Beruflich Dokumente
Kultur Dokumente
Abstract: In this study, the authors investigated an approach to calculate the standardized streamflow index (SSI), which allows accurate
spatial and temporal comparison of the hydrological conditions of a stream or set of streams. For this purpose, the capability of six three-
Downloaded from ascelibrary.org by Hankyong National University on 04/18/20. Copyright ASCE. For personal use only; all rights reserved.
parameter distributions (lognormal, Pearson Type III, log-logistic, general extreme value, generalized Pareto, and Weibull) and two different
approaches to select the most suitable distribution the best monthly fit (BMF) and the minimum orthogonal distance (MD), were tested
by using a monthly streamflow data set for the Ebro Basin (Spain). This large Mediterranean basin is characterized by high variability
in the magnitude of streamflows and in seasonal regimes. The results show that the most commonly used probability distributions for flow
frequency analysis provided good fits to the streamflow series. Thus, the visual inspection of the L-moment diagrams and the results of the
Kolmogorov-Smirnov test did not enable the selection of a single optimum probability distribution. However, no single probability distri-
bution for all the series was suitable for obtaining a robust standardized streamflow series because each of the distributions had one or more
limitations. The BMF and MD approaches improved the results in the expected average, standard deviation, and the frequencies of extreme
events of the SSI series in relation to the selection of a unique distribution for each station. The BMF and MD approaches involved using
different probability distributions for each gauging station and month of the year to calculate the SSI. Both approaches are easy to apply
and they provide very similar results in the quality of the obtained hydrological drought indexes. The proposed procedures are very flexible
for analyses involving contrasting hydrological regimes and flow characteristics. DOI: 10.1061/(ASCE)HE.1943-5584.0000433. © 2012
American Society of Civil Engineers.
CE Database subject headings: Droughts; River flow; Streamflow; Precipitation; Runoff; Low flow.
Author keywords: Hydrological drought; River flows; Standardized streamflow index; Standardized precipitation index; Low flows;
Runoff index; Flow index.
(SPEI) is based on the log-logistic distribution (Vicente-Serrano Stahl (2001) developed the regional deficiency index (RDI)
et al. 2010a). Several studies have shown that these distributions to characterize hydrological droughts. The RDI uses daily flow
are able to fit the time series of precipitation and water balance over data and removes the influence of streamflow seasonality on
a wide range of climate regions (Guttman 1999; Vicente-Serrano droughts. Nevertheless, the index does not indicate the magnitude
et al. 2010b). of a drought–it only quantifies the duration from a binary time
Hydrological drought refers to a decrease in surface or ground series–and it quantifies drought severity by the surface affected.
water resources-usually river flows, reservoir storages, and aquifers This makes comparisons of hydrological droughts at different sta-
(Tallaksen and Van Lanen 2004). Hydrological droughts can have tions impossible.
widespread impacts by reducing or eliminating water supplies, Two hydrological drought indexes have recently been developed
deteriorating water quality, restricting water for irrigation and caus- that follow the approach commonly used for climate drought
ing crop failure, reducing power generation, disturbing riparian indexes. They are the streamflow drought index (Nabaltis and
habitats, limiting recreation activities, and affecting a diversity Tsakiris 2009) and the standardized runoff index (Shukla and Wood
of economic and social activities (Mishra and Singh 2010). The 2008). These indexes have the same theoretical background because
origin of hydrological droughts is commonly climate droughts they derive the hydrological drought index by transforming monthly
(Zaidman et al. 2001; Hisdal and Tallaksen 2003; Beersma and streamflows into z-scores. The problem with this approach is that
Buishand 2004; Vasiliades and Loukas 2009; Edossa et al. 2010; selecting the most suitable probability distribution to calculate
Hannaford et al. 2011; Lorenzo-Lacruz et al. 2010; Vidal et al. the index and the consequences of the selection on the final series
2010) however, the quantification of hydrological droughts as have not been tested in depth. In developing the SPI, various prob-
independent phenomena has also received much attention in the ability distributions were carefully tested to select the most suitable
scientific community. This is because usually no direct spatial (McKee et al. 1993; Guttman 1999). The same approach was fol-
or temporal relationship exists between climate and hydrological lowed in developing the SPEI (Vicente-Serrano et al. 2010a). These
droughts (Vicente-Serrano and López-Moreno 2005; Tallaksen studies showed the suitability of the Pearson Type III and log-
et al. 2009; Hannaford et al. 2011; Lorenzo-Lacruz et al. 2010; logistic distributions to fit precipitation and water balance series,
Vidal et al. 2010). Moreover, the analysis of hydrological droughts respectively, across a wide range of climate regimes and time scales.
allows direct quantification of the deficits in usable water sources. Streamflow commonly shows a greater spatial variability than
In contrast to climate droughts, the quantification of hydrolog- climatic variables that are used to derive drought indicators. This is
ical droughts is not usually on the basis of indexes,but on the theory because of the influence of a number of factors, including topog-
of runs (Yevjevich 1967). A drought event is defined as a period raphy, lithology, vegetation, and human management; it is also a
during which the hydrological variable is below a predetermined consequence of the spatial aggregation of the flows, which changes
truncation level. By using this approach, the duration of drought the statistical properties of the series downstream (Mudelsee 2007).
is the time when the streamflow is below the truncation level, Therefore, a high degree of spatial variability occurs in the prob-
and drought severity is the cumulative deviation below the trunca- ability distributions that best fit the monthly streamflow data (Riggs
tion level during that period. Therefore, the selection of the trun- 1973; Kroll and Vogel 2002; Yue and Wang 2004; Yue and Pilon
cation level is critical in defining a drought and can use the average 2005; McMahon et al. 2007; Yue and Hashino 2007), which makes
of a series (Dracup et al. 1980), a percentile (Zelenhasic and Salvai it difficult to select the most appropriate distribution to calculate a
1987), or relative values with respect to the average (Clausen and streamflow drought index over a wide area.
Pearson 1995). Some studies have improved the identification of In this study, the authors tested the performance of several prob-
droughts by considering a mutual dependence among drought ability distributions (assuming that each month may fit different
events (Tallaksen et al. 1997; Fleig et al. 2006). This approach probability distributions), to calculate a standardized streamflow
has the advantage of considering the real low-flow periods, during index (SSI). This has enabled the authors of this paper to propose
which the availability of water is unable to meet demand, with the an accurate procedure to obtain a hydrological drought index that is
subsequent environmental and socioeconomic impacts (Smakhtin useful for making spatial and temporal comparisons over a wide
2001). Given the contrasting river regimes and flow magnitudes variety of river regimes and flow characteristics.
that can occur among neighboring basins, the spatial comparison
of drought severity and the development of drought maps are
impossible when using the method of runs. This is because the Methods
same drought magnitude may have different implications, depend-
ing on the basin. In addition, the seasonality of river flows must be
Data Set
taken into account when quantifying hydrological droughts. In riv-
ers with high seasonality, the low-flow periods usually correspond The authors used the Ebro River basin (in northeast Spain) to
to reduced water availability in summer. A reduction in discharges calculate and test the performance of the SSI. This basin is char-
during high flow periods can nevertheless have negative effects acterized by large spatial variability in the river regimes and flow
Fig. 1. Location of the study area and spatial distribution of the gauging stations used in the study
off series (Hosking and Wallis 1993; Bobee et al. 1996; Burn 1997;
Yue and Wang 2004; Modarres and Sarhadi 2010; Shi et al. 2010).
The use of a unique probability distribution function for different
river sectors and months nevertheless requires high spatial homo-
geneity among the series, which is unlikely to be found in rivers
with physical characteristics, climatic, hydrological regimes, and
hierarchy that differ, even over very short distances. It can therefore
be assumed that the most suitable probability distribution to fit
individual streamflow series will vary. Large seasonal variations
in streamflow characteristics can similarly cause differences in
the 12 monthly series at a single station. For example, Fig. 2 shows
the frequency histograms of the streamflow series for January,
March and May at the Sástago station in the Ebro River; Fig. 2
also indicates the theoretical probability distributions that show
the best fit to the data. This example illustrates how fitting the
monthly runoff series from the same location to different distribu-
tions can affect the procedure for calculating hydrological drought
indexes.
In this study, the authors applied two approaches to using prob-
ability distributions to obtain a standardized streamflow index.
Using a Unique Distribution of Probability
In the first approach, the authors used a unique probability
distribution to fit the monthly streamflow series with independent
fitted parameters at each site and month. The authors selected six
three-parameter distributions that are widely used in hydrological
analysis (Chow et al. 1988; Bobée and Ashkar 1991; Vogel et al.
1993; Rao and Hamed 2000): the general extreme value (GEV),
Pearson Type III (PIII), log-logistic, lognormal, and generalized Pa-
reto and Weibull distributions. For example, Fig. 3 shows the May
series for the Sástago station (Ebro River) with the six theoretical
Fig. 2. Frequency histograms of the runoff series at the Sástago station
distributions that fit the data. The flexibility of having a wide range
(Ebro River) and the theoretical curves of the log-logistic, Weibull, and
of distributional shapes, given the three-parameters of the distribu-
Pearson Type III distributions: (a) January; (b) March; (c) May
tions, allows a good fit to the observed frequencies. Therefore, for
each streamflow series the authors calculated six SSI series, corre-
sponding to each of the six probability distributions.
Table 1 shows the cumulative distribution functio (Fx) for each
of the six selected distributions. The calculation of FðxÞ is essential
for obtaining the SSI. The table includes the equations used to cal-
culate the parameters of the six probability distributions. These
used the L-moment method (see below). Once FðxÞ is calculated,
the SSI (in z-scores) can easily be determined by following the
classical approximation of Abramowitz and Stegun (1965). For
example
C0 þ C1 W þ C2 W 2
SSI ¼ W ;
1 þ d1 W þ d2 W 2 þ d3 W 3
where
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
W¼ 2 lnðPÞ for P ≤ 0:5
Fig. 3. Frequency histogram of the May runoff series at the Sástago
P is the probability of exceeding a determined x value, and
station (Ebro River) and the curves of the six theoretical distributions
P ¼ 1 FðxÞ. If P > 0:5, P is replaced by 1 P and the sign
α ¼ λ2 ð1 þ κÞð2 þ κÞ
ε ¼ λ1 λ2 ð2 þ κÞ
α β 1 1 w0
Log-logistic FðxÞ ¼ ½1 þ ðxγÞ β ¼ 6w12w
w0 6w2 Singh et al. (1993)
ðw0 2w1 Þβ
α ¼ Γð1þ1=βÞΓð11=βÞ
Downloaded from ascelibrary.org by Hankyong National University on 04/18/20. Copyright ASCE. For personal use only; all rights reserved.
distribution function h i
λ2
μ ¼ ln erfð σ2 erf is the Gauss error
2
σ
2Þ
pffiffiffi σ2
function: erfðσ2 Þ ¼ 2Φðσ2 2Þ 1 a ¼ λ1 eμþ 2
Rx xγ β1 ðxγ
Pearson Type III FðxÞ ¼ αΓðβÞ
1
γð α Þ e αÞ If τ 3 ≥ 1=3, then τ m ¼ 1 τ 3 : Hosking (1990)
pffiffiffi ΓðβÞ
α¼ πλ2 Γðβþ1=2Þ
γ ¼ λ1 αβ
xμ 1
GEV FðxÞ ¼ e½1κð α Þκ κ ¼ 7:859C þ 2:9554C 2 Hosking et al. (1985)
C ¼ 3þτ
2
3
0:6309
λ2 κ
α ¼ Γð1þκÞð12 κ Þ
μ ¼ λ1 þ ακ ½Γð1 þ κÞ 1
C ¼ 3τ
2
3
0:6309
λ2
a¼ 1
Γ 1þ1b 12 b
m ¼ λ1 aΓ 1 þ 1b
of the resultant SSI is reversed. The constants are C 0 ¼ 2:515517; distribution for each series, the authors tested the capability of
C 1 ¼ 0:802853; C 2 ¼ 0:010328; d 1 ¼ 1:432788; d 2 ¼ 0:189269; two different approaches:
and d 3 ¼ 0:001308. If the probability distribution is suitable for 1. The selection of the most suitable distribution according to
fitting the monthly streamflow series, the average value of the L-moment ratio diagrams (Hosking 1990; Peel et al. 2001),
SSI and the standard deviation must equal 0 and 1, respectively. which gives a visual indication of how well the probability dis-
The SSI is a standardized variable and can therefore be compared tributions fits the data.
with other SSI values across time and space. The L-moment ratios τ 3 and τ 4 are calculated, as follows:
Using Different Distributions for Each Monthly Series λ3 λ4
The second approach used the selection of the most suitable τ3 ¼ τ4 ¼ ;
λ2 λ2
probability distribution for each monthly streamflow series. By us-
ing this approach an SSI was created for each monthly streamflow where λ2 , λ3 , and λ4 are the L-moments of the river flow or
series, whereby the 12 monthly streamflow series could be fitted reservoir storage series, and are obtained from probability-
by different probability distributions. To select the most suitable weighted moments (PWMs) by using the formulas
Fig. 5. KS D statistic obtained for the six probability distributions from the monthly series of 98 stations
Fig. 6. Orthogonal distance between the sample L-moments at site i and the L-moment relationship for a specific distribution obtained for the six
probability distributions from the monthly series of 98 stations
showed the best fit in the highest percentage (28 to 32%) of stations (611 of the 1176 monthly series), the selected distribution differed
in late autumn and early winter; and the log-logistic distribution between the BMF and the MD approaches.
showed the best fit in the highest percentage (28 to 36%) of stations The reasons of why certain distributions fit better than others for
in the summer and autumn. The other six distributions did not different months can be driven by physical or climatic factors. Late
show any seasonal dominance. The pattern is similar for the spring and early summer are characterized by high flows and by
MD approach because the generalized Pareto showed the minimum high interannual differences, given the large climate variability that
orthogonal distance for the 23% of the cases and the seasonal characterizes the region. This could explain why during these
percentages are very close to those obtained by using the BMF. months the generalized Pareto distribution showed the best fit
The percentages are comparable; however, in 52% of the cases in a high percentage of the cases because this is an extreme value
Fig. 7. Spatial pattern of the probability distribution showing the best fit to each monthly runoff series
distribution widely used for modeling excesses over high thresh- identified. At the Sástago station, the main droughts were recorded
olds. By contrast, the precipitation events commonly have a frontal in 1948–1949, 1955, and in the decades of 1980, 1990, and 2000.
origin during late fall and early winter; they are not very extreme in There was some agreement with the drought episodes recorded at
their intensity over short periods and flashy high flows are more the Puigcerdá station, primarily in the decades of 1940 and 2000.
rare. This behavior could explain why log-logistic or lognormal On the basis of visual inspection, it was very difficult to distin-
distributions showed the best fit in several stations. In any case, guish among the series calculated by using the various probability
the variability found was very high, and given the high spatial vari- distributions or the BMF/MD approaches. Some common features
ability that Figs. 7 and 8 show, it is very difficult to identify patterns were nevertheless evident: 1) the PIII distribution tended to over-
of dominant distributions to fit the data in some months and areas. estimate the magnitude of the negative anomalies; 2) the general-
On the basis of high spatial and seasonal differences in the stat- ized Pareto distribution tended to generate higher values than
istical characteristics of the streamflow series and large variability the other distributions; 3) the generalized Pareto and the Weibull
in the probability distributions that provided the best fit to the data, distributions did not show a solution for some low/high streamflow
it does not seem feasible to select a single probability distribution values (because the values of the index tend to ∞ or þ∞, given
for the entire data set, or for specific regions or months. An alter- the shape of the probability density functions). In these cases, the
native approach was to seek the most suitable probability distribu- values had to be truncated to the corresponding probability of 1 in
tion for each station and monthly series. Figs. 9 and 10 show 100 years to have a value in the index. This explains why some
representative examples of the evolution of the SSI at two stations contiguous months had the same negative value. The SSI obtained
with very different characteristics: the Ebro River at the Sástago from the lognormal, GEV and log-logistic distributions, and the
station (in the middle course of the main river, with an average an- BMF and MD approaches showed similar characteristics in the
nual flow of 7;549 hm3 =year) and the Segre River at the Puigcerdá magnitude of the most extreme values, highlighting the difficulty
station (in the headwaters of a mountain river, with an average of selecting the most appropriate distribution by visual inspec-
annual flow of 126 hm3 =year). The figures represent seven SSI tion alone.
series for each station, corresponding to the six probability distri- Because the authors could not reject (for most stations) the hy-
butions and the BMF and MD approaches. They also show that pothesis that the 12 monthly series came from at least one of the six
irrespective of whether the probability distribution or BMF/MD analyzed distributions (Table 3), therefore the authors made a stat-
approach is used, the major dry and moist episodes were clearly istical comparison on the basis of the expected statistical properties
Fig. 8. Spatial pattern of the probability distribution showing the minimum orthogonal distance to each monthly runoff series
Table 2. Percentage of the Monthly Series in which Each Probability Distribution Shows the Best Fit and the Minimum Orthogonal Distance
Distribution January February March April May June July August September October November December Annual
(a) Best monthly fit
GEV 7 16 6 12 6 15 13 14 17 13 17 21 13
Pearson Type III 13 15 15 9 18 5 7 12 11 10 11 11 12
Generalized Pareto 27 17 23 23 30 35 20 14 14 14 23 17 22
Lognormal 21 19 17 11 15 13 12 9 15 23 32 28 18
Log-logistic 12 16 16 13 13 17 20 36 28 29 7 9 18
Weibull 19 15 21 31 16 14 27 14 14 10 9 13 17
(b) Mimimum distance
GEV 10 12 7 13 12 9 10 17 16 19 9 18 13
Pearson Type III 19 14 14 19 24 21 12 16 11 8 14 13 16
Generalized Pareto 25 24 28 26 25 21 22 13 19 14 32 25 23
Lognormal 14 30 18 14 13 15 17 9 13 24 22 27 18
Log-logistic 5 9 9 7 10 18 18 35 30 21 8 7 15
Weibull 27 11 23 20 16 14 19 9 10 13 14 10 16
of the SSI. Because the SSI is a standardized variate, the long-term of some distributions, the indexes were deficient in the expected
average of the index is expected to equal 0 and its standard average and standard deviation. For example, the mean of the
deviation is expected to equal 1. Fig. 11 shows two box-plots of average values was negative for the lognormal and the PIII distri-
the average and standard deviation of the SSI by using the six prob- butions, whereas the mean of the average SSI was close to zero
ability distributions and the BMF and MD approaches. On the basis for the other distributions and the BMF and MD approaches.
Fig. 9. Evolution of the SSI obtained by using the six selected probability distributions and the BMF approach for the Sástago station in the Ebro
River
The range of average values among the stations was very high marked differences among the eight approaches. Therefore, the log-
for the lognormal and PIII distributions, which introduced a normal and the PIII distributions generated average values much
substantial limitation in the SSI series obtained from these distri- greater than 1, which suggests against their use for the SSI calcu-
butions. Some of the series were biased; it was therefore difficult lation. The other distributions generated values of approximately 1,
to establish reliable spatial comparisons because a value of 0 and the standard deviation for the BMF and MD approaches was
would not represent normal conditions in all cases. The GEV and closest to 1.
log-logistic distributions and the BMF approach provided SSI The comparisons among the indexes included the expected fre-
series with averages closest to zero, and low variability among quency of the most extreme SSI values, and the average and stan-
the stations. The standard deviations of the SSI series also showed dard deviations. On the basis of a normal standard distribution, one
Fig. 10. Evolution of the SSI obtained by using the six selected probability distributions and the BMF approach for the Puigcerdá station in the
Segre River
Table 3. Percentage of Cases that Reject the Hypothesis that the Monthly Runoff Series Come from the Corresponding Probability Distribution
Distribution January February March April May June July August September October November December Annual
GEV 3.1 4.1 1.0 2.0 1.0 2.0 3.1 2.0 3.1 5.1 3.1 4.1 2.8
Pearson Type III 3.1 5.1 3.1 2.0 1.0 2.0 4.1 2.0 3.1 5.1 3.1 4.1 3.1
Generalized Pareto 2.0 2.0 1.0 1.0 5.1 2.0 7.1 11.2 5.1 1.0 2.0 4.1 3.7
Lognormal 2.0 3.1 1.0 2.0 1.0 0.0 4.1 6.1 2.0 1.0 3.1 3.1 2.4
Log-logistic 6.1 4.1 3.1 3.1 2.0 2.0 3.1 10.2 6.1 3.1 9.2 5.1 4.8
Weibull 5.1 3.1 2.0 2.0 2.0 1.0 3.1 5.1 3.1 3.1 7.1 4.1 3.4
Fig. 11. Average and standard deviation values of the SSI at the 98 gauging stations, showing the lower quartile, median, upper quartile, and mean
(dashed line); the whiskers represent the 90th (top) and 10th (bottom) percentiles for GEV, lognormal (LN), Pearson Type III (PIII), generalized Pareto
(G-PAR), log-logistic (L-LOG), and Weibull (WEI)
Fig. 12. Frequency of low (< 1:65 and < 2:33) and high values (> 1:65 and > 2:33) of the SSI at the 98 gauging stations, showing the lower
quartile, median, upper quartile, and mean (dashed line); the whiskers represent the 90th (top) and 10th (bottom) percentiles for GEV, lognormal (LN),
Pearson Type III (PIII), generalized Pareto (G-PAR), log-logistic (L-LOG), and Weibull (WEI)
value is expected to be less than 1:65 or more than 1.65 every These distributions also showed the highest variability in the ob-
20 years because these values represent 5% of the cumulative dis- served frequencies among the stations. The most important differ-
tribution function. An SSI event exceeding 2:33 similarly corre- ences were found for the low values. The GEV and log-logistic
sponds to a return period of 100 years. Given the length of the series distributions tended to underestimate the expected frequency of
(i.e., 61 years × 12 months), an average of 36.6 SSI events above values below 1:65. The lognormal distribution moreover showed
1.65 (or below1:65) or 7.3 events above 2.33 (or below2:33) a high variability among the stations, and in some cases yielded
would be expected. The authors compared the observed and ex- very high frequencies of low SSI values. The BMF and MD ap-
pected frequency of the values above and below these thresholds proaches showed the closest frequency to the expected values
and summarized the results in a boxplot (Fig. 12). The observed and the lowest variability among the 98 stations. The same pattern
frequency of positive extreme values tended to follow the expected was observed for the frequency of values less than 2:33, for
frequency, independent of the approach used for the calculation. which the BMF provided the best result.
The GEV and log-logistic distributions nevertheless tended to show Table 4 summarizes the main limitations in the calculation of
a higher frequency of values greater than 1.65 and a lower fre- the SSI on the basis of each of the six probability distributions
quency of values greater than 2.33 than did the other distributions. and the BMF and the MD approaches. If a unique distribution
Note: Cells without checkmarks show that a particular distribution does not have this limitation.
of probability is selected to fit the 12 monthly streamflow series suitable probability distribution to fit the empirical streamflow
of a gauging station, the most suitable and flexible probability dis- data is usually performed by the visual inspection of L-moment
tributions were the GEV and the log-logistic. The generalized diagrams (Hosking 1990) or by a goodness-of-fit test. The authors
Pareto and Weibull distributions showed no solution for some neg- found that neither of these approaches provided a sufficient basis
ative values, which is a critical limitation for calculating the index. for choosing a single probability distribution for the data set, even
However, these distributions also had very large differences among if the analyses were made by regions, streams, or months. A visual
stations in the average values of the SSI and underestimated the inspection of the L-moment diagrams showed that none of the six
expected frequency of extreme negative SSI values. The PIII and probability distributions was best able to fit the streamflow series.
lognormal distributions also had large limitations, primarily in re- A goodness-of-fit test moreover did not help identify a unique
lation to the expected average and standard deviations of the SSI probability distribution because most of the cases could not reject
series. This made spatial comparisons of the SSI values difficult. the null hypothesis (i.e., the series follows at least one of the six
The PIII distribution also showed an excessive frequency of ex- distributions). Therefore, neither approach could help select a sin-
treme negative values. The results indicate that, rather than using gle probability distribution for computing the SSI. When the SSI
a unique probability distribution for calculating the SSI over a series obtained were compared, they nevertheless showed marked
range of streams with different flow magnitudes and regimes, it differences in their statistical properties (i.e., the mean and standard
is better to select the most suitable distribution for each monthly deviation) and in the estimation of extreme quantiles.
series in accordance wth the BMF of MD approaches on the basis Previous attempts to calculate an SSI have used a unique prob-
of the probability distribution showing the best fit to each of the ability distribution (Zaidman et al. 2001; López-Moreno et al.
12 streamflow series at each station for the BMF or the MD 2009; Mo 2008; Nabaltis and Tsakiris 2009; Lorenzo-Lacruz et al.
approaches. Applying the BMF or MD approaches overcame the 2010). Shukla and Wood (2008) nevertheless warned of the need to
limitations of the two most suitable distributions (i.e., GEV and use different probability distributions to fit the streamflow series
log-logistic) because no underestimation of the most negative when calculating an SSI. The results also provide evidence that us-
SSI values or overestimation of the extreme positive SSI values ing a unique probability distribution for each gauging station does
was observed. not obtain a reliable index because of the large variability in the
statistical properties of the monthly series. If a unique probability
distribution is to be used to obtain the SSI, the results in this paper
Discussion and Conclusions suggest that the GEV or the log-logistic distribution should be se-
A key recommendation of this study is to assess and, if necessary, lected, although both tended to underestimate the frequency of the
to use different probability distributions for each gauging station most extreme negative values in the data set. None of the other four
and month of the year for calculating an SSI. The authors compared distributions the authors tested is recommended for using in a sin-
the performance of this calculation procedure by means of two gle distribution approach to obtaining an SSI because the derived
different approaches (the BMF and the MD) by using a single indexes were unsatisfactory. For example, the generalized Pareto
probability distribution for an entire region over the 12 months and Weibull distributions were undetermined for some values,
of a year. The analysis was conducted in the Ebro River basin, and the PIII and lognormal distributions did not generate reliable
an area that exhibits large variability among river regimes, the average and standard deviation values because they commonly di-
frequency characteristics of streamflows, and flow magnitudes. verged from 0 and 1, respectively. Large differences can moreover
The BMF and the MD approaches had the greatest flexibility in occur among stations, making spatial comparability of the series
accounting for the variety of conditions found within the basin, very difficult. Whereas some previous studies have recommended
and these approaches overcame the problems associated with using using the lognormal distribution to fit monthly streamflow data
a single distribution. Therefore, both approaches provide a robust (Kroll and Vogel 2002; Yue and Wang 2004; Chen et al. 2006;
index that guarantees the spatial and temporal comparability of McMahon et al. 2007) and to obtain standardized streamflows
drought conditions. They provide very similar results in the ex- (Zaidman et al. 2001; Nabaltis and Tsakiris 2009), the authors
pected average, variance, and the occurrence of extremes in the of this paper demonstrated that using this distribution is inappro-
series of the drought indexes. priate in a region with high variability in the river regimes and
The procedure for obtaining the SSI was by converting the con- streamflow characteristics. The PIII has also been widely used
version of the streamflow magnitudes to standardized anomalies for flow frequency analyses (Bobee and Rasmussen 1995; Saf
(i.e., z-scores). For this purpose, the authors fit various probability 2009). The results of the KS test for the streamflow series in
distributions to the available streamflow data. Selecting the most the present study indicates that in most cases (> 95% for each
research into climate drought indicators has aimed to develop frequency analysis.” Rev. Geophys., 33(Supplement 2), 1111–1116.
procedures that ensure that the drought conditions quantified by Burn, D. H. (1997). “Catchment similarity for regional flood frequency
analysis using seasonality measures.” J. Hydrol. (Amsterdam),
a drought index are comparable among different regions (Wells
202(1–4), 212–230.
et al. 2004; Vicente-Serrano et al. 2010a). The BMF or the MD
Chen, Y. D., Huang, G., Shao, Q. X., and Xu, C.-Y. (2006). “Regional
approaches allow the development of a reliable standardized index, analysis of low flow using L-moments for Dongjiang Basin, South
with the expected averages, standard deviations, and frequencies China.” Hydrol. Sci. J., 51(6), 1051–1064.
of extreme positive and negative values to be obtained. Therefore, Chow, V. T., Maidment, D. R., and Myas, L. W. (1988). Applied hydrology,
the SSI obtained by using either approach at a particular gauging MacGraw-Hill, New York.
station can be compared with the SSI obtained at another station. Clausen, B., and Pearson, C. P. (1995). “Regional frequency analysis of
The values of the index are spatially comparable and they are tem- annual maximum streamflow drought.” J. Hydrol. (Amsterdam),
porally independent of the streamflow regimes. 173(1–4), 111–130.
By using the SSI, it is also possible to characterize the magni- Cuadrat, J. M., Saz, M. A., and Vicente-Serrano, S. M. (2007). Atlas
tude and duration of droughts, and to apply the widely used low- climático de Aragón, Gobierno de Aragón, Zaragosa, Spain, 229.
flow frequency analysis on the basis of the run theory (Yevjevich Dracup, J. A., Lee, K. Y., and Paulson, E. G. (1980). “On the definition of
droughts.” Water Resour. Res., 16(2), 297–302.
1967; Smakhtin 2001). However, some advantages accrue with re-
Edossa, D. C., Babel, M. S., and Gupta, A. D. (2010). “Drought analysis in
spect to other procedures because the magnitude of the identified the Awash River Basin, Ethiopia.” Water Resour. Manage., 24(7),
droughts can be compared spatially, and the truncation levels can be 1441–1460.
defined in accordance with the theoretical return periods. Because Fink, A. H., Brücher, T., Krüger, A., Leckebush, G. C., Pinto, J. G., and
the SSI is measured in the same units that are currently used for Ulbrich, U. (2004). “The 2003 European summer heatwaves and
other climate drought indexes, the comparisons between hydrolog- drought—synoptic diagnosis and impacts.” Weather, 59(8), 209–216.
ical and climate droughts are possible–thereby aiding the study of Fleig, A. K., Tallaksen, L. M., Hisdal, H. Y., and Demuth, S. (2006). “A
the complex interactions between the two phenomena. global evaluation of streamflow drought characteristics.” Hydrol. Earth
A main objective of drought indexes is drought monitoring, Syst. Sci., 10(4), 532–552.
early warning, and improving drought preparedness and mitigation Guttman, N. B. (1999). “Accepting the standardized precipitation index: A
(Svoboda et al. 2002 and 2004). Current drought monitoring calculation algorithm.” J. Am. Water Resour. Assoc., 35(2), 311–322.
systems are largely based on climate drought indexes. Including Hannaford, J., Lloyd-Hughes, B., Keef, C., Parry, S., and Prudhomme, C.
(2011). “Examining the large-scale spatial coherence of European
additional hydrological drought indexes (such as the SSI) could
drought using regional indicators of precipitation and streamflow
improve the performance of these systems. The capacity of an deficit.” Hydrol. Processes, 25(7), 1146–1162.
SSI to add reliable and direct information about the spatial extent Havens, A. V. (1954). “Drought and agriculture.” Weatherwise, 7(3), 51–55.
and severity of hydrological drought conditions will provide a con- Heim, R. R. (2002). “A review of twentieth-century drought indices used in
venient way to quantify the level of risk in a way that is accurate the United States.” Bull. Am. Meteorol. Soc., 83(8), 1149–1165.
and easily understood by end-users. Hirsch, R. M. (1982). “A comparison of four streamflow record extension
techniques.” Water Resour. Res., 18(4), 1081–1088.
Hisdal, H., and Tallaksen, L. M. (2003). “Estimation of regional meteoro-
Acknowledgments logical and hydrological drought characteristics: A case study for
Denmark.” J. Hydrol. (Amsterdam), 281(3), 230–247.
This work has been supported by research projects grant numbers Hosking, J. R. M. (1986). “The theory of probability weighted moments.”
CGL2008-01189/BTE and CGL2006-11619/HID financed by the Research Rep. RC12210, IBM Research Div., Yorktown Heights, NY.
Spanish Commission of Science and Technology and FEDER; Hosking, J. R. M. (1990). “L-moments: Analysis and estimation of distri-
EUROGEOSS grant number (FP7-ENV-2008-1-226487), and butions using linear combinations of order statistics.” J. R. Stat. Soc.
Ser. B, 52(1), 105–124.
ACQWA grant number (FP7-ENV-2007-1-212250) financed by
Hosking, J. R. M., and Wallis, J. R. (1993). “Some statistics useful in
the VII Framework Programme of the European Commission; regional frequency analysis.” Water Resour. Res., 29(2), 271–281.
and “Las sequías climáticas en la cuenca del Ebro y su respuesta Hosking, J. R. M., Wallis, J. R., and Wood, E. F. (1985). “Estimation of the
hidrológica” and “La nieve en el Pirineo aragonés: Distribución generalized extreme-value distribution by the method of probability-
espacial y su respuesta a las condiciones climáticas” financed by weighted moments.” Technometrics, 27(3), 251–261.
“Obra Social La Caixa” and the Aragón government. Kroll, C. N., and Vogel, R. M. (2002). “Probability distribution of low
streamflow series in the United States.” J. Hydrol. Eng., 7(2), 137–146.
López-Moreno, J. I., Vicente-Serrano, S. M., Beguería, S., García-Ruiz,
References J. M., Portela, M. M., and Almeida, A. B. (2009). “Downstream propa-
gation of hydrological droughts in highly regulated transboundary
Abramowitz, M., and Stegun, I. A. (1965). Handbook of mathematical rivers: The case of the Tagus River between Spain and Portugal.” Water
functions, Dover Publications, New York. Resour. Res., 45(2), W02405.
ecological impacts, and model projections.” Bull. Am. Meteorol. rural development, land, desertification and drought, Department of
Soc., 81(3), 413–416. Economic and Social Affairs, New York.
Mishra, A. K., and Singh, V. P. (2010). “A review of drought concepts Vasiliades, L., and Loukas, A. (2009). “Hydrological response to meteoro-
review article.” J. Hydrol. (Amsterdam), 391(1–2), 202–216. logical drought using the palmer drought indices in Thessaly, Greece.”
Mo, K. C. (2008). “Model-based drought indices over the United States.” Desalination, 237(1–3), 3–21.
J. Hydrometeorol., 9(6), 1212–2130. Vicente-Serrano, S. M. (2006). “Differences in spatial patterns of drought
Modarres, R., and Sarhadi, A. (2010). “Frequency distribution of extreme on different time scales: An Analysis of the Iberian Peninsula.” Water
hydrologic drought of southeastern semiarid region, Iran.” J. Hydrol. Resour. Manage., 20(1), 37–60.
Eng., 15(4), 255–264. Vicente-Serrano, S. M., Beguería, S., and López-Moreno, J. I. (2010a).
Mudelsee, M. (2007). “Long memory of rivers from spatial aggregation.” “A multi-scalar drought index sensitive to global warming: The stand-
Water Resour. Res., 43, W01202. ardized precipitation evapotranspiration index—SPEI.” J. Clim., 23(7),
Nabaltis, I., and Tsakiris, G. (2009). “Assessment of hydrological drought 1696–1718.
revisited.” Water Resour. Manage., 23(5), 881–897. Vicente-Serrano, S. M., Beguería, S., López-Moreno, J. I., Angulo,
Nicholson, S. E. (2001). “Climatic and environmental change in Africa M., and El Kenawy, A. (2010b). “A new global 0.5° gridded dataset
during the last two centuries.” Clim. Res., 17(2), 123–144. (1901–2006) of a multiscalar drought index: Comparison with current
Obasi, G. O. P. (1994). “WMO’s role in the international decade for natural drought index datasets based on the Palmer drought severity index.”
disaster reduction.” Bull. Am. Meteorol. Soc., 75(9), 1655–1661. J. Hydrometeorol., 11(4), 1033–1043.
Ollero, A., Pellicer, F., and Sánchez-fabre, M. (2004). “La crecida de Vicente-Serrano, S. M., Lanjeri, S., and López-Moreno, J. I. (2007).
febrero de 2003 en el curso medio del Ebro: análisis de su evolución “Comparison of different procedures to map reference evapotranspira-
espacio-temporal.” Aportaciones geográficas en homenaje al profesor tion using geographical information systems and regression-based tech-
A. Higueras Arnal, M. C. Faus Pujol, ed., Departamento de Geografía, niques.” Int. J. Climatol., 27(8), 1103–1118.
Universidad de Zaragoza, Zaragoza, Spain, 243–256. Vicente-Serrano, S. M., and López-Moreno, J. I. (2005). “Hydrological re-
Peel, M. C., Wang, Q. J., Vogel, R. M., and McMahon, T. A. (2001). “The sponse to different time scales of climatological drought: An evaluation
utility of L-moment ratio diagrams for selecting a regional probability of the standardized precipitation index in a mountainous Mediterranean
distribution.” Hydrol. Sci. J., 46(1), 147–155. basin.” Hydrol. Earth Syst. Sci., 9(5), 523–533.
Rao, A. R., and Hamed, K. H. (2000). Flood frequency analysis, Vidal, J. P., et al. (2010). “Multilevel and multiscale drought reanalysis over
CRC Press, Boca Raton, FL, 350. France with the Safran-Isba-Modcou hydrometeorological suite.”
Redmond, K. T. (2002). “The depiction of drought: A commentary.” Bull. Hydrol. Earth Syst. Sci., 14(3), 459–478.
Am. Meteorol. Soc., 83(8), 1143–1147. Vogel, R. M., and Stedinger, J. R. (1985). “Minimum variance streamflow
Riggs, H. C. (1973). “Regional analyses of streamflow characteristics.” record augmentation procedures.” Water Resour. Res., 21(5), 715–723.
Techniques of Water-Resources Investigations of the United States Vogel, R. M., Thomas, W. O., Jr., and McMahon, T. A. (1993). “Flood
Geological Survey, Book 4, Chapter B3, U.S. Dept. of the Interior, frequency model selection in southwestern United States.” J. Water
Washington, DC., 15. Resour. Plann. Manage., 119(3), 353–366.
Saf, B. (2009). “Regional flood frequency analysis using L-moments for the Wells, N., Goddard, S., and Hayes, M. J. (2004). “A self-calibrating Palmer
west Mediterranean region of Turkey.” Water Resour. Manage., 23(3), drought severity index.” J. Clim., 17(12), 2335–2351.
531–551. Wilhite, D. A. (1993). Drought assessment, management and planning:
Shi, P., Chen, X., Qu, S. M., Zhang, Z.-C., and Ma, J.-L. (2010). “Regional Theory and case studies, Kluwer, Boston.
frequency analysis of low flow based on L moments: Case study in Yevjevich, V. M. (1967). “An objective approach to definition and inves-
Karst area, Southwest China.” J. Hydrol. Eng., 15(5), 370–377. tigation of continental hydrologic droughts.” Hydrologic Paper 23,
Shukla, S., and Wood, A. (2008). “Use of a standardized runoff index for Univ. of Colorado, Fort Collins, CO.
characterizing hydrologic drought.” Geophys. Res. Lett., 35, L02405. Yue, S., and Hashino, M. (2007). “Probability distribution of annual,
Siegel, S., and Castelan, N. J. (1988). Nonparametric statistics for the seasonal and monthly precipitation in Japan.” Hydrol. Sci. J., 52(5),
behavioral sciences, McGraw-Hill, New York. 863–877.
Singh, V. P., and Guo, F. X. Y. (1993). “Parameter estimation for Yue, S., and Pilon, P. (2005). “Probability distribution type of Canadian
3-parameter log-logistic distribution (LLD3) by Pome.” Stoch. Hydrol. annual minimum streamflow.” Hydrol. Sci. J., 50(3), 427–438.
Hydraul., 7(3), 163–177. Yue, S., and Wang, C. Y. (2004). “Possible regional probability distribution
Smakhtin, V. U. (2001). “Low flow hydrology: A review.” J. Hydrol. type of Canadian annual streamflow by L-moments.” Water Resour.
(Amsterdam), 240(3–4), 147–186. Manage., 18(5), 425–438.
Stahl, K. (2001). “Hydrological drought—A study across Europe.” Zaidman, M. D., Rees, H. G., and Young, A. R. (2002). “Spatio-temporal
Doctoral dissertation, Geowissenschaftlichen Fakultät, Albert-Ludwigs development of streamflow droughts in north-west Europe.” Hydrol.
Universität Freiburg, Freiburg, Germany. Earth Syst. Sci., 5(4), 733–751.
Svoboda, M., et al. (2002). “The drought monitor.” Bull. Am. Meteorol. Zelenhasic, E., and Salvai, A. (1987). “A method of streamflow drought
Soc., 83(8), 1181–1190. analysis.” Water Resour. Res., 23(1), 156–168.