You are on page 1of 14

Journal of Hydrology 373 (2009) 204217

Contents lists available at ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

Rainfallrunoff modeling in a small hyper-arid catchment


Yonatan Bahat, Tamir Grodek, Judith Lekach, Efrat Morin *
Department of Geography, Hebrew University of Jerusalem, Jerusalem, Israel

a r t i c l e i n f o s u m m a r y

Article history: Quantitatively estimating rainfallrunoff relations in extremely arid regions is a challenging task, mainly
Received 24 September 2008 because of lack of in situ data. For the past 40 years, rain and oods have been monitored in the Nahal
Received in revised form 30 March 2009 Yael catchment (0.5 km2) in southern Israel, providing a unique data set of runoff hydrographs and rain-
Accepted 26 April 2009
fall in a hyper-arid region. Here we present an exploratory study focusing on rainfallrunoff modeling
issues for a small (0.05 km2) sub-catchment of Nahal Yael. The event-based model includes the compu-
This manuscript was handled by tation of rainfall excess, hillslope and channel routing. Two model parameters of the inltration process
K. Georgakakos, Editor-in-Chief, with the were found by calibration. A resampling methodology of calibration group composition is suggested to
assistance of Ana P. Barros, Associate Editor derive optimal model parameters and their uncertainty range. Log-based objective functions were found
to be more robust and less sensitive than non-log functions to calibration group composition. The t
Keywords: achieved between observed and computed runoff hydrographs for the calibration and validation events
Rainfallrunoff model is considered good relative to other modeling studies in arid and semi-arid regions. The study indicates
Hyper-arid catchment that, under the calibration scheme used, a lumped model performs better than a model representing the
Objective function catchment division into three sub-catchments. In addition, the use of rain data from several gauges
Uncertainty analysis improves runoff prediction as compared to input from a single gauge. It was found that rainfall uncer-
Model calibration
tainty dominates uncertainties in runoff prediction while parameter uncertainties have only a minor
Israel
effect.
2009 Elsevier B.V. All rights reserved.

Introduction Research related to rainfallrunoff modeling in arid and semi-


arid catchments has focused on several questions:
Rainfall in arid and semi-arid climate regions is typically mea-
gre, irregular and highly variable (Sharon, 1972; Goodrich et al., (1) What level of model complexity is appropriate? Theoreti-
1995; Ahrens, 2003). Runoff generation in these regions is domi- cally, the high rainfall variability typical of dry environments
nated by an inltration excess mechanism with a short time to - supports the utilization of high-resolution complex hydro-
nal inltration rates and a fast response due to steep hillslopes logical models that represent small-scale processes in the
with shallow soils, exposed rocks and lack of vegetation (Wheater, catchment. In practice, however, the complex models often
2002; Greenbaum et al., 2006). Some of arid and semi-arid catch- do not provide more accurate runoff predictions relative to
ments have alluvial channel beds through which ood water inl- simpler and less detailed ones. For example, Michaud and
trates during ow and ood peaks and thus volumes are reduced Sorooshian (1994a), for the 150-km2 semi-arid Walnut
(Wheater, 2002). For catchments with rocky channel beds, this Gulch experimental watershed in Arizona, found similar per-
process is not signicant. Rainfallrunoff modeling in arid and formance of simple (Soil Conservation Service (SCS), 1964)
semi-arid catchments is a challenging task for several reasons. Rain and complex Kineros (Smith et al., 1995) calibrated distrib-
intensity data for short durations are important input in such mod- uted models and poor performance of the lumped (SCS)
els for the computation of inltration and rainfall excess, but rain- model. Al-Qurashi et al. (2008) modeled a 734-km2 arid
fall spatial distribution over the catchment is often not well watershed in Oman with the complex Kineros model but
represented by rain-gauge networks (Faures et al., 1995). In addi- performance was poor and inferior to that of a simple
tion, many of the hydrological models developed for more humid regression model.
areas are tuned to a saturation excess mechanism and not to the (2) What is the appropriate model-calibration strategy? In most
inltration excess mechanism that often dominates in dry regions. models, some parameters require calibration. A calibration
The process of transmission loss is also missing from many of the strategy includes the selection of calibration data, parame-
humid-region models. ters to calibrate, objective functions, and an optimization
algorithm (Sorooshian and Gupta, 1995). These choices can
* Corresponding author. Tel.: +972 2 5883020; fax: +972 2 5820549. have a great effect on the models performance. For example,
E-mail address: msmorin@mscc.huji.ac.il (E. Morin). in a modeling study of an arid catchment in central

0022-1694/$ - see front matter 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.jhydrol.2009.04.026
Y. Bahat et al. / Journal of Hydrology 373 (2009) 204217 205

Australia, Costelloe et al. (2005) claimed inconsistency of Study area and data
model parameter sets found by manual calibration between
medium and large oods. Yatheendradas et al. (2008) Nahal Yael is a 0.5-km2 catchment located at the edge of the
reached a similar conclusion in another study using an auto- southern Negev Desert in Israel, near the town of Eilat (Fig. 1a
matic calibration. Al-Qurashi et al. (2008) tested different and b) in a hyper-arid climate region. The catchment has been
calibration strategies that included six objective functions, equipped with rain gauges, hydrometric stations and other devices
reducing the number of calibrated model parameters, and since 1965 by Prof. Schick (1970, 2000). This record provides a rare
excluding problematic events from the calibration, however and unique data set allowing investigation of rainfallrunoff pro-
they reported that none of these actions improved model cesses in a small catchment located in a hyper-arid climate regime.
performance. In the current work, we focused on modeling hillslope inltration
(3) What are the major sources of uncertainty in runoff predic- and hillslope- and channel-routing processes. We therefore ana-
tion? Possible sources of uncertainty are rainfall and runoff lyzed an instrumented sub-catchment of Nahal Yael located in its
data, model structure, parameters and natural uncertain- uppermost part, the 0.05-km2 Yael05 catchment (Fig. 1b and c),
ties. Yatheendradas et al. (2008) showed, for a sub-catch- which has a rocky channel bed in which transmission losses
ment of the Walnut Gulch watershed, that the (occurring downstream of this sub-catchment) are minor (Green-
uncertainty due to bias in the radar rainfall estimates baum et al., 2003).
almost completely dominated the uncertainty in the mod-
eled response. They also identied a greater impact on run- Hydrometeorological characteristics
off predictions of uncertainties in the hillslope model
parameter values than of uncertainties in the channel Climate in the studied region is classied as hyper-arid, mean-
parameters. McMichael et al. (2006) applied the MIKE- ing that precipitation is less than 3% of the potential evapotranspi-
SHE model (Abbott et al., 1986) to a 34-km2 semi-arid ration (FAO, 1989). The mean temperature is 25 C but can reach
watershed in California and computed the uncertainty 50 C in summer and the mean annual potential evapotranspira-
bounds. They found that observed values outside the tion is 2800 mm. Mean annual rainfall is 28 mm with large year-
uncertainty bounds are generally associated with large to-year variability (Fig. 2). Most of the rain events are local, and
rainfall and wildre events, indicating model or data de- originate in small-diameter convective cells (Sharon, 1972) devel-
ciency. Both studies, as along with many others, analyzed oped in association with intrusion of the Red Sea Trough synoptic
uncertainties in the Generalized Likelihood Uncertainty systems (Kahana et al., 2002; Dayan and Morin, 2006) northerly
Estimation (GLUE) framework (Beven and Binley, 1992) extension of the Sudanese low, mainly in spring and fall. The rain
that explicitly represents the concept of equinality, i.e., from these convective rain cells is short and intense and has very
the coexistence of alternative model parameter sets that high spatial variability. The remaining rain events are associated
provide equally good model results. Rainfall has been with east Mediterranean lows that occasionally penetrate south-
shown to be the major source of uncertainty in hydrologi- ward far enough to reach the southern Negev and Sinai (Kahana
cal modeling of semi-arid and arid catchments in studies et al., 2002). These events cover larger areas and are of longer dura-
examining the effect of rainfall sampling errors on runoff tion, occur mainly in winter and are typically characterized by low-
prediction. Michaud and Sorooshian (1994b) found for the er rain intensities (Dayan and Sharon, 1980).
Walnut Gulch watershed that approximately half of the dif-
ference between observed and simulated runoff peaks was
due to rainfall-sampling errors. Faures et al. (1995) came to Physical characteristics
the same conclusion looking at a much smaller scale in the
same region and Yatheendradas et al. (2008) showed The morphology of the Yael05 catchment is relatively steep
with an altitude difference of 60 m and hillslope gradients between
through a global sensitivity analysis that in the case of
radar rainfall sampling errors, the difference can be as high 15 and 20. The area is dominated by bare rock of amphibolites
(41%), schist (18%) and gneiss (10%), and the rest is covered by col-
as 80100%.
luvium (31%) (Greenbaum et al., 2003). The bedrock is well frac-
There are other examples of hydrological models applied to tured and its density varies between lithology types: schist
semi-arid and arid environments with different foci. For example, (0.130.9 cm/cm2), amphibolites (0.080.2 cm/cm2) and gneiss
(0.040.08 cm/cm2). Yael05 is a third-order catchment and is an
Shamir et al. (2007) presents a model framework for water-re-
assemblage of three smaller sub-catchments (Fig. 1c). The steep
source management and planning for the arid upper Santa Cruz
bare slopes cause the initiation of short rst-order channels at
River in southern Arizona. CRUM is a distributed, dynamic model
about 1020 m from the headwaters with channel lengths of
for semi-arid catchments described in Reaney et al. (2007) and
130200 m. The symmetrical arrangement of the sub-catchment
Lange et al. (1999) presents a non-calibrated distributed model,
channels results in quick runoff convergence. The colluvium
developed for the 1400-km2 arid catchment of Nahal Zin, Israel.
The current modeling study was conducted in the small hyper- weathered, pebble-sized materialpartly covers the bedrock and
partly creates colluvial slopes.
arid experimental catchment of Nahal Yael in southern Israel,
which offers a unique long record of rainfall and runoff data from
Previous studies
a climate regime for which rainfallrunoff processes are rarely
recorded.
The experimental design of Nahal Yael has enabled a large num-
The objectives of the current study were to:
ber of geomorphological and hydrological investigations over the
years. Most of these have focused on hillslope processes and runoff
1. construct a rainfallrunoff model for the study catchment
2. examine calibrationvalidation strategies in terms of event generation (Schick, 1970; Yair and Klein, 1973; Salomon and
Schick, 1980; Greenbaum et al., 2003), the origin and movement
selection and objective functions
3. examine and formulate the uncertainty characteristics of rain- of sediment (suspended load and bed load) during oods (Schick
and Sharon, 1974; Schick, 1977; Lekach and Schick, 1982; Lekach
fall data and model parameters and apply this formulation to
estimate uncertainties in runoff prediction. and Schick, 1983; Schick and Lekach, 1993), channel losses, water
206 Y. Bahat et al. / Journal of Hydrology 373 (2009) 204217

Fig. 1. Regional map (a), Nahal Yael catchment map (b), and Yael05 catchment map (c). Intensity gauges and hydrometric stations are indicated.

Fig. 2. Annual rain depth measured in Elat for the years 1950/512006/07. Dashed line represents the mean annual rainfall for this period.

budget in the alluvial bed (Schwartz, 1986) and the formation of ferent plots and sediment loads were not found to be correlated
the uvio-pedogenic unit (Lekach et al., 1998). with the hillslope gradient as in humid regions, but to depend on
The rst work to deal with hillslope processes was carried out the surface characteristics. The highest ratios and maximum sedi-
by Yair and Klein (1973). Three experimental lots (4060 m2) were ment load were found for low-gradient lots with small grain size.
built on colluvial amphibolite hillslopes in the Yael05 catchment. Salomon and Schick (1980) and Greenbaum (1986) studied the
The plots differed in their percentage of colluvial coverage, hill- formation of runoff on hillslopes. Inltration rates were deter-
slope gradient and colluvium grain size. For each plot, inltration mined by rain-simulation experiments on several lots with an area
rates were measured and turned out to be correlated with gradient of 0.25 m2. Values for initial loss and nal inltration rate were de-
and grain size. The runoff-to-rainfall ratios obtained from the dif- rived for different kinds of lithology, hillslope gradients and collu-
Y. Bahat et al. / Journal of Hydrology 373 (2009) 204217 207

vial coverage with an approximate range of 0.55 mm and 0.5


30 mm/h for initial loss and nal inltration rate, respectively
(Greenbaum et al., 2003).
Greenbaum (1986) measured high inltration values on collu-
vial hillslopes, which detained runoff generation. Yair and Lavee
(1985) indicated that uphill runoff might inltrate into the down-
hill colluvium. Therefore, not all of the hillslope area contributes
runoff to the channels.
The relatively dense rain gauge network at Nahal Yael catch-
ment allowed spatial patterns of rainfall to be investigated in this
hyper-arid environment. Sharon (1972) describes the spottiness of
rainfall in the region with estimates from the catchment indicating
typical rain cell diameters of 45 km. In another study, Sharon
(1970) examined the effect of local topography on rain areal pat-
terns in the catchment and reports an average increase of 40
50% in total rain depth over the channel area as compared to the
water divide. It must be emphasized, however, that this behaviour
was found for a topographic transaction in the middle part of Nahal
Yael further downstream to the current studied sub-catchment Fig. 4. Coefcient of determination (R2) between runoff peak discharge and
(Yael05) where the topographical differences and gradients are maximal rain intensity for different durations as computed for the 16 analyzed
rainfallrunoff events.
much higher. This effect is not expected to be signicant in the
Yael05 sub-catchment that is located in the uphill part of Nahal
Yael (Fig. 1b) as evident also in storm rain maps (Sharon, 1970; were installed in the catchment recording rain accumulations or
Figs. 3 and 4). ow level vs. time by rotating a drum with a paper chart attached
Although an extensive and long record of rainfall and runoff controlled by a mechanical clock. A trace of a pen on the chart is
data is available for the catchment, it is only recently that an at- then digitized to derive the data in a digital form. One common
tempt has been made to apply a rainfallrunoff model (Thormah- implication of mechanical gauges and the digitization process is
len, 2003). The current study focuses on rainfallrunoff modeling inaccurate timing of the data that could reach a few tens of min-
of the Yael05 catchment. utes in some cases. To demonstrate this problem, Fig. 3 presents
observed data for the storm of 26/12/1980 including rain intensity
Instrumentation and data data for two rain gauges and observed runoff data at the Yael05
outlet. The rst rain gauge is located within the Yael05 catchment
Since 1965 there has been a dense network of various gauges in less than 200 m to the catchment outlet. The second rain gauge
Nahal Yael: 13 intensity-recording rain gauges, 42 accumulation (Gauge 2 in Fig. 3) is located further north about 800 m from the
rain gauges, 5 hydrometric stations, 2 wind gauges and 7 samplers rst gauge. In this example the timing of Gauge 1 seems suspicious
for suspended matter (Fig. 1b). This extensive network was oper- because of the large rainfallrunoff lag time (40 min between rain-
ated for about 10 years and later, the rain gauge network was re- fall centre of mass and runoff peak discharge) as well as the large
duced to between 2 and 5 intensity-recording rain gauges. time differences between the rainfall data at the two rain gauges
Within and near the Yael05 catchment there were four intensity (40 min between time of maximal rain intensity). The source of
rain gauges and a hydrometric station at the catchment outlet this error is probably the time reference in the digitization process
(Fig. 1c). Not all rain gauges, however, worked during all of the of Gauge 1 data. Such errors can be corrected (by identifying the
years of the record. In the rst two decades, mechanical gauges correct time reference) but was not possible to do so in this case

Fig. 3. Observed rain intensity from two gauges and runoff data at Yael05 outlet for the storm of 26/12/1980. Gauge 1 is located within Yael05 catchment less than 200 m
travel distance to the catchment outlet and Gauge 2 is located further north in 800 m distance from the rst gauge. In this example the timing of Gauge 1 seems suspicious
because of the large rainfallrunoff lag time as well as the large time differences between the rainfall data at the two rain gauges (see Instrumentation and data).
208 Y. Bahat et al. / Journal of Hydrology 373 (2009) 204217

Table 1
List of selected rainfallrunoff events.

Date Rain depth Range Max. 20-min rain Rain duration Number of Runoff volume Peak discharge Calib (C)/valid
(mm) (mm)a intensity (mm/h) (min) gauges (m3) (m3/s) (V)
24/11/1968 6.4 1.0 12.1 185 3 12.6 0.02 C
25/11/1968 8.1 2.3 10.7 304 3 25.7 0.10 C
21/01/1969 18.6 7.8 17.1 400 2 241.7 0.35 C
25/03/1971 5.2 0.3 12.2 51 2 36.1 0.03 C
22/12/1971 12.3 10.1 273 1 91.9 0.03 C
20/02/1975 32.0 10.6 629 1 888.4 0.15 C
11/12/1978 11.3 0.1 9.3 301 2 105.6 0.07 V
09/02/1979 6.3 2.4 6.2 190 2 72.4 0.06 C
26/12/1980 30.8 23.7 599 1 545.8 0.63 V
30/10/1981 20.5 2.7 38.7 67 2 696.8 0.69 C
22/03/1985 6.3 3.4 17.2 182 2 46.1 0.18 V
16/10/1987 8.3 2.7 12.2 239 2 16.2 0.10 C
29/11/1989 3.4 10.2 16 1 24.5 0.03 V
23/10/1990 11.3 33.6 40 1 177.4 0.52 C
01/01/1994 5.9 11.4 124 1 12.2 0.05 C
18/10/1997 25.4 35.1 94 1 670.1 0.62 C
a
Range of observed rain depth for events with more than one gauge data: maximum minus minimum observed rain depth.

and is generally recognized as timing errors of the data. In the late capacity parameter (mm/h). When these conditions are met, the
1980s, digital tipping-bucket rain gauges were installed and timing rainfall excess (in mm/h) is the difference between rain intensity
accuracy was considerably improved to minutes at most. and the inltration capacity; otherwise, it is zero. In this inltra-
Intensive quality control was conducted for the rainfall and run- tion model the constant inltration capacity represents the nal
off data records. From the complete record of rainfallrunoff inltration rate which is attained relatively quickly in arid regions
events, we chose 16 good-quality storms for the analysis. Their (Greenbaum et al., 2006; Morin et al., in press). Rainfall excess data
main rainfallrunoff characteristics are presented in Table 1 with are computed for rain data resolution, i.e. 1 min.
rainfall information based on the weighted average of available Rainfall excess ows over the hillslope toward the channel and
gauges for each event. The range of rainfall observations (third col- enters as lateral ow into the channel segment. The routing of the
umn) is the maximal minus minimal observed value. As can be rainfall excess is applied by the one-dimensional kinematic wave
seen in the table, runoff is observed in the catchment from storms model. This model is a simplied form of the continuity and
as small as 3.4 mm. The table presents the maximal rain intensity momentum equations neglecting local acceleration, convective
for 20-min duration because the correlation between maximal rain acceleration and pressure terms in the latter and thus assuming
intensity and runoff peak discharge is highest (R2 = 0.87) at dura- that the friction and surface slopes are equal (Bras, 1990). The kine-
tions of 1530 min (Fig. 4). A high correlation (R2 = 0.86) also exists matic wave equation for surface ow is
between storm rain depth and runoff volume. The large variation p
in event rainfalls is amplied when runoff characteristics are @yt; x s @yt; x
myt; xm1 Ie t 1
examined. For example, the coefcient of variation (CV) of rain @t n @x
depth and 20-min maximal intensity is 0.6 and 0.7, respectively,
where y(t, x) is the water depth (m) at time t and distance x from
while the CV of runoff volume and peak discharge is 1.3 and 1.1,
the slope top, m = 5/3, s is the hillslope gradient, n is the hillslope
respectively. Specic peak-discharge values, computed as the peak
Manning roughness coefcient, and Ie(t) is the uniform rainfall ex-
discharge in m3/s divided into the catchment area in km2, are very
cess over the hillslope at time t (m/s).
high for some of the events (maximum of 13.5 m3/s km2 for the
The lateral inow from the hillslope bottom into the channel
October 1981 event)a typical phenomenon for small arid catch-
q(t) in discharge per unit length (m2/s) is computed by
ments. Rainfallrunoff ratios are in the range of 467%. p
s
qt yl; tm 2
Rainfallrunoff model n
where l is the hillslope length (m).
The event-based hydrological model represents the major rain- Water enters the channel laterally from hillslopes or from up-
fallrunoff processes in the Yael05 catchment: rainfall over the stream channels. Channel-ow routing is also computed by the
hillslopes, generation of rainfall excess, routing of water over hill- kinematic wave model:
slopes that enter laterally into the channel, and ow routing in
p
channels toward the outlet. As explained above, transmission @At; x sc @At; x
mc At; xmc 1 qt 3
losses are neglected in the rocky channel bed of the studied catch- @t nc p2=3 @x
ment. The basic modeling unit is a sub-catchment that includes
two equal hillslopes and a channel. Headwaters are neglected be- where A(t, x) is the wetted cross-sectional area (m2), mc = 5/3, sc the
cause of their short length in the studied catchment (see Physical channel gradient, nc the channel Manning roughness coefcient,
characteristics). and p is the cross-sectional wetted perimeter (m). Discharge Q(t)
Rainfall is assumed to be uniform over each sub-catchment and in m3/s at the channel outlet is computed by
is computed as the weighted average of 1-min rain intensities p
sc
measured by gauges within or near each sub-catchment with Qt Alc ; tmc 4
nc p
weights applied using the Thiessen polygon method.
Rainfall excess in the model is generated when the accumulated where lc is channel length.
event rainfall depth is larger than the initial loss parameter value The above kinematic ow equations are solved by a backward
(mm) and the rain intensity is higher than the constant inltration numerical scheme (Bras, 1990), assuming dry initial conditions.
Y. Bahat et al. / Journal of Hydrology 373 (2009) 204217 209

The numerical routing components are computed in a time step of (2) provide a distribution of parameter values representing the
4 s, and spatial steps of 10 and 50 m for the hillslope- and channel- uncertainty associated with the data sample, and
routing components, respectively. Most of the model parameters (3) examine the dependency of the optimal parameter values on
were found from eld and topographic data. The values of the hill- the size of the sub-group and the specic events included in
slope and channel Manning roughness coefcients are xed as 0.08 it.
and 0.025, respectively, according to surface characteristics.
Four model combinations with two levels of catchment detail- Several objective functions are examined; all are based on the
ing and two levels of rainfall detailing are examined. The catch- root mean square difference (RMSD) function:
ment is rst considered as one sub-catchment (i.e., lumped r
model), and then divided into three sub-catchments (3-Sub) whose 1 Xn
RMSDV Voi  Vci 2 5
channels join at the end. Rainfall is applied once with all gauges n i1
r
(one to three gauges) available for each storm and once with only 1 Xn
RMSDQ Qoi  Qci 2 6
one gauge, as close as possible to the catchment center. Model n i1
r
parameters are assumed uniform in all combinations. The 3-Sub 1 Xn
model allows for differences in rainfall between the three sub- RMSDLV logVoi 1  logVci 12 7
n i1
catchments, obviously only relevant if data from more than one r
1 Xn
gauge are available. Differences between the lumped and 3-Sub RMSDLQ logQoi 0:001  logQci 0:0012 8
models are also attributed to the routing procedures. n i1

where V is runoff volume (m3), Q is runoff peak discharge (m3/s),


Calibrationvalidation strategies the o and c postxes stand for observed and computed, respectively,
i is an event index and n is the number of events in the calibration
Special attention was paid in this study to the calibration and group.
validation strategies. Two model parameters were found by cali- Constants that are considerably smaller than all observed val-
bration: initial loss (mm) and constant inltration capacity (mm/ ues are added to the logarithmic objective functions (Eqs. (7) and
h). The rainfallrunoff event record (Table 1) was divided into cal- (8)) to prevent analysis of zero values in a log function; 1 m3 and
ibration and validation groups (last column in Table 1). While it is 0.001 m3/s for the volume and peak discharge, respectively.
generally advisable that both groups will be large enough and will The calibration is done by resampling sub-groups of different
represent the evident high variability of event magnitudes, the di- sizes from the calibration group and nding the optimal parameter
rect role of the calibration group in determining model parameter values for each sub-group and objective function by scanning the
values puts more importance to this group. Therefore the calibra- parameter space at constant intervals and nding the parameter
tion group contains 12 events and the validation group contains set with the minimal objective function value. Fig. 5 presents the
four events; both groups are composed from small, medium and median (circle) and the 95% range (vertical bars) of parameter val-
large events. The events in the validation group are not included ues found as a function of sub-group size for the lumped model
in the calibration procedure and are used to assess independently with all gauges included. As can be seen, for small sub-group sizes
the calibrated model performance. the parameter range in all cases is quite large, indicating a high
Our goals in the calibration process are to: uncertainty in the calibrated parameters. For sub-groups of nine
events or more, the RMSDLV and RMSDLQ objective functions pres-
(1) identify optimal model parameter values based on one or ent a higher stability of parameter values while the RMSDV and
more objective functions that are best suited to the problem, RMSDQ functions continue to range substantially.

Fig. 5. Optimal initial loss and inltration capacity parameters found for the all combinations of sub-groups of 12 events as a function of sub-group size. Circles represent the
median of the values found and bars represent the 95% range of the values. The analysis is shown for the four objective functions examined.
210 Y. Bahat et al. / Journal of Hydrology 373 (2009) 204217

The difference between the standard and log-based objective Table 2


functions is that the former are more sensitive to large events in Optimal parameter values for the lumped model with all-gauge data and performance
criteria for calibration and validation events.
the calibration group. For example, the optimal parameters based
on RMSDQ for all 12 events together are 9 mm and 1 mm/h for Calibration events Validation events
the initial loss and inltration capacity, respectively. If the largest Initial loss 3 mm 3 mm
runoff volume event of February 20, 1975, is removed from this Inltration capacity 3 mm/h 3 mm/h
group, the calibration yields values of 6 mm and 8 mm/h. The RMSDV 163.5 m3 (70%) 114.7 m3 (64%)
RMSDQ 0.15 m3/s (71%) 0.04 m3/s (20%)
parameters obtained for the RMSDLV objective function are RMSDT 27.3 min (19%) 23.5 min (12%)
4 mm and 3 mm/h for groups with or without the February 20 BiasV 94.1 m3 (40%) 76.5 m3 (42%)
event. The response surfaces for this example are shown in BiasQ 0.05 m3/s (23%) 0.02 m3/s (9%)
Fig. 6. Although the RMSDV response surface is smoother than that BiasT 8.6 min (6%) 8.6 min (4%)
NashSutcliffeV 0.71 0.71
of RMSDLV, the minimum area shifts when the large event is re-
NashSutcliffeQ 0.60 0.96
moved from the group. The response surface of RMSDLV is less NashSutcliffeT 0.95 0.96
smooth but the optimal region is more stable.
V runoff volume.
Similar behavior was found for other objective functions, such
Q runoff peak discharge.
as the mean absolute difference of volume or peak discharge. They T runoff peak time.
are relatively sensitive to the presence of a single large event in the Number in parenthesis are percentage of average value.
calibration group, while their logarithmic versions are consider-
ably more stable.
Examination of the results from the four objective functions Instrumentation and data, above). It should be emphasized, how-
(Fig. 5) led to selection of the RMSDLQ objective function for cali- ever, that these differences are averaged to 19% of the mean ob-
bration. To obtain an optimal parameter set as well as an assess- served (Table 2) and are considered satisfactory. It is also
ment of the uncertainties in these parameters, the analysis of recognized that, in general, the falling limb of the observed hydro-
nine event sub-groups is used. The median of the 220 combina- graph decays faster than the falling limb of the computed hydro-
tions (selection of nine events out of twelve) is used as optimum graph (Fig. 7a). This may indicate that the effective value of the
and the distribution of parameter values describes their uncertain- Manning roughness coefcient needs to be lower than the xed
ties (Uncertainty analysis, below). (uncalibrated) value determined in this study.
Table 2 (rst column) presents the optimal parameters for the The optimal parameter values are applied in the simulation of
lumped model with all rain gauges and values of several perfor- the four validation events (Fig. 7b) and the t achieved is as good
mance criteria for the 12 events in the calibration group, including as for the calibration events with relatively close peak discharge,
the RMSD, Bias and NashSutcliffe efciency score for the volume, and volume and hydrograph shape. However, there are some time
peak discharge and time of peak. Fig. 7a presents observed and discrepancies that we speculate to be caused by errors in the ob-
computed runoff hydrographs for four calibration events. A rela- served data. Table 2 (second column) presents the performance cri-
tively good t is obtained in both magnitude and hydrograph teria for the validation events. Observed and computed peak
shape. Differences in timing may well be related to the time accu- discharge and volumes of all the analyzed events are presented
racy of rain and runoff data obtained from mechanical stations (see in Fig. 8a and b. From qualitative (Figs. 7 and 8) and quantitative

Fig. 6. Response surfaces of the RMSDV (a and b) and RMSDLV (c and d) objective functions based on all 12 events (a and c) and when the extreme event of February 20 1975
is excluded (b and d).
Y. Bahat et al. / Journal of Hydrology 373 (2009) 204217 211

Fig. 7. Observed (thick line) and computed (thin line) runoff hydrographs for (a) four calibration events (out of 12, see Table 1), and, (b) the four validation events. Note the
different x-axis and y-axis scales in the plots.

(Table 2) assessments of the t between observed and computed regions supports this conclusion. For example, Michaud and Soro-
runoff, we found that the model performance was relatively good. oshian (1994a) applied the Kineros model with 11 parameter cali-
Comparison with previous modeling studies in semi-arid and arid bration and derived RMSD and Bias criteria based on runoff
212 Y. Bahat et al. / Journal of Hydrology 373 (2009) 204217

Fig. 8. Comparison of observed and computed runoff peak discharge (a and c) and runoff volume (b and d) for the spatially lumped (a and b) and 3-Sub (c and d) models.
Circles show events in the calibration group and stars show events in the validation group.

volume, peak discharge, and peak time for calibration and valida- bration data in terms of the RMSDLQ criteria. A similar result,
tion events. While the performance for the calibration events is although less pronounced, is found in a cross-validation scheme
generally better in Michaud and Sorooshian (1994a), the current where the optimal parameters for each sub-group of nine events
study performance is better for the validation events. are applied to the other three events in the calibration group.
Fig. 9c shows that in most cases (173 out of 220 combinations),
Catchment discretization and rainfall resolution effect the RMSDLQ of the lumped model is lower than that of the 3-Sub
model.
The two optimal parameters for the four model combinations The superiority of the lumped vs. 3-Sub model lies in those
(lumped vs. 3-Sub and all gauges vs. one gauge) were found with cases in which the model overestimates the peak discharge, which
the above-described procedure, i.e., median of all optimal parame- contributes a relatively large portion of the error value. The
ters for each sub-group of nine out of twelve events using the lumped model has a larger attenuating effect and the computed
RMSDLQ objective function. The calibrated parameter values for runoff discharge values are generally lower, in particular the over-
the 3-Sub model are 3 mm and 5 mm/h for the initial loss and inl- estimation is less pronounced (compare Fig. 8a and c) and there-
tration capacity parameters, respectively. Fig. 8c and d compares fore the lumped model results in better scores. It must be
the observed and computed peak discharge and volume based on emphasized, however, that the better performance of the lumped
the calibrated 3-Sub model. The level of spatial rainfall information model was obtained with a uniform assumption of Manning
does not affect the optimal parameter values for the two catch- roughness coefcient and without a calibration of this parameter.
ment resolutions and the same optimum is found based on one Hence this result may be a consequence of the calibration scheme
gauge or all gauges. used here.
For each of the 220 calibration event sub-groups, the minimal The advantage of the detailed rainfall information (observed in
RMSDLQ values achieved are compared between the lumped and Fig. 9a) probably results from the better accuracy of the spatially
3-Sub models (Fig. 9a) and between all-gauge and one-gauge rain- averaged rain-intensity data when computed from several gauges
fall information (Fig. 9b). Fig. 9 clearly shows that the lumped as compared to a single gauge. Although in principle the model
model and all-gauge information provide a better t for the cali- could also benet from the more detailed information on the spa-
Y. Bahat et al. / Journal of Hydrology 373 (2009) 204217 213

Fig. 9. The optimum value of RMSDLQ obtained for each of the 220 sub-groups of nine calibration events: (a) using data from a single rain gauge vs. from all gauges. (b) Using
the lumped model vs. 3-Sub model. (c) Cross-validation: the RMSDLQ values obtained when parameters were optimized for a sub-group of nine events are tested on the other
three events. Points under the line are better (lower value) in terms of the y-axis model and points above the line are better in terms of the x-axis model.

tial distribution of the rainfall, this is not the case here as no where the PDFs of the sources are formulated based on prior
improvement is achieved with the 3-Sub model as compared to knowledge or by making some assumptions on their form, and
the lumped model. To better demonstrate this, we compared the the runoff PDFs are estimated with Monte Carlo simulation utiliz-
RMSDQ achieved by the 3-Sub model with: (1) distributed rainfall ing the sources PDFs. In the current study, we focus on two sources
(i.e., each of the three sub-catchments receives different rainfall in- of uncertainty: model parameters and rainfall input data. The
put), (2) spatially averaged rainfall (all sub-catchments receive the lumped model is applied because it provides a better t to the ob-
same rainfall input, which is the averaged catchment rainfall), and served data (Catchment discretization and rainfall resolution
(3) rainfall from a single gauge. This analysis includes only those effect).
events with available data from more than one gauge (9 out of Uncertainties in model parameters are related to the optimiza-
16 events). The RMSDQ value obtained was: 0.057 m3/s, tion procedure, events making up the calibration group, data er-
0.056 m3/s and 0.082 m3/s for the above three types of rainfall dis- rors, and model structural errors, among others. The concept of
tributions, respectively. The difference between distributed and equinality claims that there is more than one set of optimal
lumped rainfall is very small as compared to the difference of parameters that can provide an adequate t of observed and com-
one vs. two or three gauges. It is postulated that an improvement puted data (Beven and Binley, 1992).
associated with spatial distribution of rainfall can be achieved In the current study, the information on model parameter
but with a considerably larger density of rain gauges within the uncertainties is obtained in the afore-described calibration process.
study area. The set of parameter values found to be optimal for each sub-group
of nine out of twelve calibration events form the basic distribution
Uncertainty analysis sample. Often independence is assumed between the model
parameters when applying the Monte Carlo simulation; however,
Uncertainty analyses are aimed at quantifying the effect of it is well known that a substantial degree of correlation often exists
uncertainties that exist in different sources of the modeling pro- between the parameters. For this reason, in this study we tted a
cess on uncertainties in runoff predictions. Uncertainties are com- bivariate normal distribution to the sample data, resulting in the
monly represented with probability-distribution functions (PDFs), following PDF (shown also in Fig. 10a):
214 Y. Bahat et al. / Journal of Hydrology 373 (2009) 204217

Fig. 10. Uncertainty of model parameters and rainfall input: (a) the bivariate normal distribution tted to the two model parameters, (b) the normal distribution of the log 10
of ratios of the 1-min rain intensity between one-gauge rainfall and all-gauge rainfall, and (c) autocorrelation and partial autocorrelation graphs of the of the log 10 of ratios of
the 1-min rain intensity between one-gauge rainfall and all-gauge rainfall.


1 1 Obviously this is done for the events with data from more than one
f p1 ; p2 p  exp 
2prp1 rp2 1  q2 21  q2 gauge (9 out of 16 events) and the analysis includes only those
!! cases with intensities of more than 5 mm/h to avoid the very high
p1  lp1 2 2qp1  lp1 p2  lp2 p2  lp2 2
  9 or very low ratios that could be obtained for low intensity values.
r2p1 rp1 rp2 r2p2 Our basic assumption is that this distribution roughly represents
lp1 3:00; lp2 3:18; rp1 0:32; rp2 0:83; q 0:65 the distribution of ratio values between the catchment-averaged
1-min rain intensity value computed from the available gauges
where p1 is the initial loss parameter (mm) and p2, the inltration and the true, unknown, 1-min rain intensity. A normal distribution
capacity parameter (mm/h). Fig. 10a presents the tted PDF with is tted to the base-10 logarithm of the ratio (Fig. 10b):
the mean vector and covariance matrix. The matrix includes the
variance of each parameter: r2p1 and r2p2 and the covariance rp1 p2 x  N0:032; 0:324 10
 
rp p
note that q rp 1rp2 . As expected, there is a relatively high corre- where x is the log 10 of the ratio.
1 2

lation (q) between the two parameters. The temporal structure of the ratio time series (log 10-trans-
Uncertainties in rainfall data are known to be a major source of formed) was examined and shown to behave like a rst-order
uncertainty in runoff predictions (Michaud and Sorooshian, autoregressive process (Chateld, 2004), as indicated by the typical
1994b). They result mainly from inadequate spatial sampling by exponential decay of the autocorrelation graph and by the cutting
the rain gauge network of the rain intensitys spatial distribution. to zero of partial autocorrelations of lag-2 and on (Chateld, 2004)
To quantify these uncertainties, a dense gauge network is required shown in Fig. 10c.
to provide relatively accurate rain information. Although the gauge The derived autoregressive model is
network in our study area is considered very dense for such a small
xt 1 0:7  xt zt
area, it is still limited to between one and three gauges for the dif- 11
ferent events. To estimate the rainfall uncertainties, we examined
zt  N0; 0:164
the distribution of the derived ratio values between 1-min rain where x(t) is the log 10 of the ratio at time step t, and z(t) is a white-
intensity values based on a single gauge and catchment-averaged noise time series. The coefcient of determination of the model is
values based on two or three gauges (as available for each event). R2 = 0.49. Low and high threshold values of 1 and 1 (representing
Y. Bahat et al. / Journal of Hydrology 373 (2009) 204217 215

Fig. 11. Uncertainty in runoff peak discharge prediction resulting from the uncertainties in model parameters and rainfall input shown in Fig. 9. Circles represent the median
of the uncertainty bound and the gray area the 95% uncertainty range. Triangles are observed peak discharge. (a) Combined effect of the two sources, (b) effect of parameter
uncertainty only, and (c) effect of rainfall uncertainty only.

ratios of 0.1 and 10, respectively) are applied to the model to pre- deviation of computed peak discharge is 0.96), which may be
vent unrealistically low or high intensities. These values are rare be- viewed as heteroscedasticity of the model prediction (Sorooshian,
cause they are about three standard deviations distant from the 1981). As expected, the upper limit of the uncertainty range is
mean. more variable than the lower limit, resulting from the non-linear
Monte Carlo simulation is applied to estimate the uncertainties hydrological response of the catchment. For all cases, the observed
in runoff predictions resulting from uncertainties in model param- peak discharge lies within the 95% uncertainty range of the pre-
eters and rainfall input. These two sources are assumed to be inde- dicted peak discharge, although this range is quite large in some
pendent of one another. At each iteration and for each event: (1) of the cases. The ratio of the mean to standard deviation, referred
model parameters are sampled from the bivariate normal distribu- to as the index of reliability, is in the range of 0.62.5. For a peak
tion representing the parameter uncertainties (Eq. (9)) and (2) an discharge of 0.05 m3/s or more, the reliability index is in the range
autoregressive time series of logs of 1-min ratios is sampled of 12.5.
according to Eq. (11) with a length based on the analyzed rain
event and with the rst term sampled from the normal distribution Discussion
of Eq. (10). The catchment-averaged rain intensity data of each 1-
min event are divided by the 10th power of the sampled time ser- In this study, we attempted to model rainfallrunoff for a small
ies. One thousand simulations are performed for each event and (0.05 km2) hyper-arid catchment in southern Israel. A unique re-
the generated PDFs of peak discharge and runoff volumes are cord of rainfall and runoff data is available for this catchment that
examined. This analysis is performed once with the two included goes back to the late 1960s. The event-based model includes the
sources of uncertainty (parameters and rainfall data) and once computation of rainfall excess, hillslope and channel routing. A rel-
for each source separately. atively good t between observed and computed runoff in terms of
Fig. 11a shows the observed peak discharge with the median peak discharge and volume was found after calibration. The study
and the 95% uncertainty range of the computed peak discharges also showed that the lumped model is preferable to a three-sub-
for each of the analyzed events. Fig. 11b and c presents the uncer- catchment model and that using data from several rain gauges per-
tainties for each source separately, and indicates that the main forms better than using data from a single gauge.
source of uncertainty in the runoff prediction is the uncertainty The study emphasizes the following aspects of modeling: (1)
in rain intensities (Fig. 11c), while parameter uncertainties account calibrationvalidation strategies, (2) effect of model and rainfall
for a much smaller portion of the prediction uncertainty (Fig. 11b). detailing, and (3) effect of uncertainties in model parameters and
It is also observed that as the peak discharge increases, so does the rainfall input on uncertainties in the runoff prediction. The impli-
uncertainty range (the correlation between the mean and standard cations of the results found in this work are discussed below.
216 Y. Bahat et al. / Journal of Hydrology 373 (2009) 204217

Almost all hydrological models require calibration, but it has concluded that inltration rates are considerably lower under nat-
been long recognized that this is not a straightforward process ural rainfall.
(Sorooshian and Gupta, 1995). The dependency of parameter iden- Our study indicates better model performance when a lumped
tication on data quality and quantity (e.g. Gupta and Sorooshian, model with all gauge data is used rather than data from a single
1985) is further amplied in the face of the large hydrological var- gauge. These two model characteristics appear to contradict each
iability typical to arid areas. We show a range of optimal parameter other because the lumped model cannot represent the spatial het-
values for different groups of calibration events (Fig. 5) and the ef- erogeneity of the rainfall over the catchment. We suggest that the
fect that a single extreme event has on the parameter response sur- advantage of the all-gauge data lies in the less biased rainfall input
face (Fig. 6a and b). The same phenomenon was also noted by applied to the model as compared to data from a single gauge. In
Costelloe et al. (2005) in a modeling study of an arid catchment theory, the spatial distribution of the rainfall is important, however
in central Australia, raising the following question: what events it is not necessarily well-represented by the two or three rain
should compose the calibration group? The following strategy is gauges and rainfall biases seem to have a larger effect on runoff
suggested: rather than use all calibration events in a single group, prediction. The importance of unbiased rainfall input for accurate
the model should be calibrated for several combinations of sub- runoff prediction was also demonstrated for a different scale and
groups and a sample distribution of model parameters formed. climate regime by Habib et al. (2008), where bias adjustment of
The central values of the distribution can be assigned as the opti- the radar rainfall estimation error was found to provide the most
mal parameters and the distribution can be further used in the signicant improvement in runoff simulation.
uncertainty analysis. The parameter covariance structure can be Uncertainties in model parameters and rainfall input were for-
extracted from this distribution to provide a more realistic analysis mulated in this study, accounting for correlation of model param-
of uncertainties that does not make the common assumption of eters and temporal dependency of rainfall uncertainties. Rainfall
independence of model parameters. The above analysis also pro- uncertainties were found to be far more important than uncertain-
vides information on the robustness of different objective func- ties in model parameters in terms of runoff prediction uncertain-
tions and on the preferred size of the calibration sub-groups. We ties, a result which is in agreement with other studies (Michaud
believe that this strategy can also be useful for other climate re- and Sorooshian, 1994b; Faures et al., 1995; Al-Qurashi et al.,
gimes, but it is especially recommended where there is large vari- 2008; Yatheendradas et al., 2008).
ability in rainfallrunoff events causing a strong dependency of
optimal parameters on the composition of the calibration group. Conclusions
The degree of t between observed and computed runoff hydro-
graphs should be assessed in the context of arid regions. Wheater The main conclusions from the study are:
(2002) emphasizes the difculties in deriving reliable long-term
hydrological records in arid regions due to sparse populations, lim-  Runoff generation at the study catchment can be well simulated
ited economic resources, harsh climate and infrequent hydrologi- by the presented rainfallrunoff model as compared to hydro-
cal events. In addition to the above limitations, the problem of logical models in arid catchments.
inadequate spatial sampling of rainfall is often recognized as the  A spatially lumped model performs better than a model
most signicant source of error in computed runoff (Michaud accounting for the catchments division into three sub-catch-
and Sorooshian, 1994b; Faures et al., 1995) due to the convective ments under the calibration scheme used in this study.
nature of rainfall in arid areas (Wheater, 2002). The performances  Log-based objective functions are more robust and less sensitive
of the current model in terms of both qualitative resemblance and to calibration group composition.
objective criteria are at least as good as the performances pre-  It is recommended to calibrate the model for different sub-
sented in the few publications that deal with hydrological models groups of calibration events in order to get a sample distribution
in semi-arid and arid catchments. For example, most of the perfor- of optimal parameters.
mance criteria for the validation events are better in this study  The use of rain data from several gauges improves runoff
compared to validation results in Michaud and Sorooshian prediction.
(1994a). Also, the NashSutcliffe values reported in Costelloe  Rainfall uncertainty dominates uncertainties in runoff predic-
et al. (2005) are considerably smaller than those reported here (Ta- tion while parameter uncertainties have a relatively minor
ble 2). It is therefore concluded that the model performances in the effect.
current study are satisfactory for arid region hydrological
modeling.
The parameter values found by calibration were 3 mm for the Acknowledgements
initial loss and 3 mm/h for the inltration capacity. The latter value
is relatively low compared to other studies in the same catchment The research project was funded by the Israel Science Founda-
that reported nal inltration rates in the range of 130 mm/h tion (Grant No. 880/04) and by the Israel Ministry of Science and
depending on the hillslope characteristics (Schick, 1970; Yair and Technology. This study benets from the TerraDB-Nahal Yael Data-
Klein, 1973; Greenbaum, 1986; Greenbaum et al., 2003). This dis- base. The authors thank Tamar Soffer for her help in preparing the
crepancy might be due to the parameters spatial scale and the maps, Camille Vainstein for her editing work and the journal edi-
method in which they were obtained. Most of the above investiga- tors and anonymous reviewers for their valuable comments.
tions were based on rainfall simulations over small plots of up to
2 m2, while the parameters found in the current study were ob- References
tained by calibration and represent a model hillslope with an area
of approximately 25,000 m2. These calibrated parameters should Abbott, M.B., Bathurst, J.C., Cunge, J.A., Oconnell, P.E., Rasmussen, J., 1986. An
introduction to the European hydrological system Systeme Hydrologique
be referred to as effective parameters and they represent an
European, SHE, 1. History and philosophy of a physically-based, distributed
integration of small-scale heterogeneity and processes (Grayson modeling system. Journal of Hydrology 87 (12), 4559.
and Bloschl, 2001). Differences were also indicated by Yair and Ahrens, C.D., 2003. Meteorology Today: An Introduction to Weather, Climate, and
Klein (1973): they compared total runoff collected from hillslope the Environment. Brooks/Cole, Thomson Learning.
Al-Qurashi, A., McIntyre, N., Wheater, H., Unkrich, C., 2008. Application of the
strips under natural rainfall with the runoff that was expected Kineros2 rainfallrunoff model to an arid catchment in Oman. Journal of
using inltration rates obtained by inltration experiments and Hydrology 355 (14), 91105. doi:10.1016/j.jhydrol.2008.03.022.
Y. Bahat et al. / Journal of Hydrology 373 (2009) 204217 217

Beven, K., Binley, A., 1992. The future of distributed models model calibration and Michaud, J.D., Sorooshian, S., 1994b. Effect of rainfall-sampling errors on
uncertainty prediction. Hydrological Processes 6 (3), 279298. simulations of desert ash oods. Water Resources Research 30 (10), 2765
Bras, 1990. Hydrology: an Introduction to Hydrologic Science. Addison Wesley 2775.
Publishing Company. 643 pp.. Morin, E., Jacoby, Y., Navon, S. Bet-Halachmi, E., in press. Towards ash ood
Chateld, C., 2004. The Analysis of Time Series: An Introduction. Chapman and Hall/ prediction in the dry Dead Sea region utilizing radar rainfall information.
CRC. 333 pp.. Advances in Water Resources. doi: 10.1016/j.advwatres.2008.11.011.
Costelloe, J.F., Grayson, R.B., McMahon, T.A., 2005. Modelling stream ow for use in Reaney, S.M., Bracken, L.J., Kirkby, M.J., 2007. Use of the Connectivity of Runoff
ecological studies in a large, and zone river, central Australia. Hydrological Model (CRUM) to investigate the inuence of storm characteristics on runoff
Processes 19 (6), 11651183. doi:10.1002/hyp.5558. generation and connectivity in semi-arid areas. Hydrological Processes 21 (7),
Dayan, U., Morin, E., 2006. Flood-producing rainstorms over the Dead Sea basin. In: 894906. doi:10.1002/hyp.6281.
Enzel, Y., Agnon, A., Stein, M. (Eds.), New Frontiers in Dead Sea Salomon, O., Schick, A.P., 1980. Inltration tests. In: Schick, A.P. (Ed.), Arid Zone
Paleoenvironmental Research. Geological Society of America, pp. 5362. Geosystems, a Research Report. The Hebrew University of Jerusalem, pp. 55
Dayan, U., Sharon, D., 1980. Meteorological parameters for discriminating between 115.
widespread and spotty storms in the Negev. Israel Journal of Earth Sciences 29 Schick, A.P., 1970. Desert oods: interim results of observation in the Nahal Yael
(4), 253256. research watershed, Southern Israel, 19651970. In: IAHS-UNESCO
FAO, 1989. Arid zone forestry: a guide for eld technicians, www.fao.org/docrep/ Symposium: Representative and Experimental Basins, Wellington, NZ, pp.
T0122E/t0122e00.htm. 478493.
Faures, J.M., Goodrich, D.C., Woolhiser, D.A., Sorooshian, S., 1995. Impact of small- Schick, A.P., 1977. A tentative sediment budget for an extremely arid watershed in
scale spatial rainfall variability on runoff modelin. Journal of Hydrology 173 (1 the Southern Negev. In: Doehring, D.O. (Ed.), Geomorphology of Arid Regions.
4), 309326. SUNY, Binghamton, pp. 139163.
Goodrich, D.C., Faures, J.M., Woolhiser, D.A., Lane, L.J., Sorooshian, S., 1995. Schick, A.P., 2000. A brief summary of a third of a century: Nahal Yael in retrospect,
Measurement and analysis of small-scale convective storm rainfall variability. with some ideas for the future of catchment research. In: Hassan, M.A.,
Journal of Hydrology 173 (14), 283308. Slaymaker, O., Berkowicz, S.M. (Eds.), The HydrologyGeomorphology Interface:
Grayson, R.B., Bloschl, G., 2001. Spatial modelling of catchment dynamics. In: Rainfall, Floods, Sedimentation, Land Use. IAHS Press, pp. 59.
Grayson, R.B., Bloschl, G. (Eds.), Spatial Patterns in Catchment Hydrology: Schick, A.P., Lekach, J., 1993. An evaluation of 2 10-year sediment budgets, Nahal-
Observations and Modelling. Cambridge University Press, Cambridge, pp. 51 Yael, Israel. Physical Geography 14 (3), 225238.
81. Schick, A.P., Sharon, D., 1974. Geomorphology and Climatology of Arid Watersheds.
Greenbaum, N., 1986. Runoff in an extremely arid region. Inltration-runoff tests on Dept. of Geog, Hebrew University, Jerusalem. 161 pp..
small plots in the Southern Arava and their hydrological and pedological Schwartz, U., 1986. Water in the alluvial ll in an area of extreme aridity M.Sc.
implications. M.Sc. Thesis, Hebrew University of Jerusalem, Israel, 206 pp. (in Thesis, Hebrew University of Jerusalem, Israel, 100 pp. (in Hebrew).
Hebrew). Shamir, E., Meko, D.M., Graham, N.E., Georgakakos, K.P., 2007. Hydrologic model
Greenbaum, N., Salmon, O., Schick, A.P., 2003. Geomorphological implications and framework for water resources planning in the Santa Cruz River, southern
hydrological applications of inltration tests in a hyperarid region. In: Arizona. Journal of the American Water Resources Association 43 (5), 1155
Greenbaum, N., Lekach, J., Inbar, M. (Eds.), Asher Schick Volume, Current 1170.
Aspects of RainfallRunoff-Sediment Relations in Israel. University of Haifa, Sharon, D., 1970. Areal pattern of rainfall in a small watershed as affected by wind
Haifa, pp. 4469 (in Hebrew). and other meteorological conditions. In: Toebes, C. (Ed.), Results of Research on
Greenbaum, N., Ben-Zvi, A., Haviv, I., Enzel, Y., 2006. The hydrology and Experimental Basins. Int. Assoc. Hydrol. Sci. Publication No. 96, pp. 311.
paleohydrology of the Dead Sea tributaries. In: Enzel, Y., Agnon, A., Stein, M. Sharon, D., 1972. The spottiness of rainfall in the desert area. Journal of Hydrology
(Eds.), New Frontiers in Dead Sea Paleoenvironmental Research. Geological 17, 161175.
Society of America, pp. 6393. Smith, R.E., Goodrich, D.C., Woolhiser, D.A., Unkrich, C.A., 1995. KINEROS: a
Gupta, V.K., Sorooshian, S., 1985. The relationship betweendata and the precision of kinematic runoff and erosion model. In: Singh, V.P. (Ed.), Computer Models of
parameter estimates of hydrologic-models. Journal of Hydrology 81 (12), 57 Watershed Hydrology. Water Resources Publication, Highlands Ranch, pp. 697
77. 732.
Habib, E., Aduvala, A.V., Meselhe, E.A., 2008. Analysis of radar-rainfall error Soil Conservation Service (SCS), 1964. Hydrology, SCS National Engineering
characteristics and implications for streamow simulation uncertainty. Handbook. US Department of Agriculture, Washington, DC.
Hydrological Sciences Journal Journal Des Sciences Hydrologiques 53 (3), Sorooshian, S., 1981. Parameter-estimation of rainfallrunoff models with
568587. heteroscedastic streamow errors the non-informative data case. Journal of
Kahana, R., Ziv, B., Enzel, Y., Dayan, U., 2002. Synoptic climatology of major oods in Hydrology 52 (12), 127138.
the Negev Desert, Israel. International Journal of Climatology 22 (7), 867882. Sorooshian, S., Gupta, H.V., 1995. Model calibration. In: Singh, V.P. (Ed.), Computer
Lange, J., Leibundgut, C., Greenbaum, N., Schick, A.P., 1999. A noncalibrated rainfall Models of Watershed Hydrology. Water Resources Publications, Colorado, pp.
runoff model for large, arid catchments. Water Resources Research 35 (7), 2368.
21612172. Thormahlen, A.C., 2003. Hydrological modelling in a small hyperarid catchment
Lekach, J., Schick, A.P., 1982. Suspended sediment in desert oods in small Nahal Yael, Israel, runoff generation and transmission losses. University of
catchments. Israel Journal of Earth Sciences 31, 144156. Freiburg. in German.
Lekach, J., Schick, A.P., 1983. Evidence for transport of bedload in waves: analysis of Wheater, H.S., 2002. Hydrological processes in arid and semi arid areas. In:
uvial sediment samples in a small upland desert stream channel. Catena 10, Wheater, H.S., Al-Weshah, R.A. (Eds.), Hydrology of Wadi Systems. UNESCO,
267280. Paris.
Lekach, J., Amit, R., Grodek, T., Schick, A.P., 1998. Fluvio-pedogenic processes in an Yair, A., Klein, M., 1973. The inuence of surface properties on ow and erosion
ephemeral stream channel, Nahal Yael, southern Negev, Israel. Geomorphology processes on debris covered slopes in an arid area. Catena 1, 128.
23 (24), 353369. Yair, A., Lavee, H., 1985. Runoff generation in arid and semi-arid zones. In:
McMichael, C.E., Hope, A.S., Loaiciga, H.A., 2006. Distributed hydrological modelling Anderson, M.G., Burt, T.P. (Eds.), Hydrological Forecasting. Wiley and Sons, pp.
in California semi-arid shrublands: MIKE SHE model calibration and uncertainty 183220.
estimation. Journal of Hydrology 317 (34), 307324. doi:10.1016/ Yatheendradas, S., Wagener, T., Gupta, H., Unkrich, C., Goodrich, D., Schaffner, M.,
j.jhydrol.2005.05.023. Stewart, A., 2008. Understanding uncertainty in distributed ash ood
Michaud, J., Sorooshian, S., 1994a. Comparison of simple versus complex distributed forecasting for semiarid regions. Water Resources Research 44 (5), 12.
runoff models on a midsized semiarid watershed. Water Resources Research 30 doi:10.1029/2007wr005940.
(3), 593605.

You might also like