Retraction Notice

Title of retracted article: Revisiting Systems Type Black-Box Rainfall-Runoff Models for
Flow Forecasting Application
Author(s): Dereje Tesfahun Mengistu, Semu Ayalew Moges, Asgeir Sorteberg

* Corresponding author. Email: dertes.fiker@gmail.com

Journal: Journal of Water Resource and Protection

Year: 2016
Volume: 8
Number: 1
Pages (from - to): 65 - 83
DOI (to PDF): http://dx.doi.org/10.4236/jwarp.2016.81006
Paper ID at SCIRP: 9402764
Article page: http://www.scirp.org/Journal/PaperInformation.aspx?PaperID=62835

Retraction date: 2016-01-29

Retraction initiative (multiple responses allowed; mark with X):

XAll authors
Some of the authors:
Editor with hints from Journal owner (publisher)
Institution:
Reader:
Other:
Date initiative is launched: 2016-01-29

Retraction type (multiple responses allowed):

Unreliable findings
Lab error Inconsistent data Analytical error Biased interpretation
Other:
Irreproducible results
Failure to disclose a major competing interest likely to influence interpretations or recommendations
Unethical research

Fraud
Data fabrication Fake publication Other:
Plagiarism Self plagiarism Overlap Redundant publication *
Copyright infringement Other legal concern:

Editorial reasons
Handling error Unreliable review(s) Decision error Other:

X Other:

Results of publication (only one response allowed):

are still valid.
Xwere found to be overall invalid.

Author's conduct (only one response allowed):

honest error
academic misconduct
Xnone (not applicable in this case e.g. in case of editorial reasons)

* Also called duplicate or repetitive publication. Definition: "Publishing or attempting to publish substantially the same
work more than once."

1
History
Expression of Concern:
yes, date: yyyy-mm-dd
Xno

Correction:
yes, date: yyyy-mm-dd
Xno

Comment:

Free style text with summary of information from above and more details that can not be
expressed by ticking boxes.

This article has been retracted according to COPE's Retraction Guidelines. Since authors have their personal reasons,
they have to withdraw this paper from Journal of Water Resource and Protection.

2
Journal of Water Resource and Protection, 2016, 8, 65-83
Published Online January 2016 in SciRes. http://www.scirp.org/journal/jwarp
http://dx.doi.org/10.4236/jwarp.2016.81006

Revisiting Systems Type Black-Box

Rainfall-Runoff Models for Flow
Forecasting Application
RE
Dereje Tesfahun Mengistu1,2*, Semu Ayalew Moges3, Asgeir Sorteberg2,4
1
Arba Minch Institute of Technology, Arba Minch University, Arba Minch, Ethiopia
2
Geophysical Institute, University of Bergen, Bergen, Norway
3
School of Civil and Environmental Engineering, AAiT-AAU, Addis Ababa, Ethiopia
4
Bjerknes Centre for Climate Research, University of Bergen, Bergen, Norway
TR
Received 18 December 2015; accepted 16 January 2016; published 19 January 2016

Copyright 2016 by authors and Scientific Research Publishing Inc.

This work is licensed under the Creative Commons Attribution International License (CC BY).
http://creativecommons.org/licenses/by/4.0/
A

Abstract
CT
Often we tend to spend huge amount of time and resources to setup and use complex hydrological
models for simple goal of flow estimation. Running complex models becomes even more difficult
when the amount of available data is scarce as we usually face in many parts of Africa. The aim of
this study is to evaluate and revitalize the systems type black box model against complex hydro-
logical models for easy flow estimation application. Six systems type black box models, the Simple
Linear Model (SLM), Non-Parametric Simple Linear Model (NP-SLM), Linear Perturbation Model
(LPM), Non-Parametric Linear Perturbation Model (NP-LPM) and Linearly Varying Gain Factor
ED
Model (LVGFM), a non-linear black box type artificial Neural Network model (ANN) are compared
with three complex hydrological models of those under SMAR, HBV and SWAT. The models are
compared based on daily rainfall and stream flow data (1980-2000) on Gilgel Abbay watershed.
Event-based analysis was also conducted using 100 selected runoff events. In terms of the event
rainfall-runoff relationship, it was indicated that the event runoff is largely a function of the
amount of rainfall. The event rainfall-runoff relationships explained as much as 62% for the wet
periods without the integration of the evaporating demands. Although rainfall intensity, duration
and catchment characteristics play a role, in this watershed, rainfall amount affects substantial
part of the runoff response consolidating that a simple rainfall-runoff relationship can describe
the runoff in this watershed. Comparison of systems type black box and complex hydrological
models in the study area indicates that the LPM and the ANN models perform better than the com-
plex hydrological models such as SMARG, HBV and SWAT in terms of R2 and Nash Sutcliffe Effi-
ciency (NSE) criteria. This confirms that simpler models (that take only rainfall as input) can sur-
pass their complex counterparts in performance for continuous simulation and reproducing the
*
Corresponding author.

How to cite this paper: Mengistu, D.T., Moges, S.A. and Sorteberg, A. (2016) Revisiting Systems Type Black-Box Rainfall-
Runoff Models for Flow Forecasting Application. Journal of Water Resource and Protection, 8, 65-83.
http://dx.doi.org/10.4236/jwarp.2016.81006
D. T. Mengistu et al.

hydrographs or flow estimation. There is a strong justification, therefore, for the claim that in-
creasing the model complexity, thereby increasing the number of parameters, does not necessar-
ily enhance the model performance. It is suggested that, in practical hydrology, the simpler models,
may still play a significant role as effective simulation tools, and countries with scarce hydrologi-
cal data should revitalize application of such systems type black box modelling schemes that de-
pend only on rainfall and runoff data sets which could be easily available.

Keywords
Black Box Models, Distributed Models, Rainfall-Runoff Relationships, Event
RE
1. Introduction
Rainfall-runoff relationships are complicated processes which may be highly non-linear and exhibit both tem-
poral and spatial variations. Understanding their relationship is essential for practical basin management prac-
tices. Several hydrological models have been developed to simulate rainfall-runoff relationships across the
world [1]. These can be classified as empirical black box models, conceptual and physically based distributed
models [1]. Black box models are empirical, involving mathematical equations that have been assessed, not
TR
from the physical processes in the catchment, but from analysis of concurrent input and output time series. Con-
ceptual (lumped) models treat the catchment as a single unit, with state variables that represent average values
over the catchment area, such as storage in the saturated zone. Another approach to hydrological processes
modeling is the attempt to construct models based on the governing equations describing all the surface and
subsurface flow processes in the catchment called physically distributed models. Each of these types of models has
their own advantages and limitations [1]. For instance, in areas where getting sufficient hydro-meteorological data
are problematic or the purpose of hydrological modeling is limited to flow estimation, applications of linear
A
systems theoretic models (black box models) are inevitably important for water related development. However,
to choose between the various available hydrological models to suit a practical demand and find the most ap-
propriate model for the specified basin is a big challenge. Many models are in practice simple linear system
theoretic models (black box models) [1] which do often not represent the non-linear dynamics, which are inhe-
CT
rent in the process of rainfall-runoff transformation.
[2] and [3] observed that the rainfall information alone is not sufficient to calculate the runoff from a catch-
ment as the initial state (such as amount of soil moisture and orographic features) of the catchment plays an im-
portant role in determining the runoff rate behavior. The rainfall-runoff relationship in mountainous regions is
influenced by the steep gradient profiles (i.e. inter flow and sheet flow) and less influenced by soil composition
[4]. Nevertheless, soil composition in less steep environment plays a major role in runoff generation due to the
presence of very to moderately drained soils [4]-[7]. Therefore, higher streamflows and runoff coefficients (R/P,
ED
where R is runoff and P precipitation) are usually associated with mountainous area [8], while smaller R/P ratios
are expected for low-topographic gradient watersheds [8]-[10] argue that runoff in lower land plain watersheds
have a much larger variability than upland watersheds because of a wider range of variable source areas, includ-
ing ephemeral water storage in depressions in a low gradient terrain.
Evapotranspiration is another factor that affects the hydrological processes of the watershed in shallow water
tables [11]. It is mainly influenced by humidity gradients, solar energy, wind speed, soil properties and vegeta-
tion type [12] [13]. Other studies have found that depending on the soil moisture status, lowland watersheds
were highly responsive to rainfall by producing more frequent and greater amounts of runoff, with peak flow
rates also depending on the surface depressional storage [8]. Furthermore, some rainfall-runoff simulation mod-
els have demonstrated that the degree of water saturation in the soil prior to a precipitation event (the antecedent
soil moisture condition, AMSC) correlates with the portioning of the event rainfall into infiltration and stream
flow [14]-[16].
Seasonal climate variability affects both the soil moisture and the characteristics of the storm events that in
turn affect the runoff generation pattern [11]. Some of these characteristics are rainfall intensity, frequency, du-
ration, and direction [17]. Although the antecedence soil moisture condition of the watershed influences water

66
 D. T. Mengistu et al.

available for runoff, evapotranspiration and infiltration via soil water storage, it is highly variable and difficult to
measure [18].
The main objective of this study is to evaluate the performance of nine rainfall-runoff models (from simple to
complex) whether model complexity is important for flow estimation in the context of data scarce areas in Afri-
ca. There is a tendency to use complex models such as SWAT for simple purpose of flow estimation in many
African watersheds. The amount of spatial and temporal data sets required to calibrate complex models such as
SWAT doesnt warranty the purpose if the purpose of the modeling is simply to estimate flow for water re-
sources development application. Therefore, this study attempts to revitalize application of simple rainfall-
runoff hydrological models for use in water resources application. The study is conducted on Gilgel Abbay cat-
chment of Blue Nile basin (Ethiopia) using 21 years (1980-2000) historical data of daily rainfall, temperature
and stream flow. Finally we compare the results from the individual rainfall-runoff models with several methods
of combining the outputs to investigate if there is added value in making ensemble means.
RE
2. The Study Area
The Gilgel Abbay catchment (4051 km2) is one of the largest among the four main sub-catchments in Lake Tana
sub basin of the upper Blue Nile basin, Ethiopia (Figure 1), providing about 60% of the lake inflow. It is located
TR
A
CT

(a)
ED

(b)

67
D. T. Mengistu et al.
RE

(c)
TR
Figure 1. (a) Distribution of meteorological stations and gauging station, (b) land use and (c) soil map in the
Gilgel Abbay catchment.

at 1044'N latitude and 3723'E longitude. The catchment includes the two gauged sub-basins; Upper Gilgel
Abbay (1664 km2) and Koga (307 km2), see Figure 1, with elevation ranging from 1787 m to 3518 m. The to-
pography is rugged in the southern part of the catchment and the periphery to the west and southeast, while the re-
maining part is a typical plateau with gentle slopes. The soil is dominated by clays and clay loams (Figure 1(c)).
A
The dominant land use units are agricultural (65%) and agro-pastoral land (33%) [19], among this rainfed agricul-
ture is the predominant cover of the Upper Gilgel Abbay (74%) and Koga (64%) sub-catchments (Figure 1(b)).
Upper Gilgel Abbay has its main rainy season between June and September, receiving about 70% to 90% of
the annual rainfall during this season [20]-[22]. Annual area average rainfall (1964-2005) ranged from 834 to
CT
2106 mm, with July representing the wettest month (356 mm) and January representing the driest month (3 mm)
on average. Rainfall observations indicate significant spatial variability in rainfall following the topography,
with decreasing amounts from south to north. Long-term (1980-2000) minimum and maximum annual air tem-
perature values recorded at Merawi (2020 m.as.l.) and Dangla (2180 m.as.l.) stations ranged from 11C to
37.7C, respectively with temporal temperature variations being small throughout the year [19].

3. Data Assessments and Method

ED
3.1. Historical Data
We used daily rainfall and temperature data (1980-2000) from seven and four stations, respectively (Figure
1(a)). Daily Stream flow data for Gilgel Abbay is taken from the then Ministry of Water and Energy (now Min-
istry of Water, Irrigation and Electricity). The meteorological data of rainfall and temperature is obtained from
National Meteorological Agency (NMA). Missing values in rainfall and temperature were treated using the
SWAT built-in weather generator developed by [23]. The daily area representative precipitation was calculated
using thiessen polygon method.

3.2. Event Selection

We selected storm events based on discharge rates greater than 1.84 m3/s (minimum mean discharge separated
from base flow per day),where the base flow is calculated by using variable storage method, total mean daily area
averaged rainfall values greater than 2.3 mm; and a period of 48 hours or more in between rain events (Table 1). A
rain event is defined as the rainfall amount which contributes runoff event in a period of 48 hours [11]. These
criteria were selected to identify detectable single peak events and minimize influence of prior rainfall on

68
 D. T. Mengistu et al.

Table 1. Basic hydrological characteristics for analyzed runoff events (n = 100) and t-test results calculated for peak rate,
runoff, rainfall, R/P, rain previous 5- and 30-days, SD = standard deviation.

Date Begin flow (m3/s) Rain in mm Runoff (mm) R/P Q peak Rain 5-day Rain 30-day
11-Feb-80 0.45 4.4 0.363 0.08 7.001 0.00 4.4
17-Apr-80 0.00 37.5 0.531 0.01 6.17 12.20 16.1
22-Jul-80 96.90 167.3 19.83 0.12 412.617 40.4 309
8-Aug-80 44.63 149.9 16.72 0.11 321.92 41.4 350.5
8-Oct-80 11.92 34 3.53 0.10 93.251 27.1 106.8
9-Jun-81 4.22 17 1.43 0.08 49.143 16 57.2
19-Aug-81 0.48 154.7 17.14 0.11 330.124 40.1 546.4
RE
13-Sep-81 46.39 54.21 11.62 0.21 239.536 28.2 183.01
3-Oct-81 21.58 60.6 8.88 0.15 171.042 2.9 119.51
25-Nov-81 1.68 5.7 0.569 0.10 18.932 0.20 8
5-Jan-82 0.21 19.9 0.320 0.02 6.44 0.00 19.9
20-Aug-82 13.69 141.8 14.68 0.10 282.703 68.2 342.9
8-Sep-82 23.73 71.7 9.89 0.14 230.509 35.1 269.5
TR
12-Oct-82 32.34 39 5.41 0.14 104.096 8.5 117.8
11-Mar-83 0.11 3.32 0.53 0.16 12.112 1.21 3.32
18-Jun-83 6.03 50.6 3.15 0.06 60.661 6.4 68.9
17-Jul-83 45.14 78.6 12.68 0.16 248.767 25.9 151.9
23-Aug-83 42.88 185.6 16.86 0.09 324.641 95.9 478.5
21-Mar-84 0.10 10 0.142 0.01 3.256 0.50 11.6
31-May-84 4.13 42.91 2.94 0.07 56.664 20.7 57.31
A
12-Jul-84 51.74 116 13.91 0.12 267.847 28.7 412.8
8-Aug-84 19.15 146.3 13.78 0.09 330.124 60.1 378.6
11-Sep-84 11.57 113.4 13.66 0.12 263 25.1 258.6
CT
27-Mar-85 0.00 12.6 0.199 0.02 3.833 0.00 12.6
8-Sep-85 25.80 114.21 13.16 0.12 253.46 29.4 302.6
12-Nov-85 0.89 10.6 0.531 0.05 13.335 0.00 50.2
5-Apr-86 0.25 74.6 0.142 0.00 2.733 3.30 19.5
14-Jun-86 14.80 40.2 2.37 0.06 62.72 31.1 61.4
17-Jul-86 21.41 85.3 7.76 0.09 241.825 55.9 264.3
ED
17-Aug-86 26.06 92 10.63 0.12 230.509 29.4 276.9
22-Oct-86 5.61 28.2 2.37 0.08 142.6 21.1 76.6
20-Nov-86 1.17 17.1 0.589 0.03 16.46 0.00 17.1
14-Jan-87 0.14 3.2 0.199 0.06 5.157 3.20 3.2
5-Mar-87 0.10 44.2 0.320 0.01 3.635 0.00 0
29-Apr-87 0.43 13.2 0.232 0.02 4.466 0.00 27.7
14-Jun-87 31.11 82.5 7.49 0.09 144.289 46.4 147.5
1-Jul-87 41.68 89.1 7.49 0.08 253.46 46.7 172.5
19-Aug-87 25.90 127.2 11.97 0.09 234.997 34.2 264.9
15-Sep-87 53.33 73.4 12.20 0.17 234.997 46.8 224.7
24-Feb-88 0.17 25.6 0.268 0.01 5.157 0.00 30.3
12-Apr-88 0.08 2.3 0.096 0.04 2.118 0.00 2.3
12-Aug-88 46.81 107.4 14.42 0.13 277.699 32.4 357.5

69
D. T. Mengistu et al.

Continued
11-Sep-88 11.95 83.8 10.84 0.13 219.511 47.6 235.9
9-Oct-88 3.62 98.2 5.55 0.06 140.922 51 197.4
13-Mar-89 0.00 16.6 0.255 0.02 3.442 0.00 2.8
17-Jul-89 27.75 97.4 19.83 0.20 406.365 62.8 161.9
13-Aug-89 17.98 137.3 16.30 0.12 438.188 54.2 441.6
7-Sep-89 25.72 62.6 12.09 0.19 234.997 28.4 226.7
7-Nov-89 4.23 7.8 0.930 0.12 26.332 3.10 6.7
15-Mar-90 0.20 18.9 0.133 0.01 2.901 12.80 19.9
14-Jun-90 0.35 9.9 0.96 0.10 90.648 0.8 24.1
RE
22-Jul-90 38.67 136.3 11.29 0.08 248.767 55.7 255.8
13-Sep-90 20.90 70.1 8.03 0.11 277.699 26.3 171
7-Oct-90 22.55 54.2 4.57 0.08 121.692 2.3 133.9
11-Aug-91 23.44 155.3 16.86 0.11 341.251 62.6 416
22-Sep-91 13.15 110.7 12.68 0.11 352.596 83 265.4
6-Oct-91 7.55 79.9 4.57 0.06 101.319 55.6 254.1
TR
12-May-92 0.16 13.7 2.28 0.17 43.911 13.7 152.4
7-Jul-92 56.32 80 10.41 0.13 200.503 15.6 93.4
14-Oct-92 13.94 72.2 7.85 0.11 151.163 36.5 183.4
8-Nov-92 6.89 49.7 4.443 0.09 85.573 6.10 75.9
7-Dec-92 2.00 17.9 0.671 0.04 19.451 0.00 20.6
5-Sep-93 18.47 121.2 19.83 0.16 381.915 74.8 198.3
A
9-Oct-93 49.80 73.9 10.41 0.14 200.503 23.2 136.3
7-Nov-93 8.98 75.3 1.867 0.02 40.613 0.00 38.7
27-Feb-94 0.21 20.4 0.210 0.01 4.038 20.40 20.4
11-Aug-94 32.74 72.5 16.02 0.22 308.515 17.5 333.7
CT
5-Oct-94 9.75 17.3 1.83 0.11 44.759 4.2 85.2
10-Dec-94 1.51 3.8 0.609 0.16 11.72 0.00 13
25-May-95 0.15 15.5 1.79 0.12 66.96 13.1 76.7
7-Jun-95 26.23 92 6.81 0.07 131.1 32.1 148.5
7-Sep-95 0.00 77.3 9.58 0.12 290.308 14.1 278
17-Oct-95 3.00 33.8 1.94 0.06 50.964 33.8 90.9
ED
3-Nov-95 2.72 17.2 1.122 0.07 21.609 1.50 18.7
23-Jan-96 0.43 6.2 0.199 0.03 4.69 0.00 6.2
3-Feb-96 0.44 2.5 0.125 0.05 3.076 0.00 2.5
6-Jun-96 5.18 57.4 5.19 0.09 190.377 35.1 224.2
3-Aug-96 60.60 124 16.72 0.13 355.466 38.1 324.7
11-Sep-96 18.51 135.8 11.51 0.08 253.46 31.2 273.6
28-Nov-96 0.00 85.5 3.257 0.04 97.236 5.30 108.3
20-Jun-97 8.65 127.2 6.72 0.05 142.6 44.4 253.7
9-Jul-97 3.77 92 9.58 0.10 258.204 77.5 233.4
15-Aug-97 33.47 100.8 18.61 0.18 358.35 54.9 263.2
22-Sep-97 29.26 124.2 8.78 0.07 202.565 59.5 180.1
4-Aug-98 20.95 146.5 14.16 0.10 298.031 42.7 358.1
6-Sep-98 47.86 90.1 10.63 0.12 241.825 27.7 227

70
 D. T. Mengistu et al.

Continued
12-Oct-98 28.68 154.5 11.51 0.07 221.685 72.5 269.31
5-Jan-99 0.38 26.3 0.221 0.01 4.466 0.00 26.3
27-Apr-99 0.00 14.6 0.379 0.03 7.291 12.30 16.5
12-Jun-99 16.60 137.3 6.32 0.05 174.812 38.9 127.21
11-Jul-99 19.92 126.2 12.80 0.10 285.225 43.2 188.9
30-Aug-99 20.56 115.3 14.29 0.12 298.031 60.1 264.6
5-Sep-99 18.44 120.9 9.78 0.08 190.377 33.9 260.9
4-Oct-99 36.74 83.8 9.58 0.11 234.997 29.1 140.8
20-Nov-99 3.43 26.6 1.151 0.04 40.613 0.00 26.6
RE
11-Dec-99 1.57 3.3 0.629 0.19 12.112 0.60 16.4
23-Jul-00 14.51 83.7 11.29 0.13 223.872 62.6 255.5
17-Aug-00 26.63 154.5 19.83 0.13 381.915 56.1 314.6
10-Oct-00 15.02 91.1 10.30 0.11 251.107 46.7 179.1
14-Dec-00 2.04 7.11 0.550 0.08 13.335 0.00 9.11
Mean 69.77 7.18 0.09 157.80 26.90 159.12
TR
Median 72.95 6.77 0.09 147.73 26.70 148.00
Standard Deviation 49.76 6.21 0.05 128.55 24.00 131.89
Skewness 0.28 0.39 0.20 0.25 0.58 0.57
Range 183.30 19.73 0.22 436.07 95.90 546.40
Minimum 2.30 0.10 0.00 2.12 0.00 0.00
Maximum 185.60 19.83 0.22 438.19 95.90 546.40
A
Confidence interval (95.0%) 9.87 1.23 0.01 25.51 4.76 26.17

several peaks. Multiple peak events were excluded from this analysis in order not to complicate the identifica-
tion of storm duration and total storm volume. The same method was used by [7] [8] [24]-[26]. Single event
CT
peak discharge can be modeled easier as described by [27]. The start of the runoff event is the rainfall available
for runoff after infiltration and other abstractions have been accounted for.
The antecedent precipitation index (API) was calculated as a measure of the available soil moisture content
(ASMC):
N
API=
t Pt + K t i API t i (1)
i =1
ED
where Pt is the area averaged precipitation at day t, N the number of days prior to the start of the runoff event
and K t is the recession constant calculated as the product for three individual constants, i.e. K = K s K i K g
where Ks = 1.0, Kj = 0.43 and Kg = 1.33 are recession constants associated with surface storage, inflow and
ground water flow respectively.
Total event runoff volume is calculated by dividing the daily runoff by the watershed area to obtain runoff
depth in mm. The runoff volume was considered to be the area under the hydrograph from the start of the event
(defined above) until it reaches base flow level or until the next rise starts. Using the above definition of an
event, the average duration of the storm events was 12 days. Since the rainfall data was in daily basis, there was
sometimes challenge to identify an exact amount of rainfall amounts resulting in a particular runoff event. The
begin flow for each event was calculated as the flow at which direct runoff starts. R/P was calculated using the
accumulated runoff amount and corresponding accumulated area averaged rainfall over the runoff event.
Based on the above criteria, we selected 100 storm events with various parameters showing the hydrologic cha-
racteristic of the study area (Table 1) [11]. Used 51 storms for only a 10-year (1964-1973) period that included
events with multiple peaks also. In other studies [25] used 75 events, [28] used 29 events, [7] used 23 events,
[26] used 4 - 9 events for each watershed, and [24] used 55 storm events for various study areas. We, therefore,

71
D. T. Mengistu et al.

believe that analysis of 100 storm events is adequate for testing event based rainfall runoff relationship.

3.3. Regression Analysis

A simple linear regression analysis was used to determine the relationship between runoff and other variables.
Equations of the linear regression lines and their parameters were tested for statistical significance at the 5%
level ( = 0.05) using a two-tailed t-test. In addition, a standard stepwise regression analysis was used to ex-
amine the effects of different factors on the runoff.

3.4. The Galway River Flow Forecasting System (GFMFS)

The Galway Flow Modeling Forecasting System (GFMFS) is software packages developed at Department of
Engineering Hydrology, National University of Ireland, Galway [29]. A brief descriptions of the GFMFS soft-
RE
ware package may be found in [30]-[44]. The GFMFS models may be run in updating mode or simulation mode,
depending on the choice of model. In updating mode the models use the lagged observed discharge along with
precipitation input to simulate the streamflow simulation hydrograph to the corresponding observed hydrograph.
In contrast, non-updating (simulation) mode uses the input of precipitation and excluding the use of the recently
observed discharge as model inputs.
The five major hydrological applications of the GFMFS packages are 1) Modeling by calibration and valida-
tion for simulation of continuous river flow, 2) Estimation of river flow i.e. generation of synthetic flow series,
TR
using inflow data and a calibration model. 3) Modeling by calibration and verification, for lead-time forecasting
in absence of QPF (quantitative precipitation forecasts), 4) Modeling by calibration and verification, for lead-
time forecasting using QPFs, 5) Real time flow forecasting using models and techniques chosen in step 1).
and/or 3) or 4).
Modeling by calibration and validation for simulation of continuous river flow in step 1 was applied for this
study. The models implemented in this study (from the GFMFS package) are six rainfall runoff models that de-
pend only on rainfall and runoff relationships (single input) and one complex hydrological models that uses
more than one rainfall input and several conceptual parameters in the model formulation. These models are the
A
Simple Linear Model (SLM), Non-Parametric Simple Linear Model (NP-SLM), Linear Perturbation Model
(LPM), Non-Parametric Linear Perturbation Model (NP-LPM) and Linearly Varying Gain Factor Model
(LVGFM), a non-linear black box type artificial Neural Network model (ANN). The complex hydrological
models accessed from GFMFS package include Soil Moisture Accounting and Routing Model conceptual model
CT
(SMAR, we tested the three variants, namely the SMARG, SMAR-NC1 and SMAR-NC2 versions, but only the
SMARG is reported on here since the results were similar for the two other versions) (see Table 2).

3.5. Other Complex Hydrological Models

In addition to the above hydrological models we further apply a conceptual semi distributed (HBV) and a phys-
ically distributed (SWAT) model in the study. The HBV model [45]-[48] is a rainfall-runoff model, which in-
ED
cludes conceptual numerical descriptions of hydrological processes at the catchment scale. The general water
balance can be described as:
d
Rday ETa Qsurf= SW + Wseep + Qgw (2)
dt
where SW is the soil water content (mm), Rday is the daily precipitation, Qsurf is the amount of surface runoff/

Table 2. Descriptive statistics results the run off coefficient for R/P for wet and dry events and wet and dry conditions on
both 5-day prior rainfall-values correspond to the significant difference in periods and conditions.

Parameters n. (no. of events) R/P ratio ranges Mean R/P SD () COV P-value
Wet period 69 0.05 - 0.22 0.11 0.04 0.34 0.25
Dry period 31 0.0 - 0.19 0.05 0.05 0.97 0.03
Wet condition (5-day prior) 73 0.0 - 0.22 0.1 0.05 0.44 0.01
Dry condition (5-day prior) 27 0.01 - 0.19 0.07 0.05 0.74 0.46

72
 D. T. Mengistu et al.

streamflow, ETa is the amount of actual evapotranspiration, Wseep is the amount of water entering the vadose
zone from the soil profile and Qgw is the amount of ground flow.
The HBV model can be used as a semi-distributed model by dividing the catchment into sub basin. Each sub
basin is then divided into zones according to altitudes and the elevation zones which are further divided into
different vegetation zones (e.g. Lakes, forested and non-forested areas).
The model is normally run on daily values of rainfall and air temperature, and daily or monthly estimates of
potential evaporation. Observed streamflow data were used for calibration of the model through optimizing the
embedded parameters.
SWAT (Soil and Water Assessment tool, version SWAT2005) is a physically based, distributed parameter
model which operates on daily time step and uses physiographical data such as elevation, land use and soil
properties as well as meteorological data and, stream flow data for calibration. It is computationally efficient for
use in large watersheds, and is capable of simulating the impact of land management practices [49].
RE
The effects of spatial variations in topography, land use, soil and other characteristics of watershed hydrology
is incorporated by dividing a basin into several sub-basins based on drainage areas of tributaries and is further
divided the sub-basins into a number of hydrological response unit (HRUs) within each sub-basin, based on land
cover and soils. Each HRU is assumed spatially uniform in terms of land use, soil, topography and climate. The
subdivision of the watershed enables the model to reflect differences in evapotranspiration for various crops and
soils. All model computations are performed at the HRUs level [50].
The fundamental hydrology of a watershed in SWAT is based on the following water balance equation [50].
TR
SW
= Rday Qsurf ETa Wseep Qgw (3)
t
where SW is the soil water content (mm), Rday is the amount of precipitation on (mm), Qsurf is the amount of sur-
face runoff/streamflow (mm), ETa is the amount of actual evapotranspiration (mm), Wseep is the amount of water
entering the vadose zone from the soil profile (mm), and Qgw is the amount of ground flow (mm). Detail descrip-
tions of the different model components can be found in [51] [52]. Like HBV model SWAT used observed
A
streamflow data for calibration purpose to optimize high to very high sensitive parameters.
Table 5 rank the models according to complexity from simple to complex. SLM is the simplest followed by
the LPM in their non-parametric and parametric forms and the LVGFM. Non parametric and parametric assume
the observations must be independent, the observations in non-parametric forms must be drawn from normally
CT
distributed populations, these populations must have the same variances however in parametric form variable
distributions have been.
The three models (SLM, LPM and LVGFM) are system-linear model in structure, and an ordinary least
squares solution is used for estimation of the pulse response function except for the parametric forms where the
parameters were optimised. HBV and SWAT models are the most complex with a complicated mathematical
procedure to be processed during simulation.

3.6. Combination of Outputs

ED
A particular rainfall-runoff model may have been selected from among a number of competing alternative mod-
els, based, perhaps, on its accuracy, its familiarity to the user, its ease of use, the type of the catchment, and the
available data. However, there may be a potentially danger in relying entirely on one substantive rainfall-runoff
model (a sample of one) since it is unlikely to perform satisfactorily at all time or under all conditions (e.g. per-
haps not all of its structural assumptions are valid or the conditions under which it is assumed to operate are not
entirely fulfilled. The method of combination of outputs from each of the model applied for the study area are
carried out in the concept that the individual model assumed to capture some physical characteristics of the
study area. We use three different methods of combining outputs (MOCT): The Simple Average Method (SAM),
the Weighted Average Method (WAM), and the Neural Network Method (NNM).

3.6.1. The Simple Average Method (SAM)

The simple average method (SAM) is the simplest method of combining the outputs of different individual
models. Given the estimated discharges from N rainfall-runoff models, a combined estimate of the discharge of
the ith time period, using the SAM, is given by

73
D. T. Mengistu et al.

1
Q Ci = j =1 Q ji
N
(4)
N
where Q Ci : is the combined estimate of the discharge of the ith time period, N is the number of rainfall runoff
models and Q ji the average simulate discharge for time period I from rainfall-runoff model j.

3.6.2. The Weighted Average Method (WAM)

When some of the individual models selected for combination appear to be consistently more accurate than oth-
ers, in which case the use of the simple average method for combination can be quite inefficient [53], the use of
a weighted average would be considered.
The weighted average method (WAM) for combining the estimated model outputs, in the case of N rain-
fall-runoff models, may be expressed as [54]
RE
= j =1 a j Q ji + ei
N
Qi (5)

where Qi is the combined discharge and aj is the weight assigned to the jth model estimated discharge. ei is the
combination error term.
The above equation may alternatively be expressed in matrix notation as
=
Q PA + E (6)
TR
where P is the input matrix defined by
Q1,1 Q2,1 QN 1,1 QN ,1

Q1,2 Q2,2 QN 1,2 QN ,2


1, k 1
Q Q2, k 1 QN 1, k 1 QN , k 1
Q Q QN 1, k QN , k
1, k 2, k
A
Q = ( Q1 , Q2 , Q3 , , Qk 1 , Qk ) is the output vector, A = ( a1 , a2 , a3 , , ak 1 , ak ) is the weight vector and
T

E = ( e1 , e2 , e3 , , ek 1 , ek ) is the combination error vector, T denotes the transpose of the vector and k is the total
number of observations.
The preceding equation can be perceived as a multiple linear regression model. Thus, it can be readily shown
CT
that the ordinary least squares estimate of the weight vector is given by;

( )
1
A = PT P P T Q. (7)

In the WAM, the sum of the weights ai is normally constrained to be equal to unity, that is

i =1 ai = 1.
N
(8)
ED
The main rationale behind constraining the sum of the weights to unity is that if the models included in the
weighted average are unbiased, i.e. having a zero mean output error term, then the weighted average combined
forecast is likewise unbiased [53].
In the case where the sum of the weights is constrained to equal unity, it can be shown using the method of
constrained least squares (CLS) that the estimate of the weights vector
Acls (9)
is given by
1 1
=Acls p T P p T Q + b (10)
2
where b is the unit vector (i.e. all of its scalar components are unity) having the same dimension as the parame-
ters vector A and is the Lagrangian multiplier which is given by

( ) (1 b ( p p ) )
1
( )
1 1
=A 2 bT p T p b T T
pTQ . (11)

74
 D. T. Mengistu et al.

Alternative techniques, other than least squares, for estimating the weights ai have also been used, e.g. by
considering the covariance of the forecast errors of the individual models being considered.

3.6.3. The Neural Network Method (NNM)

The neural network method (NNM) provides an alternative to the simple average (SAM) and the weighted av-
erage methods (WAM) for combining outputs from different models. Neural networks are applied, in the
GFMFS package, in the context of providing a non-linear function mapping of the simulated flows. Using a
multi-layer feed forward neural network [44].

4. Results and Discussion

4.1. Relationship between Precipitation and Runoff
RE
A linear regression analysis revealed a significant ( = 0.05) correspondence between mean event runoff and
mean event rainfall (Table 2). The coefficient of determination (R2) was 0.62 and 0.33 for the wet (69 events) or
dry (31 events) (Table 3). A dry event is defined as an event where precipitation is zero the day before the event
started. Figure 2(a) and Figure 2(b) shows the scatter plot between the mean event runoff and mean event rain-
fall for the wet and dry case. The average runoff coefficient (R/P) was 0.11 and 0.05 for the wet and dry events
respectively. R/P ranged from 0.01 to 0.19 with a coefficient of variation (CV, which is the ratio of the standard
deviation to the mean) of 0.97 for the dry events. This was almost three times higher than during the wet events.
TR
The higher relative variability observed during the dry period may be explained by the soil being close to satura-
tion for the wet events, thus the relative variability between the different events will be small. Thus, event aver-
aged precipitation is well correlated with event averaged runoff in the wet case, but not in the dry case. The
mean monthly water balance plot for the study period for the Gilgel Abbay watershed shows the cycle of rainfall
and runoff in relation to PET as estimated by the Penman Monteith method (Figure 3). A seasonal cycle is also
seen in PET rates which start increasing in September and peaks during the months of March. The plot in Fig-
ure 3 suggests that the difference between rainfall and runoff is close to the PET values during the wet periods
with unlimited soil moisture calculated for the watershed (Figure 3). Thus, it is important to examine alternative
A
relationships that include other important variables, such as PET or water table depth as a surrogate for the
ASMC, and their interactions for understanding rainfall-runoff dynamics.

4.2. Relationship between the Antecedent Precipitation Index (API) and Runoff
CT
Using the precipitation information only on the day before the beginning of the event to classify the event into

Table 3. Regression statistics results for runoff-rainfall relationships for wet and dry periods, and wet and dry conditions
based on 5 prior rainfall.
Parameters Regression equation r2 P-value Intercept (P-value) Slope (P-value)
Wet period Runoff = 1.853P + 25.265 0.62 0.0 0.16 0.0
ED
Dry period Runoff = 0.024P + 0.15 0.33 0.0 0.45 0.0
Wet conditions (5-day prior) Runoff = 0.156P + 4.026 0.38 0.0 0.0 0.0
Dry conditions (5-day prior) Runoff = 0.062P + 0.463 0.31 0.0 0.0 0.0

Figure 2. Event rainfall-runoff relationship for (a) wet (May-October = 69) and (b) dry (November-April = 31) periods.

75
D. T. Mengistu et al.

wet and dry may means that we are not taking into account information about the soil moisture content prior to
the event. A crude way of getting information about this is to use the API (Equation (3.1) to define wet and dry
cases instead of the precipitation the day before the start of the event. Table 3 shows the regression between the
runoff and the 5-day API. API 5-day prior to event are correlated with event averaged runoff both for dry (27
events) and wet (73 events) cases. Results shows a significant (R2 = 0.38) relationship between runoff and 5-day
API (R2 = 0.38 and R2 = 0.31 for the wet and dry conditions, respectively, Table 3 and Figure 4).
R/P is ranging from 0.01 to 0.22 for the 5-day wet conditions with a CV of 0.44 and a CV of 0.74 for the dry
period with the ratios ranging from 0.01 to 0.19 (Table 2). These results suggested that seasonal event rain-
fall-runoff dynamics in the watershed may have been complicated due to other factors such as rainfall intensity
and its aerial variability, spatial distribution (for watershed of this scale) of soil type and their properties, and
depth to water table (i.e. soil water storage volume). These situations have been clearly examined through detail
investigation of eight storm events among 100 ones.
RE
Events selected either have very small or high R/P. For example, on day 17-Apr-80 a rainfall amount of 37.7
mm produced a runoff response of only 0.531 mm (R/P = 0.01), whereas for the storm event at day 2-Oct-94
17.3 mm produced 1.83 mm runoff (R/P = 0.11). The later event occurred directly following the return to base
flow condition of the prior event and had a higher peak flow value, perhaps caused by surface runoff and shal-
low subsurface flow for a larger ASMC (already saturated conditions). For the event on day 17-Apr-80 with no
previous rain 5 days prior, it is likely that the high ET rate during the dry period caused a large decrease in
stream flow. Compare this with the event on day 2-Oct-94 where only 17.3 mm rain (5-day prior) produced a ra-
TR
tio of 0.11, whereas a ratio of 0.01 was generated for day 17-Apr-80 with no rain in the 5-day prior. Apparently
the near-term soil moisture condition played a larger role in determining the runoff response during the dry pe-
riod rather than a longer-term condition (30-day prior rainfall). The amount and rate of runoff will be dependent
upon these key controlling factors (i.e. soil hydrologic properties, soil moisture storage, and rainfall) that vary
spatially and temporarily [7]. However, we hypothesized that additional information on rainfall intensity, water
table positions, and PET would help to more accurately determine the runoff and change in soil water storage
processes.
A
4.3. Relationship between the Antecedent Precipitation Index (API) and Begin Flow
As described in Section 4.2 above, the equation of the simple linear regression model showed the 5-day API not
significantly related to the begin flow during either the wet or dry period. Thus the 5-day API is unrelated to the
CT
begin flow as shown in Table 4.
ED

Figure 3. Mean monthly rainfall-runoff and PET for 1980-2000 period. PET was calculated using Penman monieth method.

Figure 4. Rainfall-Runoff relationship for events wet (n = 73) based on 5-day API.

76
 D. T. Mengistu et al.

Table 4. Results of stepwise regression analysis evaluated for all rainfall events (n = 100), and wet and dry periods.

Variable Begin flow Rainfall 5-day rainfall

All storm events(n = 100)
Runoff r2 = 0.51, P = 0.0 r2 = 0.76, P = 0.0 r2 = 0.23, P = 0.0
2 2
R/P r = 0.26, P = 0.0 r = 0.08, P = 0.0 P > 0.07
Wet Period
Runoff r2 = 0.29, P = 0.0 r2 = 0.62, P = 0.0 r2 = 0.12, P = 0.004
R/P r2 = 0.09, P = 0.01 P > 0.25 P > 0.69
Dry period
Runoff r2 = 0.41, P = 0.0 r2 = 0.33, P = 0.0 P > 0.57
RE
R/P P > 0.10 r2 = 0.13, P = 0.05 P > 0.30

4.4. Multilinear Relationship between Begin Flow, API, Event Precipitation and Runoff
Results from the above analysis showed runoff was significantly related to rainfall amount and the API the ini-
tial flow rate (begin flow) (see Table 4). The 5-day API prior to the event had impact on runoff generation, but
this was not strong. In the wet period, the runoff was correlated with initial flow due to previous rainfall condi-
tion and event rainfall amount. However, we did not find this was the case for the dry period at 5-day prior rain-
TR
fall.

4.5. Model Results and Performance

In the above analysis we had seasonal relationship using event based analysis. From the result, it has been noted
a big difference in relationship of rainfall runoff which are explained by various variables such as mean begin
flow, API and soil moisture. The importance of doing simulation of various models in this section is that to de-
A
rive the advantage of individual model outputs from their particular consideration of the study area. GFMFS
software packages and other hydrological models have been applied to simulate rainfall-runoff relationship us-
ing observed daily rainfall and streamflow data for a period of 1980-2000.
The comparison among each model output is made using three evaluators: The coefficient of variation (R2),
CT
the Nash Sutcliffe Efficiency (NSE) and the bias (simulation minus observations divided by the observations
in %). The 1980-1992 data were used for model calibration and the remaining data from 1993-2000 used for
validation.

4.5.1. Simulation Mode

The performance of the SLM is inferior to that of all other models. The LVGFM, which is a modification of the
SLM, incorporating an element of linear variation of the gain factor (Gt, see appendix) with the catchment wet-
ED
ness index at each time step, performs consistently better than the SLM where the surface storage of the catch-
ments might have affected the results. As Gilgel Abbay is characterized by strong seasonality, the LPM in
simulation mode, with its inherent component of seasonal variation, outperforms the LVGFM and SLM. From
Table 6 and Figures 5(a)-(i), we see that in simulation mode, the performance of the ANN model is clearly the
best followed by LPM during calibration and validation with R2 of 87.8% and 76.3%, respectively. This implic-
itly shows the non-linearity of rainfall-runoff relationships can be well be handled by systems type black box
models without using complex conceptual or physically based models. The performance of the ANN model is
R2 89% and 85% during calibration and validation period, respectively. In the case of SMAR model, the pa-
rameters lumping applied to the study catchment which has diverse topographic variations may have influenced
the performance of the model. Following LPM and ANN, the SMAR of SMARG variant explains the rain-
fall-evaporation-runoff relationships with 80.5% and 70.7% of R2 during calibration and validation, respectively.
As Table 5, the overall performance of the systems type black box models is comparable and in the case of
ANN better than the conceptual or physically based hydrological models. Therefore, as far as estimation flow
either in continuous or event based is concerned, systems type black box model with simple rainfall and runoff
input can be adequate for water resources development purposes.

77
D. T. Mengistu et al.
RE
TR
A
CT
ED

Figure 5. Scatter plots for various models [from (a) to (i)] for upper Gilgel Abbay river basin.

78
 D. T. Mengistu et al.

Table 5. List of models ranking from simple to complex in terms of increased mathematical procedures involved in the
model to be processed.

Complexity
Model Type mode Description
ranking
Empirical Simulation and Non-parametric simple linear model. A linear time-invariant
NP-SLM 9
black-box non parametric relationship between the total rainfall Ri and the total discharge
Parametric simple linear model.
Empirical Simulation and
P-SLM 8 The linear transfer function type representation of the transformation
black-box parametric
process of the input series to the output series for discrete data intervals
Empirical Simulation and Non-parametric linear perturbation model. The model uses the seasonal
NP-LPM 7
black-box non parametric information of the observed rainfall and discharge series.
Parametric linear perturbation model. The linear transfer function type
representation of the transformation process of the departures of the
RE
Empirical Simulation and
P-LPM 6 values of the input series from their respective seasonal means to the
black-box non parametric
departure of the values of the output series from their respective
seasonal means for discrete data intervals
Linearly varying gain factor model.
Empirical Simulation and
LVGFM 5 The model is non-linear, can be viewed as a multiple linear regression
black-box non parametric
model
Artificial neural network model.
Empirical The model is the multi-layer feed forward network consists of an input
TR
ANN 4 Simulation
black-box layer, an output layer and only one hidden layer between the input and
the output layers.
Physically
Soil moisture accounting and routing model.
inspired Simulation and
SMAR 3 It is rainfall-evaporation-runoff model with three variants; SMARG,
conceptual non parametric
SMAR-NC1, SMAR-NC2
model
Hydrologiska Byrans Vattenbalans-Avedlning (Hydrological Bureau
Conceptual Water balance-section). It is considered as semi-distributed conceptual
HBV 2 Simulation
A
model model and possible to run the mode separately for several sub basins
and then add the contributions from all sub basins.
Physical
Soil and Water Assessment Tool. It is physically based distributed
SWAT distributed 1 Simulation
parameter model which operates on daily time step.
model
CT
4.5.2. Updating Mode
In updating mode, LPM consistently performed the best of all other models. It accounted for more than 90% and
85% of the initial variance during calibration and validation period, respectively. Even, the simple models like
P-SLM and P-LPM performs better than HBV and SWAT shown in Figures 5(a)-(i).
Generally, the updating models are better than the simulation mode models (with some exceptions) and the
updating mode models has a lower reduction in R2 and ENS in the validation period compared to the calibration
ED
period (Table 7).

4.5.3. Combining Outputs from Different Models

The method of combined outputs was used to the results of the five substantive models included in GFMFS
software packages, both in simulation and updating mode. Three techniques for combining the estimated outputs
of different models were conducted namely, the simple average method (SAM), the weighted average method
(WAM) and the neural network method (NNM).
SAM, WAM and NNM account with R2 values of 83.4, 90.4 and 90.42 percent during calibration and 79,
85.5 and 84.8 percent during validation period, respectively in simulation mode; and in the case of updating
mode the values of R2 are 90.2, 91.5 and 91.4 percent during calibration and 85, 85.3 and 85 percent during
validation, respectively (Table 6 and Table 7). Therefore, combining the outputs of the systems type black box
model has shown increased improvement in the accuracy of the results.

5. Summary and Conclusions

We believe application of complex conceptual or physically based models for simple flow estimation may not

79
D. T. Mengistu et al.

Table 6. The model efficiencies (%) in simulation mode.

Calibration Validation
Model Method 2 2
R Bias ENS R Bias ENS
NP-SLM OLS 64.17 7.81 0.65 51.00 7.66 0.50
NP-LPM OLS 86.94 0.02 0.87 77.98 5.70 0.78
LVGFM OLS 73.00 0.78 0.73 57.00 1.99 0.57
SMAR OLS 80.49 3.69 0.81 70.72 4.44 0.71
ANN OLS 89.70 2.35 0.90 85.00 3.44 0.85
HBV Parametric 86.0 12.82 0.70 87.0 10.25 0.71
SWAT Parametric 66.32 0.02 0.62 59.4 9.47 0.55
RE
SAM Non-parametric 83.37 79.16
MOCT WAW Non-parametric 90.41 85.53
NNM Non-parametric 90.42 84.82

Table 7. The model efficiencies (%) in updating mode.

Calibration Validation
Model Method
TR
2 2
R Bias ENS R Bias ENS
P-SLM Parametric 89.81 0.91 0.898 85.33 0.91 0.85
P-LPM Parametric 91.35 0.00 0.91 84.90 2.49 0.85
SAM Parametric 90.21 85.06
MOCT WAW Parametric 91.46 85.32
NNM Parametric 91.42 85.01
A
be always feasible especially in scarcely gauged locations in Africa. On that basis, we compared systems type
black box rainfall-runoff models and other complex models that require inputs beyond rainfall such as SMAR,
HBV and SWAT. The models were compared on the basis of long-term rainfall and stream flow data (1980-
2000) and for 100 selected runoff events for the Gilgel Abbay watershed.
CT
Event runoff is largely a function of rainfall amount. The event rainfall-runoff relationships explained as
much as 62% for the wet periods without incorporating the evaporative demands. Although the relationship
between runoff and rainfall was significant ( = 0.05) for wet and dry periods, it was not as strong as ex-
pected. This suggests that about 38% of the runoff response in this watershed is influenced by other factors
such as its intensity and duration, and to the near-term soil moisture conditions created by accumulated eva-
potranspiration and precipitation balances. Event rainfall-runoff relationships were also affected by 5-day
prior rainfall under wet conditions, suggesting that soil moisture condition is an important element dictating
ED
the hydrologic dynamics of this watershed. However, this was not the case for the dry period indicating that
the rainfall-runoff dynamics was more complex and variable in this system with shallow moderately drain
soil. We argued that this variability is most likely related to rainfall characteristics such as intensity and du-
ration. We confirm that event R/P were significantly higher during the wet period than for the dry periods.
Although peak flow rate relationships with rainfall for both wet and dry periods were also significant, the
wet period relationship was found to be stronger. It was also concluded that the rainfall amount and ASMC
represented by the initial base flow rate were the main controlling factors for event runoff. The results of this
study showed that all event variables (runoff, R/P, and peak flow rates) were controlled by rainfall amounts
and available soil water storage. Future studies should further investigate other hydrologic indicators that af-
fect the runoff response, such as spatial and temporal water table dynamics determined by balances of rain-
fall and ET. Information on depth to water table along with soil drainage porosity is necessary to determine
available subsurface storage and, therefore, the ASMC at given times. Additionally, event rainfall intensity
data are necessary not only to characterize the peak flow rates but also to accurately determine the rainfall
amount responsible for event runoff regeneration and duration of storm events at different seasons and pe-
riods.

80
 D. T. Mengistu et al.

Using continuous simulation models described in Section 3.5 and 3.6, we compared systems type black box
and complex hydrological models in Gilgel Abbay catchment. Though the performance of the nave SLM is
clearly inferior to that of all other models (from systems type black box models), models such as the LPM
and the ANN perform better than the complex hydrological models such as SMARG, HBV and SWAT. For
instance it is shown in Table 6 that the performance of LPM and ANN (both systems type black box) mod-
els evaluated with NSE criteria gives better hydrograph response than the complex models such as HBV and
SWAT. Therefore it confirms that simpler models (that takes only rainfall as input) can surpass their com-
plex counterparts in performance for continuous simulation and reproducing of hydrographs or flow estima-
tion. There is a strong justification, therefore, for the claim that increasing the model complexity, thereby in-
creasing the number of parameters, does not necessarily enhance the model performance. It is suggested that,
in practical hydrology, the simpler models, may still play a significant role as effective simulation tools, and
countries with scarce hydrological data should revitalize application of such systems type black box model-
RE
ling schemes that depend only on rainfall and runoff data sets which could be easily available.
As a concluding statement, results of this study site may be of great importance for regional water manage-
ment and water quality studies, for that matter designing the water related structures such as detention ponds and
restoration efforts. These data will also provide useful insight to explain the variability in storm runoff response
observed for the dry period, for example. Additionally, future rainfall-runoff event analysis study at this site
should take advantage of current monitoring of rainfall intensity data, water table depths, solar radiation, and
other hydro meteorological data, as well as modeling studies for accurately estimating soil moisture and actual
TR
ET that would help to explain the variability in runoff generation.

References
[1] Rajurkar, M.P., Kothyari, U.C. and Chaube, U.C. (2004) Modeling of the Daily Rainfall-Runoff Relationship with Ar-
tificial Neural Network. Journal of Hydrology, 285, 96-113. http://dx.doi.org/10.1016/j.jhydrol.2003.08.011
[2] Minns, A.W. and Hall, M.J. (1996) Artificial Neural Networks as Rainfall-Runoff Models. Hydrological Sciences
Journal, 41, 399-417. http://dx.doi.org/10.1080/02626669609491511
A
[3] Campolo, M., Andreussi, P. and Soldati, A. (1999) River Flood Forecasting with a Neural Network Model. Water Re-
sources Research, 35, 1191-1197. http://dx.doi.org/10.1029/1998WR900086
[4] Markewich, H., Pavich, M. and Buell, G. (1990) Contrasting Soils and Landscapes of the Piedmont and Coastal Plain,
Eastern United States. Geomorphology, 3, 417-447. http://dx.doi.org/10.1016/0169-555X(90)90015-I
CT
[5] Skaggs, R., Gilliam, J. and Evans, R. (1991) A Computer Simulation Study of Pocosinhydrology. Wetlands, 11, 399-
416. http://dx.doi.org/10.1007/BF03160759
[6] Amatya, D., Chescheir, G., Skaggs, R. and Fernandez, G. (2002) Hydrology of Poorly Drained Coastal Watersheds in
Eastern North Carolina. ASAE/CIGR Congress Annual International Meeting, Chicago, 27-29 July 2002, Paper No.
022034.
[7] Slattery, C., Gares, A. and Phillips, D. (2006) Multiple Models of Storm Runoff Generation in a North Carolina Coast-
al Plain Watershed. Hydrological Processes, 20, 2953-2969. http://dx.doi.org/10.1002/hyp.6144
ED
[8] Amatya, D., Skagss, R., Gregory, J. and Hermann, R. (1997) Hydrology of a Drained Forested Pocosin Watershed.
Journal of the American Water Resources Association, 33, 535-546.
http://dx.doi.org/10.1111/j.1752-1688.1997.tb03530.x
[9] Amatya, D., Gregory, J. and Skaggs, R. (2000) Effects of Controlled Drainage on Storm Event Hydrology in a Loblol-
ly Pine Plantation. Journal of the American Water Resources Association, 36, 175-190.
http://dx.doi.org/10.1111/j.1752-1688.2000.tb04258.x
[10] Sun, G., Boggs, J., McNulty, S., Amatya, D., Trettin, C., Dai, Z., Vose, J., La Torre Torres, I. and Callahan, T. (2008)
Hydrologic Processes of Forested Headwater Watersheds across a Physiographic Gradient in the Southeastern United
States. Proceedings of the South Carolina Water Resources Conference, Charleston, 14-15 October 2008.
[11] La Torre Torres, I., Amatya Devendra, M., Sun, G. and Callahan Timothy, J. (2011) Seasonal Rainfall-Runoff Rela-
tionships in a Lowland Forested Watershed in the Southeastern USA. Hydrological Processes, 25, 2032-2045.
http://dx.doi.org/10.1002/hyp.7955
[12] Amatya, D. and Trettin, C. (2007) Annual Evapotranspiration of a Forested Wetland Atershed, SC. ASABE Annual In-
ternational Meeting, Minneapolis, 17-20 June 2007, Paper No. 072222.
[13] Sun, G., Noormets, A., Gavazzi, M.J., McNulty, S.G., Chen, J., Domec, J.-C., King, J.S., Amatya, D.M. and Skaggs,
R.W. (2010) Energy and Water Balance of Two Contrasting Loblolly Pine Plantations on the Lower Coastal Plain of

81
D. T. Mengistu et al.

North Carolina, USA. Forest Ecology and Management, 259, 1299-1310.

http://dx.doi.org/10.1016/j.foreco.2009.09.016
[14] Ye, W., Bates, B.C., Viney, N.R., Sivapalan, N. and Jakeman, A.J. (1997) Performance of Conceptual Rainfall-Runoff
Models in Low-Yielding Ephemeral Catchments. Water Resources Research, 33, 153-166.
http://dx.doi.org/10.1029/96WR02840
[15] Wei, L., Zhang, B. and Wang, M. (2007) Effects of Antecedent Soil Moisture on Runoff and Soil Erosion in Alley
Cropping Systems. Agricultural Water Management, 94, 54-62. http://dx.doi.org/10.1016/j.agwat.2007.08.007
[16] Tramblay, Y., Bouvier, C., Martin, C., Didon-Liscot, J., Todorovik, D. and Domergue, J. (2010) Assessment of Initial
Soil Moisture Conditions for Event-Based Rainfall-Runoff Modeling. Journal of Hydrology, 387, 176-187.
http://dx.doi.org/10.1016/j.jhydrol.2010.04.006
[17] Singh, V. (1997) Effect of Spatial and Temporal Variability in Rainfall and Watershed Characteristics on Stream Flow
Hydrograph. Hydrological Processes, 11, 1649-1669.
http://dx.doi.org/10.1002/(SICI)1099-1085(19971015)11:12<1649::AID-HYP495>3.0.CO;2-1
RE
[18] Sun, G., McNulty, S., Amatya, D., Skaggs, R., Swift Jr., L., Shepard, J. and Riekerk, H. (2002) A Comparison of the
Watershed Hydrology of Coastal Forested Wetlands and the Mountainous Uplands in the Southern US. Journal of Hy-
drology, 263, 92-104. http://dx.doi.org/10.1016/S0022-1694(02)00064-1
[19] (1999) BCEOM: Abay River Basin Integrated Master Plan. Main Report, MoWR, Addis Ababa.
[20] Tarekegn, D. and Tadege, A. (2005) Assessing the Impact of Climate Change on the Water Resources of the Lake Ta-
na Sub-Basin Using the WATBAL Model. CEEPA, Republic of South Africa.
[21] Kebede, S., Travi, Y., Alemayehu, T. and Ayenew, T. (2006) Water Balance of Lake Tana and Its Sensitivity to Fluc-
TR
tuations in Rainfall, Blue Nile Basin, Ethiopia. Journal of Hydrology, 316, 233-247.
http://dx.doi.org/10.1016/j.jhydrol.2005.05.011
[22] Abdo, K.S., Fiseha, B.M., Rientjes, T.H.M., Gieske, A.S.M. and Haile, A.T. (2009) Assessment of Climate Change
Impacts on the Hydrology of Gilgel Abbay Catchment in Lake Tana Basin, Ethiopia. Hydrological Processes, 23,
3661-3669.
[23] Nicks, A.D. (1974) Stochastic Generation of the Occurrence, Pattern and Location of Maximum Amount of Daily
Rainfall. In: Statistical Hydrolgy, US Governmental Print Office, Washington, DC, 154-171.
A
[24] Swindel, B., Lassiter, C. and Riekerk, H. (1983) Effects of Different Harvesting and Site Preparation Operations on the
Peak Flows of Streams in Pinus elliottii Flat Woods Forests. Forest Ecology and Management, 5, 77-86.
http://dx.doi.org/10.1016/0378-1127(83)90059-2
[25] Swank, W., Vose, J. and Elliott, K. (2001) Long-Term Hydrologic and Water Quality Responses Following Commer-
CT
cial Clear-Cutting of Mixed Hardwoods on a Southern Appalachian Catchment. Forest Ecology and Management, 143,
163-178. http://dx.doi.org/10.1016/S0378-1127(00)00515-6
[26] Sheridan, J.M. (2002) Peak Flow Estimates for Coastal Plain Watersheds. Transactions of the ASAE, 45, 1319-1326.
http://dx.doi.org/10.13031/2013.11069
[27] Descroix, L., Nouvelot, J.-F. and Vauclin, M. (2002) Evaluation of an Antecedent Precipitation Index to Model Runoff
Yield in Western Sierra Madre (North-West Mexico). Journal of Hydrology, 263, 114-130.
http://dx.doi.org/10.1016/S0022-1694(02)00047-1
[28] Amatya, D., Callahan, T., Trettin, C. and Radecki-Pawlik, A. (2009) Hydrologic and Water Quality Monitoring on
ED
Turkey Creek Watershed, Francis Marion National Forest, SC. Proceedings of the ASABE Annual International Meet-
ing, Reno, 21-24 June 2009, Paper No. 095999.
[29] OConnor, K.M., Goswami, M., Liang, G.C., Kachroo, R.K. and Shamseldin, A.Y. (2001) The Development of the
Galway Real-Time River Flow Forecasting System (GFFS). In: Subtopic 1: Minimisation of Impacts of Extreme
Hydrological Events in Europe: Volume C of the Proceedings, Paper No. 035 (38 Pages) of the 19th European Con-
ference on Sustainable Use of Land and Water of the International Commission on Irrigation and Drainage (ICID),
Brno & Prague, 4-8 June 2001
[30] OConnell, P.E., Nash, J.E. and Farrell, J.P. (1970) River Flow Forecasting through Conceptual Models. Part 2. The
Brosna Catchment at Ferbane. Journal of Hydrology, 10, 317-329. http://dx.doi.org/10.1016/0022-1694(70)90221-0
[31] Nash, J.E. and Foley, J.J. (1982) Linear Models of Rainfall-Runoff Systems. In: Singh, V.P., Ed., Rainfall-Runoff Re-
lationship: Proceedings of the International Symposium on Rainfall-Runoff Modelling, Water Resources Publications,
Littleton, 51-66.
[32] Nash, J.E. and Barsi, B.I. (1983) A Hybrid Model for Flow Forecasting on Large Catchments. Journal of Hydrology,
65, 125-137. http://dx.doi.org/10.1016/0022-1694(83)90213-5
[33] Kachroo, R.K., Liang, G.C. & OConnor, K.M. (1988) Application of the Linear Perturbation Model (LPM) to Flood

82
 D. T. Mengistu et al.

Routing on the Mekong River. Journal of the International Association of Hydrological Sciences (IAHS), 33, 193-224.
http://dx.doi.org/10.1080/02626668809491238
[34] Kachroo, R.K. and Natale, L. (1992) Non-Linear Modelling of the Rainfall-Runoff Transformation. Journal of Hy-
drology, 135, 341-369. http://dx.doi.org/10.1016/0022-1694(92)90095-D
[35] Kachroo, R.K., Sea, C.H., Warsi, M.S., Jemenez, H. and Saxena, R.P. (1992) River Flow Forecasting. Part 3. Applica-
tions of Linear Techniques in Modelling Rainfall-Runoff Transformations. Journal of Hydrology, 133, 41-97.
http://dx.doi.org/10.1016/0022-1694(92)90148-O
[36] Liang, G.C. and Guo, Y.F. (1994) Observed Seasonal Hydrological Behaviour Used in Flow Forecasting on the
Yangtze River above Hankou. Journal of Hydrology, 154, 383-402. http://dx.doi.org/10.1016/0022-1694(94)90226-7
[37] Liang, G.C., OConnor, K.M. and Kachroo, R.K. (1994) A Multiple-Input Single-Output Variable Gain Factor Model.
Journal of Hydrology, 155, 185-198. http://dx.doi.org/10.1016/0022-1694(94)90164-3
[38] Ahsan, M. and OConnor, K.M. (1994) A Reappraisal of the Kalman Filtering Technique as Applied in River Flow
RE
Forecasting. Journal of Hydrology, 161, 197-226. http://dx.doi.org/10.1016/0022-1694(94)90129-5
[39] Ahsan, M. and OConnor, K.M. (1994) A Simple Non-Linear Rainfall-Runoff Model with a Variable Gain Factor.
Journal of Hydrology, 155, 151-183.
[40] Tan, B.Q., OConnor, K.M. and Liu, Z.C. (1996) Application of a Distributed form of the SMAR Model. Proceedings
of the International Conference on Water Resources and Environment Research: Towards the 21st Century (Volume 1),
Kyoto, 29-31 October 1996.
[41] hamseldin, A.Y. (1997) Application of a Neural Network Technique to Rainfall-Runoff Modeling. Journal of Hydrol-
ogy, 199, 272-294. http://dx.doi.org/10.1016/S0022-1694(96)03330-6
TR
[42] Shamseldin, A.Y., OConnor, K.M. and Liang, G.C. (1997) Methods for Combining the Outputs of Different Rainfall-
Runoff Models. Journal of Hydrology, 197, 203-229. http://dx.doi.org/10.1016/S0022-1694(96)03259-3
[43] Xiong, L., Shamseldin, A.Y. and OConnor, K.M. (2001) A Non-Linear Combination of the Forecasts of Rainfall-
Runoff Models by the First-Order Takagi-Sugeno Fuzzy System. Journal of Hydrology, 245, 196-217.
[44] Goswami, M. and OConnor, K.M. (2005) Application of a Conceptual Rainfall-Runoff Simulation Model to Three
European Catchments Characterised by Non-Conservative System Behaviour. Proceedings of the International Confe-
rence on Hydrological Perspectives for Sustainable Development, Roorkee, 23-25 February 2005, 117-130.
A
[45] Bergstrm, S. (1976) Development and Application of a Conceptual Runoff Model for Scandinavian Catchments. Bull.
Series A52, University of Lund, Lund, 134 p.
[46] Bergstrm, S. (1997) Development and Test of the Distributed HBV-96 Hydrological Model. Journal of Hydrology,
201, 272-288. http://dx.doi.org/10.1016/S0022-1694(97)00041-3
CT
[47] Bergstrm, S. (1995) The HBV Model. In: Singh, V.P., Ed., Computer Models of Watershed Hydrology, Water Re-
sources Publications, Highlands Ranch, Colorado, 443-476.
[48] Bergstrm, S. (1992) The HBV-ModelIts Structure and Applications. SMHI Reports RH No. 4, Norrkping.
[49] Arnold, J.G. and Allen, P.M. (1996) Estimating Hydrologic Budgets for Three Illinois Watersheds. Journal of Hydrol-
ogy, 176, 57-77. http://dx.doi.org/10.1016/0022-1694(95)02782-3
[50] Mengistu, D.T. and Sorteberg, A. (2012) Sensitivity of SWAT Simulated Streamflow to Climatic Changes within the
Eastern Nile River Basin, Hydrol. Earth System Science, 16, 391-407. http://dx.doi.org/10.5194/hess-16-391-2012
ED
[51] Arnold, J.G., Srinivasan, R., Muttiah, R.S. and Williams, J.R. (1998) Large Area Hydrologic Modeling and Assess-
ment, Part 1: Model Development. Journal of the American Water Resources Association, 34, 73-89.
http://dx.doi.org/10.1111/j.1752-1688.1998.tb05961.x
[52] Neitsch, S.L., Arnold, J.G., Kiniry, J.R., Williams, J.R. and King, K.W. (2002) Soil and Water Assessment Tool:
Theoretical Documentation, Version 2000. http://www.brc.tamus.edu/swat/
[53] Granger, C.W.J. and Ramanathan, R. (1984) Improved Methods of Combining Forecasts. Journal of Forecasting, 3,
197-204. http://dx.doi.org/10.1002/for.3980030207
[54] Armstrong, J.S. (1989) Combining Forecasts: The End of the Beginning or the Beginning of the End? International
Journal of Forecasting, 5, 585-588. http://dx.doi.org/10.1016/0169-2070(89)90013-7

83