Physics and Chemistry of the Earth xxx (2014) xxx–xxx

Contents lists available at ScienceDirect

Physics and Chemistry of the Earth

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

Estimation of design floods in ungauged catchments using a regional

index flood method. A case study of Lake Victoria Basin in Kenya
Joel Nobert a,⇑, Margaret Mugo b, Hussein Gadain b
a
University of Dar es Salaam, Dar es Salaam, Tanzania, Nairobi, Kenya
b
Food and Agricultural Organization of the United Nations (FAO), Somalia Water and Land Information Management (SWALIM), Nairobi, Kenya

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

Article history: Reliable estimation of flood magnitudes corresponding to required return periods, vital for structural
Available online xxxx design purposes, is impacted by lack of hydrological data in the study area of Lake Victoria Basin in Kenya.
Use of regional information, derived from data at gauged sites and regionalized for use at any location
Keywords: within a homogenous region, would improve the reliability of the design flood estimation. Therefore,
Flood magnitude the regional index flood method has been applied. Based on data from 14 gauged sites, a delineation
Regionalization of the basin into two homogenous regions was achieved using elevation variation (90-m DEM), spatial
annual rainfall pattern and Principal Component Analysis of seasonal rainfall patterns (from 94 rainfall
stations). At site annual maximum series were modelled using the Log normal (LN) (3P), Log Logistic Dis-
tribution (LLG), Generalized Extreme Value (GEV) and Log Pearson Type 3 (LP3) distributions. The param-
eters of the distributions were estimated using the method of probability weighted moments. Goodness
of fit tests were applied and the GEV was identified as the most appropriate model for each site. Based on
the GEV model, flood quantiles were estimated and regional frequency curves derived from the averaged
at site growth curves. Using the least squares regression method, relationships were developed between
the index flood, which is defined as the Mean Annual Flood (MAF) and catchment characteristics. The
relationships indicated area, mean annual rainfall and altitude were the three significant variables that
greatly influence the index flood. Thereafter, estimates of flood magnitudes in ungauged catchments
within a homogenous region were estimated from the derived equations for index flood and quantiles
from the regional curves. These estimates will improve flood risk estimation and to support water man-
agement and engineering decisions and actions.
Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction Three approaches as proposed by Wiltshire (1986) have been

considered globally for regional flood frequency analysis; (1) aver-
Estimation of design floods in Kenya is generally based on age parameter approach (2), index flood approach, and (3) specific
empirical and deterministic models such as the rational method, frequency approach. With the average parameter approach, some
unit hydrograph and the SCS model (Strategy for Flood Manage- parameters are assigned average values based upon regional anal-
ment for Lake Victoria Basin, 2004). While direct statistical analysis yses, such as the log-space skew or standard deviation. Other
and regional techniques are advocated by Alexander (1990), little parameters are estimated using at-site data or regression on basin
effort has been devoted to the use of regional information in most characteristics, based on real or log-space means.
basins in Kenya. According to Kachroo et al. (2000) there are no The index flood method is a special case of the average param-
universally accepted methods of regionalization as a result of com- eter approach. It is a simple regionalization technique that involves
plexity in understanding the factors which affect flood generation. establishing a relationship based on regression models, between
However, a region can be considered homogeneous if sufficient the index-flood, which is often the Mean Annual Flood (MAF)
evidence can be established that data at different sites in the region and the corresponding catchment characteristics. The specific fre-
are drawn from the same parent distribution (Xu, 1999). quency approach employs regression relationships between catch-
ment characteristics and particular quantiles of a flood frequency
distribution.
⇑ Corresponding author. Tel.: +255 752546259. Hosking and Wallis (1997) mentioned that of all the stages in
E-mail address: [email protected] (J. Nobert). Regional Flood Frequency Analysis (RFFA) involving many sites,

http://dx.doi.org/10.1016/j.pce.2014.02.001
1474-7065/Ó 2014 Elsevier Ltd. All rights reserved.

Please cite this article in press as: Nobert, J., et al. Estimation of design floods in ungauged catchments using a regional index flood method. A case study of
Lake Victoria Basin in Kenya. J. Phys. Chem. Earth (2014), http://dx.doi.org/10.1016/j.pce.2014.02.001
2 J. Nobert et al. / Physics and Chemistry of the Earth xxx (2014) xxx–xxx

the identification of homogeneous regions is usually most difficult into homogeneous regions. Secondly, the derivation of a method
and requires the greatest amount of subjective judgment. The for determining index flood values for ungauged catchments, based
grouping into homogeneous regions can be done by the identifica- on catchment characteristics. Lastly, derivation of regional flood
tion of geographically contiguous regions. Geographical proximity frequency curves for each homogenous region delineated, so that
does, however, not guarantee hydrological similarity (Cunderlik design floods for an ungauged catchment can be estimated from
and Burn, 2001; Patil and Stieglitz, 2012). Therefore, it is better catchment specific index flood values and the corresponding
to define similarity between sites based on catchment characteris- regionalized quantile estimate obtained from the Regional Flood
tics or statistical flow characteristics. Different sites are only con- Frequency Curves (RFFC’s).
sidered to belong to different homogeneous regions when the
hypothesis of complete homogeneity between the stations is
rejected by the statistical hypothesis testing. Another approach
may be based on correlations whereby sites are grouped when 2. Description of the study area
they have similar catchment characteristics, e.g. flood. An unga-
uged (or gauged) site can then be assigned to a region based on The study area as shown in Fig. 1, serves as a major water
catchment characteristics alone. Several authors (Kumar et al., resource mainly for irrigation of about 200,000 ha out of
2003; Kjeldsen et al., 2002; Noto and La Loggia, 2009; Capesius 539,000 ha in Kenya. The basin has a total area of 194,000 km2
and Stephens 2009) have proposed methods for forming groups shared between five countries – Uganda, Rwanda, Burundi, Kenya
of similar sites for use in RFFA. and Tanzania, with a total basin area of 46,229 km2 within the
The use of regional information to estimate flood magnitudes at Kenyan territory. The Sio, Nzoia, Yala, Nyando and Sondu rivers,
sites with little or no observed data has become increasingly and several other streams drain into Lake Victoria. The Nzoia River
important since many projects which require design flood informa- with a catchment area of 12,790 km2 and a river length of 334 km
tion are located in areas where observed flood data are either not to its outlet into Lake Victoria is the largest among them.
available or in adequate. In the Lake Victoria Basin in particular, The Nzoia, Yala and Nyando rivers cause extensive floods in
use of regional information derived from data at gauged sites and their lower reaches especially the Budalangi Division of Busia
regionalized for use at any location within a homogenous region district and the Kano Plains. The Sondu and Kuja rivers inundate
in the basin has experienced major setbacks due to the lack of tools low-lying areas in their outfall reaches. Floods in the basin affect
and techniques. The availability of such tools would improve the parts of six districts; these are Busia, Kisumu, Nyando/Ahero,
reliability of flood risk estimation and to support water manage- Bondo, Migori and Siaya Districts, which fall under the Rift Valley,
ment and engineering decisions and actions within the basin. Western and Nyanza provinces. Out of a total area of 8770 km2 in
In this study, the Regional Index Flood method has been applied the six districts, more than a quarter are water bodies or is covered
to the Lake Victoria Basin in Kenya. In its application, three main with swamps (Strategy for Flood Management for Lake Victoria
components were undertaken. Firstly, the delineation of the basin Basin, 2004).

Fig. 1. Location Map of Lake Victoria Basin in Kenya.

Please cite this article in press as: Nobert, J., et al. Estimation of design floods in ungauged catchments using a regional index flood method. A case study of
Lake Victoria Basin in Kenya. J. Phys. Chem. Earth (2014), http://dx.doi.org/10.1016/j.pce.2014.02.001
 J. Nobert et al. / Physics and Chemistry of the Earth xxx (2014) xxx–xxx 3

Table 1
Flow gauging stations used.

Station Gage station name Longitude Latitude Catchment name Catchment area (km2) Data type
1EF01 Nzoia at Ruambwa Ferry 34.09 0.124 Nzoia 12763 Discharge (m3/s)
1EE01 Nzoia 34.255 0.158 Nzoia 12456.7
1DA02 Nzoia 34.806 0.569 Nzoia 8336.96
IDD01 Nzoia 34.487 0.372 Nzoia 11507.5 Water levels (m)
1FE02 Yala 34.936 0.183 Yala 1364.38 Discharge (m3/s)
1FF02 Zaaba 34.628 0.058 Yala 273.68
1GB03 Ainamatua 35.096 0.072 Nyando 971.97 Discharge (m3/s)
1GD07 Nyando 35.164 0.164 Nyando 1681.86
1GD03 Nyando 34.96 0.125 Nyando 271.13
1JG03 Miriu Sondu 34.794 0.347 Sondu 3597.57 Water levels (m)
1KB05 Gucha Migori 34.207 0.947 Gucha 6444.33 Discharge (m3/s)
1KC03 Migori 34.471 1.063 Gucha 2555.99 Water levels (m)
1LA03 Nyagores 35.347 0.786 Mara 662.97 Water levels (m)
1LB02 Amala 35.438 0.897 Mara 789.25

3. Methodology where Q i is the mean flow of site i (m3/s), Q ij is flow rate of station i
in region j (m3/s), ri is standard deviation and CV i is the coefficient
3.1. Data availability of variance.
However, the regional mean and regional standard deviation
Much of the rainfall and stream flow data on Lake Victoria Basin were calculated by the following Eqs. (2) and (3) respectively:
covers 1969 to 2000s. While some stations have records dating X
N
CV i
back to the first quarter of the 1900s, other stations hold records Regional mean is CV ¼ ð2Þ
until the end of 2010, Gadain (2013). However, due to high number i¼1
N
of missing values not all data were used in the study. A base period vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
from 1971–2000 for both stream flow and rainfall data was uP  2
u N
selected. Daily rainfall data from 94 stations were obtained from t i¼1 CV i  CV
rcv ¼ ð3Þ
the Kenya Meteorological Department (KMD). Daily stream flow N
data from 14 gauging stations were obtained from the Water Re-
where N is number of sites within a homogeneous region.
sources Management Authority (WRMA) with gauges located as
According to Sine and Ayalew (Standard deviation of CVs for
shown in Table 1. Of the data collected from the gauging stations,
rivers within the study region, 2004), a region is declared to be
9 stations contained mean daily flows and 5 stations contained
homogeneous if Eq. (4) is satisfied
mean daily water levels. Stage-discharge relationship needed to
be developed to determine the flow series from these stations. A rcv
CC ¼ 6 0:3 ð4Þ
90-m DEM was obtained from the USGS. CV

3.2. Delineation of homogeneous regions 3.2.3. Discordance measurement using L – moment approach
According to Hamed and Rao (2000), the discordance measure
3.2.1. Principal Component Analysis (PCA) is intended to identify those sites that are grossly discordant with-
Parameters such as perimeter, mean elevation, area and mean in the group as a whole. It estimates how far a given site is from the
annual rainfall were used as basin attributes for homogeneity centre of the group. L moments are developed for this purpose as
analysis. In this study seasonal rainfall was used as the input for shown in equation
classification using factor analysis. Data from all the 94 rainfall    9
stations was prepared in terms of the two rainfall seasons experi- ^ 10k ¼ 1 PN
M
Ni N1 >
xi >
>
N i1 k k =
enced in Lake Victoria Basin in Kenya. The two seasons are a long
XN    ð5Þ
rain season, from March to May (MAM) and a short rain from i1 N1 >
^ 1j0 ¼ 1
M xi > >
;
October to December (OND). Linear combinations of the observed N j j
i1
variables were generated and Eigen values (non-zero vectors)
greater than one were selected as the factors to retain for use in where i is rank of observed flow data in ascending order, k and j = 0,
representing the data. Factors per station were then weighted 1, 2,. . .N.
using the percentage of total variance explained by each factor The first L moments are given by the following Eq. (6)
and the aggregation of the weighted factors for each season were k1 ¼ M 0
plotted to display rainfall zones with the basin.
k2 ¼ 2M 1  M 0
ð6Þ
3.2.2. Homogeneity test k3 ¼ 6M 2  6M1 þ M 0
For each site in the delineated regions; the mean, standard devi- k4 ¼ 20M 3  30M 2 þ 12M 1  M 0
ation and coefficient of variance were calculated given by Eq. (1):
Finally, according to Hosking (1990), L moment ratio estimators
Pni 9 are defined from Eq. (6) above as;
Q ij >
1Þ Q i ¼ j¼1 >
>
ni
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi >
= t ¼ k2 =k1  L coefficient of variation
Pni 2
ð Q ij Q i Þ ð1Þ
2Þ ri ¼ j¼1
>
> t3 ¼ k3 =k2  L Skewness
ni 1 >
>
3Þ CV i ¼ Qri ;
i
t4 ¼ k4 =k2  L kurtosis

Please cite this article in press as: Nobert, J., et al. Estimation of design floods in ungauged catchments using a regional index flood method. A case study of
Lake Victoria Basin in Kenya. J. Phys. Chem. Earth (2014), http://dx.doi.org/10.1016/j.pce.2014.02.001
4 J. Nobert et al. / Physics and Chemistry of the Earth xxx (2014) xxx–xxx

Table 2
Homogeneity test results for the flow stations in OND rainfall regions.

Station Station name Qi ri CV i

Region 1
IEF01 Nzoia at Ruambwa Ferry 566.08 199.58 0.35
1EE01 Nzoia 419.80 184.71 0.44
1JG03 Miriu Sondu 306.14 97.96 0.32
1KC03 Migori 332.00 167.19 0.50
1KB05 Gucha Migori 412.13 170.41 0.41
Region 2
1DA02 Nzoia 217.85 127.23 0.58
1FF02 Zaaba 113.17 68.18 0.76
IDDO1 Nzoia 342.45 169.24 0.30
1GB03 Ainamatua 102.88 61.80 0.62
1GD07 Nyando 128.95 85.63 0.72
1GD03 Nyando 208.67 106.31 0.51
1FE02 Yala 155.25 94.55 0.61
1LA03 Nyagores 93.37 89.05 0.95
1LB02 Amala 83.26 67.33 0.81

Region No. of stations (N) CV rcv CC

Calculation of CC
1 4 0.40 0.07 0.18
2 10 0.63 0.19 0.29

ðiÞ ðiÞ T
If U i ¼ ½t ðiÞ ; t 3 ; t 4  is the vector containing the t, t3, t4 values for site
(i), then the group average for N sites within the region is given by
Eq. (7) as:

1X N
U¼ Ui ð7Þ
N i¼1

The regional standard deviation is given by Eq. (8) as:

X
N
T
S ¼ ðN  1Þ ðU i  UÞðU i  UÞ ð8Þ
i¼1

The discordance measure is defined by Eq. (9):

Fig. 2. Delineated homogenous regions of Lake Victoria Basin in Kenya.

1 T
Di ¼ ðU i  UÞ S1 ðU i  UÞ ð9Þ
3 3.3. Identification of the most appropriate flood frequency model
According to Semu and Abebe (2004), a site is declared to be
unusual if Di is large. A suitable criterion to classify a station as dis- This involved the determination of the best flood frequency
cordant is that Di should be greater than or equal to 3 (Di P 3). model, which can be fitted to the available historical data. Two

Table 3
Results of discordance measure for the regions delineated.

Station Station name Mean flow (m3/s) LCv LCs LCk Di

Region 1
IEF01 Nzoia at Ruambwa Ferry 566.08 0.21 0.05 0.07 0.01
1EE01 Nzoia 419.80 0.25 0.17 0.15 0.08
1JG03 Miriu Sondu 306.14 0.18 0.12 0.13 0.10
1KC03 Migori 332.00 0.27 0.28 0.22 0.40
1KB05 Gucha Migori 412.13 0.24 0.15 0.03 0.26
Region 2
1DA02 Nzoia 217.85 0.31 0.32 0.19 0.53
1FF02 Zaaba 113.17 0.39 0.30 0.23 0.63
IDDO1 Nzoia 342.45 0.16 0.23 0.23 0.20
1GB03 Ainamatua 102.88 0.35 0.20 0.13 0.40
1GD07 Nyando 128.95 0.39 0.27 0.13 0.58
1GD03 Nyando 208.67 0.28 0.20 0.19 0.31
1FE02 Yala 155.25 0.32 0.27 0.17 0.45
1LA03 Nyagores 93.37 0.47 0.38 0.22 0.97
1LB02 Amala 83.26 0.19 0.09 0.04 0.12

Note: LCv – L coefficient of variation.

LCs – L coefficient of skewness.
LCk – L coefficient of kurtosis.

Please cite this article in press as: Nobert, J., et al. Estimation of design floods in ungauged catchments using a regional index flood method. A case study of
Lake Victoria Basin in Kenya. J. Phys. Chem. Earth (2014), http://dx.doi.org/10.1016/j.pce.2014.02.001
 J. Nobert et al. / Physics and Chemistry of the Earth xxx (2014) xxx–xxx 5

commonly used approaches were applied in this study, graphical catchment slope (MCS) and mean altitude (ALT). The mean annual
curve fitting by visual assessment and objective methods of curve rainfall (MAR) was estimated from the rainfall data.
fitting. Fitting of the curves to the observed data, for each of the
four distributions was done by the method of Probability Weighted 3.5. Derivation of Mean Annual Flood (MAF)
Moment (PWM). After selecting the best model, the next step was
to derive basin and climatic characteristics termed as the indepen- Since a common period of 30 years was used, the MAF for each
dent variables, which were used to deduce relationships with the station was obtained by aggregating the flood data from each
index flood. station and dividing it by the number of record years (ni = 30) as
shown in Eq. (10) below.
3.4. Derivation of catchment and climatic characteristics Pni
j¼1 Q ij
Qi ¼ ð10Þ
For each station a number of catchment characteristics were ni
derived from the sub-basin upstream of the station under consid-
eration. These constitute a subset of the characteristics listed by where Q i ¼ MAF of site i and Q ij ¼ Flood data of station i in region j.
Alexander (1990) as being of importance to flood hydrology.
Furthermore, the selection was based on the assumption that they 3.6. Relationship between Index flood and catchment characteristics
could accurately be derived from the datasets used in this study
(90-m DEM and rainfall data). The catchment characteristics The key assumption underlying the index-flood method is that
derived from the 90-m DEM were catchment area (AREA), Mean annual maximum series of floods from different sites in the

Fig. 3. Fitting of the statistical distributions to the observed data.

Please cite this article in press as: Nobert, J., et al. Estimation of design floods in ungauged catchments using a regional index flood method. A case study of
Lake Victoria Basin in Kenya. J. Phys. Chem. Earth (2014), http://dx.doi.org/10.1016/j.pce.2014.02.001
6 J. Nobert et al. / Physics and Chemistry of the Earth xxx (2014) xxx–xxx

Table 4
Independent and dependent variables.

STN ID X – Independent variables Y – Dependent variables

X1 X2 X3 X4 Y
AREA (km2) MAR (mm) MSC (%) ALT (m) MAF (m3/s)

Region1
1EF01 12763 1386 6.19 1894.7 566.08
1EE01 12456.7 1256 6.75 1911.37 419.80
1JG03 3597.57 1413 8.88 1938.18 306.14
1KB05 6444.33 1549 6.94 1601.48 412.13
1KC03 2555.99 1483 6.08 1714.04 332.00
Region 2
1DA02 8336.96 1714 7.38 2195.65 217.85
IDDO1 11507.5 1922 6.82 1961.25 342.45
1FE02 1364.38 1971 7.35 2101.58 155.25
1FF02 273.68 1712 6.82 1559.05 113.17
1GB03 971.97 1451 13.89 2199.89 102.88
1GD07 1681.86 1393 11.68 2005.85 128.95
1GD03 271.13 1376 4.38 1221.42 208.67
1LA03 662.97 1375 12.07 2380.32 93.37
1LB02 789.25 1321 11.26 2395.37 83.26

3.7. Ungauged catchments flood quantiles estimation

Table 5
Derived regression equations for Regions 1 and 2. Flood quantile estimation forms the basis for flood frequency
Region Regression equations analysis, where quantile magnitudes Q corresponding to the re-
quired return periods T are estimated. From the analysis of several
1 MAF1 = 0.028 [AREA] + 0.868 [MAR] + 0.260 [ALT]  1503.78
2 MAF2 = 0.018 [AREA] + 0.016 [MAR]  0.093 [ALT] + 268.96 distributions the General Extreme Value (GEV) was selected as the
most appropriate model for the observed data, for all the stations.
The model parameters for GEV estimated for each station were
then used to compute standardized quantile estimates correspond-
ing to the return periods T = 2, 5, 10, 20, 25, 50, 100, 200. Plots of Q/
homogeneous regions are identically distributed except for a scale
MAF against the reduced variate yT, known as growth curves, were
parameter, termed the index-flood. In this study the MAF was
generated for each station and were used in the derivation of the
defined as the index-flood. Using Microsoft Excel under data anal-
regional curves.
ysis, the function LINEST was used to calculate the statistic using
the least squares method. The derived catchment characteristics
and the MAF were used as the input data for the LINEST function 3.8. Derivation of regional curves
so as to derive a regression equation for index flood. From the
established relationship between index floods and catchment For each of the homogeneous regions identified, a regional
characteristics flood quantiles in ungauged sites were estimated; curve was derived. The approach used was based on the growth
by first estimating the magnitude of the index flood and then curves determined for all the stations in each region. The average
multiplying it by the ratios read from the regional frequency of the growth curves was determined to represent the frequency
curves for the required return periods. curve for each region. The event with return period, T (XT) for

Table 6
Comparison between estimated and observed MAF.

Station Station name Observed MAF – yi (m3/s) ^i (m3/s)

Estimated MAF – y Relative error

Region 1
IEF01 Nzoia at Ruambwa Ferry 566.08 545.86 0.036
1EE01 Nzoia 419.80 428.86 0.022
1JG03 Miriu Sondu 306.14 325.78 0.064
1KC03 Migori 412.13 435.50 0.057
1KB05 Gucha Migori 332.01 299.36 0.098

Region 2
1DA02 Nzoia 217.85 243.80 0.119
IDDO1 Nzoia 342.45 326.319 0.047
1FE02 Yala 155.25 130.49 0.159
1FF02 Zaaba 113.17 136.92 0.210
1GB03 Ainamatua 102.88 105.84 0.029
1GD07 Nyando 128.95 135.77 0.053
1GD03 Nyando 208.67 182.78 0.124
1LA03 Nyagores 93.37 91.27 0.022
1LB02 Amala 83.26 82.28 0.012

Please cite this article in press as: Nobert, J., et al. Estimation of design floods in ungauged catchments using a regional index flood method. A case study of
Lake Victoria Basin in Kenya. J. Phys. Chem. Earth (2014), http://dx.doi.org/10.1016/j.pce.2014.02.001
 J. Nobert et al. / Physics and Chemistry of the Earth xxx (2014) xxx–xxx 7

the discordance test was performed on Regions 1 and 2 and the

results are as shown in Table 3. Based on the results obtained, none
of the stations considered in the homogenous Regions 1 and 2 were
classified as discordant. Having satisfied the homogeneity and
discordance measure tests, the hydrologically homogenous Regions
1 and 2 formed the basic units for regionalization as illustrated in
Fig. 2.

4.2. Selection of an appropriate flood frequency model

From the graphical curve fitting, four distributions LN (3P), LLG,

GEV and LP3 were observed to provide a near good fit to the histo-
grams of observed data for all the 14 station. However, sufficient
(a) Region 1
evidence needed to be deduced if the hypothesis regarding the four
distributional forms was rejected or accepted at the chosen signif-
icance level a of 0.01 and 0.05. This was based on the test statistic
of the Kolmogorov–Smirnov – D, Anderson Darling – A2 and
Chi-square – X2 test. The test statistics Di, A2 and X2 for the LN
(3P), LLG, GEV and LP3 distributions were accepted at the chosen
significance level a of 0.01; since none of the test statistics values
was greater than the critical values obtained from a table. The null
hypothesis was valid that all the candidate distributions could be
acceptable in modelling the observed data. The plots of the fitting
of the curves to observed data for each station, for the LN (3P), LLG,
GEV and LP3 distributions are as presented in Fig. 3. From the curve
plots of the LN (3P), LLG, GEV and LP3 distributions, it was noted
(b) Region 2 that the curves plotted close to each other in all the stations and
Fig. 4. Scatter plot of estimated and observed MAF for Regions 1 and 2.
thus it would be valid to say that all the four distributions would
be acceptable models for the at site data. However an assessment
of the individual curves from each distribution, revealed a slightly
ungauged catchment within the homogeneous region was then higher divergence from the Gringorten values of the observed data
estimated as: for the LN (3P), LLG and LP3 distributions compared to the GEV
distribution. The GEV model was thus selected as the most
X T ¼ li Z T ð11Þ
appropriate model by both the graphical and objective curve fitting
where li is a MAF at site i and Z T is a quantile in the regional growth for the flood data for all the 14 stations.
curve.
4.3. Relationship between Index flood and catchment characteristics
4. Results and discussions
It was noted that, floods within the basin are greatly influenced
4.1. Delineated homogeneous regions by the area, rainfall and the variations in elevation. Based on pre-
vious studies by Mkhandi and Kachroo (1997) and similar studies
Homogeneity test results for the flow stations in the seasonal from other parts of the world. AREA is the single most important
October–November–December (OND) rainfall regions are shown variable and should always be included in the models.
in Table 2. Both regions in OND seasonal rainfall satisfied the homo- Region 1 which represents the lower reaches of most of the riv-
geneity condition of having CC values less than 0.3 and are thus ers, experiences high flood generated from the highlands, slopes
hydrologically homogenous. Having satisfied the homogeneity test, and plains further upstream. The Nzoia, Yala, and Nyando rivers

Region 1 Region 2
Fig. 5. Regional growth curves plotted with the at-site growth curves for Regions 1 and 2.

Please cite this article in press as: Nobert, J., et al. Estimation of design floods in ungauged catchments using a regional index flood method. A case study of
Lake Victoria Basin in Kenya. J. Phys. Chem. Earth (2014), http://dx.doi.org/10.1016/j.pce.2014.02.001
8 J. Nobert et al. / Physics and Chemistry of the Earth xxx (2014) xxx–xxx

Table 7 observed and estimated MAF for all the sites and thus the equa-
Regional quantiles estimates. tions could be deemed reasonable for use to predict the MAF at
Return period-T Reduced variate yT Quantiles estimates – zT ungauged sites.
Region 1 Region 2 Finally, from the derived regional frequency curve, Region 1
exhibited higher quantile estimates than Region 2, for the same
2 0.37 1.619 1.456
5 1.5 2.989 2.446
return period interval. This could be attributed to the variability
10 2.25 3.960 3.026 in their flood regimes and the corresponding contributing areas.
20 2.97 5.110 3.536 The method will be of great importance in flood risk estimation
25 3.2 5.532 3.690 and to support water management and engineering decisions
50 3.9 6.897 4.136
and actions in ungauged catchments.
100 4.6 8.536 4.551
200 5.29 10.454 4.933
500 6.213 13.693 5.410 Acknowledgments

This work was supported through financial support from

cause extensive floods in their lower reaches especially the Budal- WaterNet. The authors would also like to acknowledge the support
angi division of Busia district and the Kano Plains. The Sondu and of facilities provided by the Water Resources Engineering Depart-
Kuja rivers inundate low-lying areas in their outfall reaches. ment, University of Dar es Salaam, Tanzania and, the Somalia office
Dependent and independent variables and the derived regression of the Food and Agriculture Organization (FAO).
equations for the both regions are as shown in Tables 4 and 5
respectively. The relative errors and the scatter plots between esti- References
mated mean annual flow (MAF) from the regression equation and
Alexander, W.J.R., 1990. Flood Hydrology of Southern Africa. South African
the observed mean annual flows are presented in Table 6 and Committee for Large Dams, Pretoria, RSA.
Fig. 4, respectively. It was observed that the relative error was Capesius, J.P., Stephens, V.C., 2009. Regional regression equations for estimation of
natural streamflow statistics in Colorado: U.S. Geological Survey Scientific
small and thus the derived equations could accurately be applied
Investigations, Report 2009–5136, p. 46.
in the estimation of MAF in ungauged sites. Cunderlik, J.M., Burn, D.H., 2001. The use of flood regime information in regional
flood frequency analysis. Hydrol. Sci.-J.-des Sci. Hydrol. 47 (1).
4.4. Flood quantile estimation Gadain, H.M., 2013. Development and Application of Decision Support Tools for
Lake Victoria Sub-Catchments, PhD Thesis, University of Dar es Salaam,
Tanzania.
Plots of the Q/MAF against the reduced variate YT (for EVI distri- Hamed, K.H., Rao, A.R., 2000. Flood Frequency Analysis. CRC Press LLC, Florida.
bution), known as the growth curves, for the averaged growth Hosking, J.R.M., 1990. ‘‘L – Moments’’ Analysis and estimation of distribution using
linear combination of order statistics. J.R. Stat., Ser B 52 (1), 105–124.
curve (GC) and the at-site growth curves for Regions 1 and 2 are Hosking, J.R.M., Wallis, J.R., 1997. Regional Frequency Analysis: ‘‘Approach Based on
presented in Fig. 5. From the figure, it is observed that there is large L-Moments’’. Cambridge Univ. Press, UK, pp. 224.
variation between individual curves, which could be explained by Kachroo, R.K., Mkhandi, S.H., Parida, B.P., 2000. Flood frequency analysis of southern
Africa: ‘‘I. Delineation of homogenous regions’’. Hydrol. Sci. J. 45 (3), 437–448.
the differences in catchment characteristics. However, curves both Kjeldsen, T.R., Smithers, J.C., Schulze, R.E., 2002. Regional flood frequency analysis in
in Regions 1 and 2 plot nearly close to each other indicating that the KwaZulu-Natal province, South Africa, using the index-flood method. J.
the stations in the study area could be assumed to constitute single Hydrol. (Amst) 255, 194–211.
Kumar, R., Chatterjee, C., Kumar, S., Lohani, A.K., Singh, R.D., 2003. Development of
homogenous regions. This finding suggests that a single theoretical
Regional Flood Frequency Relationships Using L-moments for Middle Ganga
extreme distribution can be used to derive a regional frequency Plains Subzone 1(f) of India. Water Resour. Manage. 17, 243–257.
curve for each of the regions within Lake Victoria Basin in Kenya. Mkhandi, S., Kachroo, S., 1997. Regional flood frequency analysis for Southern
Africa. In: Southern African FRIEND, Technical Documents in Hydrology No. 15,
Higher quantile estimates are recorded for Region 1. This can be
UNESCO, Paris, France.
attributed to the variability in their flood regimes. The extent of the Noto, L.V., La Loggia, G., 2009. Use of L-moments approach for regional flood
variability of the flood regimes depends to a larger extent on frequency analysis in sicily, Italy. Water Resour Manage 23, 2207–2229.
the meteorological phenomena generating the flood events and Patil, S., Stieglitz, M., 2012. Controls on hydrologic similarity: role of nearby gauged
catchments for prediction at an ungauged catchment. Hydrol. Earth Syst. Sci. 16
the size of contributing area. Quantiles in the regional frequency (551–562), 2012.
curve corresponding to return periods T2  T500 years are as Semu, A., Abebe, S., 2004. Identification and delineation of hydrological
presented in Table 7. homogeneous regions-the case study of blue Nile river basin. Arba Minch
University, Hydraulic Engineering department, Ethiopia.
Sine, A., Ayalew, S., 2004. Identification and Delineation of Hydrological
5. Conclusion Homogeneous Regions-The case study of Blue Nile River Basin, In:
Proceedings, Lake Abaya Research Symposium, 2004.
Strategy for Flood Management for Lake Victoria Basin in Kenya., 2004. Prepared
The study has demonstrated regional flood frequency analysis under Associated Programme on Flood Management (APFM).
in Lake Victoria Basin in Kenya, using the index flood method. From Wiltshire, S.E., 1986. Regional flood frequency analysis I: homogeneity statistics.
Hydrol. Sci. J. 31, 321–333.
the deduced regression equations it was noted that; mean annual
Xu, B.X., 1999. Regional flood frequency analysis for South-western Alberta,
flood within a homogenous region is greatly influenced by the Published Master’s thesis, University of Calgary.
contributing area, rainfall and the variations in elevation. Based
on the relative error, there is a small variation between the

Please cite this article in press as: Nobert, J., et al. Estimation of design floods in ungauged catchments using a regional index flood method. A case study of
Lake Victoria Basin in Kenya. J. Phys. Chem. Earth (2014), http://dx.doi.org/10.1016/j.pce.2014.02.001