of RSR, [57] Overall, annual precipitation estimates for the CRB
indicating that a good correlation (0.81 < R < 85) was are in the range of satisfactory to good for Model I at ‘‘t þ 1’’
achieved between the measured and estimated precipitation when using all the four oscillation indices. Additionally,
for the Upper Basin climate divisions and a satisfactory the estimates serve as a better predictor than the ‘‘climatol-
value (R > 0.7) for the climate divisions in the Lower ogy.’’ Although, Model I results indicated that all four
Basin. In case of NSE error statistics, the climate divisions oscillations have some influence on the hydroclimatology
that generate approximately 60% of the Upper Basin flow of the CRB, this needs to be further investigated to evaluate
had good precipitation estimates (NSE > 0.65). The the coupled and individual effects of each oscillation mode.
remaining Upper Basin divisions had satisfactory estimates Moreover, the coupled and individual impacts of the oscil-
(0.51–0.65), and the Lower Basin’s divisions were in the lations in relation to precipitation may vary within the
satisfactory range, except for Division 1. Overall, the SVM Upper and Lower Basins.
model was able to provide satisfactory estimates for all the
climate divisions, except for Division 1. It should be noted 6.2. Coupled and Individual Response of Oscillation in
that even in the case of Climate Division 1, results are only Relation to Annual Precipitation
slightly below or above the satisfactory levels, with RSR, [58] To analyze the coupled and individual response of
R, and NSE values of 0.71, 0.71 and 0.49, respectively. oscillation in relation to annual precipitation, separate
Figure 5. Spatial maps showing the range of performance measures for 17 climate divisions for Model
I: (a) RSR, (b) R, and (c) NSE.
11 of 24
Table 4. Performance Measures for SVM Model I Outputa the Upper Basin were in the range of ‘‘good’’ whereas satis-
factory estimates were obtained for the Lower Basin.
Climate
Division RMSE ME RSR R NSE Pbias SK [61] Dropping NAO and using combination of PDO,
AMO, and ENSO as input, the three error statistics showed
1 2.38 0.05 0.71 0.71 0.49 1.07 55.1 deterioration for majority of the climate division in the
2 2.03 0.01 0.63 0.80 0.60 1.01 58.1
3 3.09 0.03 0.69 0.74 0.52 0.86 57.0
Upper Basin (five); however, no significant improvements
4 3.50 0.04 0.66 0.78 0.56 0.53 61.3 were noticed for the climate divisions of the Lower Basin.
5 1.29 0.10 0.66 0.78 0.56 1.37 54.3 Dropping PDO and using combination of NAO, AMO, and
6 2.04 0.04 0.62 0.80 0.61 0.71 62.7 ENSO as inputs, significant improvement in error statistics
7 1.98 0.01 0.61 0.81 0.62 1.30 60.3 was noticed for Lower Basin climate divisions. Nine out of
8 1.53 0.01 0.57 0.83 0.67 0.50 62.6
9 1.47 0.01 0.57 0.83 0.67 0.70 59.5 ten divisions showed improvement in all the three perform-
10 1.64 0.02 0.60 0.81 0.63 0.37 57.9 ance measures (Figure 6). Precipitation estimates for the
11 1.53 0.10 0.66 0.78 0.57 2.18 53.0 Lower Basin were ‘‘good’’, compared to ‘‘satisfactory’’
12 2.14 0.02 0.59 0.83 0.65 1.61 63.8 Model I estimates.
13 1.57 0.03 0.59 0.81 0.65 0.60 61.6
14 2.57 0.03 0.64 0.78 0.58 0.16 55.9
[62] To quantify the improvement in performance meas-
15 1.29 0.03 0.63 0.79 0.59 0.70 57.3 ures, Table 5 is presented that highlights the percent change
16 1.38 0.03 0.60 0.82 0.64 0.21 63.3 in NSE error statistics compared to Model I results. Table 5
17 1.37 0.02 0.59 0.82 0.65 0.26 62.4 clearly indicated that overall the best estimates for the
a
The RMSE and ME values are in inches. The Pbias and SK are in Upper Basin were obtained using a combination of PDO,
percentage. NAO, and AMO, with six climate divisions showing
improvement compared to Model I results. By coupling
NAO, AMO, and ENSO precipitation estimates improved
SVM models (Models II–IV) were created for each climate for the Lower Basin for a 1 year lead time, with nine cli-
division. Precipitation estimates obtained using Model I are mate divisions showing improvement (Table 5). The worst
used as a baseline to compare precipitation estimates precipitation estimates for the upper basin were obtained when
obtained from Models II–IV. This will help to better under- NAO is dropped from the model whereas; dropping AMO
stand the coupled and individual responses of oscillation resulted in worst estimates for the lower basin (Table 5).
modes in relation to precipitation within the two regions of Based on Table 2, ‘‘very good’’ precipitation estimates were
the CRB. Similar to Figure 5 spatial maps depicting the obtained for the Upper Basin, and ‘‘good estimates’’ were
three performance measures RSR, R, and NSE were created obtained for the Lower Basin climate divisions. Very good
for Models II–IV. precipitations estimates within the UCRB may be helpful
[59] In Model II, one oscillation was dropped every time, for the water managers because, on average, 90% of the
and the remaining three oscillations were used to predict streamflow is generated in the Upper Basin above Lees
annual precipitation. This resulted in four models for each Ferry. These results also indicated that statistically ENSO
climate division; the results are shown in Figure 6. Drop- has a weak association with Upper Basin precipitation and
ping ENSO and using a combination of PDO, NAO, and NAO has a strong association (Table 5). Moreover, remov-
AMO as input resulted in an improvement in RSR (Figure ing PDO from the model may provide better precipitation
6a), R (Figure 6b), and NSE (Figure 6c) for five climate estimates for the Lower Basin; whereas deterioration in pre-
divisions in the Upper Basin and eight climate divisions in cipitation estimates is witnessed when AMO is dropped
the Lower Basin, compared to Model I results. The divi- from the model (Table 5).
sions showing improvement in the Upper Basin generated [63] The spatial map (Figure 6) shows the error statistics
approximately 79% of the flow in the Colorado River, based between the measured and estimated precipitation for
on Figure 1c. The climate divisions showing improvement Model II, based on the performance ratings established in
in Lower Basin encompassed the northern and central por- Table 2. Figure 7 demonstrates the capability of the pro-
tion, generating approximately 70% of the flow [Thomas, posed modeling framework in relation to capturing high
2007] in the Lower Basin. Deterioration in RSR, R, and and low annual precipitation values by using the best
NSE error statistics was noted only for Climate Division 13 Model II estimates, with PDO, NAO, and AMO indices, for
in the Upper Basin, generating approximately 2% of the the Upper Basin. The divisions selected for visualizing the
flow. Climate Division 14 showed no change in error statis- results were Climate Divisions 8, 9, and 16. These climate
tics compared to Model I results. Based on the flow contri- divisions, which contribute approximately 73% flow to the
bution, the precipitation estimates were in the range of Colorado River, indicated good-to-very good annual pre-
‘‘very good’’ for the Upper Basin and good for the Lower cipitations estimates (Figure 6). The scatterplot shows a
Basin. close match between the measured and estimated annual
[60] Dropping AMO and using the combination of PDO, precipitations (Figures 7a–7c). The majority of the points
NAO, and ENSO as input resulted in improvement in RSR, are saturated around the 45 bisector, indicating good
NSE and R for five climate divisions in the Upper Basin. model fit. Higher R value was achieved. A positive Pbias
Whereas, for the Lower Basin, RSR and NSE improved in of less than a half percent was achieved for Climate Divi-
two climate divisions and R in one climate division (Cli- sion 8, and a negative Pbias of less than a half percent was
mate Division 10). The divisions showing improvement in achieved for Climate Divisions 9 and 16, indicating very
the Upper Basin generated approximately 27% of the flow good estimates (Table 2). A smaller value of ME is
in the Colorado River. No significant improvements were achieved indicating a good match between the measured
noted for the Lower Basin climate divisions. The results for and estimated precipitation values.
12 of 24
Figure 6. Spatial maps showing the range of performance measures for 17 climate divisions for Model II:
(a) RSR, (b) R, and (c) NSE.
Table 5. Percentage Change in NSE Error Statistic for Model II and Model III Compared to Model I Resultsa
Model II Oscillations Model III Oscillations
Climate
Division 123 124 134 234 12 13 14 23 24 34
1 15.3 4.1 15.6 35.3 11.1 9.3 29.2 27.3 21.7 20.7
2 9.2 5.7 0.3 6.0 4.7 12.7 20.8 12.9 2.1 30.4
3 25.2 4.1 7.4 14.0 20.4 11.4 11.1 19.2 20.3 36.5
4 10.6 0.3 4.8 20.1 12.3 2.0 5.3 5.6 22.3 13.9
5 21.1 6.5 5.7 11.1 7.7 16.1 2.7 2.4 14.7 11.8
6 5.3 7.4 1.7 13.0 3.6 10.8 8.8 4.9 9.4 21.9
7 2.1 7.1 2.8 14.2 0.9 1.9 6.3 4.6 14.7 5.4
8 15.2 0.5 2.7 0.8 0.5 2.3 23.6 0.6 4.7 1.7
9 11.3 3.0 2.3 8.5 12.7 6.9 36.9 7.5 12.8 0.4
10 8.2 10.9 6.0 13.7 19.2 9.5 16.1 2.7 14.7 16.8
11 9.4 1.7 2.2 17.7 5.1 27.6 1.7 6.0 12.7 23.4
12 3.8 14.4 15.5 6.4 5.7 10.6 17.2 7.2 2.8 8.4
13 2.9 13.9 7.5 6.0 1.2 12.9 17.1 4.0 1.1 0.7
14 6.8 28.4 4.9 3.6 17.2 7.8 21.7 9.8 22.4 14.4
15 20.0 17.1 7.5 21.7 21.0 10.5 9.4 10.4 7.0 6.6
16 10.9 6.3 9.1 0.4 6.7 2.6 34.7 0.6 0.8 1.7
17 0.7 10.9 2.7 1.4 5.3 2.0 1.2 11.4 22.2 6.8
a
The negative sign implies decrease in NSE value. 1, 2, 3, and 4 refer to PDO, NAO, AMO, and ENSO.
13 of 24
Figure 7. (a–c) scatterplot, (d–f) box plot, and (g–i) nonexceedance probability plots between the
measured and estimated precipitation for selected climate divisions in the UCRB using PDO, NAO and
AMO oscillation indices for Model II. The horizontal dotted line in the nonexceedance probability plots
shows an error value of 10%.
[64] The variability (Figures 7d–7f) and probabilistic was noticed that for approximately 60% of the predictions,
error (Figures 7g–7i) of the estimated precipitation values, a negligible error—close to 1%—is achieved. Approxi-
compared to the measured values, also were analyzed. Fig- mately 80% predictions have a 10% error. Based on past
ures 7d–7f shows the box plot between measured and esti- research that addressed difficulties in estimating precipita-
mated annual precipitation for three climate divisions : 8, 9, tion [Bell, 1987; Olsson, 1998; Guenni and Bardossy,
and 16. The span of the box represents interquartile range 2002], an error of less than 10% for approximately 80% of
(25th–75th percentile) with horizontal line inside the box the estimates shows the robustness of the SVM approach in
indicating the median (50th percentile) value. The whiskers capturing the variability in annual precipitation in relation
extend from the box to 5th and 95th percentile on the lower to large-scale climate patterns. Furthermore, it was seen
and upper side, respectively. It was noticed that the model that a SK value of greater than 10% was achieved for the
was able to capture the extreme (high and low values) for selected divisions, indicating ‘‘skillful’’ forecasts. Overall,
Climate Divisions 9 (Figure 7e) and 16 (Figure 7f), but the model performed better in capturing the low values
missed the high values for Climate Division 8 (Figure 7d). compared to the high values.
The model performed reasonably well in representing the [66] In Model III, oscillations were dropped in pairs, and
variability exhibited by annual precipitation. This was evi- the remaining two oscillation modes were used to predict an-
dent by the similarity in the interquartile range of the meas- nual precipitation. The three performance measures are shown
ured and estimated precipitation values. Overall, the model on a spatial map in Figure 8. Based on RSR (Figure 8a),
performed satisfactorily in capturing the low values com- R (Figure 8b), and NSE (Figure 8c) error statistics, three
pared to the high values. climate divisions in the Upper Basin—14, 15, and 17—
[65] Probabilistic cumulative error between the measured showed improvements compared to the base case when
and estimated annual precipitation was computed for the using a combination of NAO and ENSO indices. These
selected divisions in the Upper Basin (Figures 7g–7i). It divisions accounted for approximately 21% of the flow in
14 of 24
Figure 8. Spatial maps showing the range of performance measures for 17 climate divisions for Model
III: (a) RSR, (b) R, and (c) NSE.
15 of 24
the Colorado River. The estimates were in the range of Basin climate divisions are obtained by using a combina-
‘‘good’’ to ‘‘very good.’’ tion of AMO and ENSO, compared to Model I results. All
[67] Additionally, coupling NAO and AMO also showed three performance measures showed improvement in the
improvement for five climate divisions in the Upper Basin, climate divisions covering the north central portion of
compared to Model I results. These divisions generate LCRB, which region generates 70% of the flow in the
approximately 27% of the flow in the Upper Basin. Overall, lower basin. The estimates for the majority of the climate
coupling NAO and AMO resulted in ‘‘good’’ precipitation divisions were in the range of ‘‘very good,’’ compared to
estimates for the majority of the Upper Basin climate divi- ‘‘satisfactory’’ estimates for Model I and ‘‘good’’ estimates
sions. For the majority of the Upper Basin climate divi- for the best combination of Model II. This implies that a
sions, all other combinations resulted in deterioration of combination of AMO and ENSO had statistically a stronger
error statistics compared to Model I results. In indicating influence on precipitation in the LCRB compared to any
that NAO has a stronger presence in the Upper Basin, other combination of indices. Additionally, it was noticed
Model III results agreed with Model II results. This was that a combination of PDO and ENSO resulted in unsatis-
evident by NAO being one of the inputs in the best Model factory estimates for the majority of the Lower Basin divi-
III estimates for the Upper Basin, based on the three error sions, as compared to the Upper Basin (Table 5).
statistics. The best predictions were obtained using a com- [69] The visual inspection of results of the best Model III
bination of NAO and AMO, whereas combination of PDO precipitation estimates using AMO and ENSO for the Lower
and ENSO resulted in worst predictions for majority of the Basin are shown in Figure 9. The divisions selected to show
upper basin climate divisions (Table 5). Overall, none of the results are Climate Divisions 2, 3, and 6. The scatterplot
the combinations resulted in better predictions for the ma- shows a match between measured and estimated precipita-
jority of the climate divisions, compared to Model II best tion for the selected climate divisions (Figures 9a–9c). The
results using PDO-NAO-AMO for the Upper Basin. majority of the points for the selected divisions follow the
[68] All the three error statistics were in agreement in 45 bisector, indicating good model fit. The sample points
indicating that best precipitation estimates for the Lower are aligned vertically along the diagonal, suggesting different
Figure 9. (a–c) scatterplot, (d–f) box plot, and (g–i) nonexceedance probability plots between the
measured and estimated precipitation for selected climate divisions in the LCRB using AMO and ENSO
oscillation indices for Model III.
16 of 24
estimated precipitation values for different years, which have few Lower Basin divisions using AMO as the sole input in
different measured precipitation values. Higher correlation is the model. It is noticed that none of the oscillations, when
achieved between measured and estimated precipitation. A used individually, resulted in improved precipitation esti-
positive Pbias ranging from 0.67% to 1.84% and smaller ME mates for CRB, as compared to Model I results. This is in
is achieved for the selected divisions indicating a good match agreement with findings from previous studies, where
between the measured and estimated precipitation. researchers showed that a qualitative understanding of the
[70] The model is able to perform well for low values, as relationship between oscillations and precipitation within the
the 5th percentile whiskers of estimated values match the CRB could be enhanced by evaluating the coupled response
measured values better compared to such high values as the of oscillation indices rather than by using an individual oscil-
95th percentile (Figures 9d–9f). The model has negligible lation mode [Kim et al., 2006, 2008; Canon et al., 2007;
error for approximately 60% of the predictions (Figures Hidalgo and Dracup, 2003; McCabe et al., 2007].
9g–9i); approximately 80% of the predictions have an error [72] To evaluate and show the robustness of the pro-
close to 10%. This is similar to the Upper Basin results and posed SVM approach, a bootstrap cross validation tech-
shows the robustness of the modified SVM modeling nique was used. Input data was resampled 100 times, and
approach in capturing the variability in precipitation in for each sample the SVM model was fitted and then tested
relation to oceanic-atmospheric oscillations. on the entire data set. The final estimated value was the
[71] In Model IV, each oscillation was used individually mean of these 100 ensembles. Figure 11 shows the time se-
to estimate precipitation for each climate division. Figure 10 ries plot between the measured, estimated and bootstrapped
shows the spatial map representing the three performance precipitation for the selected Upper Basin (Figure 11a) and
measures, i.e., RSR (Figure 10a), R (Figure 10b), and NSE Lower Basin (Figure 11b) climate divisions for the best
(Figure 10c). A satisfactory correlation (Figure 10b) between model combinations. The time series plots show that esti-
measured and estimated precipitation was achieved for ma- mated and boot strapped precipitation values were close to
jority of the Upper Basin climate divisions and also for a the measured values for the selected climate divisions. The
Figure 10. Spatial maps showing the range of performance measures for 17 climate divisions for
Model IV: (a) RSR, (b) R, and (c) NSE.
17 of 24
Figure 11. Precipitation time series plot for selected climate divisions in the (a) UCRB (using PDO,
NAO, AMO indices) and (b) LCRB (using AMO and ENSO indices) showing the measured, estimated
by SVM using actual data, and bootstrapped data.
estimated and bootstrapped precipitation followed a trend association with precipitation in the western United States,
similar to the measured precipitation. It is also evident that particularly to the CRB [Kim et al., 2008]. Our results indi-
the model missed the extremes in the early part of the 20th cate that NAO, coupled with other indices, can improve the
century (around 1910) and later for few years between precipitation estimates in the UCRB. Furthermore, Webb
1970 and 1990. Additionally, it was seen that the model et al. [2004] indicated that a combination of AMO and
performed adequately in capturing high and low precipita- PDO may help to explain more thoroughly the long-term
tion values during the severe sustained drought within the fluctuations in streamflow within the Colorado River Basin.
CRB, occurring from 2000–2008. Both the estimated and AMO usually reflects the conditions in the Atlantic Ocean
bootstrapped values were in agreement in indicating that that may affect the climate in North America [Enfield
the model does fairly well in capturing the low values com- et al., 2001]. The findings of the current study are in partial
pared to the high values. agreement with the other studies that showed that PDO and
[73] The cross validation results highlighted the robust- AMO in combination with other indices can serve as useful
ness of proposed SVM framework. The results showed predictors, to some extent, for extending lead times of dif-
that the model is stable and findings are not altered using ferent hydroclimatic variables—in this case, precipitation
different training and testing data sets. The estimated and within the Upper Basin [Hidalgo and Dracup, 2003; Webb
bootstrapped time series have a close match and exhibit et al., 2004; Kim et al., 2006, 2008; McCabe et al., 2007].
similar pattern. [75] In the case of the Lower Basin, best estimates are
[74] Overall, based on Model II–IV results and boot- obtained using the combination of AMO and ENSO (Model
strapping cross validation analysis, it can be inferred that a III). This finding is in agreement with the available litera-
combination of PDO, NAO, and AMO (Model II) has stat- ture, indicating that ENSO effects are more pronounced in
istically the strongest association with the annual precipita- the Lower Basin than in the Upper Basin [Redmond and
tion for a 1 year lead time for the majority of the UCRB. Koch, 1991; Webb and Betancourt, 1992; Kahya and Dra-
Interdecadal hydroclimatic variations in the UCRB that are cup, 1993; Piechota and Dracup, 1996]. ENSO has been
related to possible PDO influences have been investigated linked to the occurrence of floods in the Lower Basin
by Hidalgo and Dracup [2003]. Their study indicated that [Webb and Betancourt, 1992]. Additionally, Thomas
the shifts in the mean of precipitation and streamflow [2007] identified that AMO—both individually and in com-
within UCRB are related with decadal PDO changes. Simi- bination with PDO and ENSO—can explain the streamflow
lar to our findings, Hidalgo and Dracup [2003] concluded variability in the Lower Basin. Furthermore, Ellis et al.
that ENSO associations are not always consistent, and may [2010] indicated that much of drought variance in the
not be linked with the hydrologic fluctuations in UCRB. Lower Basin can be explained using AMO compared to
On the other hand, documented literature has shown the other indices. The dominance of AMO over PDO in influ-
linkages between NAO and precipitation over the Europe encing CRB droughts have become more in phase during
and the Mediterranean basin [Hurrell, 1995; Qian et al., the later half of 20th century. The three teleconnections,
2000]; however, little attention has been given to NAO’s AMO, PDO, and ENSO, can be used to explain much of
18 of 24
the hydrologic variability within the Lower Basin com- (Figure 12a) and MLR (Figure 12b) models show that for
pared to the Upper Basin [Ellis et al., 2010]. This is in par- the majority of the climate divisions, the model estimates
tial agreement with the current findings. The physical the mean of the measured data. Low correlation values
significance of the combined effect of AMO and ENSO on were obtained for all climate divisions, indicating the
the hydrologic conditions in the Lower Basin is yet to be inability of the model to capture the extremes. In general,
explored. the model estimates are parallel to the x axis instead of fol-
[76] It is evident from the literature that various features lowing the bisector line, showing poor prediction capabil-
of climate are interrelated in a complicated fashion [Webb ity. The LEPS score SK was less than 10%, for majority of
et al., 2004; McCabe et al., 2007]. Their interrelationship the climate divisions indicating an unsatisfactory forecast.
depends on the variability both in space and time. There This implies that model estimates are worse than the
may be additional, yet unidentified, factors that may contrib- ‘‘climatology’’.
ute to the interrelationship of climate patterns [California [78] The performance measures obtained from ANN and
Dep. of Water Resources, 2005; Webb et al., 2004; McCabe MLR are in the unsatisfactory range (Table 6). Although
et al., 2007]. Though many studies have demonstrated that the SVM model outperforms both the ANN and MLR mod-
ENSO and PDO are mainly teleconnnected with monthly, els, all three models perform comparatively better for the
seasonal, and annual precipitation variability in the U.S., it climate divisions within the Upper Basin compared to the
still is necessary to investigate the influence of other climate Lower Basin divisions. Similar to Model I, Model II–IV
indices [Ropelewski and Halpert, 1986; Ropelewski and (results not shown) was also created, and the results
Jones, 1987; Redmond and Koch, 1991; Kahya and Dracup, showed a better performance of SVM over both the ANN
1993; Webb et al., 2004]. and MLR models.
[79] Based on the results, it was noticed that SVM out-
6.3. Comparison of SVM With ANN and MLR Models performed both MLR and ANN models. There has been
[77] The SVM model results also were compared with sufficient evidence from other studies in different fields of
ANN and MLR model results. The scatter plots between hydrology that show the superiority of SVM over the regu-
measured and estimated precipitation for ANN and MLR lar ANN and MLR modeling approaches [Lin et al., 2009,
models, using all four oscillations indices, are shown in 2010; Ahmad et al., 2010; Kalra and Ahmad, 2009; Gill
Figure 12. Table 6 shows the comparison of different per- et al., 2006; Asefa et al., 2006; Dibike et al., 2001]. This is
formance measures between the measured and estimated because the SVM model has a better ability to generalize,
precipitation for ANN and MLR. The scatter plots for ANN relating the input to the desired output. In addition, the
Figure 12. Scatterplot between measured and estimated precipitation for (a) ANN and (b) MLR model
for 17 climate divisions for Model I. Dashed diagonal line is the 45 bisector.
19 of 24
Table 6. Comparison of Performance Measures for ANN and MLR Outputs for Model Ia
ANN MLR
Climate
Division RMSE ME RSR R NSE Pbias SK RMSE ME RSR R NSE Pbias SK
1 3.40 0.13 1.02 0.22 0.05 0.18 2.4 3.39 0.12 1.02 0.04 0.04 0.04 1.3
2 3.09 0.04 0.96 0.29 0.08 0.09 10.7 3.05 0.04 0.94 0.33 0.10 0.02 13.7
3 4.39 0.08 0.98 0.20 0.03 0.08 7.3 4.36 0.08 0.97 0.24 0.04 0.01 10.6
4 5.20 0.08 0.97 0.20 0.04 0.22 4.4 5.19 0.08 0.97 0.24 0.04 0.01 9.0
5 1.95 0.23 1.00 0.10 0.01 0.41 0.9 1.93 0.20 0.99 0.20 0.02 0.12 4.9
6 3.21 0.10 0.97 0.23 0.05 0.28 7.1 3.21 0.10 0.97 0.25 0.05 0.03 9.1
7 3.06 0.04 0.94 0.32 0.10 0.22 9.1 3.10 0.04 0.96 0.29 0.07 0.02 11.4
8 2.67 0.03 1.00 0.10 0.00 0.05 1.9 2.68 0.03 1.00 0.14 0.01 0.05 4.0
9 2.49 0.05 0.97 0.23 0.05 0.15 6.5 2.49 0.05 0.97 0.25 0.05 0.00 9.8
10 2.69 0.05 0.99 0.16 0.01 0.12 2.8 2.62 0.04 0.96 0.27 0.06 0.00 7.0
11 2.37 0.25 1.01 0.09 0.04 0.12 1.6 2.29 0.23 0.98 0.22 0.03 0.03 4.7
12 3.52 0.09 0.97 0.24 0.05 0.13 4.9 3.53 0.09 0.97 0.25 0.05 0.00 5.7
13 2.66 0.04 1.00 0.12 0.00 0.19 3.4 2.66 0.04 1.00 0.16 0.00 0.05 4.8
14 3.97 0.05 1.00 0.34 0.00 0.00 0.1 4.08 0.05 1.02 0.03 0.05 0.04 0.0
15 2.06 0.07 1.01 0.03 0.03 0.33 0.6 2.05 0.07 1.00 0.11 0.02 0.10 3.4
16 2.37 0.07 1.02 0.16 0.05 0.11 0.0 2.33 0.07 1.01 0.11 0.02 0.04 5.0
17 2.28 0.06 0.99 0.15 0.02 0.28 3.5 2.29 0.05 0.99 0.16 0.00 0.07 6.2
a
The RMSE and ME values are in inches.
optimization algorithm used in SVM is more robust than years [California Dep. of Water Resources, 2005; Regonda
the one used in regular ANN models. In case of MLR, the et al., 2006]. During ENSO years, the precipitation forecast
models are based on the assumption of normality, and can is higher in the southern part of the basin (LCRB) and has
be used efficiently to relate simple processes. In case of no skill in the headwaters that generate majority of the run-
hydrological processes where the data does not follow the off in the Colorado River [Redmond and Koch, 1991].
usual normal distribution, MLR models fail to capture the Therefore, obtaining accurate estimates of precipitation
variability. The feed-forward back-propagation algorithm within the CRB is a formidable challenge. However, the
used for ANN in the current analysis is simple and widely advancement along several scientific fronts has opened doors
used. There are other ANN architectures and activation for statistical forecast possibilities. In an attempt to address
functions [see Dibike et al., 1999] that may be able to better this challenge, we evaluated the link between individual and
capture the relationship between precipitation and ocean- coupled oceanic-atmospheric indices and temporal variabili-
atmospheric indices. An exhaustive comparison of methods ty in precipitation, and developed a data driven model to
was not the focus of our work. estimate annual precipitation with a lead time of 1 year. It
should be noted that the precipitation estimates obtained in
7. Summary and Conclusion the current study are not compared with the CPC estimates
[80] In this study we explored the association between as both have different temporal resolution. The current work
individual and coupled oceanic-atmospheric indices and contributes to the existing literature on the use of statistical
precipitation in the Colorado River Basin. We used an approaches for estimating precipitation.
AI-type model to capture the relationship between oceanic- [82] Multiple SVM models incorporating individual and
atmospheric indices and precipitation, and used this model coupled oceanic-atmospheric oscillations were developed.
to extend the lead time for precipitation estimation up to Results indicate that coupled PDO, NAO, and AMO have a
1 year. For this purpose, a modified Support Vector Machine statistically stronger association with precipitation in the
predictive framework incorporating oceanic-atmospheric oscil- Upper Basin. This combination resulted in very good pre-
lations was constructed for the 17 climate divisions encom- cipitation estimates for climate divisions that contribute
passing the Colorado River Basin. The oceanic-atmospheric approximately 67% (two-thirds) of the flow in Colorado
oscillations used in this study were PDO, NAO, AMO, and River. Coupled AMO and ENSO have a statistically stron-
ENSO. ger association with precipitation in majority of the Lower
[81] Annual precipitation within CRB is variable both at Basin. The estimates for the majority of the Lower Basin
the temporal and spatial scales. It is difficult to construct a divisions were in the range of ‘‘very good;’’ these divisions
single precipitation time series that is representative of the account for generating approximately 70% of flow in the
entire basin. For this reason, monthly time series data for Lower Basin. The findings of the SVM approach were con-
the climate divisions were used that extend more than a firmed using a bootstrapped cross validation approach. The
century in record and spatially cover the entire basin. Cur- cross validation analysis indicated that the results are stable
rently, CPC issues 3 month forecast for lead times of 0.5 to and do not change with different training and testing data
12.5 months with modest skill for 3–9 month lead time sets. Overall, the results showed that the SVM approach
based on ENSO and its indices. Forecasts are termed ‘‘skill- does better in capturing the low precipitation values com-
ful’’ if they show improvement over the long-term averages pared to the higher values. The inability of SVM model in
of the precipitation record used in the analysis. In general, estimating some high precipitation values may be attributed
they have no skill for summer precipitation during ENSO to fitting of outliers in the training phase. This drawback
years and no skill for winter precipitation during non-ENSO associated with SVM under the condition of non-Gaussian
20 of 24
outliers has been indicated in prior studies [Asefa et al.; relations to any hydrologic variable, understanding the dy-
2006; Gill et al., 2006; Twarakavi et al., 2006]. Therefore, namics becomes even more challenging. Therefore, it
removing the outliers may result in improving the model should be noted that the results presented in this study are
performance. Also, there are other SVM machines such as statistical in nature, and the physical mechanisms that drive
the simple Least Square Support Vector Machine (LS-SVM) these relationships are not fully understood at this time.
or robust LS-SVM that have shown to improve model per- [86] The results from the current research help in statisti-
formance in case of noisy data and outliers [Suykens, 2001; cally understanding the association between different oce-
Twarakavi et al., 2006]. The results also indicated that the anic-atmospheric indices and precipitation in the Upper
SVM approach performed better in capturing the interaction and Lower CRB. Using the modified SVM predictive
of oscillation indices and precipitation when compared with framework, cumulative precipitation totals for the current
ANN and MLR. year can be made available as early as 1 January of that
[83] The major contributions of this research are as fol- year. The annual precipitation values can be disaggregated
lows. First, there is no single oscillation that can be used to into seasonal or monthly resolution depending on the need
explain the climate variability within CRB. It is evident of end user [Lall and Sharma, 1996; Lall et al., 1996;
that various oscillations are interrelated and can be used in Sharma et al., 1997; Rajagopalan and Lall, 1999; Ahmad
combination to improve annual precipitation forecasting and Prashar, 2010; Kalra and Ahmad, 2011] for other
with a 1 year lead time within CRB [Webb et al., 2004; applications. The predictive framework incorporates global
Kim et al., 2008]. Second, NAO coupled with other indices climate information therefore can be potentially used in
can improve the precipitation estimates in the UCRB. This other catchments and river basins. However, the predictive
requires further attention and should be investigated. Third, potential of model will vary depending on the strength of
the modified SVM predictive framework showed statisti- connections between input climate indices and precipita-
cally significant results; conclusions drawn from this anal- tion in the region. Besides precipitation, the proposed SVM
ysis are cross validated and are not specific to any framework may be applied to estimate other hydrological
particular period. The approach uses the entire data and variables such as streamflow, groundwater levels, and soil
may be used in situation when limited data is available. moisture.
Fourth, the analysis performed is comprehensive in nature [87] Water managers have considered past climate vari-
and spatially covers the entire CRB. There have been ability over different time scales for water resources plan-
attempts to estimate precipitation within CRB, but no other ning and management [Vedwan et al., 2008]. With the
study has attempted to utilize a data driven model coupled projections of future climate being uncertain, multiple plau-
with major Pacific and Atlantic Ocean climate patterns to sible options need to be considered. Simple, robust, and
estimate annual precipitation for the entire Colorado River parsimonious statistical techniques can serve as good pre-
Basin. dictor of hydroclimatology. Overall, the SVM model used
[84] There are also some limitations associated with the in this study provides very good precipitation estimates that
current work. The study used climate division precipitation have the potential to improve water management within the
data and does not differentiate between precipitation as basin.
rainfall and snowfall. Second, the current research used the
four most common climate indices ; there are other indices, [88] Acknowledgments. This work was supported by the National
such as IPO, PNA, and predefined SST regions [Kalra Oceanic and Atmospheric Administration (NOAA-SARP) Award #
et al., 2012] that potentially can be used to improve the NA07OAR4310324 and National Science Foundation (NSF) Award
CMMI 0846952.
results. Although, the proposed modeling framework does
not suffer from the so-called ‘‘curse-of-dimensionality,’’ if
more predictors and a longer data set are involved, the References
trade-off in computation time and accuracy will need to be Ahmad, S., and D. Prashar (2010), Evaluating municipal water conserva-
considered. tion policies using a dynamic simulation model, Water Resour. Manage.,
[85] The current study was able to successfully estimate 24(13), 3371–3395.
Ahmad, S., and S. P. Simonovic (2000), System dynamics modeling of res-
precipitation using SVM approach, there is still some unex- ervoir operations for flood management, ASCE J. Comput. Civ. Eng.,
plained variability, which cannot be addressed using the 14(3), 190–198.
statistical approach. These variations were seen with a par- Ahmad, S., and S. P. Simonovic (2001), Integration of heuristic knowledge
ticular predictor combination resulting in different hydro- with analytical tools for selection of flood control measures, Can. J. Civ.
logical response among adjacent climate divisions. For Eng., 28(2), 208–221.
Ahmad, S., and S. P. Simonovic (2004), Spatial system dynamics: A new
example, a combination of NAO and ENSO had stronger approach for simulation of water resources systems, ASCE J. Comput.
associations with Climate Divisions 14 and 17 precipitation Civ. Eng., 18(4), 331–340.
but not for Climate Divisions 8 and 15. The cause for this Ahmad, S., and S. P. Simonovic (2005), An artificial neural network model
variability may be somewhat explained through Table 3, for generating hydrograph from hydro-meteorological parameters, J.
Hydrol., 315(1–4), 236–251.
which highlighted that climate indices are correlated differ- Ahmad, S., and S. P. Simonovic (2006), An intelligent decision support sys-
ently with the CRB precipitation. The indices individually tem for management of floods, Water Resour. Manage., 20(3), 391–410.
show different hydrologic response to climate divisions in Ahmad, S., A. Kalra, and H. Stephen (2010), Estimating soil moisture using
higher elevations compared to lower elevation. Also, this remote sensing data: A machine learning approach, Adv. Water Resour.,
response varies among regions having snowfall as their pri- 33(1), 69–80.
ASCE Task Committee (2000a), Artificial neural networks in hydrology I:
mary precipitation compared to rainfall. Besides climate Preliminary concepts, J. Hydrol. Eng., 5(2), 115–123.
other topographic features can also alter this response. Fur- ASCE Task Committee (2000b), Artificial neural networks in hydrology II:
thermore, when coupled response of indices is evaluated in Hydrologic applications, J. Hydrol. Eng., 5(2), 124–137.
21 of 24
Asefa, T., M. Kemblowski, M. McKee, and A. Khalil (2006), Multi-time Dracup, J. A., and E. Kahya (1994), The relationship between US stream-
scale stream flow predictions: The support vector machines approach, J. flow and La Nina, Water Resour. Res., 30(7), 2133–2141.
Hydrol., 318, 7–16. Efron, B. (1979), Bootstrap methods: Another look at the jacknife, Ann.
Ashok, K., Z. Guan, and T. Yamagata (2001), Impact of the Indian Ocean Stat., 7, 1–26.
dipole on the relationship between the Indian monsoon rainfall and ENSO, Ellis, A. W., G. B. Goodrich, and G. M. Garfin (2010), A hydroclimatic
Geophys. Res. Lett., 28(23), 4499–4502, doi:10.1029/2001GL013294. index for examining patterns of drought in the Colorado River Basin, Int.
Ashok, K., Z. Guan, and T. Yamagata (2003), Influence of the Indian Ocean J. Climatol., 30, 236–255.
Dipole on the Australian winter rainfall, Geophys. Res. Lett., 30(15), Enfield, D. B., A. M. Mestas-Nunez, and P. J. Trimble (2001), The Atlantic
1821, doi:10.1029/2003GL017926. multidecadal oscillation and its relation to rainfall and river flows in the
Ashok, K., S. K. Behera, S. A. Rao, H. Weng, and T. Yamagata (2007), El continental U.S., Geophys. Res. Lett., 28, 2077–2080.
Niño Modoki and its possible teleconnection, J. Geophys. Res., 112, Feng, S., and Q. Hu (2007), Changes in winter snowfall/precipitation ratio
C11007, doi:10.1029/2006JC003798. in the contiguous United States, J. Geophys. Res., 112, D15109,
Barlow, M., H. Cullen, and B. Lyon (2002), Drought in central and south- doi:10.1029/2007JD008397.
west Asia: La Nina, the warm pool, and Indian Ocean precipitation, J. Forsee, W. J., and S. Ahmad (2011), Evaluating urban stormwater infra-
Clim., 15(7), 697–700. structure design in response to projected climate change, J. Hydrol. Eng.,
Bell, T. L. (1987), A space-time stochastic model of rainfall for satellite 16(11), 865–873, doi:10.1061/ASCE-HE.1943-5584.0000383.
remote-sensing studies, Geophys. Res. Lett., 92(D8), 9631–9643. Geisser, S. (1975), The predictive sample reuse method with applications,
Bjerknes, J. (1966), A possible response of the atmospheric Hadley circula- J. Am. Stat. Assoc., 70, 320–328.
tion to equatorial anomalies of ocean temperature, Tellus, 18, 820–829. Gershunov, A., and T. P. Barnett (1998), Interdecadal modulation of ENSO
Boser, B., I. Guyon, and V. Vapnik (1992), A training algorithm for optimal telecommunications, Bull. Am. Meteorol. Soc., 79, 2715–2726.
margin classifiers, in Annual Conference on Computational Learning Giannini, A., M. A. Cane, and Y. Kushnir (2001), Interdecadal changes in
Theory, pp. 144–152, ACM Press, Pittsburgh, Pa. the ENSO teleconnection to the Caribbean region and the North Atlantic
Brito-Castillo, L., A. Leyva-Contreras, A. V. Douglas, and D. Lluch-Belda oscillation, J. Clim., 14, 2867–2879.
(2002), Pacific decadal oscillation and the filled capacity of dams on the riv- Gill, M. K., T. Asefa, M. Kemblowski, and M. McKee (2006), Soil mois-
ers of the Gulf of California continental watershed, Atmosfera, 15, 121–138. ture prediction using support vector machines, J. Am. Water Resour.
California Dep. of Water Resources (2005), Colorado River Basin Climate: Assoc., 42(4), 1033–1046.
Paleo, Present, Future, Spec. Publ. of Assoc. Calif. Water Agencies and Grantz, K., B. Rajagopalan, M. P. Clark, and E. A. Zagona (2005), A tech-
Colorado River Water Users Assoc. Conf., Calif. Dep. Water Res., Sacra- nique for incorporating large-scale climate information in basin-scale
mento, Calif., 80 pp. ensemble streamflow forecasts, Water Resour. Res., 41, W10410,
Cancelliere, A., D. Mauro, B. Bonaccorso, and G. Rossi (2007), Investigat- doi:10.1029/2004WR003467.
ing the potential of NAO index to forecast droughts in Sicily, in Hydrol- Guenni, L., and A. Bardossy (2002), A two steps disaggregation method for
ogy Days, edited by J. A. Ramirez, pp. 75–86, Colorado State Univ., Fort highly seasonal rainfall, Stoch. Environ. Res. Risk Assess., 16, 188–206.
Collins, Colo. Gupta, H. V., S. Sorooshian, and P. O. Yapo (1999), Status of automatic
Canon, J., J. Gonzalez, and J. Valdes (2007), Precipitation in the Colorado calibration for hydrologic models: Comparison with multilevel expert
River Basin and its low frequency associations with PDO and ENSO sig- calibration, J. Hydrol. Eng., 4(2), 135–143.
nals, J. Hydrol., 333, 252–264. Guttman, N. B., and R. G. Quayle (1996), A historical perspective of U.S.
Carrier, C., A. Kalra, and S. Ahmad (2011), Using proxy reconstructions climate divisions, Bull. Am. Meteorol. Soc., 77, 293–303.
for streamflow forecasting, in World Environmental and Water Resour- Gutzler, D. S., D. M. Kann, and C. Thornbrugh (2002), Modulation of
ces Congress: Bearing Knowledge for Sustainability, edited by R. E. ENSO-based long-lead outlooks of Southwestern U.S. winter precipitation
Beighley II and M. W. Killgore, pp. 3124–3133, Palm Springs, Calif. by the pacific decadal oscillation, Bull. Am. Meteorol. Soc., 17, 1163–1172.
Cayan, D. R., M. D. Dettinger, H. F. Diaz, and N. E. Graham (1998), Deca- Guyon, I., B. Boser, and V. Vapnik (1993), Automatic capacity tuning of
dal climate variability of precipitation over western North America, J. very large VC-dimension classifiers, in Advances in Neural Information
Clim., 11(12), 3148–3166. Processing Systems, edited by S. J. Hanson, J. D. Cowan, and C. L. Giles,
Cayan, D. R., K. T. Redmond, and L. G. Riddle (1999), ENSO and hydro- pp. 147–155, Morgan Kaufmann Publ., San Francisco, Calif.
logic extremes in the western United States, J. Clim., 12, 2881–2893. Hamlet, A. F., P. W. Mote, M. P. Clark, and D. P. Lettenmaier (2005),
Chowdhury, S., and A. Sharma (2009), Multisite seasonal forecast of arid Effects of temperature and precipitation variability on snowpack trends
river flows using a dynamic model combination approach, Water Resour. in the western United States, Am. Meteorol. Soc., 18, 4545–4561.
Res., 45, W10428, doi:10.1029/2008WR007510. Haykin, S. (2003), Neural Networks: A comprehensive foundation, 4th edi-
Colle, B. A. (2004), Sensitivity of Orographic precipitation to changing am- tion, Pearson Educ., Singapore.
bient conditions and terrain geometries: An idealized modeling perspec- Hidalgo, H. G., and J. A. Dracup (2003), ENSO and PDO effects on hydro-
tive, J. Atmos. Sci., 61, 588–606. climatic variations of the Upper Colorado River Basin, J. Hydrome-
Coulibaly, P. (2006), Spatial and temporal variability of Canadian seasonal teorol., 4, 5–23.
precipitation (1900–2000), Adv. Water Resour., 29(12), 1846–1865. Higgins, R. W., W. A. Leetmaa, Y. Xue, and A. Barnston (2000), Dominant
Coulibaly, P., F. Anctil, P. Rasmussen, and B. Bobee (2000), A recurrent neu- factors influencing the seasonal predictability of U.S. precipitation and
ral networks approach using indices of low-frequency climatic variability surface air temperature, J. Clim., 13, 3994–4017.
to forecast regional annual runoff, Hydrol. Processes, 14, 2755–2777. Hsu, K.-L., H. V. Gupta, and S. Sorooshian (1995), Artificial neural net-
Dawadi, S., and S. Ahmad (2012), Changing Climatic Conditions in the work modeling of the rainfall-runoff process, Water Resour. Res.,
Colorado River Basin: Implications for Water Resources Management, 31(10), 2517–2530.
J. Hydrol., 430–431, 127–141, doi:10.1016/j.jhydrol.2012.02.010. Hsu, K. L., X. Gao, S. Sorooshian, and H. V. Gupta (1997), Precipitation
Diaz, H. F., and G. N. Kiladis (1992), Atmospheric teleconnection associ- estimation from remotely sensed information using artificial neural net-
ated with the extreme phase of the Southern Oscillation, in El Nino: His- works, J. Appl. Meteorol., 36, 1176–1190.
torical and Paleoclimatic Aspects of the Southern Oscillation, edited by Hu, Q., and S. Feng (2001), Variations of teleconnection of ENSO and
H. F. D. A. V. Markgraf, pp. 7–28, Cambridge Univ. Press, New York. interannual variation in summer rainfall in the central United States, J.
Diaz, H. F., M. P. Hoerling, and J. K. Eischeid (2001), ENSO variability, Clim., 14, 2469–2480.
teleconnections and climate change, Int. J. Climatol., 21, 1845–1862. Hurrell, J. W. (1995), Decadal trends in the North Atlantic Oscillation: Re-
Dibike, Y. B., D. Solomatine, and M. B. Abbott (1999), On the encapsula- gional temperatures and precipitation, Science, 269(5224), 676–679.
tion of numerical hydraulic models in artificial neural network, J. Kahya, E., and J. A. Dracup (1993), U.S. streamflow patterns in relation to
Hydraul. Res., 37(2), 147–161. the El Nino/southern oscillation, Water Resour. Res., 29(8), 2491–2503,
Dibike, Y. B., S. Velickov, D. Solomatine, and M. B. Abbott (2001), Model doi:10.1029/93WR00744.
induction with support vector machines: Introduction and application, J. Kalra, A. (2012), Association of oceanic-atmospheric oscillations and
Comput. Civ. Eng., 15(3), 208–216. hydroclimatic variables in the Colorado River Basin, Ph.D. dissertation,
Dickson, R. R., T. J. Osborn, J. W. Hurrell, J. Meincke, J. Blindheim, 220 pp., Univ. of Nev., Las Vegas, Nev.
B. Adlandsvik, T. Vinje, G. Alekseev, and W. Maslowski (2000), The Kalra, A., and S. Ahmad (2009), Using oceanic-atmospheric oscillations
Arctic Ocean response to the North Atlantic oscillation, J. Clim., 13, for long lead time streamflow forecasting, Water Resour. Res., 45,
2671–2696. W03413, doi:10.1029/2008WR006855.
22 of 24
Kalra, A., and S. Ahmad (2011), Evaluating changes and estimating sea- Moriasi, D. N., J. G. Arnold, M. W. Van Liew, R. L. Bingner, R. D. Harmel,
sonal precipitation for Colorado River Basin using stochastic nonpara- and T. L. Veith (2007), Model evaluation guidelines for systematic quan-
metric disaggregation technique, Water Resour. Res., 47, W05555, tification of accuracy in watershed simulations, Trans. Am. Soc. Agric.
doi:10.1029/2010WR009118. Biol. Eng., 50(3), 885–900.
Kalra, A., W. P. Miller, K. W. Lamb, S. Ahmad, and T. Piechota (2012), Mosquera-Machado, S., and S. Ahmad (2007), Flood hazard assessment of
Using large-scale climatic patterns for improving long lead time stream- Atrato river in Colombia, Water Resour. Manage., 21(3), 591–609.
flow forecasts for Gunnison and San Juan River Basins, Hydrol. Proc- Nash, J. E., and J. V. Sutcliffe (1970), River flow forecasting through concep-
esses, doi:10.1002/hyp.9236, in press. tual models: Part 1. A discussion of principles, J. Hydrol., 10(3), 282–290.
Kane, R. P. (1999), El Nino timings and rainfall extremes in India, South- Olsson, J. (1998), Evaluation of a scaling cascade model for temporal rain-
east Asia and China, Int. J. Climatol., 19, 653–672. fall disaggregation, Hydrol. Earth Syst. Sci., 2(1), 19–30.
Karl, T. R., and R. W. Knight (1997), Secular trends of precipitation Piechota, T. C., and J. A. Dracup (1996), Drought and regional hydrologic
amount, frequency, and intensity in the United States, Bull. Am. Mete- variation in the United States: Association with the El Nino-Southern
orol. Soc., 79(2), 231–241. Oscillation, Water Resour. Res., 32(5), 1359–1373.
Karl, T. R., C. N. Williams Jr., and P. J. Young (1986), A model to estimate Potts, J. M., C. K. Folland, I. T. Jolliffe, and D. Sexton (1996), Revised
the time of observation bias associated with monthly mean maximum, ‘‘LEPS’’ scores for assessing climate model simulations and long-range
minimum, and mean temperatures for United States, J. Clim. Appl. Mete- forecasts, J. Clim., 9, 34–53.
orol., 25, 145–160. Prairie, J., and R. Callejo (2005), Natural Flow and Salt Computation Meth-
Khalil, A., M. N. Almasri, M. McKee, and J. J. Kaluarachchi (2005), ods, Calendar Years 1971–1995, U.S. Dep. Interior, Bur. of Reclam.,
Applicability of statistical learning algorithms in groundwater quality Boulder City, Nev., 112 pp.
modeling, Water Resour. Res., 41, W05010, doi:10.1029/2004WR00 Pui, A., A. Lal, and A. Sharma (2011), How does the Interdecadal Pacific
3608. Oscillation affect design floods in Australia?, Water Resour. Res., 47,
Kiem, A. S., and S. W. Franks (2004), Multi-decadal variability of drought W05554, doi:10.1029/2010WR009420.
risk, eastern Australia, Hydrol. Processes, 18(11), 2039–2050. Pulwarty, R. S., and T. S. Melis (2001), Climate extremes and adaptive
Kim, T.-W., J. B. Valdes, B. Nijssen, and D. Roncayolo (2006), Quantifica- management on the Colorado River: Lessons from the 1997–1998 ENSO
tion of linkages between large-scale climate patterns and precipitation in event, J. Environ. Manage., 63, 207–324.
the Colorado River Basin, J. Hydrol., 321, 173–186. Puri, S., H. Stephen, and S. Ahmad (2011), Relating TRMM precipitation
Kim, T.-W., C. Yao, and J.-H. Ahn (2008), Influence of climate variation radar land surface backscatter response to soil moisture in the southern
on seasonal precipitation in the Colorado River Basin, Stoch. Environ. United States, J. Hydrol., 402(1–2), 115–125.
Res. Risk Assess., 22, 411–420. Qaiser, K., S. Ahmad, W. Johnson, and J. Batista (2011), Evaluating the
Kuligowski, R. J., and A. P. Barros (1998), Experiments in short-term pre- impact of water conservation on fate of outdoor water use: A study in an
cipitation forecasting using artificial neural networks, Mon. Weather arid region, J. Environ. Manage., 92(8), 2061–2068.
Rev., 126, 470–482. Qian, B., J. Corte-real, and H. Xu (2000), Is the North Atlantic Oscillation
Kushnir, Y., V. J. Cardone, J. G. Greenwood, and M. A. Cane (1997), The the most important atmospheric pattern for precipitation in Europe, Geo-
recent increase in the North Atlantic wave heights, J. Clim., 10, 2107– phys. Res. Lett., 105, 11,901–11,910.
2113. Rajagopalan, B., and U. Lall (1998), Interannual variability in western US
Lall, U., and A. Sharma (1996), A nearest neighbor bootstrap for resam- precipitation, J. Hydrol., 210, 51–67.
pling hydrologic time series, Water Resour. Res., 32(3), 679–693. Rajagopalan, B., and U. Lall (1999), A k-nearest neighbor simulator for
Lall, U., B. Rajagopalan, and D. G. Tarboton (1996), A nonparametric wet/ daily precipitation and other weather variables, Water Resour. Res.,
dry spell for resampling daily precipitation, Water Resour. Res., 32(9), 35(10), 3089–3101.
2803–2823. Raman, H., and N. Sunilkumar (1995), Multivariate modelling of water
Lau, K. M., and H. T. Wu (2001), Principal modes of rainfall-SST variabili- resources time series using artificial neural networks, J. Hydrol. Sci.,
ty of the Asian summer monsoon: A reassessment of the Monsoon- 40(2), 145–163.
ENSO relationship, J. Clim., 14, 2480–2895. Redmond, K. T., and R. W. Koch (1991), Surface climate and streamflow
Legates, D. R., and G. J. McCabe (1999), Evaluating the use of ‘‘goodness- variability in the western United States and their relationship to large
of-fit’’ measures in hydrologic and hydroclimatic model validation, scale circulation indices, Water Resour. Res., 27, 2381–2399.
Water Resour. Res., 35(1), 233–241. Regonda, S. K., B. Rajagopalan, M. Clark, and E. Zagona (2006), A multi-
Li, P. H., H. H. Kwon, L. Sun, U. Lall, and J. J. Kao (2010), A modified sup- model ensemble forecast framework: Application to spring seasonal
port vector machine based prediction model on streamflow at the Shih- flows in the Gunnison River Basin, Water Resour. Res., 42, W09404,
men Reservoir, Taiwan, Int. J. Climatol., 30, 1256–1268. doi:10.1029/2005WR004653.
Lin, G.-F., G.-R. Chen, M.-C. Wu, and Y.-C. Chou (2009), Effective fore- Ropelewski, C. F., and M. S. Halpert (1986), North American precipitation
casting of hourly typhoon rainfall using support vector machines, Water and temperature patterns associated with El-Nino-Southern Oscillation
Resour. Res., 45, W08440, doi:10.1029/2009WR007911. (ENSO), Mon. Weather Rev., 114, 2165–2352.
Lin, G.-F., G.-R. Chen, and P.-Y. Huang (2010), Effective typhoon charac- Ropelewski, C. F., and P. D. Jones (1987), An extension of the Tahiti-
teristics and their effects on hourly reservoir inflow forecasting, Adv. Darwin Southern Oscillation Index, Am. Meteorol. Soc., 115(9), 2161–
Water Resour., 33, 887–898. 2165.
Liong, S.-Y., and C. Sivapragasam (2002), Flood stage forecasting with Sax, J. L., B. H. Thompson, J. D. Leshy, and R. H. Abrams (2000), Legal
support vector machines, J. Am. Water Resour. Assoc., 38(1), 173–186. Control of Water Resources: Cases and Materials, West Group, St. Paul,
Mantua, N. J., Y. Z. Hare, J. M. Wallace, and R. C. Francis (1997), A pa- Minn., 956 pp.
cific interdecadal climate oscillation with impacts on salmon production, Schölkopf, B., K.-K. Sung, C. J. C. Burges, F. Girosi, P. Niyogi, T. Poggio,
Bull. Am. Meteorol. Soc., 78, 1069–1079. and V. Vapnik (1997), Comparing support vector machines with gaus-
McCabe, G. J., and M. D. Dettinger (1999), Decadal variations in the sian kernels to radial basis function classifiers, IEEE Trans. Signal Pro-
strength of ENSO teleconnections with precipitation in the western cess., 45(11), 2758–2765.
United States, Int. J. Climatol., 19, 1399–1410. Sharma, A., D. G. Tarboton, and U. Lall (1997), Streamflow simulation: A
McCabe, G. J., M. A. Palecki, and J. L. Betancourt (2004), Pacific and At- nonparametric approach, Water Resour. Res., 33(2), 291–308.
lantic Ocean influences on multidecadal drought frequency in the United Shrestha, E., S. Ahmad, W. Johnson, P. Shrestha, and J. Batista (2011), Car-
States, Proc. Natl. Acad. Sci., 101(12), 4136–4141. bon footprint of water conveyance versus desalination as alternatives to
McCabe, G. J., J. L. Betancourt, and H. G. Hidalgo (2007), Associations of expand water supply, Desalination, 27(1–3), 120–127.
decadal to multidecadal sea-surface temperature variability with Upper Shrestha, E., S. Ahmad, W. Johnson, and J. Batista (2012), The carbon foot-
Colorado river flow, J. Am. Water Resour. Assoc., 43(1), 183–192. print of water management policy options, Energy Policy, 42, 201–212,
Melesse, A. M., S. Ahmad, M. E. McClain, X. Wang, and Y. H. Lim doi:10.1016/j.enpol.2011.11.074.
(2011), Suspended sediment load prediction of river systems: An artifi- Simonovic, S. P., and S. Ahmad (2005), Computer-based model for flood
cial neural network approach, Agric. Water Manage., 98(5), 855–866. evacuation emergency planning, Nat. Hazards, 34(1), 25–51.
Merideth, R. (2000), A primer on climatic variability and change in the Singh, J., H. V. Knapp, and M. Demissie (2004), Hydrologic modeling of
southwest, 28 pp., Udall Cent. for Stud. in Public Policy and the Inst. of the Iroquosis River watershed using HSPF and SWAT, Rep. ISWS CR
the Study of Plant Earth, Univ. of Ariz., Tucson, Ariz. 2004-08, Ill. Water Surv., Champaign, Ill.
23 of 24
Singh, J., H. V. Knapp, J. G. Arnold, and M. Demissie (2005), Hydrologic Vedwan, N., S. Ahmad, F. Miralles-Wilhelm, K. Broad, D. Letson, and
modeling of the Iroquois River watershed using HSPF and SWAT, J. G. Podesta (2008), Institutional evolution in Lake Okeechobee management
Am. Water Resour. Assoc., 41(2), 361–375. in Florida: Characteristics, impacts, and limitations, Water Resour.
Singhrattna, N., B. Rajagopalan, M. P. Clark, and K. K. Kumar (2005), Sea- Manage., 22(6), 699–718.
sonal forecasting of Thailand summer monsoon rainfall, Int. J. Climatol., Venkatesan, A. K., S. Ahmad, W. Johnson, and J. Batista (2011a), Salinity
25, 649–664. reduction and energy conservation in direct and indirect portable water
Smola, A. J., and B. Schölkopf (2004), A tutorial on support vector regres- reuse, Desalination, 272(1–3), 120–127.
sion, Stat. Comput., 14, 199–222. Venkatesan, A. K., S. Ahmad, W. Johnson, and J. Batista (2011b), System
Smola, A. J., B. Schölkopf, and K. R. Muller (1998), The connection dynamics model to forecast salinity load to the Colorado River due to
between regularization operators and support vector kernels, Neural Net- urbanization within the Las Vegas valley, Sci. Total Environ., 409(13),
works, 11, 637–649. 2616–2625.
Stephen, H., S. Ahmad, T. C. Piechota, and C. Tang (2010), Relating surface Viles, H. A., and A. S. Goudie (2003), Interannual, decadal and multideca-
backscatter response from TRMM precipitation radar to soil moisture: dal scale climate variability and geomorphology, Earth Sci. Rev., 61,
Results over a semi-arid region, Hydrol. Earth Syst. Sci., 14(2), 193–204. 105–131.
Stone, M. (1974), Cross-validatory choice and assessment of statistical pre- Walsh, J. E., W. L. Chapman, and T. Shy (1996), Recent decrease of sea
dictions, J. R. Stat. Soc., 36, 111–147. level pressure in the Central Arctic, J. Clim., 9, 480–486.
Suykens, J. A. K. (2001), Nonlinear modeling and support vector machines, Wang, B., R. G. Wu, and X. H. Fu (2000), Pacific-East Asia teleconnec-
in Proceedings of IEEE Instrumentation and Measurement Technology tions: How does ENSO affect East Asian climate?, J. Clim., 13, 1517–
Conference, vol. 1, pp. 287–294, Budapest, Hungary, doi:10.1109/IMTC. 1536.
2001.928828. Wang, X. L., and V. R. Swail (2001), Changes of extreme wave heights in
Thomas, B. E. (2007), Climatic fluctuations and forecasting of streamflow Northern Hemisphere oceans and related atmospheric circulation
in the Lower Colorado River Basin, J. Am. Water Resour. Assoc., 46(6), regimes, J. Clim., 14, 2204–2221.
1550–1569. Webb, R. H., and J. L. Betancourt (1992), Climate variability and flood fre-
Tokar, A. S., and P. A. Johnson (1999), Rainfall-runoff modeling using arti- quency of the Santa Cruz River, Pima County, Arizona, U.S. Geol. Surv.
ficial neural networks, J. Hydrol. Eng., 4(3), 232–239. Water Suppl. Pap., 2379, pp. 40.
Tokar, A. S., and M. Markus (2000), Precipitation-Runoff modeling using artifi- Webb, R. H., G. J. McCabe, R. Hereford, and C. Wilkowske (2004), Cli-
cial neural networks and conceptual models, J. Hydrol. Eng., 5(2), 156–161. matic fluctuations, drought, and flow in the Colorado River Basin, U.S.
Tripathi, S., V. V. Srinivas, and R. S. Nanjundiah (2006), Downscaling of Geol. Surv. Fact Sheet, 3062-04, 4 pp.
precipitation for climate change scenarios: A support vector machine Wedgbrow, C. S., R. L. Wilby, H. R. Fox, and G. O’Hare (2002), Prospects
approach, J. Hydrol., 330, 621–640. for seasonal forecasting of summer drought and low river flow anomalies
Twarakavi, N. K. C., D. Mishra, and S. Bandopadhyay (2006), Prediction in England and Wales, Int. J. Climatol., 22(2), 219–236.
of arsenic in bedrock derived stream sediments at a gold mine site under Weng, H., K. Ashok, S. K. Behera, S. A. Rao, and T. Yamagata (2007),
conditions of sparse data, Nat. Resour. Res., 15(1), 15–26. Impacts of recent El Niño Modoki on dry/wet conditions in the Pacific
Twarakavi, N. K. C., J. Simunek, and M. G. Schaap (2009), Development rim during boreal summer, Clim. Dyn., 29(2–3), 113–129, doi:10.1007/
of pedotransfer functions for estimation of soil hydraulic parameters s00382-00007-00234-00380.
using support vector machines, Soil Sci. Soc. Am. J., 73(5), 1443–1452. Wilson, L. (1973), Variations in mean annual sediment yield as a function
U.S. Environmental Protection Agency (2002), Guidance for quality assur- of mean annual precipitation, Am. J. Sci., 273, 335–349.
ance project plans for modeling, EPA QA/G-5M, Rep. EPA/240/R-02/007, Xu, Z. X., K. Takeuchi, and H. Ishidaira (2004), Correlation between El
Off. of Environ. Inf., U.S. Environ. Protect. Agency, Washington, D.C. Nino-Southern Oscillation (ENSO) and precipitation in South-east Asia
Vapnik, V. (1995), The Nature of Statistical Learning Theory, Springer, and the Pacific region, Hydrol. Process., 18, 107–123.
New York. Yu, X., and S.-Y. Liong (2007), Forecasting of hydrologic time series with
Vapnik, V. (1998), Statistical Learning Theory, Wiley, New York. ridge regression in feature space, J. Hydrol., 332, 290–302.
24 of 24

Kalra Et Al-2012-Water Resources Research

Uploaded by

Copyright:

Available Formats

Kalra Et Al-2012-Water Resources Research

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Kalra Et Al-2012-Water Resources Research

Uploaded by

Copyright:

Available Formats

WATER RESOURCES RESEARCH, VOL. 48, W06527, doi:10.

Estimating annual precipitation for the Colorado River Basin

1. Introduction 2012] have been studied by many researchers. Considering

You might also like