Vous êtes sur la page 1sur 17

Hydrological Sciences Journal/Journal des Sciences

Hydrologiques

ISSN: 0262-6667 (Print) 2150-3435 (Online) Journal homepage: http://www.tandfonline.com/loi/thsj20

Regression analysis of rainfall–runoff data from an


arid catchment in Oman / Analyse par régression
de données pluie–débit d'un bassin aride d'Oman

Neil McIntyre , Aisha Al-Qurashi & Howard Wheater

To cite this article: Neil McIntyre , Aisha Al-Qurashi & Howard Wheater (2007) Regression
analysis of rainfall–runoff data from an arid catchment in Oman / Analyse par régression de
données pluie–débit d'un bassin aride d'Oman , Hydrological Sciences Journal/Journal des
Sciences Hydrologiques, 52:6, 1103-1118, DOI: 10.1623/hysj.52.6.1103

To link to this article: https://doi.org/10.1623/hysj.52.6.1103

Published online: 15 Dec 2009.

Submit your article to this journal

Article views: 3615

View related articles

Citing articles: 27 View citing articles

Full Terms & Conditions of access and use can be found at


http://www.tandfonline.com/action/journalInformation?journalCode=thsj20
Hydrological Sciences–Journal–des Sciences Hydrologiques, 52(6) December 2007 1103
Special Section: Dryland Hydrology in Mediterranean Regions

Regression analysis of rainfall–runoff data from an arid


catchment in Oman

NEIL McINTYRE, AISHA AL-QURASHI & HOWARD WHEATER


Department of Civil and Environmental Engineering, Imperial College London, London SW72AZ, UK
n.mcintyre@imperial.ac.uk

Abstract There is a pressing need to improve capability to predict the hydrological responses of arid
and semi-arid catchments. The literature indicates that physically-based rainfall–runoff models are not
yet able to meet this challenge. Simple empirical or semi-empirical models may perform equally well or
better, and provide basic but important information into catchment functioning. In this context, hourly
data from 36 rainfall–runoff events in Wadi Ahin, a 734 km2 catchment in the Sultanate of Oman, are
analysed. Runoff volumes and peaks are regressed against descriptors of rainfall characteristics and
antecedent conditions. The controls on runoff are found to be: rainfall volume, rainfall peak, rainfall
spatial location and variability, and antecedent wetness. Simple linear relationships between runoff peak
and volume, and rainfall volume produce the best predictions, and robust prediction confidence limits.
We speculate that application of physically-based models will not produce better predictions, but will
allow us to test hypotheses about the underlying hydrological processes.
Key words arid zone; rainfall–runoff; wadi; regression; modelling

Analyse par régression de données pluie–débit d’un bassin aride d’Oman


Résumé Il est nécessaire d’améliorer notre capacité à prévoir les réponses hydrologiques dans les
bassins arides et semi-arides. La littérature indique que les modèles pluie–débit à bases physiques ne
peuvent pas encore résoudre ce problème. De simples modèles empiriques ou semi-empiriques peuvent
donner des résultats similaires, voire meilleurs, et peuvent apporter des informations basiques mais
importantes sur le fonctionnement des bassins. Dans ce contexte, les données horaires de 36 événements
pluie–débit de Wadi Ahin, un bassin de 734 km2 du Sultanat d’Oman, sont analysées. Les volumes et
pics de débits sont étudiés par régression avec les caractéristiques des pluies et les conditions
antérieures. Les facteurs de contrôle de l’écoulement apparaissent être: le volume de pluie, l’intensité
maximale, la localisation et la variabilité spatiales de la pluie, et l’humidité antérieure. Des relations
simples entre le pic de débit et le volume, et le volume de précipitation produisent les meilleures
prévisions, avec des limites de confiance robuste. Nous supposons que l’application de modèles à bases
physiques ne donnera pas de meilleures prévisions, mais nous permettra de tester des hypothèses sur les
processus hydrologiques sous-jacents.
Mots clefs région aride; pluie–débit; oued; régression; modélisation

INTRODUCTION

Pressures on hydrological systems in arid and semi-arid areas are severe and
increasing. Economic development is associated with increasing flood risk (due to
development in flood plains and increasing asset values), and increasing demand for
water. Water resources are, by definition, limited, and under increasing threat from
water quality degradation, due to pollution and over-abstraction. Effects of climate
change are expected to be severe for many arid and semi-arid areas, leading to
increased flood risk and reductions in sustainable water resources. Rural land
management is changing, and both natural ecosystems and agriculture face increasing
pressures from climate change.
Arid and semi-arid areas thus face major challenges for the sustainable manage-
ment of water, yet flood and water resource management in arid zones is hampered by
the difficulty of representing arid-zone hydrology within numerical models. Many of
the hydrological characteristics of arid and semi-arid zones pose major challenges for
modelling and analysis (Wheater, 2005). These features may include, for example,

Open for discussion until 1 June 2008 Copyright © 2007 IAHS Press
1104 Neil McIntyre et al.

extreme temporal and spatial variability of rainfall, absence of baseflow, intermittent


plant cover, high transmission losses from river flows and high potential evaporation.
Furthermore, the value of observations from arid areas is often limited due to the
sparse rainfall and runoff gauge networks, difficulties in timely access to flow gauge
sites during runoff events, and the difficulty of measurement of the flow because of the
high sediment and debris loads (Rodier, 1985; Pilgrim et al., 1988; Al-Qurashi, 1995).
Nouh (2006) demonstrates the lack of data and consequent difficulty of runoff
estimation for Oman and the other Arabian Gulf states.
Although a range of models have been developed specifically for, or modified for,
arid and semi-arid areas, and a number of applications have been reported (see review
below), success in terms of model performance and applicability has been limited.
Process-based rainfall–runoff simulation models allow testing of hypotheses about
hydrological processes and analysis of complex, distributed water management issues
—but have some fundamental problems. Issues of identifiability are common to humid
and arid applications, but for arid and semi-arid areas particular issues arise, due, for
example, to the difficulty of characterising rainfall inputs, and the fact that flow data,
in the presence of transmission losses, provide a non-unique descriptor of upstream
runoff generation (Grayson et al., 1992; Michaud & Sorooshian, 1994a; Wheater,
2005). There are therefore potential benefits in adopting a simple, empirical approach
to modelling. It is less expensive, can provide basic insight into controls on runoff, and
can be applicable to water balance and flood estimation problems—although it is
restricted by the information in the data.
This paper describes the arid Wadi Ahin catchment in the Sultanate of Oman,
including presentation and analysis of hourly rainfall–runoff data from 36 events.
Regression analysis is used to identify factors governing flood generation and routing,
and to predict flow volumes and flood peaks as functions of rainfall input and
antecedent wetness. The results of the analysis, and the value of the regression in the
context of simulation modelling, are discussed. Firstly, to provide background and
guidance for the analysis and discussion, we review the literature on sub-daily arid and
semi-arid zone rainfall–runoff modelling.

REVIEW OF ARID AND SEMI-ARID ZONE RAINFALL–RUNOFF


MODELLING

Flow hydrographs in arid and semi-arid areas are commonly characterised by


extremely rapid rise, often from zero baseflow. Times-to-peak of runoff hydrographs
can be of the order of 15–30 minutes, thus to capture flood response and to define
flood peaks, sub-daily data are essential. However, the literature on arid and semi-arid
zone rainfall–runoff modelling using a sub-daily time step is limited. The main studies
have been on Walnut Gulch in Arizona; other significant results have come from
catchments in Saudi Arabia and India. Here, we review this literature, beginning with
applications of empirical models, followed by conceptual models and, finally, more
complex, physics-based rainfall–runoff models.
Significant early work on characterising semi-arid zone runoff in Walnut Gulch
was reported by Osborn & Lane (1969). They studied runoff-producing convective
storms from three years on four subcatchments with areas ranging between 2.3 and

Copyright © 2007 IAHS Press


Regression analysis of rainfall–runoff data from an arid catchment in Oman 1105

44.5 km2. They developed stepwise regression models, with dependent variables: time
to peak, peak runoff, and runoff volume, rise time and runoff duration; regressed
against 10 rainfall parameters, an index of antecedent moisture, area and vegetation
cover. They found that antecedent conditions were not generally significant and in one
catchment there was a negative correlation of antecedent moisture with flow, and they
cite earlier studies which support this finding. The total precipitation was the main
factor controlling runoff volume, explaining between 76 and 89% of runoff variation
between events for each catchment—other input variables were insignificant or only
weakly significant. The 10–15 min peak rainfall was most important on peak runoff in
three of four catchments, the duration of flow was generally most strongly correlated
to watershed length, and lag time was most strongly correlated to watershed area.
Lane et al. (1971) studied transmission losses and routing in Walnut Gulch using
regression of input hydrograph parameters and output hydrograph parameters,
excluding over-bank flow events. They found that volume and peak of outflow are
strongly correlated with input volume (as opposed to flow peak as used by some
previous investigators) and no significant correlation to antecedent wetness index.
They used their predictions of time of travel, hydrograph peak and volume to fit
three-parameter Gamma distributions to the downstream hydrograph.
Sharma & Murthy (1996) analysed 79 events from nine sub-basins (areas from 814
to 2510 km2) of the arid Luni River catchment in northwest India. Observed
hydrographs at flow gauges in the upper and lower subcatchments at 30-min resolution
were available, as well as effective rainfall estimates. Runoff event lengths were
between a few hours and two days. Observed transmission losses between the upper
and lower gauges were from 8 to 56%. In the upper gauges, time-to-peak and peak
flow were estimated using regression with catchment area and effective rainfall
volume. It is not explicit in the paper how effective rainfall was calculated. The
recessions were estimated using a two-parameter conceptual leaky reservoir model.
This simple metric-conceptual model was found to give acceptable predictions in
ungauged catchments. Transmission to the lower gauges was modelled using a two-
parameter loss and routing differential equation. The authors found that predicted
hydrographs were more sensitive to the length parameter and the initial catchment
storage, than the loss parameter. They noted that, after some travel distance, the
recessions are a function of channel properties rather than input rainfall distribution.
Sharma & Murthy (1998) studied another six basins with steep slopes in upland arid
India, again using a regression equation to estimate peak flow and time to peak from
effective rainfall and area, and a two-parameter leaky reservoir model for the
recessions. Fifty-one hydrographs and associated rainfall data sets, from 1980 to 1987,
were analysed using 30-min time steps. They found that the method worked well;
however, the appropriate parameters depended on the nature of the stream network and
in cases notably varied over events within one catchment.
Nouh (1990) applied the geomorphologic instantaneous unit hydrograph (GIUH)
to 32 catchments in southwest Saudi Arabia, with the objective of testing its suitability
for arid catchments. The GIUH model included a two-parameter, time-dependent
infiltration capacity equation, applied uniformly across the catchment. The SCS curve
number was used to predict effective rainfall. Nouh showed that infiltration rate, size
and slope were important in controlling accuracy (accuracy generally higher for low
infiltration and area < 400 km2). Accuracy was shown to be affected by rainfall

Copyright © 2007 IAHS Press


1106 Neil McIntyre et al.

temporal and spatial distribution, and the length of dry period preceding the event; the
loss of accuracy was explained mainly by limitations in the infiltration model.
Volumes were better simulated than peaks. Sorman & Abdulrazzak (1993) explored
the applicability of the GIUH using 13 catchments in Saudi Arabia, and showed it to
work well in one small (19.4 km2) catchment. Dominant parameters affecting the
hydrographs were found to be soil porosity, pore size, saturated hydraulic conductivity
(ks), flood velocity, and groundwater table level. These authors also noted the large
range of ks values needed to fit individual events, and the large differences between
these values and laboratory values.
Wheater & Brown (1989) studied the rainfall–runoff of a catchment in southwest
Saudi Arabia (area 597 km2). They developed 1-mm 10-min unit hydrographs for 11
storms from 1984 and 1985. The available data were digitised to 10-min steps. The
unit hydrographs fitted the events individually with reasonable accuracy; however,
they were very variable between storms. The runoff coefficient varied from 5.9 to
79.8%. The analysis could not explain the scatter in the rainfall–runoff volume
relationship in terms of rainfall depth, maximum intensities, temporal distribution or
antecedent conditions (one of the largest runoff events was apparently associated with
one of the smallest observed rainfalls). The authors suggested that the variation of
hydrographs and runoff coefficients was due to the highly localised rainfall and
transmission losses. However, a well-defined systematic relationship was observed
between hydrograph peak and volume, suggesting that if a peak could be predicted, for
example, by frequency analysis, the volume could also be determined with reasonable
accuracy.
El-Hames & Richards (1998) tested a physically-based semi-distributed rainfall–
runoff model in a steep 107 km2 catchment in southwest Saudi Arabia. The model used
Richards’ equation to represent infiltration, the one-dimensional (1-D) kinematic wave
for overland flow and the 1-D dynamic wave equations for channel flow. Three events
were investigated using an hourly time step. Rainfall was estimated using five gauges
in or around the catchment, although the assumptions used in estimating spatial rainfall
were different from one event to the next—it seems this was due to missing data. The
top-layer hydraulic conductivity and the Manning’s coefficient for overland flow were
manually calibrated—the model results were found to be especially sensitive to these
two parameters, as well as initial soil moisture (initial soil moisture was estimated
from field measurements). The authors cite another paper that found the same para-
meters to be important. The fits to observed hydrographs were considered good,
although no validation test was reported.
Michaud & Sorooshian (1994a) used six rainfall events for calibration and 24
events for validation on the 150 km2 Walnut Gulch catchment. They compared the
semi-distributed, physically-based Kineros model (using 2-min input rainfall data)
with lumped and distributed versions of the simple SCS model (using 5-min rainfall
data). Initial conditions were estimated using a daily model, CREAMS. The
uncalibrated Kineros parameters were estimated from physical observations or from
textbook values. Manning’s coefficient was fixed at 0.04 for all channels and 0.05 for
all planes; ks for planes was derived via SCS numbers and varied spatially from 3.61 to
8.92 mm/h, while 42 mm/h was assumed for channel alluvium. Other parameters were
also fixed at typical values from experience of users. Manual calibration was used for
Kineros, varying Manning’s coefficient, ks and the coefficient of variability of ks over

Copyright © 2007 IAHS Press


Regression analysis of rainfall–runoff data from an arid catchment in Oman 1107

planes and channels. However, it was reported that the optimum parameters may not
have been found, due to the difficulty of this task. Kineros’ performance was better
than the SCS model when no calibration was used, but was nevertheless
disappointing—for example, the validation root mean square error (RMSE) for peak
flows was 79% of the mean observed peak, with four modelled peaks overestimated by
more than 100%. The performances of the calibrated semi-distributed SCS model were
equal to the more complex Kineros model and much better than the lumped model;
however all the models were considered inadequate. The same authors (Michaud &
Sorooshian, 1994b) used the same catchment to investigate effects of rainfall sampling
errors. “Approximately half of the difference between observed and simulated peaks
was due to rainfall-sampling errors”. Even data representing a spatial aggregation to
the 4 km × 4 km pixel scale generally produced serious underestimations of the peak
flows.
Grayson et al. (1992) applied a distributed process-based model THALES to
Lucky Hills, a 0.044 km2 subcatchment of Walnut Gulch. The model grid size used
was on average 15 m2 and the minimum time resolution of input rainfall was 2 min.
The event lasted approximately 140 min, with two main rainfall bursts. The soil
parameters were distributed across the catchment based on a soil survey. The
CREAMS model was used to estimate initial soil moisture. Manning’s coefficient was
estimated as 0.03 and 0.02 for hill slopes and channels, respectively, based on surface
properties (the authors note that, in a previous calibration of Kineros, these were
adjusted to 0.066 and 0.044). At first, no calibration of THALES was done and results
were “poor”. With calibration of ks to the outlet flow gauge data, significant
improvements were found, with errors between simulated and gauged peak flows equal
to 4.5% for the main peak and up to 69% for the lesser peaks. Times to peak were
2 min early. Setting soil parameters to be uniform across the catchment at their
catchment average value, rather than distributed, resulted in very little change in
results at the outlet. On the contrary, expanding the channel network within the
catchment (i.e. changing the effective conveyance), caused large change in simulated
flows. The authors note that 2-min rainfall resolution may not have been enough to
accurately simulate hydrograph peaks, because one observed peak was found to be
higher than the rainfall rate. The authors conclude that problems of verification and
validation of distributed, process-based models are acute due to parameter estimation
problems, model structure errors and limitations in data.
The general outcomes of this review of sub-daily rainfall–runoff modelling in arid
and semi-arid regions are: (a) that complex models do not generally lead to improved
predictive ability, even when high resolution supporting data sets are available and
simple conceptual or empirical approaches may be attractive alternatives (Michaud &
Sorooshian, 1994a; Sharma & Murthy, 1996, 1998; Lane et al., 1971); (b) physically-
based parameters, which in theory can be estimated directly from physical properties
of the catchment, often in practice need to be calibrated (Michaud & Sorooshian,
1994a; Grayson et al., 1992; Nouh, 1990); (c) however, physically-based models take
advantage of topographic controls on runoff and have obvious potential advantages for
ungauged catchments (El-Hames & Richards, 1998; Grayson et al., 1992); (d) large
catchments seem to be a particular challenge (Michaud & Sorooshian, 1994a; Nouh,
1990); (e) the importance of initial soil moisture in controlling runoff is not clear: in

Copyright © 2007 IAHS Press


1108 Neil McIntyre et al.

some cases a wetter catchment increases runoff (El-Hames & Richards, 1998),
sometimes it does not appear to be significant (Lane, 1971) and on occasions it reduces
runoff (Osborn & Lane, 1969); (f) channel roughness, slope and infiltration rate are the
most commonly cited principal catchment properties affecting runoff response; and
(g) rainfall spatial variability and associated rainfall estimation errors are key controls
on model results and parameter values (Wheater & Brown, 1989; Michaud &
Sorooshian, 1994b).
There are a considerable number of additional relevant studies which use daily or
monthly data (e.g. Hughes, 1995; Lange, 2005; Costelloe et al., 2006), which reinforce
the general conclusions from the sub-daily studies reviewed here. Wheater (2005)
presents a compendium of studies and reviews, which further support our conclusions.
Wheater et al. (1991a,b) developed simple stochastic models for the spatial distribu-
tion of daily rainfall in southwest Saudi Arabia, and Wheater et al. (1995) developed a
water resource model for Wadi Ghulaji, northern Oman, which included stochastic
simulation of spatial daily rainfall. It seems likely that, given the dominant effects of
rainfall spatial variability on runoff processes in arid areas, and the typical densities of
raingauge networks, that much further work into such methods will be required to
support process-based hydrological modelling.

DESCRIPTION OF WADI AHIN

Oman is bordered to the west by Saudi Arabia, Yemen and United Arab Emirates; to
the east, the Arabian Sea and Gulf of Oman define the southern and northern coasts
(Fig. 1). A mountain chain parallel to the coast divides Oman’s ephemeral rivers
(wadis) into two types; those that drain to the coast, and those that drain to the interior
desert. Wadi Ahin is one of the former, situated in the Al-Batinah Region in the north
of Oman (Fig. 1). The catchment can be subdivided into two areas: Ahin West
(upstream of the flow gauging station near Hayl, with elevation above sea level
ranging from 1300 to 300 m, area 734 km2) and Ahin East (ranging from 1300 m to
sea level, area 76 km2). In this paper we study data only from Ahin West (which we
refer to as “the catchment” from here on) because the lower catchment’s channels are
less well defined and flow gauging has been more problematic.
Geological and soil maps are available from the Ministry of Regional
Municipalities, Environment and Water Resources (MRMEWR). In the piedmont area
irregular hills and narrow ridges of ophiolite and Hawasinah Nappes, Aruma Group
rocks and Tertiary limestones, are interspersed with alluvial wadi deposits. In the
upper parts of the catchment the alluvium is typically less than 20 m thick, ranges in
composition from clay to boulder and in many instances have been weakly to strongly
cemented. The wadi channels are mainly gravel and sands. Land cover is mainly
mountainous with some sparse vegetation cover. The hillslopes are generally steep
with slopes (measured from a 90 m resolution digital elevation model) ranging from
3 to 92%. The channel slope in the catchment ranges from 0.1 to 0.32%. The width of
the main channel near the Hayl gauge is about 54 m.
In Oman there are two seasons: summer (April–October) and winter (November–
March). The maximum daily temperature varies from 26 to 29°C in winter, but is
above 40°C in summer (MWR, 1995). The wind velocity (at 2 m above ground level)

Copyright © 2007 IAHS Press


Regression analysis of rainfall–runoff data from an arid catchment in Oman 1109

55° 60°

25° Gulf of
Oman
United Arab
Emirates
Muscat
Wadi Ahin

Kingdom of
Saudi Arabia

20°

Arabian N
Sea
Yemen
100 km

Fig. 1 The location of Wadi Ahin in relation to the Sultanate of Oman.

is generally in the range of 2 to 3 m/s. The rainfall varies widely through time and
space. The long term annual average amounts vary from 50 mm in central Oman to
300 mm in the mountain areas in the north and in the south. The generation
mechanisms causing rainfall in Oman can be classified into: cold frontal troughs
(mainly in the winter and early summer), tropical cyclones (uncommon, and mainly in
May–December), advection of cold air moving from central Asia through the Arabian
Gulf to the southeast of Oman (can cause heavy rain and occurs year-round, three
times per year on average), scattered convective events over mountains (mainly June
and July), and on-shore monsoon currents.
There are nine rainfall gauges within the catchment. Five of the gauges operate
with digital data loggers. Two (Haybi and Al Ghuzayfah) are standard non-recording
gauges. The others operate using chart recorders. The locations of the seven recording
gauges, which are used in the current research, are shown in Fig. 2. The annual average
rainfall at these gauges ranges from 98 mm at Hayl Ashkariyyin (gauge no. 7 in Fig. 2)
to 143 mm at Al-Waqbah (gauge no. 3). The MRMEWR previously used isohyets,
including elevation correction, to estimate the catchment’s annual average rainfall as
152 mm. February and March have the highest monthly averages (16–29 mm),
whereas April, July and August (4–17 mm) are the wettest of the other months.
Available data (Nouh, 2006) indicate that Oman’s national annual average rainfall is
71 mm, with approximately 6% of this rainfall becoming runoff, indicating the
resource value of Wadi Ahin.
The flow gauge at Hayl is a digital stage recorder (a pressure transducer). The
stage–discharge curves are based on direct measurement, using current meter
measurements for low flows, and indirect measurement for peak flows using the slope–
area method (a channel survey was done after each flood and then the slope–area
method was used to calculate discharge assuming a Manning’s coefficient of 0.036).
The stage–discharge curves are updated if there is a shift in the channel bed. Events
with clearly erroneous flood fluctuations were not included in our data set, although it

Copyright © 2007 IAHS Press


1110 Neil McIntyre et al.

Raingauges
Hayl flow gauge

Outline of
Ahin East
7 catchment

6
5
N
3 4
1
Outline of
2 Ahin West 10 km
catchment

Fig. 2 The locations of Wadi Ahin rainfall gauges.

is noted that human, instrument or stage–discharge errors may affect the quality of the
flow estimates.
The annual average cumulative flow at the Hayl gauge is approximately 29
× 106 m3 (based on 1984–1999), falling to 0.9 × 106 m3 at the Majis Kabirah and 4.4
× 106 m3 at the Kishdah coastal gauges (based on 1985–1999). This highlights the
degree of transmission loss in parts of the catchment. At the Hayl gauge, the river is
dry for 14% of the time, as measured between 1996 and 1999.

RAINFALL–RUNOFF DATA
Thirty-six runoff-generating events from 1996–1999 were selected based on availab-
ility of the hourly rainfall and flow data, and perceived quality of data. These are
plotted as one time series in Fig. 3 and as independent events in Fig. 4 (rainfall in both
has been averaged without weighting over the seven gauges), and some statistics of the
events are given in Table 1.While there may be additional events from the same
catchment in later years, and also some non-runoff-generating events from 1996–1999,
it was not possible for us to access them, although this remains a possibility for
extending the work. Some general features of the data set (refer to Table 1) are:
(a) There is a significant flow event on 28 September 1999 with almost zero observed
rainfall (only 1 mm at one gauge) which occurred 10 hours after the flood peak.
This is omitted from all further analysis, however it has been mentioned to
illustrate that a general rainfall observation problem may exist, in particular the
limited number of gauges may mean that significant rainfall is not recorded.
(b) The time lag between peak rainfall and peak flow ranges from 2 to 15 hours,
although 26 of the 35 values are between 2 and 4 h. Of the nine events with lag
greater than 4 h, six were in June or July.
(c) Baseflow is generally small compared to peak flows, although it was zero in only
two events, and in four small events baseflow is slightly greater than 10% of peak
flow.

Copyright © 2007 IAHS Press


Regression analysis of rainfall–runoff data from an arid catchment in Oman 1111

1996−1997
2 0

1.5 5

Flow (mm/hr)

Rain (mm/hr)
1 10

0.5 15

0 20

1998−1999
2 0

1.5 5
Flow (mm/hr)

Rain (mm/hr)
1 10

0.5 15

0 20
0 100 200 300 400 500 600 700
Day no.
Fig. 3 Time series of the runoff producing events from 1996–1999.

(d) Some judgement has been made in deciding whether to treat two successive
rainfall bursts as one or two events. If the rainfall of the second burst significantly
coincides with the previous flow recession then they are treated as one event (e.g.
27 July 1996, 26 March 1997, see Fig. 4); if not, they are separated (e.g.
31 January 1998, 1 February 1998).
(e) The nine largest events in terms of rainfall volume were in the months January,
February and March. This period also had relatively uniform rainfall (a greater
proportion of the gauges had significant rainfall).
(f) There is an exceptional event on 29 March 1997, which included two peaks of
1080 and 1000 m3/s respectively (estimated return periods of 85 and 59 years). We
estimate that the 1-day rainfall on this date exceeded the 200-year return period
value at gauge no. 2 (Fig. 2). The calculated runoff coefficient for this event is
1.67. Clearly this value is due to under-measurement of rainfall or over-estimation
of flow.
(g) Excluding the 29 March 1997 event, the runoff coefficients ranged from 0.019 to
0.38, with a mean of 0.13 and a mode of approximately 0.10. There is no clear
evidence from Table 1 that the runoff coefficient or baseflow are seasonal.

REGRESSION ANALYSIS

The variables which we aim to analyse and predict by regression are peak flow (Qp),
runoff volume (Qv) and runoff coefficient (C). The predictor variables are peak gauge-
average rainfall (Rp), total event rainfall volume (Rv), rainfall spatial variability index
(Rs), baseflow (Qb) as an index of antecedent conditions, and distance of the rainfall
centroid from the flow gauge (Rd). The values of all variables for each event are listed
in Table 1. The relationships between predictor and predicted variables are shown in
Fig. 5. Further predictor variables, in particular antecedent soil wetness and plant cover
indices may have added value, however these data were not available.

Copyright © 2007 IAHS Press


1112 Neil McIntyre et al.

22 Jan 1996 24 Jan 1996 11 March 1996 26 Jun 1996


1.4 2.8 1.2 0.04 0
0.7 1.4 0.6 0.02
0 0 0 0 10

27 Jun 1996 08 Jul 1996 27 Jul 1996 07 Aug 1996


0.2 0.06 0.1 0.4 0
0.1 0.03 0.05 0.2
0 0 0 0 10

25 Jan 1997 26 Mar 1997 29 Mar 1997 22 Apr 1997


1 0.6 10 0.1 0
0.5 0.3 5 0.05
0 0 0 0 10

23 Jun 1997 03 Jul 1997 08 Aug 1997 13 Sep 1997


0.04 0.04 0.6 0.4 0
0.02 0.02 0.3 0.2
0 0 0 0 10
Flow (mm/hr)

Rain (mm/hr)
11 Oct 1997 29 Oct 1997 02 Nov 1997 04 Nov 1997
0.2 0.6 0.4 0.4 0
0.1 0.3 0.2 0.2
0 0 0 0 10

27 Jan 1998 31 Jan 1998 01 Feb 1998 21 Feb 1998


0.6 0.1 0.2 0.1 0
0.3 0.05 0.1 0.05
0 0 0 0 10

13 Jun 1998 16 Jun 1998 17 Jul 1998 20 Jul 1998


0.02 0.01 0.1 0.6 0
0.01 0.05 0.3
0 0 0 0 10

23 Jul 1998 08 Aug 1998 05 Sep 1998 06 Sep 1998


0.04 0.3 0.06 0.2 0
0.02 0.15 0.03 0.1
0 0 0 0 10

17 Sep 1998 24 Oct 1998 02 Mar 1999 28 Sep 1999


0.3 0.1 2 2 0

0.15 0.05 1 1

0 0 0 0 10
0 50 100 0 50 100 0 50 100 0 50 100
Timestep (hrs) Timestep (hrs) Timestep (hrs) Timestep (hrs)

Fig. 4 Hydrographs and associated gauge-average rainfall for 36 runoff events.

Baseflow, Qb (mm/h), is defined as the gauged flow immediately prior to the


hydrograph rise; and the runoff volume, Qv, is defined as the specific volume (mm) of
gauged flow during the observed hydrograph after Qb is deducted. In three events
(22 April 1997, 23 June 1997, and 3 July 1997), the available data did not include the
full flow recession (evident in Fig. 4), so Qv is slightly underestimated in these cases;
however, this was not considered to be an over-riding source of bias. Total rainfall, Rv,
is defined as the gauge-average rainfall (mm) totalled over the event duration; and C is
defined as Qv/Rv. The rainfall centroid distance, Rd is defined as:
∑R d
i =1, 7
i i

Rd = (1)
∑R i =1, 7
i

where Ri is the total rainfall in each event at gauge number i; and di is the straight-line

Copyright © 2007 IAHS Press


Regression analysis of rainfall–runoff data from an arid catchment in Oman 1113

Table 1 Statistics of the rainfall–runoff events.


Date RL Rv Rp Rs Rd Qp Qb Qv C (a) T (b)
(h) (mm) (mm/h) (m) (mm/h) (mm/h) (mm) (h)
22 Jan 96 40 28.5 8 0 24206 0.83 0.001 4.6 0.16 4 and 3
24 Jan 96 71 32.5 7 0 21765 2.09 0.030 10.0 0.31 3
11 Mar 96 42 20.7 6 0 20405 0.70 0.001 5.7 0.27 4
26 Jun 96 49 0.8 0.4 6 15112 0.02 0.000 0.1 0.14 4
27 Jun 96 68 4.4 4 5 18038 0.15 0.000 0.5 0.12 4
08 Jul 96 46 1.2 1 6 30914 0.04 0.003 0.2 0.13 7
27 Jul 96 94 11.1 5 0 21541 0.09 0.001 0.7 0.06 4
07 Aug 96 46 9.2 5 3 27717 0.24 0.003 1.3 0.15 3
25 Jan 97 70 34.6 7 0 23510 0.51 0.000 2.7 0.08 4
26 Mar 97 62 13.3 4 4 22756 0.49 0.001 3.7 0.28 4 and 4
29 Mar 97 80 83.5 8 0 25522 5.30 0.003 139.5 1.67 4
22 Apr 97 46 3.3 2 2 23358 0.09 0.012 0.4 0.11 3
23 Jun 97 70 4.9 2 3 27818 0.03 0.004 0.2 0.04 4
03 Jul 97 70 8.2 3 3 26816 0.04 0.004 0.4 0.05 7
08 Aug 97 70 13.0 5 2 23880 0.43 0.002 1.1 0.09 4
13 Sep 97 70 5.8 3 2 25055 0.25 0.003 0.7 0.13 4
11 Oct 97 94 4.1 2 3 19184 0.17 0.002 0.7 0.16 4
29 Oct 97 70 9.2 4 2 16259 0.41 0.004 1.4 0.15 2 and 2
02 Nov 97 29 5.7 3 3 13503 0.23 0.015 0.8 0.14 4
04 Nov 97 63 2.8 1 5 20441 0.35 0.008 1.0 0.38 4
27 Jan 98 70 22.5 6 0 22129 0.36 0.003 2.0 0.09 4
31 Jan 98 52 11.1 4 0 26171 0.08 0.004 0.4 0.04 9
01 Feb 98 35 9.2 2 0 22170 0.13 0.006 0.5 0.05 15
21 Feb 98 93 20.4 6 0 24121 0.05 0.004 0.8 0.04 3 and 7
13 Jun 98 22 5.1 1 4 23017 0.02 0.000 0.1 0.02 9
16 Jun 98 70 1.0 1 6 19426 0.01 0.001 0.0 0.02 6
17 Jul 98 70 1.9 1 5 25866 0.06 0.001 0.2 0.09 5
20 Jul 98 70 10.3 7 3 25285 0.43 0.001 1.3 0.13 4
23 Jul 98 94 3.3 1 4 19568 0.03 0.000 0.3 0.10 6
08 Aug 98 46 4.9 4 3 17061 0.18 0.001 0.4 0.09 3
05 Sep 98 34 1.3 1 6 19690 0.05 0.001 0.2 0.15 4
06 Sep 98 57 5.0 4 5 18999 0.12 0.002 0.4 0.09 4
17 Sep 98 46 2.0 2 6 19426 0.18 0.001 0.3 0.17 3
24 Oct 98 46 0.8 1 5 22273 0.07 0.001 0.2 0.26 6
02 Mar 99 46 14.4 6 4 12276 1.21 0.000 3.0 0.21 3
28 Sep 99 94 0.1 0.1 7 13803 1.16 0.000 7.6 76.00 –10
RL: rainfall duration; Rv: total rainfall; Rp: peak rainfall; Rs: number of gauges with <2mm rain; Rd: rainfall centroid
distance; Qp: peak flow; Qb: baseflow; Qv: runoff volume; C: runoff coefficient; T: time lag of peak.
(a)
Constant baseflow during event length has been deducted in the calculation of runoff volume and runoff coefficient.
(b)
Where more than one value of time lag is given, the values refer to different rainfall bursts/flow peaks within the
same event.

distance of that gauge from the outlet in metres, neglecting elevation differences.
Therefore, events which are centred further away from the flow gauge have higher Rd;
Rd is included because transmission losses and flood wave attenuation are expected to
be greater for higher Rd, all else being equal. The spatial variability index Rs is defined

Copyright © 2007 IAHS Press


1114 Neil McIntyre et al.

4
10

Qv (mm)
2
10
0
10
−2
10

1
Qp (mm/hr)

10

−1
10

−3
10

1
10
0
10
C

−1
10
−2
10
0 2 0 2 0 3 6 1 2.5 4 0 0.02 0.04
10 10 10 10
Rv (mm) Rp (mm/hr) Rs Rd (m)x 104 Qb (mm/hr)

Fig. 5 Scatter plots showing effects on flow volume, flow peak and runoff coefficient.

as the number of gauges with total rainfall less than 2 mm—so that a high value of Rs
indicates that the rainfall is more localised.
The 29 March 1997 event was found to be an outlier from all the regression
models tested and so was omitted from the analysis. Results of a stepwise linear least-
squares regression for the remaining 34 events are shown in equations (2)–(4). Only
the predictors which increase the R2 value significantly (95% significance level) are
included. The increments of R2 associated with each included predictor are shown.
Although the value of peak rainfall (Rp) is implied to be of little significance in
equation (2), this is largely because of its correlation with Rv (correlation coefficient =
0.91), not because it is physically unimportant. There is evidence in all three equations
that the more spatially varied the rainfall (higher Rs), the more runoff occurs, indi-
cating the importance of localised intense rainfall in generating flows. This
dependency between the flow variables and Rs seems contrary to the relationships in
Fig. 5; however, the plotted data reflect the high correlation of Rs with Rv (correlation
coefficient = –0.86), whereas the equations assume independence. All three flow
variables increase significantly with the antecedent wetness index (Qb), as expected.
The further the centroid of rainfall from the flow gauge (Rd), the lower the measured
runoff coefficient (C), indicating that transmission losses increase with flow travel
distance.
Q p = −0.63 + 0.027 Rv + 36.2Qb + 0.095Rs + 0.072 R p (2)
2
(R = 0.49, 0.64, 0.73, 0.76)
Qv = −2.10 + 0.21Rv + 134Qb + 0.33Rs (3)
2
(R = 0.60, 0.72, 0.77)
C = 0.084 + 5.14Qb + 0.023Rs + 0.005Rv − 0.0038Rd (4)
2
(R = 0.11, 0.20, 0.27, 0.30)

Copyright © 2007 IAHS Press


Regression analysis of rainfall–runoff data from an arid catchment in Oman 1115

Log transforms (base 10) of Qp, Qv, C, Rp and Rv were used to moderate the influence
of the largest events and to allow for some non-linearity, resulting in the significant
predictors in equations (5)–(7). This increases the apparent effect of Rd and reduces the
apparent effect of Qb, but it is difficult to draw any additional conclusions from this.
log Q p = −1.48 + 1.23 log Rv + 0.084 Rs − 0.030 Rd (5)
2
(R = 0.47, 0.51, 0.56)
log Qv = −0.945 + 1.27 log Rv − 0.019 Rd + 0.060 Rs (6)
(R2 = 0.69, 0.71, 0.73)
log C = −0.57 − 0.021Rd + 12.9Qb (7)
2
(R = 0.08, 0.14)
In order to test the ability of the equations to predict Qp and Qv, we recalculate the
regression coefficients 34 times, each time omitting a different event. Each time, the
Qp and Qv of the omitted event are predicted, as a validation of the model. The
29 March 1997 event was always omitted. Given the difficulty of the task and the
scope for data errors especially in high flow data, we consider acceptable performance
to be when predicted values are within 30% of the observed values. The significant
predictor variables are assumed not to change. The actual and predicted values of Qp
and Qv and the percentage relative error, using equations (2) and (3), are in Table 2.
Only six out of 34 and seven out of 34 results, for Qp and Qv respectively, were
“acceptable” using the 30% criterion. The total storm runoff over the 34 events was
underestimated by 6%. Table 2 shows that some events have very high relative errors;
however, these were small events and the corresponding absolute errors were small.
Using equations (5) and (6) produced very similar performance.
Fitting a linear function of Rv to Qp (equation (8)) increased the number of
acceptable Qp results to 16. Doing the same to predict Qv (equation (9)) produced 11
acceptable results and reduced the accumulated storm runoff error to 0.2%. This
illustrates that, although several explanatory variables may be statistically identifiable,
simple equations may be the most robust for predictions. This is because the regression
equations were fitted using least squares, but the performance measure was different—
a specified acceptable tolerance. The models defined by equations (8) and (9) are
illustrated in Fig. 6, together with the 90% confidence limits, which are based on the
variance of the regression line plus the variance of the residuals.
Q p = −0.00514 + 0.0315 Rv (R2 = 0.49) (8)
Qv = −0.283 + 0.172 Rv (R2 = 0.60) (9)
None of the regression models, fitted to the 34 other events, were able to predict the
29 March 1997 event to within 30% tolerance. The best effort was using equations (8)
and (9) which underestimated Qp by 50% and Qv by 90%. The flow peak of that event,
although not used in fitting the regression, is shown in Fig. 6. Its volume (Qv =
140 mm) is well off the scale and therefore omitted. It may be speculated that this
outlier is due to non-linearities in the rainfall–runoff response, the gauged rainfall not
representing the total rainfall input, and errors in the application of the slope–area
method for flow estimation.

Copyright © 2007 IAHS Press


1116 Neil McIntyre et al.

Table 2 Predictive performance using equations (4) and (5).


Qp %err in Qp %err in Qp Qv %err in Qv %err in Qv
(mm/h) (equ. (2)) (equ. (8)) (mm/h) (equ. (3)) (equ. (9))
22 Jan 96 0.72 13 8 3.94 16 1
24 Jan 96 1.80 52 66 8.61 55 60
11 Mar 96 0.60 45 8 4.89 68 46
26 Jun 96 0.02 175 4 0.10 39 238
27 Jun 96 0.13 105 14 0.44 0 10
08 Jul 96 0.04 348 19 0.13 302 155
27 Jul 96 0.08 54 280 0.57 64 148
07 Aug 96 0.21 91 21 1.16 5 3
25 Jan 97 0.44 115 152 2.34 137 147
26 Mar 97 0.42 5 16 3.20 44 48
22 Apr 97 0.08 156 10 0.33 161 25
23 Jun 97 0.03 156 371 0.16 167 208
03 Jul 97 0.03 631 631 0.35 182 180
08 Aug 97 0.37 17 6 0.96 44 79
13 Sep 97 0.22 73 30 0.64 79 4
11 Oct 97 0.14 106 26 0.58 102 38
29 Oct 97 0.36 53 32 1.22 33 8
02 Nov 97 0.20 178 25 0.70 205 16
04 Nov 97 0.30 27 80 0.91 18 86
27 Jan 98 0.31 48 107 1.70 58 90
31 Jan 98 0.07 42 361 0.37 90 291
01 Feb 98 0.11 137 123 0.43 35 168
21 Feb 98 0.05 912 1177 0.69 260 329
13 Jun 98 0.01 224 967 0.08 235 523
16 Jun 98 0.00 786 386 0.02 817 609
17 Jul 98 0.05 90 2 0.14 30 73
20 Jul 98 0.37 10 27 1.12 11 16
23 Jul 98 0.03 296 227 0.28 127 13
08 Aug 98 0.16 43 19 0.37 83 32
05 Sep 98 0.04 155 23 0.17 92 137
06 Sep 98 0.10 276 32 0.39 108 27
17 Sep 98 0.16 2 72 0.30 49 87
24 Oct 98 0.06 215 74 0.19 186 174
02 Mar 99 1.04 60 65 2.62 28 29

CONCLUSION

It is clear from the literature that arid and semi-arid zones present special challenges
for hydrological modellers, due to the variability of processes and limitations in data.
Despite significant developments in recent years, the applicability of contemporary
distributed simulation models to arid and semi-arid zones is doubtful, when using
typical available data sets. We hypothesise that well-established, inexpensive,
empirical methods may be at least as useful for predicting the main features of flow
hydrographs. The purpose of this paper is to begin to test this hypothesis by
developing and testing the capability of simple regression models, in application to a
previously unpublished data set from Wadi Ahin, in the Sultanate of Oman.
Hourly data from 36 rainfall–runoff events are presented. Two events—one
extreme high flow outlier and one with unacceptable rainfall errors—were excluded
from the regression. Stepwise least-squares regression of peak flow, flow volume and
runoff coefficient to four rainfall parameters and an antecedent wetness index

Copyright © 2007 IAHS Press


Regression analysis of rainfall–runoff data from an arid catchment in Oman 1117

6 fitted regression
90% intervals
observed data

Qp (mm/hr)
observed 29 March 1997
3

0 30 60 90
Rainfall volume, Rv (mm)

20
Qv (mm)

10

0 30 60 90
Rainfall volume, Rv (mm)

Fig. 6 Regression models (equations (8) and (9)) with observed data and 90%
confidence intervals.

(baseflow) consistently indicated that rainfall volume is the primary control on flow
volume and peak. Other effects which are significant but relatively weak are: increased
antecedent wetness increases flow volume and peak; increased spatial variability of rain-
fall increases flow volume and peak; increased distance of the rainfall centroid from the
flow gauge reduces flow volume and peak; increased peak rainfall increases peak flow.
The predictive ability of the regression procedure was tested by treating each of
the 34 events, in turn, as a validation event. The best predictive performance was
achieved by the simplest regression equation: a linear regression of flow peak and flow
volume to rainfall volume. Using this model, 16 of the 34 flow peaks were predicted to
within 30% of the observed; and 11 of the 34 flow volumes were predicted to within
30% of the observed. This is not impressive; however, the 90% confidence intervals
contained all but two of the observations of flow peaks and all but one of the observed
flow volume. The cumulative flow volume was predicted to within 0.2%. A 35th
event, the high flow outlier, could not be explained by any of the regression models.
We speculate that a distributed, process-based, simulation model will not be able
to achieve this performance in predicting flow peaks and volumes, or the same
robustness of uncertainty analysis, using the same data set plus topography, land cover
and soils data. Such a model will, however, allow us to develop our initial insights into
responses, achieved using regression, into testing hypotheses about the physical
processes behind the responses. This will be done in future research.

Acknowledgements The authors would like to thank the Ministry of Regional


Municipalities, Environment and Water Resources of the Sultanate of Oman for
provision of the Wadi Ahin data; and the Islamic Development Bank, who have partly
funded this research.

Copyright © 2007 IAHS Press


1118 Neil McIntyre et al.

REFERENCES
Al-Qurashi, A. M. (1995) Rainfall–runoff relationships in arid areas. MSc Thesis, Heriot-Watt University, Edinburgh, UK.
Beven, K. J. (2000) Rainfall–Runoff Modelling—The Primer. John Wiley & Sons Ltd, Chichester, UK.
Costelloe, J. F., Grayson, R. B. & McMahon, T. A. (2006) Modelling streamflow in a large anastomosing river of the arid
zone, Diamantina River. Aust. J. Hydrol. 323(1-4), 138–153.
El-Hames, A. S. & Richards, K. S. (1998) An integrated, physically based model for arid region flash flood prediction
capable of simulating dynamic transmission loss. Hydrol. Processes 12, 1219–1232.
Grayson, R. B., Moore, I. D. & McMahon, T. A. (1992) Physically based hydrologic modelling: 1. A terrain-based model
for investigative purposes. Water Resour. Res. 28(10), 2639–2658.
Hughes, D. A. (1995) Monthly rainfall–runoff models applied to arid and semiarid catchments for water resource
estimation purposes. Hydrol. Sci. J. 40(6), 751–769.
Lane, L. J., Diskin, M. H. & Renard, K. G. (1971) Input-output relationships for an ephemeral stream channel system.
J. Hydrol. 13, 22–40.
Lange, J. (2005) Dynamics of transmission losses in a large arid stream channel. J. Hydrol. 306(1-4), 112–126.
Michaud, J. D. & Sorooshian, S. (1994a) Comparison of simple versus complex distributed runoff models on a midsized
semiarid watershed. Water Resour. Res. 30(3), 593–606.
Michaud, J. D. & Sorooshian, S. (1994b) Effect of rainfall-sampling errors on simulations of desert flash floods. Water
Resour. Res. 30(10), 2765–2775.
MWR (Ministry of Water Resources) (1995) Water Resources of the Sultanate of Oman—An introduction guide. Ministry
of Water Resources, Sultanate of Oman.
Nouh, M. (1990) Flood hydrograph estimation from arid catchment morphology. Hydrol. Processes 4, 103–120.
Nouh, M. (2006) Wadi flow in the Arabian Gulf states. Hydrol. Processes 20, 2393–2413.
Osborn, H. B. & Lane, L. J. (1969) Precipitation–runoff relations for very small semiarid rangeland watersheds. Water
Resour. Res. 5, 419–425.
Pilgrim, D. H., Chapman, T. G. & Doran, D. G. (1988) Problems of rainfall–runoff modelling in arid and semiarid regions,
Hydrol. Sci. J. 33(4), 379–400.
Rodier, J. A. (1985) Aspects of arid zone hydrology. In: Facets of Hydrology, vol. II (ed. by J. C. Rodda), 205–247. John
Wiley & Sons Ltd, Chichester, UK.
Sharma, K. D. & Murthy, J. S. R. (1996) Ephemeral flow modelling in arid regions. J. Arid Environ. 33, 161–178.
Sharma, K. D. & Murthy, J. S. R. (1998) A practical approach to rainfall–runoff modelling in arid zone drainage basins.
Hydrol. Sci. J. 43(3), 331–348.
Sorman, A. U. & Abdulrazzak, M. J. (1993) Flood hydrograph estimation for ungaged wadis in Saudi-Arabia. J. Water
Resour. Plan. Manage. ASCE 119(1), 45–63.
Wheater, H. S. (2005) Modelling hydrological processes in arid and semi-arid areas. In: Proc. UNESCO G-WADI
Workshop on Hydological Modelling in Arid and Semi-Arid Areas (Roorkee, India) http://g-wadi.com.
Wheater, H. S. & Brown, R. P. C. (1989) Limitations of design hydrographs in arid areas-an illustration from south west
Saudi Arabia, In: British Hydrological Society National Symposium (Sheffield, Sept 1989). BHS, London, UK.
Wheater, H. S., Butler, A. P., Stewart, E. J. & Hamilton, G. S. (1991a) A multivariate spatial-temporal model of rainfall in
SW Saudi Arabia. I. Data characteristics and model formulation. J. Hydrol. 125, 175–199.
Wheater, H. S., Onof, C., Butler, A. C. & Hamilton, G. S. (1991b) A multivariate spatial-temporal model of rainfall in SW
Saudi Arabia. II. Regional analysis and model validation. J. Hydrol. 125, 201–220.
Wheater, H. S., Jolley, T. J. & Peach, D. (1995) A water resources simulation model for groundwater recharge studies: an
application to Wadi Ghulaji, Sultanate of Oman. In: Proc. Intl. Conf. on Water Resources Management in Arid
Countries (Muscat), 502–510.

Received 19 January 2007; accepted 11 October 2007

Copyright © 2007 IAHS Press

Vous aimerez peut-être aussi