Professional Documents
Culture Documents
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/245287452
CITATIONS READS
236 504
2 authors, including:
Momcilo Markus
University of Illinois, Urbana-Champaign
39 PUBLICATIONS 704 CITATIONS
SEE PROFILE
All content following this page was uploaded by Momcilo Markus on 29 November 2015.
ABSTRACT: Inspired by the functioning of the brain and biological nervous systems, artificial neural networks
(ANNs) have been applied to various hydrologic problems in the last 10 years. In this study, ANN models are
compared with traditional conceptual models in predicting watershed runoff as a function of rainfall, snow water
equivalent, and temperature. The ANN technique was applied to model watershed runoff in three basins with
different climatic and physiographic characteristics—the Fraser River in Colorado, Raccoon Creek in Iowa, and
Little Patuxent River in Maryland. In the Fraser River watershed, the ANN technique was applied to model
monthly streamflow and was compared to a conceptual water balance (Watbal) model. The ANN technique was
used to model the daily rainfall-runoff process and was compared to the Sacramento soil moisture accounting
(SAC-SMA) model in the Raccoon River watershed. The daily rainfall-runoff process was also modeled using
the ANN technique in the Little Patuxent River basin, and the training and testing results were compared to
those of a simple conceptual rainfall-runoff (SCRR) model. In all cases, the ANN models provided higher
accuracy, a more systematic approach, and shortened the time spent in training of the models. For the Fraser
River, the accuracy of monthly streamflow forecasts by the ANN model was significantly higher compared to
the accuracy of the Watbal model. The best-fit ANN model performed as well as the SAC-SMA model in the
Raccoon River. The testing and training accuracy of the ANN model in Little Patuxent River was comparatively
higher than that of the SCRR model. The initial results indicate that ANNs can be powerful tools in modeling
the precipitation-runoff process for various time scales, topography, and climate patterns.
Conceptual Models
Watbal Model
The Watbal model is a physically based model used for
forecasting water yields in areas where runoff is dominated by
snowmelt (Leaf and Brink 1973, 1976; Leaf and Alexander
1975; Markus and Baker 1994). It simulates winter snow ac-
cumulation, shortwave and longwave radiation balance, snow-
pack conditions, snowmelt, evapotranspiration, and subsequent
runoff in time and space.
Shortwave and longwave radiation represents the energy
components available for snowmelt. In the Watbal model,
shortwave radiation to snow or the ground surface beneath the
forest canopy is controlled by a transmissivity coefficient,
which varies with forest cover characteristics. The incident
FIG. 1. Artificial Neuron (xn = nth Input; w In j = Weight from n th shortwave radiation as measured on a horizontal surface is
Input to jth Neuron in I th Layer; Netj = Weighted Inputs of j th adjusted based on the slope and aspect of each hydrologic
Neuron; f ( ) = Transfer Function; and outj = Output of j th Neuron) response unit. Longwave radiation is computed by the Stefan-
Boltzmann equation (Leaf and Brink 1973, 1975).
Snowpack reflectivity varies with precipitation form, air
temperature, and the energy balance. During the winter
months, temperatures within the snow cover are simulated us-
ing the unsteady heat flow theory. The snowpack will yield
snowmelt only when it becomes isothermal at 0⬚C, and its free
water holding capacity is reached.
Evapotranspiration is derived from the Hamon equation
(Leaf and Brink 1973, 1975) for potential evapotranspiration
reduced in proportion to the radiation actually received. The
soil water content is computed for each time increment. The
critical point, at which the soil water content begins to limit
evapotranspiration, varies with time and forest tree species.
Forest cover density plays an important role in the simulation
model. It is the major descriptive parameter of the form, struc-
ture, and arrangement of forest stands, and therefore controls
the energy balance, interception, and evapotranspiration.
Reflectivity is described as a function of the forest stand
FIG. 2. Schematic of Back-Propagation Network reflectivity, forest opening reflectivity, and natural forest cover
JOURNAL OF HYDROLOGIC ENGINEERING / APRIL 2000 / 157
density. Interception is given as a function of forest cover. The parameter that distributes the surface runoff. The schematic of
combined evaporation from snow surfaces and intercepted the model is presented in Fig. 3. Daily precipitation, air tem-
snow is a function of potential evapotranspiration, intermedi- perature, and streamflow data are employed for the calibration
ate forest cover density, and forest cover density. Many other of the model. The objectives are to approximate a combination
processes, such as snow redistribution, tree cutting, ground- of the following: total measured runoff, peak discharges, tim-
water flow, etc., that are significant factors influencing runoff ing of peak discharges, soil moisture state at the beginning of
are also modeled in the Watbal model. the record, recession of the storm events, and bias in the flow
during dry periods. Since there are multiple objective func-
SAC-SMA Model tions, the calibration can be accomplished by the subjective
optimization technique.
The SAC-SMA model is a conceptually based, lumped,
rainfall-runoff model that represents spatially heterogeneous CASE STUDIES
runoff processes for river basins at scales ranging from a few
hundred to thousands of square kilometers. The SAC-SMA The numerical algorithm based on ANNs was applied to
model is an operational rainfall-runoff model that is used in three watersheds in the continental United States (Table 1).
the National Weather Service river forecast system (‘‘United Historical measurements of precipitation (P), temperature (T),
States’’ 1972). The model has six soil moisture states and 16 snow-water equivalent (SWE), and stream discharge (Q) are
parameters, not counting the 12 monthly adjustment factors of available for the Fraser River, the Raccoon River, and the Lit-
potential evaporation. Most of the parameters have to be cal- tle Patuxent River watersheds.
ibrated using historical hydrometeorological data.
In this model, the watershed consists of two areas—a per- Fraser River
vious area that produces runoff only when precipitation ex-
ceeds a specific amount, and an impervious area that produces The ANN technique was used to develop a one-step,
runoff regardless of the amount of precipitation. In pervious monthly flow forecast model for the Fraser River in central
areas, the soil profile is divided into two vertical zones—an Colorado. The Fraser River is a mountainous, subalpine river,
upper zone and a lower zone. The upper zone represents the with steep slopes, coniferous vegetation, and large snowpack
amount of water held in the interception and upper-soil layer, accumulation in wintertime. The discharge of the Fraser River
and the lower zone represents the amount of water retained in is dominated by snowmelt in the late spring and early summer.
the lower zone and ground-water storage. Both zones are di- The Fraser River flows into the Colorado–Big Thompson res-
vided into two types of storage-—tension water and free water ervoir system. Therefore, timely and accurate forecasts of
storage. Tension water represents the field capacity, which is monthly river volumes result in significant economic benefits.
the amount of water held in the soil zones due to binding with In the Fraser River watershed, data are available for the
soil particles. Once the tension water is satisfied, water starts periods of 1951–83 and 1987–93. In calibration of the ANN
accumulating in free water storage, and from there it moves and Watbal models, only data for high-flow months, May,
either laterally or vertically. In the upper zone, free water June, and July, were employed. The Watbal model was cali-
moves laterally to the streams as interflow and vertically to brated using the data for the entire record, and tested using
the lower zone as percolation. The free water from the lower data for the period from 1981 to 1983 and 1987–93. However,
zone either constitutes the base flow or feeds to the ground- the ANN model was trained using only data extending from
water storage. 1951 to 1980, and then tested for the period from 1981 to
1983 and 1987–93. A standard back-propagation algorithm
SCRR Model and the sigmoid transfer function were applied in training of
the networks. A one-hidden-layer network with two hidden
The conceptual rainfall-runoff model, SCRR, that was de- neurons was used, and the number of iterations for the training
veloped by McCuen and Snyder has three storage components varied from several hundred to several thousand, depending
—surface, ground water, and the unsaturated zone (McCuen on the particular month. The watershed runoff was modeled
and Snyder 1986). The rainfall first enters to the surface and as follows using ANN:
ground-water systems depending on the model parameter that
defines the fraction of infiltrated water into the ground. Water Q(t) = f {Q(t ⫺ 1), P(t ⫺ 1), SWE(t ⫺ 1), SWE(t ⫺ 2), T(t ⫺ 1)}
that enters to the surface storage contributes to the stream sys- (2)
tems using a unit hydrograph with a time base of four time
where Q(t) and Q(t ⫺ 1) = streamflows at times t and t ⫺ 1,
intervals. The unit hydrograph ordinates are based on a model
respectively; P(t ⫺ 1) = precipitation at time t ⫺ 1; SWE(t ⫺
1), SWE(t ⫺ 2) = snow water equivalents at times t ⫺ 1 and
t ⫺ 2, respectively; T(t ⫺ 1) = air temperature at time t ⫺ 1;
Q(t) = f {P(t), P(t ⫺ 1), P(t ⫺ 2), P(t ⫺ 3), P(t ⫺ 4), T(t)} (5)
Three years were selected based on the mean annual daily
Q(t) = f {P(t ⫺ 1), P(t ⫺ 2), P(t ⫺ 3), T(t), Q(t ⫺ 1)} (6) discharge values—average, dry, and wet years. As illustrated
in Table 4, the SAC-SMA model provided slightly higher ac-
where Q(t) = streamflow at time t; P(t ⫺ 1) = precipitation at curacy for the average year compared to the ANN models.
time t ⫺ 1; and T(t) = temperature at time t; note that t is in However, the SAC-SMA and models 1–3, with high Se to Sy
days. The calibration period for both the ANN and the SAC- ratios, fail to approximate the dry-year data as well as model
SMA extends from 1978 to 1993. The results are presented 4. In addition, model 4 provided the lowest Se to Sy ratio for
for the entire calibration period and for average-, dry-, and the wet year. As shown in Table 4, model 4 also had the high-
wet-year data sets, separately shown in Tables 3 and 4. est accuracy for the three years compared to the SAC-SMA
Model 4 has a slightly lower Se to Sy ratio, about 0.40, when and models 1–3. These results are explained by the addition
compared to the SAC-SMA model. Models 1–3 had higher Se of Q(t ⫺ 1). In model 4, the values of Q(t ⫺ 1) provided
to Sy ratios, ranging from 0.64 to 0.67, compared to both information that was not contained in precipitation and tem-
model 4 and the SAC-SMA model. As shown in Table 3, the perature.
SAC-SMA model reproduced observed standard deviation
fairly well, while the ANN models provided less variation of Little Patuxent River
streamflow values.
The SAC-SMA model underestimated the minimum and The Little Patuxent River is located in Maryland. It is a
maximum streamflows in the record, at about 69% and 11% rainfall-dominated stream exhibiting the behavior of a stable
of their observed values, respectively. While the ANN models ground-water and perennial runoff stream type (Poff et al.
closely approximated the maximum streamflow in the record, 1996). The streamflow regime in the Little Patuxent River con-
they significantly overestimated the low flow values. A poor sists of high flows from December to May. The highest pre-
performance of the ANNs in forecasting low flows is consis- cipitation variations are observed during the dry months, June
tent with the results reported in recent literature (Tokar 1996; through September, that are characterized by flash flood
Markus 1997). events. The ANN technique was used to model daily rainfall-
The sensitivity of model training accuracy to the content of runoff in the watershed. Networks defined by various combi-
data was assessed using three data sets within the training data. nations of rainfall, temperature, and discharge at present and
JOURNAL OF HYDROLOGIC ENGINEERING / APRIL 2000 / 159
TABLE 5. Statistical Indices for Training and Testing Using CONCLUSIONS
ANN and SCRR Models for Little Patuxent River, Md.
In the literature, the ANN methodology has been reported
Se /Sy Se /Sy R2 R2 to provide reasonably good solutions for circumstances having
Model Training Testing Training Testing complex systems that may be poorly defined or understood
(1) (2) (3) (4) (5) using mathematical equations, problems that deal with noise
ANN 0.46 0.42 0.78 0.83 or involve pattern recognition, and input data that are incom-
SCRR model 0.66 0.67 0.56 0.55 plete and ambiguous by nature. Because of these characteris-
Note: Se and Sy are root-mean-square error and standard deviation of tics, it was believed that ANN could be applied to model pre-
observed discharge, respectively. R 2 represents coefficient of determina- cipitation-runoff relationships. The ANN rainfall-runoff
tion between predicted and observed discharges. models exhibit the ability to extract patterns in the training
data.
The ANNs were applied to forecast monthly streamflows in
TABLE 6. Statistical Indices for Training and Testing Average,
the Fraser River basin, and daily streamflows in the Raccoon
Dry, and Wet Years Using ANN (Model 4) and SCRR Models for River basin (calibration only) and the Little Patuxent River
Little Patuxent River Watershed, Md. basin (calibration and testing). For the Fraser River, the ANN
provided more accurate monthly streamflow forecasts than did
Training Testing a physically based model. Similarly, for the Raccoon River,
Parameter Wet Dry Average Wet Dry Average the best-fit ANN model has reasonable calibration accuracy
(1) (2) (3) (4) (5) (6) (7) when compared to that of a conceptual model. Such perfor-
(a) SCRR Model mance is a result of adding the streamflow at t ⫺ 1 as an input
Se /Sy 0.68 0.64 0.63 0.68 0.74 0.63
to the ANN model. Finally, in the Little Patuxent River basin,
R2 0.54 0.59 0.60 0.54 0.45 0.60 the best-fit ANN model provided higher training and testing
accuracy compared to that of the simple conceptual rainfall-
(b) ANN Model
runoff model.
Se /Sy 0.44 0.84 0.47 0.37 0.66 0.39 These examples demonstrated that ANNs can accurately
R2 0.83 0.54 0.79 0.86 0.56 0.85
model nonlinear relationships between hydrologic inputs, rain-
Note: Se and Sy are root-mean-square error and standard deviation of fall, snow water equivalent, temperature, and output stream-
observed discharge, respectively. R 2 represents coefficient of determina- flow. The ANN models provided a systematic approach and
tion between predicted and observed discharges.
shortened time spent on training of models compared to the
conceptual models.
Although the models were applied to only three watersheds,
previous time periods were trained and tested to simulate the results presented here were encouraging, and demonstrated
streamflow using different ANN configurations. Based on av- a high potential for the application of neural networks to var-
erage precipitation, three data sets representing average, wet, ious precipitation-runoff modeling scenarios.
and dry years were selected for training and testing in order
to identify the rainfall-runoff behavior in the Little Patuxent APPENDIX. REFERENCES
River watershed. For training of the networks, 1979, 1980, and Caudill, M. (1987). ‘‘Neural network primer: Part I.’’ AI Expert, (Decem-
1984 were selected as the wet, dry, and average years. Simi- ber), 46–52.
larly for testing, 1989, 1991, and 1992 were chosen as the French, M. N., Krajewski, W. F., and Cuykendall, R. R. (1992). ‘‘Rainfall
wet, dry, and average years. The testing data were not directly forecasting in space and time using a neural network.’’ J. Hydro., 137,
1–13.
used in the training; however, in order to avoid overtraining
‘‘The Great Flood of 1993.’’ (1994). NOAA Natural Disaster Survey Rep.,
of the networks, networks were tested using a test set. For National Weather Service, Office of Hydrology, Silver Spring, Md.
different combinations of the input variables, networks were Hecht-Nielsen. (1990). Neurocomputing. Addison-Wesley, Reading,
trained using a one-hidden-layer network (Tokar 1996). The Mass.
training of networks was accomplished using the back-prop- Hsu, K, Gupta, V. H., and Sorooshian, S. (1995). ‘‘Artificial neural net-
agation algorithm of NeuralWorks (Professional 1993). The work modeling of the rainfall-runoff process.’’ Water Resour. Res.,
hyperbolic tangent was selected as an activation function in 31(10), 2571–2530.
Leaf, C. F., and Alexander, R. R. (1975). ‘‘Simulating timber yields and
training of the networks. The selection of the best-fit model hydrologic impact resulting from timber harvest on subalpine water-
was accomplished by a trial-and-error process (Tokar 1996). sheds.’’ USDA Forest Service Res. Paper RM-133, U.S. Department of
The best-fit model was the network with 10 hidden neurons, Agriculture, Washington, D.C.
and is given below Leaf, C. F., and Brink, G. E. (1973). ‘‘Hydrologic simulation model of
Colorado subalpine forest.’’ USDA Forest Service Res. Paper RM-107,
Q(t) = f {P(t), P(t ⫺ 1), T(t)} (7) U.S. Department of Agriculture, Washington, D.C.
Leaf, C. F., and Brink, G. E. (1975). ‘‘Land use simulation model of the
subalpine coniferous forest zone.’’ USDA Forest Service Res. Paper
where t is in days. The SCRR model for the Little Patuxent RM-135, U.S. Department of Agriculture, Washington, D.C.
River watershed (Tokar 1996) was calibrated under the super- Markus, M. (1997). ‘‘Application of neural networks in streamflow fore-
vision of R. H. McCuen, who originally developed the SCRR casting,’’ doctoral dissertation, Colorado State University, Fort Collins,
model (McCuen and Snyder 1986). For the conceptual model, Colo.
the Se /Sy values are 0.66 and 0.67 for calibration and verifi- Markus, M., and Baker, D. (1994). ‘‘The Fraser River: Streamflow fore-
casting and simulation computer package.’’ Tech. Rep., Northern Col-
cation, respectively. As indicated in Table 5, the forecast ac- orado Water Conservancy District, Loveland, Colo.
curacy for SCRR is relatively poor compared to that of the Markus, M., Salas, J. D., and Shin, H. (1995). ‘‘Predicting streamflows
ANN model, with low Se /Sy values of 0.46 and 0.42 for train- based on neural networks.’’ 1st Int. Conf. on Water Resour. Engrg.,
ing and testing, respectively. As shown in Table 6, the ANN ASCE, New York.
model provided a significantly higher training accuracy for the McCuen, R. H., and Snyder, M. W. (1986). Hydrologic modeling: Statis-
tical methods and applications. Prentice-Hall, Englewood Cliffs, N.J.
average and wet years compared to that of the SCRR, whereas
Muller, B., and Reinhardt, J. (1990). Neural networks: An introduction.
SCRR illustrated better training accuracy for the dry year. Springler, New York.
However, the ANN model demonstrated significantly higher ‘‘National Weather Service river forecast system.’’ (1996). National
testing accuracy for the three years. Weather Service, Office of Hydrology, Silver Spring, Md.