Expert Systems with Applications 42 (2015) 2213–2223

Contents lists available at ScienceDirect

Expert Systems with Applications

journal homepage: www.elsevier.com/locate/eswa

Performance comparison of Adoptive Neuro Fuzzy Inference System

(ANFIS) with Loading Simulation Program C++ (LSPC) model for
streamflow simulation in El Niño Southern Oscillation (ENSO)-affected
watershed
Suresh Sharma a,⇑, Puneet Srivastava b,1, Xing Fang c,2, Latif Kalin d,3
a
Civil and Environmental Engineering Program, Youngstown State University, Youngstown, OH 44555, United States
b
Biosystems Engineering Department, Auburn University, Tom E. Corley Building, Auburn, AL 36849, United States
c
Civil Engineering Department, Auburn University, 229 Harbert Engineering Center, Auburn, AL 36849, United States
d
School of Forestry and Wildlife Sciences, Auburn University, Forestry and Wildlife Building, Auburn, AL 36849, United States

a r t i c l e i n f o a b s t r a c t

Article history: Suitable selection of hydrological modeling tools and techniques for specific hydrological study is an
Available online 12 October 2014 essential step. Currently, hydrological simulation studies are relied on various physically based, con-
ceptual and data driven models. Though data driven model such as Adoptive Neuro Fuzzy Inference
Keywords: System (ANFIS) has been successfully applied for hydrologic modeling ranging from small watershed
El Niño Southern Oscillation (ENSO) scale to large river basin scale, its performance against physically based model has yet to be evalu-
Adaptive Neuro-Fuzzy Inference System ated to ensure that ANFIS are as capable as any physically based model for simulation study. This
(ANFIS)
study was conducted in Chickasaw Creek watershed, which is located in Mobile County of South
Loading Simulation Program C++ (LSPC)
Hydrologic simulation
Alabama. Since adequate rain gauge stations were not available near the watershed proximity, and
also the study area was affected with the El Niño Southern Oscillation (ENSO), the sea surface
temperature (SST) and sea level pressure (SLP) were additionally incorporated in the ANFIS model.
The research concluded that ANFIS model performance was equally comparable to a physically based
watershed model, Loading Simulation Program C++ (LSPC), especially when rain gauge stations were
not adequate. Additionally, the research concludes that ANFIS model performance was equally
comparable to that of LSPC no matter whether SST and SLP in ANFIS input vector was included
or not.
Published by Elsevier Ltd.

1. Introduction several methods of modeling including process based, conceptual

Hydrologic modeling is essential for various water resources
study including streamflow forecast, water uses scenarios, climate
change impact on water resources. Selection of suitable hydrologic
modeling approach has always been a vital issue because the suit-
ability of modeling tools and techniques depends on the various
elements. Since hydrologic simulation is a complex procedure
and associated with large number of watershed parameters,
⇑ Corresponding author. Tel.: +1 (330) 941 1741.
E-mail addresses: ssharma06@ysu.edu (S. Sharma), srivapu@auburn.edu
(P. Srivastava), xing.fang@auburn.edu (X. Fang), kalinla@auburn.edu (L. Kalin).
1
Tel.: +1 (334) 844 7426.
2
Tel.: +1 (334) 844 8778.
3
Tel.: +1 (334) 844 4671.
and data driven modeling approach have been experimented in
diverse watershed conditions and reported in various articles
(Clarke, 1973). Even though different studies in the past have
reported that neither of these approaches is considered superior
(Jayawardena, Muttil, & Lee, 2006), data-driven models such as
artificial neural network (ANN) (Cochocki & Unbehauen, 1993;
Committee, 2000; Tokar & Johnson, 1999) and the fuzzy logic
approach (Zadeh, 1965) have been widely accepted and applied
for hydrological modeling and various water resources stud-
ies (Sharma, 2012; Taheri Shahraiyni, Ghafouri, Saghafian, &
Bagheri Shouraki, 2013; Tayfur, Ozdemir, & Singh, 2003; Tayfur &
Singh, 2006) due to their simplicity and user friendly
nature.
Over the years, researchers have found limitations of the
conventionally adopted data-driven models as well. Therefore,

http://dx.doi.org/10.1016/j.eswa.2014.09.062
0957-4174/Published by Elsevier Ltd.
2214 S. Sharma et al. / Expert Systems with Applications 42 (2015) 2213–2223
researchers have started combining both techniques to overcome
the limitations of individual models, and hence develop powerful
intelligent systems. As a matter of fact, Neuro-Fuzzy system such
as Adaptive Neuro-Fuzzy Inference System (ANFIS) (Jang, 1993)
has been evolved and widely used.
Nevertheless, the hydrologic modeling in ANFIS also relies on
the accuracy of input data, especially precipitation. Even though
precipitation is the most sensitive input for hydrological modeling
(Trenberth & Shea, 1987), majority of the hydrological models suf-
fer from the adequate representation of spatial variability of the
precipitation data, as the sufficient network of rain gauge locations
needed for hydrological modeling are not frequently available
(Sharma, Isik, Srivastava, & Kalin, 2012). In these circumstances,
even a highly advanced, physically based model, which are capable
to represent the watershed complexity in terms of land use, slope
and soil may not be able to simulate streamflow with satisfactory
model performance. Therefore, evaluations of proper modeling
approach for a watershed with inadequate rain gauge stations
are needed.
Although several studies in the past compared the performance
of physically based models with data driven models in various
watershed conditions (Makkeasorn, Chang, & Zhou, 2008; Morid,
Gosain, & Keshari, 2002; Talei, Chua, & Quek, 2010; Wang,
Skahill, Samaitis, & Johnston, 2002), only few studies have been
conducted in a watershed characterized with limited rain gauge
stations. In order to better represent the precipitation in the model,
few studies in the past have been conducted to surrogate the lack
of precipitation data by using sea surface temperature (SST) and
sea level pressure (SLP) of the equatorial pacific in data driven
models (Khalil, McKee, Kemblowski, & Asefa, 2005; Sharma,
2012; Tech & Center, 2009). The Changes in SST and SLP in the
equatorial pacific has a potential to bring local and global climate
variation (Ropelewski & Halpert, 1986). Since El Niño Southern
Oscillation (ENSO), which is measured in terms of SST and SLP,
has significant potential teleconnection with temperature, precip-
itation and streamflow in various regions (Barsugli, Whitaker,
Loughe, Sardeshmukh, & Toth, 1999; Chiew, Piechota, Dracup, &
McMahon, 1998; Hansen, Jones, Irmak, & Royce, 2001; Keener,
Ingram, Jacobson, & Jones, 2007; Kulkarni, 2000; McCabe &
Dettinger, 1999; Pascual, Rodó, Ellner, Colwell, & Bouma, 2000;
Piechota & Dracup, 1996; Rajagopalan & Lall, 1998; Roy, 2006),
SST and SLP can be directly applied in data driven models for
streamflow simulation in ENSO affected watershed if the rain
gauge stations in watershed proximity are scarce (Sharma,
Srivastava, Fang, & Kalin, in press). However, conceptual or physi-
cally based watershed model are not capable to incorporate SST
and SLP as inputs in their input vectors. Since the data driven
model provides a unique opportunity for additional fusion of vari-
ables, several variables including SST, SLP and trade wind index,
which are responsible to bring precipitation variation, can be addi-
tionally utilized as model inputs in data-driven models. This is
especially true for ENSO-affected watersheds. Therefore, additional
forcing of these variables may be beneficial particularly for
watersheds which are significantly affected by ENSO, and also lack
adequate rain gauge stations. In this context, this study was unique
from various past studies in two ways; (i) it compared the two
modeling approach particularly in ENSO affected watershed with
inadequate rain gauge stations, (ii) it utilized the data from
equatorial pacific to develop data-driven model, and compared
its performance with physically based model. The objective of
this research was to incorporate SST, SLP in ANFIS model, and eval-
uate the ANFIS model performance against a physically based,
watershed model, Loading Simulation Program C++ (LSPC), for
streamflow simulations in Chickasaw Creek watershed, which
was affected by ENSO and characterized with inadequate rain
gauge stations.
2. Theoretical considerations
ENSO is a coupled oceanic and atmospheric phenomenon which
operates at interannual time scales resulting due to the complex
interaction of oceanic (oceanic temperature and oceanic currents)
and atmospheric (cloud, storms and winds) phenomenon
(Kessler, 2002). ENSO is induced due to the sea surface tempera-
ture gradient and sea level pressure difference along the equatorial
pacific. Various ENSO indicators such as SST and SLP are used to
measure the ENSO Phenomenon.
2.1. ANFIS
Recently, scientists are more interested in using combined
approaches, such as ANFIS which combines artificial neural net-
work with fuzzy logic approaches. Due to its capability of combin-
ing the qualitative aspects of a fuzzy system with the quantitative
aspect of a neural network, ANFIS has been found to be a more effi-
cient modeling tool than the two independent models (i.e., ANN
and fuzzy logic) to capture inherent non-linear processes (Jang,
1993). Hence, this model has been extensively applied in hydrolog-
ical (Mukerji, Chatterjee, & Raghuwanshi, 2009; Pramanik & Panda,
2009) and water quality modeling (Yan, Zou, & Wang, 2010).
ANFIS is a multi-layer and feed-forward network, that is, the
network is constructed in such a way that the nodes are not con-
nected to the same layer but connected to the next layer which
finds relationship of an input vector to an output layer. A standard
ANFIS model structure using two inputs and one output is shown
in the Fig. 1, which shows five layers with two rules and two mem-
bership functions (MFs) associated with each input.
The ANFIS model consisting two fuzzy if-then rules can be writ-
ten as follows:
Rule 1 : If X is u1 and Y is v 1; then f 1 ¼ p1 X þ q1 Y þ r 1 ð1Þ
Rule 2 : If X is u2 and Y is v 2; then f 2 ¼ p2 X þ q2 Y þ r2 ð2Þ
where u and v are the MFs for input X and Y, respectively. Similarly,
p1, q1, r1 and p2, q2, r2 are the parameters which are needed to be
ascertained for the output function. The operation of ANFIS model
from layer 1 to layer 5 is briefly borrowed from Sharma et al. (in
press) and presented here.
Layer 1 In this layer, input is given from each node and exter-
nal signal is passed to the next layer.
Layer 2 In this layer, every node is termed as a member-
ship function. For example, lui ðXÞ represents the
membership function for input X which varies from
0 to 1.
Layer 3 In this layer, the incoming signals are multiplied and
the product is forwarded to the next layer. Each out-
put from a node characterizes a result of the prede-
cessor (firing strength) of that rule. The outputs of
this layer (O3i) can be written as follows.
O3i ¼ Z i ¼ lui ðXÞlv i ðYÞ; forðiÞ ¼ 1; 2 ð3Þ
where u and v represents the membership function for
X and Y inputs, respectively.
Layer 4 In this layer, normalized firing strength is determined
in each node using following equation.
zi
O4i ¼ Z i ¼ ð4Þ
z1 þ z2
where i = 1, 2.
Layer 5 In this layer, the following equation is used to com-
pute the model output using the input contribution
of each ith rule.
 S. Sharma et al. / Expert Systems with Applications 42 (2015) 2213–2223
Fig. 1. ANFIS architecture for two inputs (X and Y) with two rules and two membership functions (P1, P2 and Q1, Q2) for each rule (Sharma et al., in press).
X
2 as the center of the cluster. For a given set of data points from
O5i ¼ zi f i ¼ zi ðpi X þ qi Y þ r i Þ
i¼1
where zi is the output from layer 3 and [pi, qi, ri] are
the model parameters.
Layer 6 In this layer, the final output is determined by com-
bining entire receiving signals
X
2 P2
zi f
O6i ¼ zi f i ¼ Pi 2 i :
i 1 zi
Finally, ANFIS model uses hybrid algorithm to compute the
model parameters. To effectively partition the data and minimize
the fuzzy rules, various clustering approach have been recom-
mended. Two widely used fuzzy clustering method such as
subtractive clustering algorithm (SCA) (Chiu, 1996) and Fuzzy
C-Mean (FCM) (Dunn, 1973) for structure identification has been
described in the following section.
2.1.1. Subtractive and Fuzzy C-Mean clustering
For large and good quality data sets covering the wide varia-
tions in the feature space, fuzzy systems can have better general-
ization due to the use of more fuzzy rules. However, for a small
range of data sets, which is more common in streamflow data sets,
large number of rules cannot be derived from the training data as it
may easily over fit the system, and rule out the possibility of
generalization. Hence, clustering approach is required for both
the effective partition of the input space and the reduction of the
number of rules. Several clustering methods have been suggested
to organize the data and construct the rules. Some of the widely
used approaches are grid partition (Jang & Sun, 1995), SCA, and
FCM. In this analysis, both FCM and SCA were experimented.
FCM is a technique in which each data point is categorized into a
cluster based on membership function (Thirumalaiah & Deo,
2000). It divides the data from multidimensional space into a
particular cluster number. Cluster center and the membership
function were decided through the repeated iteration by minimiz-
ing the objective function which was defined as a distance between
data point and cluster center. SCA is more suitable when users do
not have a clear idea of cluster number required to be used for a
given data sets. This algorithm is based on the amount of the den-
sity points in the data space, and the approach is to find areas in
the data space having higher densities of data points. The algo-
rithm tries to find the point with the highest number of neighbors
2215
ð5Þ X1, X2, . . .. . . to . . . Xn, the density measure at a data point Xj is given
by the following equation:
X
n
kX j  X i k2
dj ¼ e  2 ð7Þ
ra
j¼1
2
where ra is a data cluster radius. The equation indicates that those
ð6Þ data points which have high neighboring data points will be consid-
ered as a cluster center.
2.2. LSPC model
The LSPC, a C++ version of the widely used Hydrologic Simula-
tion Program-Fortran (HSPF) (Bicknell et al., 2001), is a watershed
model for simulating streamflow and stream hydraulics. Though
the LSPC’s algorithms are identical to those of the HSPF, LSPC
model is more efficient, flexible and have been used for streamflow
and water quality simulation (e.g. Sharma, Srivastava, Kalin, Fang,
& Elias, 2014). The model has been considered as one of the most
advanced hydrologic and watershed loading model. The hydrologic
portion of the model is based on the Stanford Watershed Model
(Crawford & Linsley, 1966). The hydrologic process in the model
is conceptualized with the schematic diagram as shown in Fig. 2
(Tech & Center, 2009). The diagram shows the water budget dia-
gram with the water interchanging process from surface to subsur-
face including three flow path; surface, interflow and groundwater
flow.
Precipitation is the input to the system (watershed) which is
partitioned at the first decision node for the interception and water
reaching on the land surface using the parameter CEPC (Fig. 2). In
the next decision node, surface and lower zone storage is divided
using the parameter, INFILT. The surface storage contributes to
the three components: (i) upper zone storage which partly contrib-
utes to lower zone, partly to the groundwater/inactive groundwa-
ter storage, and rest to the evapotranspiration (ET), (ii) interflow
storage, which eventually contributes to the channel flow,
(iii) overland flow. The lower zone storage may partly contribute
to groundwater storage or inactive groundwater storage (deep per-
colation) or partly lost through ET. The contribution from upper
zone and lower zone storage to groundwater is divided into active
and inactive groundwater by using a coefficient (DEEPFR). Active
groundwater storage can contribute to streamflow as base flow
2216 S. Sharma et al. / Expert Systems with Applications 42 (2015) 2213–2223

Fig. 2. Schematic diagram of the Stanford Watershed Model adapted for LSPC model. Number corresponding to each circle denotes the order of removing water to satisfy ET.
(Source: Tech & Center, 2009). INFILT, infiltration parameter; INTFW, interflow parameter; CEPSC, interception storage capacity; LZSN, lower zone storage capacity; UZSN,
upper zone storage capacity; NSUR, Manning’s surface roughness; LSUR, surface runoff length; SLSUR, surface slope; IRC, interflow recession constant; AGWETP; active
groundwater ET parameter; AGWRC; active groundwater recession; BASETP; base flow ET parameter; DEEPFR, fraction to inactive groundwater; LZETP, lower zone ET
parameter.

and losses as ET. ET can occur from the entire zone except inactive
groundwater zone. The overland flow, interflow, and base flows
eventually contribute to streamflow.
3. Study site
This research was conducted in Chickasaw Creek watershed
(Fig. 3) which is located in Mobile County of Southern Alabama
in the Mobile River Basin of Southeast USA. The watershed is pri-
marily in coastal region but starts at Citronelle in North, drains into
Mobile Bay in South, and ultimately discharges to the Gulf of
Mexico. The watershed drainage area is 714 km2, and is character-
ized by dense residential, forest, agricultural and industrial land
uses. The characteristics of the watershed have been dominated
by the coastal plain geology with highest elevation range of
13.11 m and lowest 0 m above the mean sea level. The region is
characterized with higher rainfall, i.e. the average annual precipita-
tion in the watershed is fairly higher (1651 mm) than average
annual precipitation of the Alabama.
4. Data
The climatic data sets needed for model development including
Air temperature (T) and precipitation (PCP) were available from
National Climatic Data Center (NCDC). Currently, there is no rain
gauge station with in the watershed boundary. The Mobile Regio-
nal Airport (Coop ID-015478) is the closest NCDC weather station
for the watershed, which is 32 km farther from the watershed
boundary. The long term observed streamflow data needed for
hydrologic modeling were available at the USGS gage (02471001)
which drains 357 km2 of the watershed.
The precipitation and temperature of this region is associated
with ENSO, characterized with fairly higher precipitation and
lower temperature in El Niño winter, and just opposite pattern in
La Niña winter (Sharma, Srivastava, Fang, & Kalin, 2012). The long
term recorded data starting since 1950 including SST, SLP in Tahiti
and Darwin (Kessler, 2002) and trade wind index (Lukas, Hayes, &
Wyrtki, 1984), were downloaded from NOAA’s NCDC and National
Center for Environment Prediction (NCEP). These data corresponds
to the Niño 3.4 region (5°N to 5°S, 120°W to 170°W) (Trenberth &
Stepaniak, 2001) located in the equatorial pacific and presented in
Table 1.
5. Methodology
5.1. Parameter estimation
Using the generic model structure as described in Fig. 1, the
ANFIS model was developed considering several other inputs. The
model development is borrowed from Sharma et al. (in press) and
briefly presented in this article having two inputs with two rules
for illustration. In fact, this model has been extended for numerous
input variables and implemented in our research. In this study,
streamflow (SF) is assumed to be a function of SST, SLP, temperature
(T) and precipitation (PCP). That is, SF = f (SST, SLP, T, PCP). Consider-
ing u1:u2 and v1:v2 were the corresponding membership functions
for two input variables (T and PCP) and a1:b1, a2:b2 were the asso-
ciated model parameters, model output can be determined as linear
combinations of consequential parameters. The following equation
can predict streamflow (SFi) as model output for a given time t1.
SFi ¼ ðw1 TÞa1 þ ðw1 PCPÞb1 þ w1 r1 þ ðw2 TÞa2 þ ðw2 PCPÞb2 þ w2 r2
ð9Þ
 S. Sharma et al. / Expert Systems with Applications 42 (2015) 2213–2223 2217

Fig. 3. Chickasaw Creek watershed in Mobile County of South Alabama, USA, showing the USGS gaging station and land use distribution within the watershed. The land use
classification is as per NLCD land use categories.

Table 1 2 3
2 3 SF 1
Data used for the study with their sources and format (Sharma et al., in press). a1 6 SF 7
6b 7 6 27
Data Source Additional 6 17 6 7
6 7 6 SF 3 7
information 6 r1 7 6 7
6 7
Land use, NLCD 2001 www.AlabamaView.org O¼6 7
6 a 7; and Z ¼ 6 : 7;
6 27 6 7
Soil Soildatamart.nrcs.usda.gov SSURGO soil 6 7 6 : 7
4 b2 5 6 7
data base 6 7
DEM www.seamless.usgs.gov, 10-m 4 : 5
r2
www.datagateway.nrcs.usda.gov resolution SF n
Weather gage station NOAA, National Climatic Data Coop ID-
Center, http://ncdc.noaa.gov 015478
Streamflow www.al.water.usgs.gov (USGS gage
02471001) AO ¼ Z ð10Þ
Hydrologic (stream www.aces.edu/waterquality/gisdata
network) data
GIS layers/tiger files www.AlabamaView.org where O is a matrix of unknown model parameters which can be
determined by solving the following equation.

1
O ¼ ðAT AÞ AT Z ð11Þ
Equations to predict the streamflow time series (SF1, SF2, . . . , SFn)
can be arranged in matrix form as follows:
2 3 where A1 and AT are the inverse and the transpose of the matrix
w1 T 1 w1 PCP 1 w1 w2 T 1 w2 PCP 1 w2
6 7 ‘‘A’’, respectively. ANFIS uses hybrid algorithm to solve this matrix.
6 w1 T 2 w1 PCP 2 w1 w2 T 2 w2 PCP 2 w2 7 Hybrid algorithm (Jang, Sun, & Mizutani, 1997) is a combination of
6 7
A¼6
6 : : : : : : 7
7 least-square method and back propagation algorithm which trains
6 7 the data fairly rapid with speedy convergence. Readers can find
4 : : : : : : 5
w1 T n w1 PCP n w1 w2 T n w2 PCP n w2 detail mathematical description of the hybrid approach in various
articles (Jang, 1993).
2218 S. Sharma et al. / Expert Systems with Applications 42 (2015) 2213–2223
5.2. Input data
Appropriate selections of suitable membership function with
proper number of fuzzy rules were essential for ANFIS model
development. Selection of appropriate and optimum number of
inputs is needed in fuzzy inference system for effective mapping
of the input and output relationship. Therefore, correlations of
several relevant inputs with the outputs were examined.
Since streamflow is a collective form of quick response (surface
runoff) and the slow response (base flow) of the watershed, it is
always essential to separate the streamflow into surface runoff
(SR) and baseflow (BF) components in order to adequately repre-
sent the watershed characteristics. Hence, we used baseflow filter
program (Lim et al., 2005) to partition the streamflow. There are
many reasons of baseflow separation. For example, BF is linked
with the precipitation of previous time steps whereas SR is entirely
contributed by precipitation of the current time step. Recursive
Digital Filter method was utilized to develop the separate SR and
BF data sets and corresponding models from given streamflow
data. The input variables for SR and BF models are listed in Table 2.
The model inputs for SR were temperature, SST, SLP and ‘‘estimated
surface runoff’’ at different time step (Table 2), where ‘‘estimated
surface runoff (SRE)’’ was determined using SCS curve number
equation (Bosznay, 1989). The parameter needed for SCS curve
number equation including curve number and hydrological soil
groups were estimated based on the 2001 land cover data set
(NLCD, 2010), and the SSURGO soil (SSURGO, 2010) database,
respectively. However, the model inputs for BF model were simply
SST, SLP, temperature and precipitation of various time step
(Table 2).
In fact, Optimum sets of best inputs were decided after evaluat-
ing the parameter significant test in multiple linear regressions
(MLR) model. Since precipitation and temperature were affected
by SST and SLP, we wanted to evaluate whether all these inputs
at various lead times were statistically significant (P-value <0.1).
The analysis depicted that both SST and SLP were statistically
significant for the seasons, when ENSO had a distinct signature
with streamflows especially for winter and spring. Inputs having
no significant contribution for the improvement of model perfor-
mance were ignored.
5.3. ANFIS model development
After investigating optimum inputs, ANFIS models were devel-
oped for SR as well as BF. In addition, ANFIS models, without SST
and SLP in their input vectors, were developed to identify the
differences in model simulation with and without SST and SLP.
Since we wanted to evaluate model performance in daily scale,
ANFIS model was extended to a daily scale after experimenting
with further input datasets. We did not find considerable differ-
ence in inputs for SR model whether we simulate in monthly or
daily scale. However, inputs for base flow in daily scale were differ-
ent to that of monthly scale. This is mainly because the current
level of base flow may rely on the precipitation pattern that was
encountered immediately few days or week ahead due to delay
in groundwater contribution. Therefore, the best sets of inputs
Table 2
Input variables for ANFIS model (monthly data) (Sharma et al., in press).
Model Model inputs
SR Model SST(t-1) SLP(t) T(t-1) T(t) SRE(t-1) SRE(t)
BF Model SST(t) SLP(t) T(t-1) P(t-1) P(t)
Note: SRE, T(t), P(t), and SST(t), SLP(t) denote ‘‘estimated surface runoff’’, temper-
ature, precipitation, sea surface temperature, and sea level pressure difference
between Tahiti and Darwin at the month ‘‘t’’, respectively.
for baseflow simulation in daily and monthly simulation were
different.
The ANFIS models were developed using 50 years of observed
data. The data were segregated into model training (80%), valida-
tion (10%) and testing (10%). Initially, both FCM and SCA were
experimented for clustering the data. Finally, we found that SCA
was more suitable to cluster the given datasets. Parameters needed
for SCA clustering including the membership function, range of
influence of the cluster center, acceptance and rejection ratio
were determined through a recurrent trial and error method by
minimizing the root mean square error and maximizing R2.
Similarly, suitable numbers of epochs required for ANFIS model
training were optimally decided using trial and error from the
visual inspection of error curve for training and validation. The
overtraining of the data was eluded by evaluating its performance
against checking data sets and controlling proper epoch number.
5.4. Performance evaluation
The performance of ANFIS model was evaluated through R2
(coefficient of determination), Nash–Sutcliffe efficiency, NSE
(Nash & Sutcliffe, 1970), mass balance error (MBE), and the root
mean square error (RMSE). The detail descriptions about the
modeling protocols (Engel, Storm, White, Arnold, & Arabi, 2007)
statistical parameters to measure the model performance are
available in many articles (Moriasi et al., 2007). Additionally, the
performance of the ANFIS model was further compared with
another independent watershed model, LSPC (Shen, Parker, &
Riverson, 2005; Tech & Center, 2009).
5.5. Comparison with LSPC model
The modeling paradigms in two models, ANFIS and LSPC, are
completely different. ANFIS is based on the learning approach,
which is accomplished through model training between the known
set of association between inputs and output data sets, whereas,
LSPC is based on physical processes related with rainfall input
and its transformation into runoff at the watershed outlet. The
brief description of the LSPC model application in Chickasaw Creek
watershed is explained in the following section.
5.6. LSPC model configuration
Hydrological processes in LSPC model are governed by certain
hydrological parameters. Watershed parameters such as length,
slope, stream network and sub-watershed area were determined
through the interactive watershed delineation using high resolu-
tion (10 m) Digital Elevation Model (DEM) (DEM, 2010). Land use
and soil-related parameters were extracted using Land cover
data set of year 2001 (NLCD, 2010) and high resolution soil
data (SSURGO) (SSURGO, 2010), respectively. LSPC model was
configured to simulate a series of hydrologically connected sub-
watersheds, with defined geometry, soil and land use characteris-
tics. Each sub watershed area contributed runoff to their respective
reach, where the cumulative flow was routed downstream to the
watershed outlet. LSPC model utilizes different sets of hydrologic
parameters for surface and subsurface hydrologic analyses in dif-
ferent sub watersheds based on the soil type and land use catego-
ries. Chickasaw Creek watershed was predominantly characterized
by hydrological soil groups A, B and D. The watershed land cover
was dominated by forest. The land use was classified as low,
medium, and industrial urban (13%), deciduous, evergreen and
mixed forest (47.4%), woody and herbaceous wetland (18.6%),
range shrubland, grassland herbaceous and hay (19.4%), and the
rest as water, south western range and agricultural land (1.6%).
 S. Sharma et al. / Expert Systems with Applications 42 (2015) 2213–2223 2219

The information about the land use data, soil data and USGS gage is Table 4
reported in Table 2. Performance of the ANFIS model in different modeling stages (monthly simulation
from 1952–2002) (Sharma et al. (in press)).

5.7. Hydrologic model calibration and validation Statistics Streamflow simulation

Training Validation Testing
The streamflow calibrations were carried out using the (1952–1989) (1990–1996) (1997–2002)

observed streamflow recorded from 1/1/1990 to 12/31/1996 at a
USGS gage station (station ID02471001). Since 2001 land cover
data sets were used, the calibrated model parameters were applied
to an independent time period (1/1/1997–12/31/2002) for model
validation. This period was selected as a validation period to
evaluate the model performance for the latest land use condition.
The simulation was started from 1/1/1985, permitting long spin
up period for the model in order to minimize the effect of unknown
initial moisture conditions, and to stabilize the hydrological
conditions.
The watershed parameters, difficult to measure, were calibrated
within a physically possible range, through the repeated trial and
error procedure until the simulated streamflows closely approxi-
mated with the observed streamflows. The model parameters
adjusted during hydrologic model calibration are presented in
Table 3. Once the model was calibrated and validated, the model
was run for another 40 years to simulate the long term streamflow.
LPSC was run in hourly time step, and simulated flows were
aggregated into daily and monthly time scale for model compari-
sons by computing corresponding statistical error parameters.
Once the LSPC model was simulated for long term, we com-
pared the performance of LSPC model with ANFIS in different ENSO
events. ANFIS models was compared with LSPC model in two ways.
In the first step, ANFIS model was developed using long term
dataset starting since 1950, and its performance was compared
with the same consecutive period of LSPC model calibration and
validation. In other words, ANFIS model was developed from the
long term datasets but ANFIS simulated result from 1990 to
2002, which is a period of LSPC model calibration and validation
was compared with LSPC model performance. In the second step,
ANFIS model was developed with same period of datasets from
1990 to 2002, a period of LSPC model calibration and validation.
This was mainly because we wanted to see the response of ANFIS
model for shorter-term datasets and make realistic comparison
with the LSPC model using datasets of the same period. In addition,
we wanted to test the performance of ANFIS model without SST
and SLP. For this, we removed the SST and SLP from its input
vector. The model performance excluding SST and SLP was com-
pared with the model performance including SST and SLP.
6. Result and discussion
6.1. Model performance
6.1.1. ANFIS
Table 4 shows the ANFIS model performance for streamflow
simulation (combined SR and BF) compared with observed data.
The ANFIS model performance was consistently better in all stages
RMSE (m3/s) 2.63 2.37 3.7
NSE 0.8 0.85 0.74
MBE 0% 3.8% 8.8%
Table 5
Model simulation at daily scale and performance comparison (1990–1996).
Statistics LSPC ANFIS
RMSE (m3/s) 9.9 7.0
NSE 0.49 0.68
MBE 2% 0.6%
of training, validation, and testing with reasonable accuracy
(Fig. 4), as indicated through the visual inspection of the observed
and simulated streamflow time series. Additionally, the ANFIS
model was simulated to a daily scale, separately for SR and BF.
Table 5 shows the ANFIS model performance for combined SF
simulation in a daily scale. Our analysis indicates that that SLP,
SST and SST(t-1) were the sensitive inputs in daily scale as well.
6.1.2. LSPC
The performance of the LSPC model was promising for monthly
simulation and satisfactory for daily simulation demonstrating
that hydrological parameters were able to capture the dynamics
of the system. The following section further compares the perfor-
mance of the LSPC model with the ANFIS Model.
6.2. Model comparison
Model results after calibration were assessed through the
comparison of simulated and observed streamflow in terms of
water budget, storm flow comparison with respect to volume
and peak, low flow and high flow period (Table 6), etc. The statis-
tical parameters, such as, NSE, R2 and MBE are listed in Table 7. The
comparisons of observed and simulated streamflow from cali-
brated and validated LSPC are shown in Fig. 5 and Fig. 6, respec-
tively. LSPC model performance during calibration and validation
period is also compared with the ANFIS model (Figs. 5 and 6).
Even though ANFIS model was developed using long term data-
sets from 1950, the two model performance was compared for
the same period of model simulation. That is, we used ANFIS sim-
ulated streamflow from 1990 to 2002 to compare with the LSPC
model calibration (1/1/1990–12/31/1996), and model validation
(1/1/1997–12/31/2002) period. The model comparisons in those
consecutive periods suggested that the performance of the ANFIS
model was comparable to the performance of the LSPC model
(Table 7).

Table 3
The calibrated LSPC model parameters.

Parameter Description Units Typical Calibrated Remarks

value
LZSN Lower zone nominal soil moisture Inches 2.0–15.0 9 An increase will provide more chances for ET and decreases
storage flow
1
INFILT Index to infiltration capacity Inches hr 0.001–0.50 0.01–0.08 An increase will cause shift of surface runoff to base flow
KVARY Variable ground water recession Inch1 0.0–0.5 0.25 An increase will cause quick ground water depletion
AGWRC Base ground water recession None 0.850– 0.992–0.998 An increase will cause the flattened base recession
0.999
2220 S. Sharma et al. / Expert Systems with Applications 42 (2015) 2213–2223

Table 6
Performance comparison of two models at monthly time scale (period 1990–1996).

Errors (simulated-observed) ANFIS LSPC Recommended criteria*

Error statistics (%) Error statistics (%)
Error in total volume 0.18 1.82 10
Error in 50% lowest flows 19.79 39.56 10
Error in 10% highest flows 18.64 26.65 15
Seasonal volume error – summer 20.63 12.95 30
Seasonal volume error – fall 28.79 2.35 30
Seasonal volume error – winter 9.99 5.58 30
Seasonal volume error – spring 13.84 6.19 30
Error in storm volumes: 0.65 54.82 20
Error in summer storm volumes 80.93 72.54 50
*
Kim, Benham, Brannan, Zeckoski, and Doherty (2007).

60
Observed flows ANFIS-Simulated flows

40

20

0
Feb-52

Aug-54

Feb-57

Aug-59

Feb-62

Aug-64

Feb-67

Aug-69

Feb-72

Aug-74

Feb-77

Aug-79

Feb-82

Aug-84

Feb-87

Aug-89

Feb-92

Aug-94

Feb-97

Aug-99

Feb-02
Fig. 4. Monthly streamflow simulated for 52 years using ANFIS model compared with observed monthly streamflow.

ANFIS model for the same period was compared, and found that
Table 7
the ANFIS model is relatively better than LSPC model (Table 8).
Performance comparison of ANFIS and LSPC model during the period of LSPC model
calibration and validation (monthly simulation). In addition, we analyzed median observed and median simulated
flow for each month and compared using the box plot diagram
Statistics LSPC ANFIS
(Fig. 7). It is noteworthy to mention that the response of the two
LSPC model calibration (1990–1996) models varied significantly from season to season. The median
NSE 0.83 0.85 observed and simulated flows from ANFIS model closely matched
MBE 1.8% 3.8%
R2 0.83 0.86
from January to August, and demonstrated significant difference
for other months (September to December). However, the response
LSPC model validation (1997–2002)
NSE 0.71 0.74
of LSPC model is almost opposite leading to the significant differ-
MBE 4.0% 8.8% ence from February to June, and close resemblance with the
2
R 0.71 0.76 observed streamflow for other months (July to December). This
indicates that ANFIS model can simulate streamflow resembling
the actual observed flows for most of the year and LSPC can simu-
In the next step, the LSPC model simulation (50 years) was late better for low flow period. Because SST and SLP were directly
extended back to 1947 using 5 years of warm up period. The model implemented into a mathematical model (ANFIS), it was interest-
performance corresponding to long term simulation against the ing for us to explore how this model would simulate streamflow

40
Observed flows
Streamflows, m3/s

30
LSPC-Simulated flows

20 ANFIS-Simulated flows

10

0
Jan-94
Mar-91

Jul-91

Mar-93

Jun-94
Dec-89

May-90

Oct-90

Dec-91

May-92

Oct-92

Aug-93

Nov-94

Apr-95

Sep-95

Fig. 5. LSPC model calibration and its performance comparison with ANFIS model from 1990 to 1996 (monthly streamflow).
 S. Sharma et al. / Expert Systems with Applications 42 (2015) 2213–2223 2221

40
Observed flows
30 LSPC-Simulated flows

Streamflows, m3/s
ANFIS-Simulated flows
20

10

0
Jan-98

Jan-99

Jan-00

Jan-01

Jan-02

Jan-03
Dec-96

Fig. 6. LSPC model validation and comparison with ANFIS model during LSPC model validation phase.

ANFIS demonstrated its competent performance over the LSPC

Table 8
Performance of LSPC and ANFIS model during 50 years simulation at monthly time
step.
Statistics LSPC ANFIS
(with SST and SLP)
NSE 0.68 0.79
MBE 1% 0.60%
R2 0.68 0.79
at corresponding major historic ENSO events (Table 9) since 1950.
Moreover, we wanted to see how the conventional watershed
model would simulate in a given condition, and also wanted to
evaluate the best approach for streamflow simulation in the
ENSO-affected region. The ANFIS model was compared with LSPC
model, separately for La Niña and El Niño events, and found that
ANFIS performed better than the LSPC model at different ENSO
events (Table 10) due to the training of ANFIS model with long
term data sets. The long term time series of ANFIS model and LSPC
model simulation in historic ENSO events are shown in Fig. 8.
Since SST, SLP and trade wind index are the parameters related
with basic ENSO phenomenon in the equatorial Pacific, we also
found trade wind index as a sensitive input for the model. Since
our objective was to run the model for 50 years to evaluate the
performance of the model in different ENSO events, we had to
exclude trade wind index as model inputs as these data were
available for the last 34 years.
model at daily scale as well, which implies that the model can be
extended to lower temporal scales. However, the influence of
climate variability in streamflow can be expected more distinct
ANFIS
(without SST and SLP) at monthly scales.
One of the reasons for better ANFIS model performance could be
0.74
1.27% due to the use of long period of datasets starting since 1950 for
0.74 model training for ANFIS model. Therefore, in order to make a real-
istic comparison (apples to apples) of ANFIS with LSPC, a shorter
period of datasets was used. That is, the ANFIS model is trained,
validated and tested using datasets from an exact period of
LSPC model calibration and validation. ANFIS model performance
in training and validation period was slightly better than the
corresponding LSPC model calibration (NSE = 0.85 for ANFIS and
NSE = 0.83 for LSPC) and validation (NSE = 0.73 for ANFIS and
NSE = 0.71 for LSPC).
The model comparison between ANFIS and LSPC at different
ENSO events and different scale suggest that ANFIS can simulate
streamflow equally or better than the LSPC model, partly due to
the explicit inclusion of the underlying ENSO phenomenon,
but mainly due to the ANFIS model skill for a better simulation
for the coarse rainfall data. It was essential to investigate that
the SST/SLP has a contribution to partially improve the model
simulation. To confirm this, we explored two methods; (1) first,
we assessed the performance of ANFIS model including and
excluding SST and SLP from its input vector, and found that ANFIS
model performed better as we include SST and SLP (Table 8),

To Lower Bound To Lower Bound

Average Monthly Rainfall (in.)
Median Observed Flow (1/1/1990 to 12/31/1996)
Modeled (Median, 25th, 75th)
20
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
Average Monthly Rainfall (in.)
Median Observed Flow (1/1/1990 to 12/31/1996)
Modeled (Median, 25th, 75th)
0 20 0
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
Monthly Rainfall (mm)

50 50
Monthly Rainfall (mm)

15 15
Flow (m3/sec)
Flow (m3/sec)

100 100

10 10

150 150

5 5
200 200

0 250 0 250
1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12
Month Month

Fig. 7. Statistical analysis of monthly streamflow simulated using: ANFIS model (left panel) and LSPC model (right panel).
2222 S. Sharma et al. / Expert Systems with Applications 42 (2015) 2213–2223

Table 9 simulation was not adequate to address the spatial variability of

El Niño and La Niña years (Dec–April) from 1950 to 2003 (Sharma the rainfall. In this context, additional forcing of SST/SLP can surro-
et al. (in press)).
gate the precipitation as the watershed was affected by ENSO.
La Niña year El Niño year
1950 1958 7. Summary and conclusion
1951 1966
1955 1969
1956 1973
In this study, we developed a data-driven model (ANFIS) and a
1968 1983 physically based watershed model (LSPC) in Chickasaw Creek
1971 1987 watershed, located in Southern Alabama. Both models were
1974 1988 developed for same period of datasets, and the model performance
1975 1992
was investigated through comprehensive model comparison in
1976 1995
1985 1998 numerous ways, including statistical analysis for various seasons
1989 2003 and periods over 50 years of simulation period. ANFIS model was
1996 developed by directly executing the SST and SLP difference
1999 between Tahiti and Darwin. Performances of the two models were
2000
evaluated at different temporal scale. Also, the model performance
was evaluated in various historic La Niña and El Niño events,
encountered throughout the 50 years of simulation. Even though
Table 10 both models performed satisfactorily with reasonable accuracy,
Model comparison for 25 ENSO events (monthly simulation).
ANFIS inclined to perform slightly better. ANFIS models were
Statistics LPSC-simulated ANFIS-simulated developed with and without incorporating SST and SLP in its input
For La Niña and El Niño events (since 1952) vector in order to evaluate the role of SST and SLP for model devel-
NSE 0.67 0.83 opment. Even though the improvement of the model using SST and
MBE 4.02% 3.10 SLP is not significant, improvement in a particular season when
R2 0.68 0.83 ENSO signal is clear, can be expected. ANFIS is capable to capture
14 La Niña events (Table 9) the low and high flows during winter La Niña and El Niño events
NSE 0.67 0.86 suggesting the application of SST and SLP for the long term contin-
MBE 5% 6%
R2
0.68 0.87
uous streamflow simulation. ANFIS is equally competent to LSPC in
any case whether SST and SLP are incorporated or not.
11 El Niño events (Table 9)
NSE 0.61 0.76
ANFIS model proved to be more appropriate than LSPC model
MBE 4.90% 3.12% for given watershed conditions especially for two reasons; (i) first,
R2 0.66 0.77 spatially distributed precipitation data were not available for phys-
ically based model, LSPC, (ii) second, SST and SLP were additional
inputs for ANFIS, which could not be incorporated by LSPC model,
(iii) third, numerous missing data and recording error in precipita-
(2) secondly, the winter season (JFM) simulation was tested using a tion data could be surrogated by SST and SLP which were the
simple MLR model as the ENSO signal was clear during this period. additional inputs in ANFIS.
It was found that model improved in a winter season due to the This research was particularly conducted for a single watershed
direct application of SST and SLP in the model. Therefore, this with reasonable teleconnection of ENSO with local precipitation
conclusion will be true only for a period when ENSO has a better and temperature of the region. It would be appropriate to calibrate
correlation with streamflow and may not be valid when ENSO does both models using fine resolution to coarse resolution precipitation
not have a clear signature with precipitation. data sets, particularly for a strongly ENSO-affected watershed with
Lack of enough rain gauge station in a study watershed was multiple model set up. Also, it would be appropriate to conduct the
clearly manifested in both model simulations. More importantly, study in various watersheds for better generalization. In addition,
we investigated the precipitation and corresponding observed separate ANFIS model for three different seasons would have been
streamflow, and found that some of the streamflow did not look appropriate as ENSO correlation with precipitation/temperature is
like the proper match with the corresponding precipitation events different in various seasons, and particularly opposite from August
indicating that rain gauge stations considered for streamflow to October. Therefore, a single model developed for entire period

60

50

40

30

20

10

0
Jun-57
Dec-51

Nov-62

May-68

Nov-73

May-79

Oct-84

Apr-90

Oct-95

Mar-01

Fig. 8. Monthly streamflow simulation using LSPC and ANFIS models corresponding to historic La Niña and El Niño events (Table 9) since 1950. Dots and triangles represent
ANFIS and LSPC simulation during La Niña and El Niño events.
 S. Sharma et al. / Expert Systems with Applications 42 (2015) 2213–2223
cannot be expected to significantly improve the simulation by
additional forcing of SST and SLP.
Future research can be conducted in ENSO affected watershed
using ANFIS model, and incorporating trade wind index besides
SST and SLP in its input vectors, as the trade wind index is also
responsible for ENSO phenomenon. It is to be clearly mentioned
that, the benefit of ANFIS model over the LSPC or any other phys-
ically based model, can be expected significant, especially for a
watershed which is suffered from lack of adequate rain gauge sta-
tions and also affected by ENSO.
Even though this research does not conclude the supremacy of
ANFIS model over the LSPC model, research concludes that ANFIS
can be a better choice when rain gauge stations are limited. More
importantly, ANFIS provides an opportunity to consider additional
inclusions of long term SST and SLP in the equatorial Pacific in
order to explicitly utilize the ENSO pattern for hydrologic model-
ing. This might be helpful as and when spatially distributed precip-
itation datasets are not available and the watershed is affected by
ENSO.
References
Barsugli, J. J., Whitaker, J. S., Loughe, A. F., Sardeshmukh, P. D., & Toth, Z. (1999). The
effect of the 1997/98 El Niño on individual large-scale weather events. Bulletin
of the American Meteorological Society, 80, 1399–1411.
Bicknell, B., Imhoff, J., Kittle, J., Jr., Jobes, T., Donigian, A., Jr., & Johanson, R. (2001).
Hydrological simulation program–fortran: HSPF, version 12 user’s manual.
Mountain View, California: AQUA TERRA Consultants.
Bosznay, M. (1989). Generalization of SCS curve number method. Journal of
Irrigation and Drainage Engineering, 115, 139.
Chiew, F., Piechota, T., Dracup, J., & McMahon, T. (1998). El Niño/southern
oscillation and Australian rainfall, streamflow and drought: Links and
potential for forecasting. Journal of Hydrology, 204, 138–149.
Chiu, S. (1996). Method and software for extracting fuzzy classification rules by
subtractive clustering. IEEE (pp. 461–465). .
Clarke, R. (1973). A review of some mathematical models used in hydrology, with
observations on their calibration and use. Journal of Hydrology, 19, 1–20.
Cochocki, A., & Unbehauen, R. (1993). Neural networks for optimization and signal
processing. John Wiley & Sons Inc..
Committee, A. T. (2000). Artificial neural networks in hydrology. I: Preliminary
concepts. Journal of Hydrologic Engineering, 5, 115–123.
Crawford, N. H., & Linsley, R. K. (1966). Digital simulation in hydrology’s stanford
watershed model 4.
Dunn, J. C. (1973). A fuzzy relative of the ISODATA process and its use in detecting
compact well-separated clusters.
Engel, B., Storm, D., White, M., Arnold, J., & Arabi, M. (2007). A hydrologic/water
quality model applicati11. In Wiley online library.
Hansen, J., Jones, J., Irmak, A., & Royce, F. (2001). El Niño-southern oscillation
impacts on crop production in the southeast United States. ASA Special
Publication, 63, 55–76.
Jang, J. S. R. (1993). ANFIS: Adaptive-network-based fuzzy inference system.
Systems, Man and Cybernetics, IEEE Transactions on, 23, 665–685.
Jang, J. S. R., & Sun, C. T. (1995). Neuro-fuzzy modeling and control. Proceedings of
the IEEE, 83, 378–406.
Jang, J.-S. R., Sun, C.-T., & Mizutani, E. (1997). Neuro-fuzzy and soft computing-a
computational approach to learning and machine intelligence [book review].
Automatic Control, IEEE Transactions on, 42, 1482–1484.
Jayawardena, A., Muttil, N., & Lee, J. (2006). Comparative analysis of data-driven and
GIS-based conceptual rainfall–runoff model. Journal of Hydrologic Engineering,
11, 1.
Keener, V., Ingram, K., Jacobson, B., & Jones, J. (2007). Effects of El Niño/southern
oscillation on simulated phosphorus loading in South Florida. Transactions of the
ASAE, 50, 2081–2089.
Kessler, W. S. (2002). Is ENSO a cycle or a series of events? Geophysical Research
Letters, 29, 40-41–40-44.
Khalil, A. F., McKee, M., Kemblowski, M., & Asefa, T. (2005). Basin scale water
management and forecasting using artificial neural network. JAWRA Journal of
the American Water Resources Association, 41, 195–208.
Kim, S. M., Benham, B. L., Brannan, K. M., Zeckoski, R. W., & Doherty, J. (2007).
Comparison of hydrologic calibration of HSPF using automatic and manual
methods. Water Resources Research, 43, W01402.
Kulkarni, J. (2000). Wavelet analysis of the association between the southern
oscillation and the Indian summer monsoon. International Journal of
Climatology, 20, 89–104.
Lim, K. J., Engel, B. A., Tang, Z., Choi, J., Kim, K. S., Muthukrishnan, S., & et al. (2005).
Automated web Gis based hydrograph analysis tool, WHAT1. In Wiley online
library.
2223
Lukas, R., Hayes, S. P., & Wyrtki, K. (1984). Equatorial sea level response during the
1982–1983 El Niño. Journal of Geophysical Research: Oceans (1978–2012) (89),
10425–10430.
Makkeasorn, A., Chang, N.-B., & Zhou, X. (2008). Short-term streamflow forecasting
with global climate change implications – A comparative study between genetic
programming and neural network models. Journal of Hydrology, 352, 336–354.
McCabe, G. J., & Dettinger, M. D. (1999). Decadal variations in the strength of ENSO
teleconnections with precipitation in the western United States. International
Journal of Climatology, 19, 1399–1410.
Moriasi, D., Arnold, J., Van Liew, M., Bingner, R., Harmel, R., & Veith, T. (2007). Model
evaluation guidelines for systematic quantification of accuracy in watershed
simulations 10.
Morid, S., Gosain, A., & Keshari, A. K. (2002). Comparison of the SWAT model and
ANN for daily simulation of runoff in snowbound ungauged catchments. In Fifth
international conference on hydroinformatics. Cardiff, UK.
Mukerji, A., Chatterjee, C., & Raghuwanshi, N. S. (2009). Flood forecasting using
ANN, Neuro-Fuzzy, and Neuro-GA models. Journal of Hydrologic Engineering, 14,
647.
Nash, J., & Sutcliffe, J. (1970). River flow forecasting through conceptual models part
I – A discussion of principles. Journal of Hydrology, 10, 282–290.
Pascual, M., Rodó, X., Ellner, S. P., Colwell, R., & Bouma, M. J. (2000). Cholera
dynamics and El Niño-southern oscillation. Science, 289, 1766.
Piechota, T. C., & Dracup, J. A. (1996). Drought and regional hydrologic variation in
the United States: Associations with the el Niño-southern oscillation. Water
Resources Research, 32, 1359–1373.
Pramanik, N., & Panda, R. K. (2009). Application of neural network and adaptive
neuro-fuzzy inference systems for river flow prediction. Hydrological Sciences
Journal, 54, 247–260.
Rajagopalan, B., & Lall, U. (1998). Interannual variability in western US
precipitation. Journal of Hydrology, 210, 51–67.
Ropelewski, C., & Halpert, M. (1986). North American precipitation and temperature
patterns associated with the El Niño/southern oscillation (ENSO). The Monthly
Weather Review; (United States), 114.
Roy, S. S. (2006). The impacts of ENSO, PDO, and local SSTs on winter precipitation
in India. Physical Geography, 27, 464–474.
Sharma, S. (2012). Incorporating El Niño southern oscillation (ENSO)-induced
climate variability for long-range hydrologic forecasting and stream water
quality protection. Auburn University.
Sharma, S., Srivastava, P., Fang, X., & Kalin, L. (in press). Hydrologic Simulation
Approach for El Niño southern oscillation (ENSO)-affected watershed with
limited rain gauge stations. Hydrologic Sciences Journal. http://www.tandfonline.
com/doi/abs/10.1080/02626667.2014.952640#.VD_TYeK5eXg.
Sharma, S., Isik, S., Srivastava, P., & Kalin, L. (2012). Deriving spatially distributed
precipitation data using the artificial neural network and multilinear regression
models. Journal of Hydrologic Engineering, 18, 194–205.
Sharma, S., Srivastava, P., Fang, X., & Kalin, L. (2012). Incorporating climate
variability for point-source discharge permitting in a complex river system.
Transactions of the ASABE, 55, 2213–2228.
Sharma, S., Srivastava, P., Kalin, L., Fang, X., & Elias, E. (2014). Predicting total
organic carbon load with El Nino southern oscillation phase using hybrid and
fuzzy logic approaches. Transactions of the ASABE, 57(4), 1071–1085.
Shen, J., Parker, A., & Riverson, J. (2005). A new approach for a windows-based
watershed modeling system based on a database-supporting architecture.
Environmental Modelling & Software, 20, 1127–1138.
Taheri Shahraiyni, H., Ghafouri, M. R., Saghafian, B., & Bagheri Shouraki, S. (2013).
Evaluation of a novel fuzzy method and a conceptual model for a long-term
daily streamflow simulation. River Systems, 20, 249–260.
Talei, A., Chua, L. H. C., & Quek, C. (2010). A novel application of a neuro-fuzzy
computational technique in event-based rainfall–runoff modeling. Expert
Systems with Applications, 37, 7456–7468.
Tayfur, G., Ozdemir, S., & Singh, V. P. (2003). Fuzzy logic algorithm for runoff-
induced sediment transport from bare soil surfaces. Advances in Water
Resources, 26, 1249–1256.
Tayfur, G., & Singh, V. P. (2006). ANN and fuzzy logic models for simulating event-
based rainfall–runoff. Journal of Hydraulic Engineering, 132, 1321–1330.
Tech, T., & Center, S. N. A. F. (2009). Loading simulation program in C++ (LSPC)
Version 3.1 User’s Manual. Fairfax, VA.
Thirumalaiah, K., & Deo, M. C. (2000). Hydrological forecasting using neural
networks. Journal of Hydrologic Engineering, 5, 180–189.
Tokar, A. S., & Johnson, P. A. (1999). Rainfall–runoff modeling using artificial neural
networks. Journal of Hydrologic Engineering, 4, 232–239.
Trenberth, K. E., & Shea, D. J. (1987). On the evolution of the southern oscillation.
Monthly Weather Review, 115, 3078–3096.
Trenberth, K. E., & Stepaniak, D. P. (2001). Indices of El Niño evolution. Journal of
Climate, 14, 1697–1701.
Wang, P., Skahill, B., Samaitis, H., & Johnston, R. (2002). GIS-based artificial neural
network and processed-based HSPF model for watershed runoff in Sinclair and
dyes inlet, WA. Proceedings of the Water Environment Federation, 2002,
1547–1564.
Yan, H., Zou, Z., & Wang, H. (2010). Adaptive neuro fuzzy inference system for
classification of water quality status. Journal of Environmental Sciences, 22,
1891–1896.
Zadeh, L. A. (1965). Fuzzy sets⁄. Information and Control, 8, 338–353.