A monthly global dataset of a multiscalar drought index is presented and compared in terms of spatial and temporal variability with the existing continental and global drought datasets based on the Palmer drought severity index (PDSI). The presented dataset is based on the standardized precipitation evapotranspiration index (SPEI). The index was obtained using the Climatic Research Unit (CRU) TS3.0 dataset at a spatial resolution of 0.5°. The advantages of the new dataset are that (i) it improves the spatial resolution of the unique global drought dataset at a global scale; (ii) it is spatially and temporally comparable to other datasets, given the probabilistic nature of the SPEI; and, in particular, (iii) it enables the identification of various drought types, given the multiscalar character of the SPEI. The dataset is freely available on the Web page of the Spanish National Research Council (CSIC) in three different formats [network Common Data Form (netCDF), binary raster, and plain text].
One of the priorities of the climate sciences is the development of reliable datasets to analyze climate processes at spatial scales ranging from continental to global. Most effort has been devoted to the development of global gridded datasets of various climate variables, including temperature, precipitation, and pressure. Access to most of the datasets is through the Climate Explorer (available online at http://climexp.knmi.nl) of the Dutch Royal Netherlands Meteorological Institute (Koninklijk Nederlands Meteorologisch Instituut).
Notwithstanding the usefulness of these datasets, there is a need for datasets of basic climate parameters and synthetic information about moisture and dryness conditions; these are highly valued by environmental, hydrological, and global change researchers, and they are indispensable for determining the possible effects of climate variability and change. The best approach to obtaining a measure of relative wetness or dryness is the calculation of drought indices (Heim 2002; Keyantash and Dracup 2002).
Among various indices for drought detection, the Palmer drought severity index (PDSI; Palmer 1965) is one of the most widely used; the calculation procedure for this index has been described in several studies (e.g., Karl 1983, 1986; Alley 1984). This is a climatic water balance index that considers precipitation and evapotranspiration anomalies and soil water-holding capacity. Many of the PDSI deficiencies were resolved by development of the self-calibrated PDSI (sc-PDSI; Wells et al. 2004), which is spatially comparable and reports extreme wet and dry events at frequencies expected for rare conditions. All currently available gridded drought datasets at continental and global scales are based on either the PDSI or the sc-PDSI (Dai et al. 2004; Van der Schrier et al. 2006a,b).
The PDSI has a fixed temporal scale, which does not allow different drought types (e.g., hydrological, meteorological, agricultural, and socioeconomic) to be distinguished. This is an important shortcoming because drought is a multiscalar phenomenon (McKee et al. 1993); it is a phenomenon that may occur simultaneously across multiple temporal scales (e.g., a short period of particular dryness embedded within a long-term drought). Therefore, multi refers to numerous, temporal periods that may or may not overlap. The response of various hydrological systems (including soil moisture, groundwater, snowpack, river discharge, and reservoir storage) to precipitation can vary markedly as a function of time (Changnon and Easterling 1989; Elfatih et al. 1999; Pandey and Ramasastri 2001). The need for a drought index that considers the multiscalar nature of droughts explains the wide acceptance of the standardized precipitation index (SPI), which was developed by McKee et al. (1993). This index can be calculated at varying time scales to monitor droughts with respect to different usable water resources (e.g., Szalai et al. 2000; Ji and Peters 2003; Vicente-Serrano and López-Moreno 2005; Khan et al. 2008; Lorenzo-Lacruz et al. 2010).
However, the SPI has the important shortcoming that it is based only on precipitation data, and it does not consider other critical variables such as evapotranspiration, which can have a marked influence on drought conditions. Abramopoulos et al. (1988) used a general circulation model experiment to show that evaporation and transpiration can consume up to 80% of rainfall. In addition, the authors found that the fraction of drying caused to temperature anomalies is as high as that attributable to rainfall shortage. Therefore, it is preferable to use drought indices that include temperature data in the formulation of datasets such as the PDSI. However, the PDSI lacks the multiscalar character essential for determining the effect of droughts on different hydrological systems, crops, and natural vegetation, and for differentiating among various drought types. For this reason, Vicente-Serrano et al. (2010) formulated a new drought index [the standardized precipitation evapotranspiration index (SPEI)] based on precipitation and potential evapotranspiration (PET). The SPEI combines the sensitivity of the PDSI to changes in evaporation demand (caused by temperature fluctuations and trends) with the multitemporal nature of the SPI.
Here we present a new global drought dataset based on the SPEI that covers time scales from 1–48 months at a spatial resolution of 0.5° and that provides temporal coverage for the period 1901–2006. This dataset represents an improvement in the spatial resolution and operative capability of previous gridded drought datasets based on the PDSI and enables identification of various drought types.
To calculate the SPEI, we used the Climatic Research Unit (CRU) TS3 dataset (available online at http://badc.nerc.ac.uk/data/cru/). This is the most complete and updated dataset of gridded precipitation and temperature at the global scale, has a spatial resolution of 0.5°, and covers the period 1901–2006.
The SPEI is based on the climatic water balance, that is, the difference between precipitation and PET:
where P is the monthly precipitation (mm) and PET (mm) is calculated according to the method of Thornthwaite (1948), which only requires data on mean monthly temperature and the geographical location of the region of interest.
The calculated D values were aggregated at various time scales:
where k (months) is the time scale of the aggregation and n is the calculation number. The D values are undefined for k > n.
A log-logistic probability distribution function was then fitted to the data series of D, as it adapts very well to all time scales. The complete calculation procedure for the SPEI can be found in Vicente-Serrano et al. (2010).
We tested the goodness of fit between the global monthly series of Dk at time scales from 1 to 48 months and the log-logistic distribution. This was checked using the Kolmogorov–Smirnov (KS) test at a critical level, α = 0.05. The KS test is based on the KS distance statistic, which quantifies the maximum vertical distance between the empirical cumulative distribution function (ECDF) of the sample (the 1901–2006 series of Dk) and the cumulative distribution function (CDF) of the reference distribution. In this case the ECDF was calculated using the plotting position formula proposed by Hosking (1990) for highly skewed data.
We compared the multiscalar SPEI dataset with the available PDSI-based datasets at global and continental scales and assessed its capabilities. For this purpose we used the global PDSI at 2.5°, developed by the University Corporation for Atmospheric Research (UCAR; Dai et al. 2004). We also used the European and North American 0.5° sc-PDSI grids, developed by the CRU of the University of East Anglia (van der Schrier et al. 2006a,b). Both datasets were obtained using precipitation and temperature data from the CRU TS 2.1 datasets (Mitchell and Jones 2005). The datasets span the period 1901–2002 and the regions 20°–50°N, 130°–60°W for North America and 35°–70°N, 10°W–60°E for Europe.
a. Goodness of fit of the global SPEI dataset
Figure 1 shows results from the application of the KS test to the global SPEI dataset for January and July, at time scales of 1, 4, and 12 months. This enabled the log-logistic distribution for most parts of the world to be accepted, because the null hypothesis (that the data came from a log-logistic distribution) was rejected for only very few areas. In most regions the P value for the KS distance statistic was well above 0.45, indicating that the log-logistic distribution was highly suitable for fitting the Dk series. Only for the shortest time scales and for regions of poor data availability (such as northern Siberia, Greenland, and the Himalayas) did the Dk series fail to match the log-logistic distribution. For most time scales and months, the log-logistic distribution fitted the Dk series very well across most of the world. The global percentage oscillates between 85% and 97% as a function of the time scale and the month. Therefore, the selected log-logistic distribution was considered highly appropriate for the calculation of the SPEI in most regions, independent of the month and time scale of analysis.
b. Comparison with other drought-gridded datasets
Figure 2 shows the spatial distribution of the UCAR PDSI and the SPEI at selected time scales for August 1936, a month in which major drought conditions occurred in some regions of North America and Russia. Comparison of these datasets showed the greater spatial resolution of the SPEI dataset, which facilitates local and regional analysis. The SPEI dataset shows some similarity in the intensity and spatial distribution of drought conditions worldwide. Nevertheless, large differences can be extracted as a function of the time scale of analysis. For example, in South Africa and Namibia, the PDSI showed moderate drought conditions for the majority of the region; however, the SPEI clearly showed that drought conditions were particularly severe at short time scales (3 months), whereas at the longest time scales, such episodes were not recognized. A similar pattern was evident for drought conditions in Australia, which were mainly characterized by short time scales. The PDSI seemed to respond differently to varying time scales of drought in different regions of the world. For example, in northwest Canada and Alaska, the humid conditions recorded in August 1936 by the PDSI were also recorded by the SPEI at the longest time scales (24–36 months). In other regions, including Scandinavia, the normal conditions (values close to 0) recorded by the PDSI were identified by the SPEI at short time scales; at longer time scales (12–24 months), the SPEI showed very humid conditions in these areas. The opposite pattern was found for the humid conditions in August 1936 in southwest Europe, where the strongest agreement between the PDSI and the SPEI occurred at time scales of 9–12 months.
Figure 3 shows a similar comparison involving the CRU sc-PDSI dataset for Europe for November 1949. The drought pattern from the sc-PDSI showed very few similarities with the SPEI at the 3-month time scale, where the sc-PDSI indicated that the most extreme drought conditions were in the Baltic countries and eastern Germany. France, the Iberian Peninsula, and large areas of the Balkans showed humid conditions. At the 6-month time scale, the most severe drought conditions were recorded in Germany, France, and Switzerland, with very humid conditions being recorded in most of the Balkans. Thus, the very humid conditions recorded in Italy and the Balkans at time scales of 3–9 months were not identified using the sc-PDSI.
In summary, these two examples show that the existing drought datasets, based on the PDSI and the sc-PDSI, are too rigid to identify droughts of varying temporal scales (short, medium, and long term). In addition, the relative conditions of dryness and humidity indicated by the PDSI and the sc-PDSI are sometimes erroneous, as they may result from very dry conditions over short time scales and very humid conditions over long time scales (and vice versa). However, the SPEI enabled detection of droughts at numerous time scales.
Figure 4 shows the average correlation and standard deviation between the CRU sc-PDSI and the SPEI at several time scales for Europe and North America and also between the global UCAR PDSI and the SPEI. Correlations were calculated for each time series of 0.5° (for Europe and North America; sc-PDSI) and 2.5° (globally, aggregating the SPEI pixels of 0.5°–2.5°). Although the correlations differed in magnitude between the global, European, and North American datasets (the highest correlations were found for the sc-PDSI in Europe and the lowest for the global PDSI), the three datasets showed maximum correlations with the SPEI at time scales of 6–18 months, with a maximum in all cases at the time scale of 12 months.
These differences are clearly very important if relationships between PDSI-based indices and the different time scales of the SPEI are analyzed spatially. Figure 5 shows the spatial distribution of correlations between the UCAR PDSI and the SPEI at different time scales. At a time scale of 3 months, only in eastern United States and Australia were large areas recorded with correlations greater than 0.6; in most other regions, the correlation between the PDSI and the 3-month SPEI was very low. For the SPEI, the correlation increased in magnitude and spatial extent at time scales of 6–12 months; however, in some regions (Canada, Central America, central Africa, and parts of Asia), correlations between the SPEI and the PDSI were very low, independent of the time-scale analyzed. At time scales of 18–48 months, the relationship between the PDSI and the SPEI decreased. The maps show that, although at time scales of 9–18 months correlations were in general very high over large regions of the world (R > 0.8), in some regions, the PDSI could represent drought conditions at a different time scale. In Australia, for example, the PDSI tended to be reflecting shorter time scales than shown for eastern Europe and the eastern United States. Thus, very great spatial variability was evident in maps showing the time scales at which the correlation between the SPEI and the PDSI was highest.
Comparison of the CRU sc-PDSI with the SPEI also showed differences in the spatial patterns and magnitudes of correlations, and the time scale of the SPEI showing the highest correlation with the sc-PDSI. Figure 6 shows the spatial distribution of correlations between the sc-PDSI and the SPEI at different time scales in Europe. The pattern shows weak correlations at the 3-month time scale (although there were some exceptions) and an increase in the magnitude and surface extent of correlations at time scales of 6–12 months. The high correlations were maintained for the 18- and 24-month SPEI intervals, but they decreased thereafter. Most of the European continent showed correlations higher than 0.6 between the sc-PDSI and the SPEI. Nevertheless, whereas in North America maximum correlations were recorded at time scales of 9–12 months (not shown), in Europe the maximum correlations were found at time scales of 12–18 months. However, in some areas of the Mediterranean region, the highest correlations were recorded at time scales of 3–6 months.
4. Discussion and conclusions
We have described a new global gridded dataset of a multiscalar drought index, the standardized precipitation evapotranspiration index (SPEI), that considers the joint effects of temperature and precipitation on droughts. The dataset has some advantages regarding existing global and continental drought datasets. SPEI improves the spatial resolution (0.5°) of the unique global dataset, based on the Palmer drought severity index (UCAR PDSI; 2.5°). Moreover, the global gridded SPEI incorporates recent high-resolution gridded precipitation and temperature data (CRU TS3.0). Nevertheless, the main advantage of the new dataset lies in its multiscalar character, which allows discrimination between different types of drought.
With very few exceptions, for most regions worldwide a good fit was found between the log-logistic distribution and the precipitation–evapotranspiration Dk series, independent of the time scale k and the month of the year. This guarantees the robustness of SPEI calculations based on such probability distributions.
The SPEI was largely comparable to the PDSI, as both indices consider water inputs by precipitation and water outputs by evapotranspiration. However, comparison between the SPEI, the global PDSI, and the continental (North America and Europe) sc-PDSI datasets showed that the PDSI and the sc-PDSI both have a rigid time scale. When a drought condition is recorded with the SPEI on a particular time scale, it is possible to establish that the drought was caused by cumulative precipitation deficit and/or excessive evapotranspiration (relative to average conditions) during the previous time-scale period.
We showed that both the PDSI and the sc-PDSI were generally correlated with the SPEI at time scales of 12–18 months; thus, the PDSI can be considered as an index representing water deficits at these time scales. Then we conclude that the PDSI is not a reliable index for identifying either the shortest or the longest time-scale droughts, which can have greater effects on ecological and hydrological systems than droughts at the intermediate time scales represented by the PDSI. This suggests that the PDSI has a limited capacity to describe the effect of droughts on a range of natural systems. With respect to hydrological systems, Vicente-Serrano and López-Moreno (2005) found that the variability of river discharge in a mountainous area of northern Spain was highly correlated with drought—as described by the SPI—at very short time scales (2–3 months), whereas reservoir storages were more related to time scales of 8–12 months. Lorenzo-Lacruz et al. (2010) found that reserves in high-capacity reservoirs of central Spain were closely related to drought indices at long time scales (36–48 months). Moreover, river discharges in the headwaters of the basins were referable to time scales of 4–8 months, whereas runoff variability in the middle and lower reaches were better explained at longer time scales (up to 12 months), with particular reference to variability of flow (López-Moreno et al. 2009). In Australia, Khan et al. (2008) analyzed fluctuations in the water table level in different basins and related these to varying time scales of droughts. The cited authors found great spatial diversity in responses. Thus, in some basins, the clearest response occurred at the 6-month time scale; however, in others, the highest correlation was found at time scales of 12–24 months.
Cultivation and natural vegetation cover also vary markedly in response to drought, at different time scales. Vicente-Serrano (2007) showed that the vegetation activity in steppe areas and cereal drylands of semiarid regions of the Iberian Peninsula was closely related to short time scales (3–6 months), whereas forests responded more to longer time scales. Many studies have described variable responses of usable water sources, vegetation, and crops to droughts over various time scales (e.g., Szalai et al. 2000; Ji and Peters 2003; Patel et al. 2007; Hoffman et al. 2009; Sternberg et al. 2009). Therefore, only hydrological and economic systems that respond to water deficits at time scales of 9–18 months can be monitored using the PDSI or the sc-PDSI. For other systems not sensitive at these time scales, the PDSI is not useful for analysis of drought conditions or effects.
We also found that the time scale represented by the PDSI is not fixed at global or continental scales. The strongest correlation between the PDSI and the SPEI at different time scales varied noticeably among regions. This implies that the manner in which the PDSI represents water deficits at different time scales depends on the world region under consideration. Although this issue has not been identified as a limitation of the PDSI in previous studies, this is nonetheless an important drawback that makes the spatial comparability of droughts difficult using this index. This pattern has also been observed with the use of the sc-PDSI in the United States and Europe. Thus, in some regions, the PDSI provides information on short-term droughts, whereas in most areas the PDSI is a medium-term (9–18 months) or long-term drought index (e.g., in the central United States, West Africa, and eastern Europe).
In summary, because of the great complexity of drought effects on different sectors and natural systems, it is desirable to use an index that can be calculated over different time scales, and the SPEI fulfils this criterion. The strict probabilistic nature of this index makes it perfectly comparable across time and space; the index provides objective information on climatic drought conditions, as the index is not influenced by external variables, relying only on climate data. The index incorporates the role of water inputs (precipitation) and outputs (evapotranspiration), is able to identify climate change processes related to alterations in precipitation and/or temperature, and can be used to assess the possible influences of warming processes on droughts. The global gridded SPEI dataset described here (spatial resolution, 0.5°; period, 1901–2006; time scale, 1–48 months) is freely available in plain text, binary raster, and network Common Data Form (netCDF) formats in the Web repository of the Spanish National Research Council (CSIC; available online at http://sac.csic.es/spei/index.html).
This work has been supported by the research project Grants CGL2008-01189/BTE and CGL2006-11619/HID, financed by the Spanish Commission of Science and Technology; and FEDER, EUROGEOSS (FP7-ENV-2008-1-226487), and ACQWA (FP7-ENV-2007-1-212250), financed by the VII Framework Programme of the European Commission, “Las sequías climáticas en la cuenca del Ebro y su respuesta hidrológica,” and “La nieve en el Pirineo aragonés: Distribución espacial y su respuesta a las condiciones climática,” financed by “Obra Social La Caixa” and the Aragón government.
Corresponding author address: S. M. Vicente-Serrano, Instituto Pirenaico de Ecología, CSIC, Campus de Aula Dei, P.O. Box 202, Zaragoza 50080, Spain. Email: email@example.com