Professional Documents
Culture Documents
Arabzadeh 2016
Arabzadeh 2016
Components Analysis
Rezgar Arabzadeh, S.M.ASCE 1; Mohammad Mehdi Kholoosi 2; and Javad Bazrafshan 3
Abstract: Drought quantification is a decision-making approach for water resources planners and managers. There are different statistical
methods for regional quantification of droughts, depending on the event being regionally discrete or continuous. Calculating a drought index
Downloaded from ascelibrary.org by University of Exeter on 07/18/15. Copyright ASCE. For personal use only; all rights reserved.
in the measured grid points of a given region is a prerequisite for the regionalization process of drought by these methods. The aim of
this study is the regional analysis of the streamflow drought using the multivariate technique of principal components analysis (PCA).
To this end, the streamflow drought index (SDI) was calculated in the seven hydrometric stations of the Sefid-Rud basin in Iran for the
period of 1984–2013 in seasonal (winter, spring, summer, and autumn), semiannual (October–March and May–September), and annual
(October–September) time scales. Regional monitoring of the SDI was carried out using the PCA technique summarizing the SDI series
of all stations into a new series, the so-called the multivariate streamflow drought index (MSDI). The MSDI series at each time scale were
analyzed from several statistical aspects. Results showed that there are relatively high correlations between the SDI series of the stations for
given time scales. The first principal component (PC1 ) explains 58–85% of the regional variations of the SDI series at the mentioned time
scales. The MSDI series at multiple time scales follow all stations’ SDI fluctuations and appropriately monitor droughts that occurred in the
region, especially in long dry periods. The dry and wet severity classes derived from the MSDI series greatly corresponded to those of the SDI
time series on different stations. The highest percentage of correspondence between MSDI and SDIs was calculated for the semiannual time
scale and the lowest for the spring season. DOI: 10.1061/(ASCE)IR.1943-4774.0000925. © 2015 American Society of Civil Engineers.
Author keywords: Hydrological drought; Principal components analysis (PCA); Streamflow; Regional analysis; Iran.
of different indices have been developed for quantifying drought; Supply Index (SWSI) monitors abnormalities in surface water
each one has its own strengths and weaknesses. In the following supply sources such as streamflow, reservoirs, snow pack stored
Downloaded from ascelibrary.org by University of Exeter on 07/18/15. Copyright ASCE. For personal use only; all rights reserved.
paragraph, some of them are briefly discussed. in mountainous regions, and precipitation (Shafer and Dezman
The standardized precipitation index (SPI) was introduced by 1982). Like the SWSI, the Reclamation Drought Index (RDI) as-
McKee et al. (1993). The SPI is applicable for different regions sesses drought using streamflow, temperature, snow pack, and pre-
and is based on long-term rainfall data records. Moreover than pre- cipitation to calculate drought on the river basin scale. Sharma and
cipitation Palmer (1965) introduced another drought index known Panu (2010) developed the Standardized Hydrological Index (SHI).
as the Palmer Drought Severity Index (PDSI), which uses temper- The concept of SHI is analogous to SPI (Mckee et al. 1993), except
ature for assessing drought on regional scales. The Surface Water that SPI represents a standardized entity and thus implicitly inherits
Fig. 1. Locations of the hydrometric stations selected in the Sefid-Rud basin in the northwest of Iran; for the names of stations corresponding to the
numbers written in the figure refer to Table 1 (data from Iranian Nation Organization of Surveying 2013)
© ASCE
04015029-3
Fig. 2. (Continued.)
of the nonnormal characteristics of the flow series (Sharma and to transition probability analysis also has been recommended
Panu 2012). (Paulo and Pereira 2008). The purposes of this study are to ag-
Most of the indices described above are computationally diffi- gregate spatial and temporal patterns of drought for assessing
cult and data demanding in contrast the Streamflow Drought Index regional drought quantity using principal component analysis
(SDI) introduced by Nalbantis and Tsakiris (2009), which is very (PCA), which reduces the dimensionality of the SDI data set.
simple in calculation and powerful in assessing at-site drought. The In research related to drought, PCA has been used to assess the
SDI is only based on the streamflow values. spatiotemporal analysis of drought in different time and space
In drought studies, there is no deterministic time scale for scales. (Bomaccorso et al. 2003; Bordi et al. 2004; Sonmez et al.
assessing drought severity, and the majority of researchers based 2005; Raziei et al. 2011; Bazrafshan et al. 2014). Based on Wilks
on their own assumptions, or other research work with time (2011), PCA is able to summarize several variables into fewer
frames of 3, 6, 9, 12, and 24 months proposed by different groups important components.
of researchers, (Mishra and Desai 2005; Cancelliere et al. 2007) The following sections discuss about how to implement and cal-
or drought classification based on a 12-month time scale in order culate SDI and reduce it to its own major components. The paper
and the temperature in mountain areas is as low as −30°C and show nonnormal distribution of the streamflow time series between
in the planes rises to 40°C. The majority of water flowing into seasons and months. All kurtosis coefficients are also greater than
the Sefid-Rud basin is from basins situated in the Kurdistan 4.71. Fig. 1 shows the locations of the stations used in this study,
territory. Table 1 shows the geographic and hydrologic features and Fig. 2 shows the naturalized streamflows used in this study.
of stations used in the current study. The Gezil-Ozan-Nisare, Selan, During the last decade (2005–2015), two major dams have been
Remisht, and Chil-Gezi stations are situated in the upstream part of constructed in the basin, the Sang-Siyah Dam and Siyazagh Dam.
Fig. 3. Locations of dams in study area and demand sites in the upstream of hydrometric stations (data from Iranian Nation Organization of
Surveying 2013)
Table 2. The p-Values Calculated Corresponding to the Statistics Values of Kolmogorov-Smirnov and Chi-Square Goodness-of-Fit Tests
Downloaded from ascelibrary.org by University of Exeter on 07/18/15. Copyright ASCE. For personal use only; all rights reserved.
Fig. 4. Illustrative diagram of SDI calculation steps from a desired commutative probability distribution fitted on data set and a normal distribution
using principal of probability transformation
For each reference period k and ith hydrological year, the SDI is calculation of the time series at all stations for a given time scale.
calculated with Eq. (4) written as follows: In this study, a station’s SDI time series is aggregated using the
Fig. 7. Cross-correlation diagrams, representing correlation coefficient value among stations; each polygon shows correlation coefficients of a given
station with other stations
Fig. 7. (Continued.)
Table 4. Results of Bartlett’s Sphericity Test for the SDI Series at Each technique of PCA to evaluate the regional drought in the study area
Time Scale in the Study Area for a particular time scale. For example, SDI was calculated for
Chi square Chi square each station at a scale of 6 months and by using the ability of multi-
MSDI series (calculated) (critical) p-value variate techniques, these drought indices were aggregated to one
Fall 217.237 32.671 <0.0001 component that represents the majority of all stations’ droughts.
Winter 84.334 32.671 <0.0001 Having a comprehensive view of drought and the vast regional in-
Spring 395.286 32.671 <0.0001 dex, which can be generalized to the whole study area for a specific
Summer 107.799 32.671 <0.0001 time scale, is much more reasonable and also applied. Also, various
First 6 months 204.129 32.671 <0.0001 drought indices for water resource managers and planners in mak-
Second 6 months 266.093 32.671 <0.0001 ing decisions about water resource systems in this situation is dif-
Annual 265.347 32.671 <0.0002
ficult, so having a unit and representative drought index for water
© ASCE
04015029-10
Fig. 10. SDI’s biplots representing the contribution of the original SDI variants in PC1 and PC2
Fig. 11. MSDI time series (solid line) and SDI time series (dark band) for each time scale
between the stations’ SDI time series, the correlation matrices for that all stations’ SDI time series have sufficient correlation. In
the 3-, 6-, and 12-month time scales were calculated and drawn addition, the results of Bartlett’s sphericity test (Table 4) confirm
using the radar graphs. Fig. 7 shows the cross correlations of SDIs the sufficiency of the hyper cloud’s correlation at each time scale.
among stations calculated for all time scales. As shown in Fig. 7, According to Table 4, chi square statistics are all significant at the
correlations among variants are high and statistically significant 1% level.
at the 5% level. The more the polygon became larger, a polygon
presents its correspondent station correlation coefficients with
other stations as being more statistically significant. The maxi- Principal Component Analysis of the SDI Time Series
mum correlation is between Telwar-Dehgolan and Selan equal According to the stations’ correlation matrices, their eigenvalues
to 0.99 in spring and the minimum correlation of 0.29 was seen and corresponding eigenvectors were derived. Scree plots in Fig. 8
between Chil-Gezi and Telwar-Mehr-Abad in the summer. In show the eigenvalues and their cumulative variability versus PCs’
the case of winter, the maximum correlation was 0.89 between number. As shown in Fig. 8, most of the hyper cloud correlation can
Telwar-Mehr-Abad and Gezil-Ozan-Nisare and the minimum be explained just by the first PC and other components only cover
was between Telwar-Mehr-Abad and Chil-Gezi. For the fall sea- a little information of the data sets in which for the fall, winter,
son, the maximum and minimum correlations were 0.99 and 0.55 spring, summer, first 6 months, and the second 6 months and for
between Selan and Telwar-Dehgolan and Remisht and Heshta-Jift, annual periods can explain 72, 78, 80, 58, 75, 85, and 83% of the
respectively. The results from the correlation matrices show a high variations, respectively, which shows the ability of the PC1 to
dependency of variation between the stations’ SDIs and revealed model regional streamflow drought.
these biplots, both positive and large contributions (correlations) and gathers their indices into one index, especially during long dry
of variants to PC1 are conspicuous. According to the diagrams and wet periods. Also, the MSDI in the original SDI time series
Fig. 12. Percentage of correspondence of the MSDI series with all stations’ SDI series for each time scale in the study area during the
period 1984–2013
Fig. 14. Probability transformation matrices of drought classes in two subsequent season in all stations
multivariate SDI time series. A notable case is that there is a • Results obtained from MSDI are simple to illustrate and com-
high density of probability in the lower-left corner in most of the parable to different time scales and spatial zones due to its abil-
matrices except Telwar-Dehgolan. This shows that whenever the ity to be standardized; and
hydrologic system traps in a severely dry season or extremely dry • MSDI calculation is computationally simple and can be applied
period difficultly, it would be able to escape it or the system faces to any arbitrary time scale and spatial zone.
another condition where the consequences of drought or its effects The current study offers an MSDI calculation as a new drought
are lower. index for monitoring regional drought. First, a comprehensive
goodness of fit test was carried out for choosing the best distribu-
tion to model the probability function of the time series for all sta-
Conclusion tions and each time scale. Then, the correspondent standard values
of each streamflow probability were extracted using normal distri-
In this study, a new drought index is proposed based on EOFs to bution, which is equal to the SDI time series. The SDIs of the sta-
assess the hydrological drought index (SDI), representing drought tions were summarized using PCA into one PC1 and eventually
dependent on streamflow. The features making MSDI a useful tool standardized to obtain the MSDI. According to results and dis-
for drought monitoring are cussed methods, the following conclusions have been made:
• MSDI is able to reduce SDI time series and gathers them into • The principal components analysis of the SDI time series
one principal component, which is very useful in regional water showed that the first PC1 could explain a large percentage of
resources planning and management; variations in the original SDI time series in all selected stations;
• Since MSDI is basically calculated on first principal component, • The dry and wet severity classes derived from the MSDI series
it contains a major part of all other components and has appro- greatly corresponded to those of the SDI time series on different
priate accuracy in regional drought monitoring; stations;
Fig. 15. Probability transformation matrices of drought classes in two subsequent seasons
• The highest percentage of correspondence was calculated for Dai, A. (2011). “Drought under global warming: A review.” Wiley Inter-
the 6-month time scales and the lowest was for the spring discip. Rev. Clim. Change, 2(1), 45–65.
season; and Gumbel, E. J. (1963). “Statistical forecast of droughts.” Bull. Int. Assoc. Sci.
• Altogether the season’s percentage of correspondence was lower Hydrol., 8(1), 5–23.
Heim, R. R. (2002). “A review of twentieth-century drought indices used in
than the larger time scale’s percentage of correspondence.
the United States.” Bull. Am. Meteorol. Soc., 83(8), 1149–1165.
Iranian Nation Organization of Surveying. (2013). “Technical service man-
agement datasets archives: Digital elevation model.” 〈http://tsm.ncc.org
References .ir/〉 (May 1, 2014).
Keyantash, J., and Dracup, J. A. (2004). “An aggregate drought index:
Allen, P. M., Harmel, R. D., Dunbar, J. A., and Arnold, J. G. (2011). Assessing drought severity based on fluctuations in the hydrologic cycle
“Upland contribution of sediment and runoff during extreme drought: and surface water storage.” Water Resour. Res., 40(9), W09304.
A study of the 1947–1956 drought in the Blackland Prairie, Texas.” McKee, T. B., Doesen, N. J., and Kleist, J. (1993). “The relationship of
J. Hydrol., 407(1–4), 1–11. drought frequency and duration to time scales.” Proc., 8th Conf. on
Bazrafshan, J., Hejabi, S., and Rahimi, J. (2014). “Drought monitoring Applied Climatology, Anaheim, CA, 179–184.
using the multivariate standardized precipitation index (MSPI).” Water Mishra, A. K., and Desai, V. R. (2005). “Drought forecasting using stochas-
Resour. Manage., 28(4), 1045–1060. tic models.” Stochastic Environ. Res. Risk Assess., 19(5), 326–339.
Bonaccorso, B., Bordi, I., Cancielliere, A., Rossi, G., and Sutera, A. (2003). Mishra, A. K., and Singh, V. P. (2010). “A review of drought concepts.”
“Spatial variability of drought: An analysis of the SPI in Sicily.” Water J. Hydrol., 391(1–2), 202–216.
Resour. Manage., 17(4), 273–296. Nalbantis, I. (2008). “Evaluation of a hydrological drought index.” Eur.
Bordi, I., Fraedrich, K., Jiang, J. M., and Sutera, A. (2004). “Spatio- Water, 23(24), 67–77.
temporal variability of dry and wet periods in eastern China.” Theor. Nalbantis, I., and Tsakiris, G. (2009). “Assessment of hydrological drought
Appl. Climatol., 79(1–2), 81–91. revisited.” Water Resour. Manage., 23(5), 81–97.
Cancelliere, A., Di Mauro, G., Bonaccorso, B., and Rossi, G. (2007). Özger, M., Mishra, A. K., Singh, V. P. (2012). “Long lead time drought
“Drought forecasting using the standardized precipitation index.” Water forecasting using a wavelet and fuzzy logic combination model.”
Resour. Manage., 21(5), 801–819. J. Hydrometeorol., 13(1), 284–297.