Using observation-driven simulations of global terrestrial hydrology and a cluster algorithm that searches for spatially connected regions of soil moisture, the authors identified 296 large-scale drought events (greater than 500 000 km2 and longer than 3 months) globally for 1950–2000. The drought events were subjected to a severity–area–duration (SAD) analysis to identify and characterize the most severe events for each continent and globally at various durations and spatial extents. An analysis of the variation of large-scale drought with SSTs revealed connections at interannual and possibly decadal time scales. Three metrics of large-scale drought (global average soil moisture, contiguous area in drought, and number of drought events shorter than 2 years) are shown to covary with ENSO SST anomalies. At longer time scales, the number of 12-month and longer duration droughts follows the smoothed variation in northern Pacific and Atlantic SSTs. Globally, the mid-1950s showed the highest drought activity and the mid-1970s to mid-1980s the lowest activity. This physically based and probabilistic approach confirms well-known droughts, such as the 1980s in the Sahel region of Africa, but also reveals many severe droughts (e.g., at high latitudes and early in the time period) that have received relatively little attention in the scientific and popular literature.
Drought is a naturally occurring climate phenomenon that impacts human and environmental activity globally. It is among the costliest and most widespread of natural disasters (Wilhite 2000). One of the reasons for this is the usually large spatial extent of droughts and their lengthy duration, sometimes reaching continental scales and lasting for many years. Drought is generally driven by extremes in the natural variation of climate, which are forced by the internal interactions of the atmosphere and feedbacks with the oceans and land surface (e.g., Jiang et al. 2006; McCabe and Palecki 2006). These are modulated by external forcings such as variations in solar input and atmospheric composition, either natural or anthropogenic.
The scientific body of research into the processes that cause drought to develop and persist is growing rapidly. This research has highlighted a number of factors that may potentially impact drought occurrence including large-scale atmospheric mechanisms that are associated with modes of climate variability and sea surface temperature (SST) anomalies (Schubert et al. 2004; Seager et al. 2005), and evidence that land–atmosphere feedbacks play a role in their persistence (Long et al. 2000; Wang and Eltahir 2000).
However, much of this research is based on coupled land–atmosphere–ocean models, and may be model specific. Part of the reason that the research has favored model-based approaches, and for our general lack of understanding of the mechanisms that control drought development and persistence, is the dearth of detailed observational data of the occurrence and variability of droughts over large time and space scales. Alternatively, land surface models forced by surface climate observations (which generally are more available than the relevant terrestrial hydrologic variables) can provide spatially and temporally consistent derived fields of variables that are not observed directly (Mitchell et al. 2004; Sheffield and Wood 2007; Wang et al. 2009). They can also form the basis for seasonal hydrologic prediction (Wood and Lettenmaier 2006) and thus drought forecasting (Luo and Wood 2007).
The Palmer drought severity index (PDSI) (Palmer 1965) has generally been the tool of choice for observation-based indices of drought, and has been used by Cook et al. (1999) and Dai et al. (2004b), among others, for drought reconstruction. However, the PDSI has notable deficiencies, including its inability to represent the effects of snow and the absence of a sound probabilistic interpretation for the resulting index values. An alternative to the PDSI is the use of land surface models that simulate the detailed processes of water and energy transfer at the earth’s surface (see e.g., Sheffield et al. 2004a; Andreadis et al. 2005, hereafter A05).
This study follows this approach and builds on previous work into the occurrence and trends in drought during the twentieth century over the United States (Sheffield et al. 2004a; A05; Andreadis and Lettenmaier 2006) and globally (Sheffield and Wood 2007, 2008). We analyze derived soil moisture from simulations of the terrestrial hydrologic cycle using the Variable Infiltration Capacity (VIC) model to quantify the global occurrence of drought for 1950–2000. Using the severity–area–duration (SAD) analysis methods of A05 we identify major drought events and explore their spatiotemporal characteristics and relationships with large-scale climate variability.
2. Datasets and methods
Soil moisture fields are taken from observation-driven simulations of the terrestrial hydrologic cycle using the VIC model [although the VIC model is our choice owing to its heritage as outlined above, Wang et al. (2009) show that comparable results can be achieved using alternative land surface models and, in any event, the methods described here are not specific to the VIC model]. Soil moisture is a useful indicator of drought because it provides an aggregate estimate of available water from the balance of precipitation, evaporation, and runoff fluxes. It also reflects the delays in the hydrologic system caused by infiltration, drainage, snow accumulation and melt, and the impacts of variation and anomalies in the meteorological drivers (such as changes in storm frequency and intensity). In drought terminology it most closely represents agricultural drought through its control on transpiration and thus plant growth, but reflects meteorological and hydrological drought because of the intimate links through the land-surface water balance.
We calculate an index of drought as the deficit of soil moisture relative to its seasonal climatology at a location (Sheffield et al. 2004a). This allows us to compare the occurrence of drought between different locations in a consistent and meaningful way. Daily soil moisture data output from the VIC model are averaged to a monthly time step and transformed to percentiles. A drought is then defined conceptually as a sequence of spatially contiguous deficits below the 20th percentile, which represents relatively rare conditions. This value has been used in a previous application to the United States (A05) and is in line with drought thresholds used by the U.S. Drought Monitor (National Drought Mitigation Center 2003). In semiarid and arid regions, it should be noted that absolute soil moisture values are generally low and drought can be considered to be the norm in terms of available water. The 20th percentile threshold therefore represents very dry values in these regions, and our definition of drought should be regarded as conceptual rather than physical. We summarize below the hydrologic simulation and the methods for drought identification and analysis. For more details, the reader is referred to Sheffield and Wood (2007) and A05.
a. Variable Infiltration Capacity global hydrologic simulation
The VIC model (Liang et al. 1994; Cherkauer et al. 2002) simulates the land-surface water and energy balances and has been used extensively for modeling studies at regional (Maurer et al. 2002; Su et al. 2006) to global scales (Nijssen et al. 2001; Sheffield et al. 2004b) and regional climate change impact assessment (Hamlet and Lettenmaier 1999; Hayhoe et al. 2007). The simulations were run for 1950–2000 at 1.0° resolution over global land areas except Greenland and Antarctica. They were driven by a hybrid observation–reanalysis meteorological dataset (Sheffield et al. 2006) that combines gridded observations of precipitation, temperature, and radiation with data from the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) reanalysis (Kalnay et al. 1996) to produce a high-resolution and bias-corrected forcing dataset. The simulations are described in detail in Sheffield and Wood (2007) and have been validated against available terrestrial hydrologic observations by J. Sheffield and E. F. Wood (2008, unpublished manuscript).
b. Clustering algorithm for drought identification
We identified drought events in time and space using the clustering algorithm of A05. At each time step, individual drought clusters are identified as spatially contiguous areas that are allowed to merge or break up through time. A minimum cluster area threshold of 500 000 km2 was used, which is higher than the one used in A05 (25 000 km2). Initial experiments with 25 000 and 100 000 km2 thresholds showed that droughts could shrink to a few pixels and traverse continents through tenuous spatial connectivities, thus persisting, incorrectly, for multiple years (see section 3b). The 500 000 km2 threshold ensured that events only persisted if there was a reasonable number of adjoining pixels in drought. Although such long-term (decadal) drought events have been identified in the literature and are the subject of intensive research, especially for North America [see Cook et al. (2007) for a summary of recent work], these are generally based on deviations in precipitation below the mean. In this paper we focus on contiguous regions of dry soil moisture identified using a systematic clustering method and a relatively low threshold (20th percentile of soil moisture). The longest duration drought found using this method was about 4 years, thus precluding such decadal droughts from our analysis.
c. Severity–area–duration analysis
A05 developed the SAD analysis technique, based on the widely used depth–area–duration (DAD) analysis (Grebner and Roesch 1997), to simultaneously evaluate the severity and spatial extent of droughts for different durations. Severity (S) is defined as
where P is the monthly percentile of soil moisture, summed over duration t (months). Average drought severities are calculated for successively larger areas, from a minimum of about 100 000 km2 to the maximum spatial extent of each event. The process used to calculate these average severities for each predefined duration (3, 6, 12, 18, 24, 36, and 48 months) starts with calculating and ranking the severities for each pixel that “experienced” the drought event in question. The model pixels with the maximum severity are taken as the “drought centers,” corresponding to the storm centers of the DAD analysis (WMO 1969). The block neighborhood of the center is then searched for the next largest severity and that pixel is taken with the center to form an intermediate drought area. This process is continued until the maximum spatial extent of the drought event is reached, with the severity and extent recorded in area increments of 20 model pixels (∼200 000 km2). This is repeated for all overlapping time intervals (equal to a preselected duration) of the particular event. The maximum severity for each area increment is then selected to form the SAD curve for that event, thus providing a way to estimate an absolute drought magnitude without being constrained to an individual basin or area. After calculating the SAD relationships for all identified events, we formed the envelope SAD curve that represents the maximum bound, or envelope, of severities from the set of all droughts that occurred in the region. These curves allow the identification of the most severe events for each area increment and duration.
a. Global and continental drought statistics
Using the data and methods in this study, there have been 296 droughts greater than 500 000 km2 globally from 1950 to 2000. Table 1 summarizes the occurrence of drought globally and for each continent including the longest and most spatially extensive events. The longest duration drought was 49 months (4 yr) in Asia from 1984 to 1988, closely followed by the 1950–53 North American drought (44 months). The most spatially extensive was the African drought of the early 1980s, which reached its peak extent in April 1983 when it covered over 11 million square kilometers.
Table 2 lists the top five drought events for each continent ranked by duration and maximum extent. We also show time series of drought in Fig. 1 for the world and the continents in terms of the areally averaged soil moisture percentile, area in drought, and contiguous area in drought from the cluster analysis. The mean soil moisture percentile over the full time period is 50% by construct, and the mean area in drought is 20%. The mean contiguous area in drought will be less than or equal to 20% (a value of 20% can only occur if all pixels experience drought at the same time, which at large scales is unlikely). The variability of the time series for the world and Asia are damped because of averaging over very large spatial scales. The smaller continents (Europe and Oceania) show much higher variability, also a result of climate variability in those regions. The droughts of the early 1970s and mid-1980s are clearly visible in the African time series, and the early 1990s also stands out in terms of spatial area in drought. In Europe, high variability in the 1950s is associated with multiple periods of large drought extent that are not repeated until the mid-1970s. In North America, the dry spell of the 1950s shows up clearly as a prolonged sequence of large drought extent reaching 36% of the total continental area. Other notable dry periods are around 1976–77 and the late 1990s. For Oceania, there are several events with high percentage extent in drought (e.g., 82% in 1965) and also the almost complete lack of contiguous drought in the mid-1970s. Drought extent is also very low (< 20%) from 1995 to the end of the time series, which is in contrast to the subsequent extended drought conditions between 2003 and 2007 (Australian Bureau of Meteorology 2008). South America is characterized by extended dry periods in the early 1950s, 1960s, and 1990s.
b. Severity–area–duration analysis of continental drought
SAD envelope curves that represent the maximum severity for each combination of duration (3, 12, 24, and 48 months) and spatial extent are shown in Fig. 2. Spatial snapshots of some of the identified major events are plotted in Figs. 3 and 4 for each continent. The analysis was carried out separately for each continent and the results are discussed for each continent next. Droughts that span more than one continent are therefore treated separately, which may bias the analysis toward smaller spatial scale events. This is likely only problematic for droughts that cover parts of Europe and Asia that have a substantial common land border. To see the impact of separating Europe and Asia, we ran the SAD analysis for a combined Eurasian landmass. This showed only small differences in the total area in drought, but in terms of individual drought events and their characteristics the differences were more significant. The Eurasian analysis identified 116 events compared to the combined total of 126 events in Europe (40) and Asia (86). Droughts such as the 1974–77 Asian drought were much more extensive in the Eurasian analysis as they spanned into Europe and ranked a little differently in the SAD analysis.
The example maps in Figs. 3 and 4 are identified by calculating a combined metric of severity and area as the average severity of the pixels in a drought multiplied by the drought area for a particular month. The droughts are then ranked by this metric and the top three months are plotted. Note that these events are identified based on a metric calculated for an individual month and, therefore, may not show up on the SAD curves as these are based on metrics calculated over three or more months.
1) North America
The longest droughts identified for North America are in the 1950s (Cook et al. 1999; Stahle and Cleaveland 1988; Palmer 1965; Karl and Quayle 1981), with only the 1999–2000 drought (Cook et al. 2004) of comparable duration (see Table 2). It should be noted that some of these events covered high latitudes of Canada and Alaska (although they joined temporarily with U.S. droughts during their lifetime) where freezing temperatures prolong the anomalously low soil moisture values through the winter (Sheffield and Wood 2007). The most spatially extensive droughts were the 1954–57 (central United States and much of Canada), the 1952/53 (most of the United States and southern Canada), and the 1976–77 events (northern United States and central Canada) (Cook et al. 1999; Keyantash and Dracup 2004). The three largest severity-extent months (October 1952, January 1997, and June 1988) are typified by extensive coverage over the central United States and southern Canada (see Fig. 3). The 1952–53 drought started in September 1952, quickly covered most of the central and midwestern United States by the autumn, and then migrated northward before dissipating in the Canadian prairies the following spring. The 1976–77 event that spanned the United States is interesting because it was at its peak during the winter and so had relatively little impact nationally, although regionally it was the drought of record in California and the Columbia River basin. The 1954–57 drought dominates the SAD envelope curves (Fig. 2) for all durations, with the exception of shorter-term drought, whereas the 1976–77 drought was the worst drought for areas up to about 7 million square kilometers. Note that, despite the 1988 drought (Janowiak 1988; Namias 1991; Ropelewski 1988; Trenberth et al. 1988) being considered the worst U.S. drought on record in economic terms (Trenberth and Branstator 1992), it does not appear on the envelope curve, although it does rank highly in terms of our combined metric of maximum monthly severity and extent (see Fig. 3). This is because other drought events were more severe for all combinations of duration and extent, yet their actual impacts may have been lower because of their location and/or timing. The high economic impacts of the 1988 drought were a result of its location over the northern and eastern Great Plains, regions of intensive agriculture, during the spring and summer.
2) South America
In South America the longest droughts identified occurred during 1958–59, 1982–83, and 1963–64 (14 months) and the most spatially extensive events were mainly in the 1950–60s (see Table 2). The three largest severity-extent months all occurred in the 1960s: October 1963, June 1968, and September 1961 (Fig. 4). All three droughts mostly covered the Amazon Basin, with the October 1963 drought extending northeastwards and the June 1968 event stretching down into southern Brazil and Paraguay. The SAD envelope curves (Fig. 2) indicate that the 1963/64 drought (Marengo et al. 2008) was the worst drought for durations up to 12 months, especially for larger spatial extents. This drought initiated in eastern Brazil in April 1963 and extended to its maximum extent and severity in October before moving northeast and dissipating in southern Amazonia the following spring. Droughts that occurred in 1958–59, 1991–92, and 1982–83 (Marengo et al. 1998; Foley et al. 2002) (all of which extended over Peru, Colombia, Venezuela, and northern Brazil) contribute to the curves but only for extents less than about 3 million square kilometers.
The longest droughts that were identified in Europe include events during 1959–61, 1976–77, and 1975–76. These 1970s droughts were focused in western and eastern Europe, respectively, with the former initiating in western Europe and moving into Scandinavia and the Baltic states and the latter stretching from the Baltic to the Caspian Seas. The months with highest severity-extent values (Fig. 3) are all in the 1950s (Briffa et al. 1994; Lloyd-Hughes and Saunders 2002; van der Schrier et al. 2006). The 1953–54 drought covered most of central Europe and west into France and Germany, while the two early 1950s droughts (1950 and 1951–52) reached their peak in European Russia, the former more southerly than the latter. The 1953–54 drought dominates the 3-month SAD envelope with the 1950 drought (Briffa et al. 1994) being the worst for areas greater than about 4 million km2. The 1975–76 drought is the most severe for 12-month durations except for relatively small areas for which the 1959–61 event was the most severe.
African droughts are dominated by events during the mid-1970s, 1980s (Hulme 1992; L’Hôte et al. 2002, Oba et al. 2001; Dai et al. 2004a; Tarhule and Lamb 2003), and early 1990s (Rouault and Richard 2005) (Table 2). Note that the 1982–84, 1984–85, and 1985–86 events could be considered to be the same drought in terms of their spatial and temporal proximity, but the chosen thresholds for the cluster algorithm meant that they were treated separately. In terms of spatial extent, the largest droughts were in the 1980s and 1990s. The 1982–84 drought began in southern Africa in December 1982 and joined with a drought that started in central Africa to reach its peak extent in April 1983 (Fig. 4). It then migrated northwest until it settled across the Sahel region at the end of 1983, where it persisted (as the 1984–85 and 1985–86 droughts) until September 1986. The mid-1970s (1973–74) and early 1990s (1990–91) events similarly affected the Sahel region. The subsequent droughts (1991–92 and 1994–95) affected mainly southern Africa, with the former initiating over Ethiopia and then migrating southward before stalling over southern Africa for 10 months. The latter covered southern Africa, centered over Botswana, but with much smaller spatial extent. The SAD envelope curves (Fig. 2) show that the 1982–84 drought was the most severe for large extents and short durations. The 1984–85 and 1973–74 droughts were generally the worst for smaller extents.
In Oceania, the most spatially extensive events were two of the three longest droughts (1951–52, 1961–62, 1965) (Plummer et al. 1999; White et al. 2003). The 1965 drought (Mpelasoka et al. 2007; Australian Bureau of Meteorology 2008) affected the majority of Australia (82%) at its peak (Fig. 4), sparing only the extreme north and west, while the earlier 1961 drought initiated in the interior before migrating westward. The 1957 event (Fig. 4) was very intense (high severity but only 3-month duration) and spanned the entire continent. Some of the longer-duration droughts were more regional, such as the 1977 drought, which persisted for 11 months over western Australia, and the 9-month 1982–83 drought (Nicholls 2004; Mpelasoka et al. 2007) that coincided with the centers of agriculture and population in the east and southeast. Interestingly, despite its large economic impacts, the 1982–83 drought does not appear on the SAD envelope curves (Fig. 2), which are dominated by the 1965 (3-month duration) and 1951–52 (12-month duration) events, somewhat expected from the maps of drought shown. For small areas, the 1982–83 drought is comparable to these other events (although always numerically less severe) but with increasing area, its severity decreases far more rapidly, likely due to its more localized coverage over eastern parts of Australia and shorter duration (9 months).
Droughts in Asia have been studied by many authors, often on a national basis (Gruza et al. 1999; Jiang et al. 2006; Meshcherskaya and Blazhevich 1997; Patel et al. 2007; Zhou et al. 2005) but also at regional scales (Barlow et al. 2002; Min et al. 2003). In this study, the longest drought identified (1984–88) persisted over central Siberia before migrating southeast to northern China and back again. The 1974–77 event was centered over Kazakhstan during 1974–76 and moved into the Urals region of Russia during 1976–77. This event was physically connected to the concurrent eastern European drought, but is considered separately here in this continental analysis. Other long duration droughts are the 1978–80 (central Siberia), 1981–83 (Siberia), and 1950–52 (Iran–Turkmenistan–Kazakhstan region) events. Similar to the high-latitude droughts in North America, the Siberian events are prolonged by freezing temperatures over the winter. The 1997–98 drought was the most spatially extensive (Fig. 3), covering more than 8 million square kilometers from eastern China to central Asia in October 1997, which was also the peak month in terms of severity times area. However this event only lasted 12 months. Dry conditions in central-southwest Asia during 1999–2000 are highlighted in the map for May 2000, but note that this drought actually continued beyond the period of this study (Barlow et al. 2002). The early 1950s drought (1950–52) was most severe and extensive in May 1951 when it was centered on Kazakhstan. Not surprisingly, the SAD envelope curve for 48-month durations is completely derived from the 4 yr 1984–88 Siberian drought (Fig. 2). Note how the peak severity was reached at about 2 million square kilometers and not at the smallest spatial extent as would be expected, most likely because this drought consisted of multiple subdroughts.
c. SAD curves for the top five continental events
Further insight into the severity of the individual drought events can be gathered from Fig. 5, which shows 3- and 12-month SAD curves for the top five droughts in each continent (chosen by first ranking by duration and then by maximum extent). All droughts have approximately the same severity at the lowest spatial extent, which at 3-month duration is close to the maximum possible value. This is because the spatially extensive and long duration droughts considered here are likely to contain several cells that are close to maximum severity, and these are identified by the SAD algorithm and used to construct the curves. With increasing spatial extent, and therefore increasing spatial variability in soil moisture and severity, the curves tend to diverge with, in some cases, one particular drought clearly more severe (e.g., 1963–64 in South America). Elsewhere, two or more droughts are similarly severe for a large range of spatial extents (e.g., the 1950–52 and 1984–88 Asian droughts for 3- and 12-month duration). In Africa, several droughts are of comparable severity for 3-month durations, but the 1973–74 and 1982–84 droughts clearly dominate at 12-month duration for small and large extents, respectively. The 1973–74 drought did not reach extents beyond about 11 million square kilometers, so the 1982–84 drought stands out on its own for very large spatial extents. Note the shallow gradients of the African curves, which indicate that the rate of decrease in severity with increasing area is lower in Africa than in other continents, with the exception of a small number of events in North (e.g., 1954–57) and South America (1963–64).
In Asia the 1984–88 and 1950–52 droughts are inseparable in terms of severity at 3-month durations, and only at relatively large spatial extents are any of the other droughts comparable. For 12-month duration, the 1984–88 event is dominant, but other droughts come close at low and high extents. Note that the 1997–98 and 1999–2000 events that make up much of the 3-month curve in Fig. 2 are not among the top five events in terms of duration and extent and so do not appear here. Similarly, the European 3-month curve is mostly derived from the 1953–54 and 1950 droughts, but these are not among the top five events under the definition here. At 12-month duration, the 1975–76 event is by far the worst European drought at extents larger than about 2 million square kilometers. In North America, 3-month duration severities are similar for all events at low extents. At extents beyond 3 million square kilometers, the 1976–77 and 1954–57 droughts dominate. At 12-month duration the SAD envelope curve in Fig. 2 is made up solely from data from the 1954–57 drought, and this is confirmed in Fig. 4. However, at smaller spatial extents the 1950–53 and 1976–77 droughts are comparably severe, and at very large extents the 1999–2000 drought is very close. It would be interesting to see how this drought would compare if the analysis extended beyond 2000.
The top five droughts in Oceania show similar behavior in terms of the slope of severity versus spatial extent at 3-month duration. The 1965 drought dominates the envelope curve (Fig. 2) but is only slightly more severe than other events at low spatial extents. Beyond about 4 million square kilometers it is clearly the worst 3-month duration event. For 12-month duration, the 1951–52 drought is the only one that persists for 12 months and so provides all points on the curve, although the 1977, 1965, and 1961–62 droughts lasted for 11 months. In South America, the 1963–64 drought is easily the worst at either duration, except for below 4 million square kilometers where nearly all other droughts are of similar severity. Note the 1965–66 drought that unusually shows a local severity minimum at 3–4 (1) million square kilometers for 3 (12) month duration.
d. Global SAD envelope curves
In Fig. 6, we merge the SAD envelope curves for the world and identify the most severe droughts globally over the second half of the twentieth century. The droughts are listed according to the number of points that they contribute to the SAD envelope curves and are dominated by events in Asia and North America. The 1974–77 Asian drought that covered much of Kazakhstan and western Russia contributes to most of the 12- and 24-month curves. The rest of the 24-month curve comes from the 1954–57 and 1950–53 North American droughts. The 1982–84 African drought contributes to the remainder of the 12-month curve for large spatial extents and the 1978–80 Asian drought completes the curve for extents less than about 2 million km2. The 3-month curve is made up of contributions from seven different events that do not appear on the 12- and 24-month curves plus the 1982–84 African drought, the most notable of which are the 1999–2000 Asian (central and southwest) and the 1982–84 African events. Only one drought from each of Europe (1953–54), South America (1963–64), and Oceania (1951–52) contribute to the curves and for only a limited number of points.
4. Global variability in large-scale drought and teleconnections
Much work has been carried out on the covariability of climate with terrestrial states and fluxes and their extremes, including drought. Rajagopalan et al. (2000) found relatively robust relationships in the southwestern United States between PDSI and several climate indices, although these varied in time. McCabe et al. (2004) looked at the association of the Pacific decadal oscillation (PDO) and Atlantic multidecadal oscillation (AMO) with decadal variability in North American PDSI. Globally, Dai et al. (2004b) found the PDSI to be coupled to ENSO variability. Jiang et al. (2006) looked at proxy indicators for the Yangtze River for 1868–2003 and found it to be related to ENSO. McCabe and Palecki (2006) explored covariability between PDSI and global SSTs on decadal to multidecadal time scales and revealed strong relationships and a dominance of the PDO and AMO in explaining most of the variability. Sheffield and Wood (2008) explored the variability of the VIC soil moisture dataset used here and showed it was mainly related to ENSO with the NAO and the AMO possibly playing secondary roles. We extend this to look specifically at the occurrence of major drought events and large-scale climatic variability.
a. Variability of global large-scale drought occurrence
Figure 7 shows the monthly time series of the number of large-scale global droughts as defined by the cluster analysis. The number of droughts is counted separately in each month such that a single drought event will contribute to several months within the time series (the variability of droughts of different durations is explored in the next section). Also shown is the number of droughts filtered by their maximum spatial extent, where the data for each line is calculated for incremental thresholds of spatial extent, to see the how the temporal variation changes for increasingly larger drought events. The mean number of global droughts > 500 000 km2 occurring in any month is about 4.5 (or 55 yr−1) with a standard deviation of 1.6. This time series is quite variable and indicates several periods of increased global drought activity: the mid-1950s, 1960s, late 1980s to early 1990s, and late 1990s. The mid-1970s to mid-1980s are characterized by the lowest number of droughts, apart from a short burst of activity around 1976–77. The year with most drought months is 1992 (74). Table 3 shows the top and bottom five ranked years in terms of the number of months with drought extent > 500 000 km2.
When filtered by spatial extent, the number of droughts between 500 000 and 1 000 000 km2 shows similar magnitude and variability. For droughts greater than 5 000 000 km2 there is a large reduction in the number of events and a change in the temporal variation. In fact, for these very large droughts, the series shows a decadal variation with 5–10 years of activity interspersed with a few years of no drought events. The drop in the number of droughts greater than 1 000 000 and 5 000 000 km2 is an interesting result that likely reflects the geographic or climatic factors that control the size of drought events.
b. Links with large-scale climate variability
Given the influence of ENSO on global climate and soil moisture, it is tempting to compare the time series of global drought numbers with an index of ENSO. In Fig. 8 we compare the number of droughts with the time series of the Niño-3.4 SSTs (an indicator of ENSO variability). As ENSO episodes (warm El Niño or cold La Niña) and their impacts persist on time scales of several months to a year or two, the drought time series is calculated here for only those drought events of 24-month duration or less (Fig. 8, top panel). Correlation is weak (r = 0.36 for the 13-month filtered data) but the year-to-year variation is somewhat similar, excepting around the early 1950s, 1979, and 1988 when the two time series diverge briefly. This similarity is consistent with reported covariabilities between ENSO and soil moisture indices (modeled soil moisture or PDSI) mentioned before, although here the relationship with the number of contiguous large-scale dry extremes of soil moisture is somewhat less distinct. At smaller scales, the correlations are likely to be stronger in regions known to be influenced by ENSO, such as Oceania and Southeast Asia, as suggested by the high correlations found by Sheffield and Wood (2008) between ENSO and soil moisture in these regions.
Links between ENSO and large-scale drought may also be revealed using a nonlinear, categorical-type relationship rather than a linear correlative one. Figure 9 shows composites of three characteristics of drought (soil moisture, contiguous area in drought, and the number of droughts) over a 36-month window that straddles the year of an El Niño or La Niña event, helping to show the potentially lagged response of the land surface. The soil moisture data are averaged over all terrestrial land regions of the simulation. The contiguous area is derived from the cluster analysis and was shown previously as a time series in Fig. 1. The number of droughts is again filtered for droughts of duration less than 24 months. The three characteristics are first standardized to z scores and then averaged for each month in the 36-month window over all El Niño or La Niña years. ENSO years are taken from the Climate Prediction Center (CPC) oceanic Niño index (ONI), which is the 3-month running mean of extended reconstructed SST version 3 (ERSST.v3) dataset SST anomalies over the Niño-3.4 region (5°N–5°S, 120°–170°W). The figure shows results for two sets of ENSO years: 1) all ENSO years, defined as years with five consecutive overlapping anomalies greater than ±0.5°C and 2) strong ENSO years, similarly defined but for ±1.0°C. Statistical significance of the composite index was determined using the hypergeometric distribution by representing the index’s departure from zero as a binomial variable with values “success” and “failure” (Dracup and Kahya 1994). A “success” is defined as the occurrence of an index value above (below) zero. The hypergeometric distribution was then used to determine the cumulative probability that at least m successes (the number of successes occurring in La Niña or El Niño years) are obtained in n trials (the number of La Niña or El Niño years) from a finite population of size N (the total number of years) containing k successes (the total number of successes). A cumulative probability value less than 0.05 (95% confidence level) indicates that the index value is significant.
There is a clear tendency for all drought characteristics to be more pronounced in the second half of the ENSO year and the beginning of the following year. During El Niño (La Niña) years, soil moisture is more likely to be drier (wetter) globally with a contemporaneous increase (decrease) in the number of large-scale drought events. The signal is weaker for all ENSO years (the composite index stays below 0.5) but is much more robust for strong ENSO years. The response for contiguous area is not as strong, especially for all ENSO years, with the La Niña signal being particularly weak. The peak composite index for all characteristics falls around the end of the ENSO year (September–November), and the anomalies persist to about May of the following year with only the larger El Niño index values being statistically significant. This is consistent with the known impacts of ENSO on global terrestrial climate and highlights that large-scale drought occurrence in particular is impacted and that El Niño events may act to increase global drought. The soil moisture response is much shorter because the global average value reflects changes in all parts of the world, including those that are essentially unaffected by ENSO variability, and may thus act to cancel out the signal derived from ENSO-affected regions.
The bottom panel of Fig. 8 shows the time series of long-term drought events (duration > 12 months; the number of droughts longer than 24 months were too few to discern their temporal variation). We compare this with the AMO and PDO, which are the main modes of decadal climate variability in the North Atlantic and Pacific, respectively. All time series in Fig. 8 (lower panel) are smoothed using a 121-month (10 yr) running mean filter. Correlations of the smoothed and filtered number of droughts with the smoothed climate indices are 0.62 for the AMO and 0.57 for the PDO. As the smoothed time series are serially correlated, we determined the statistical significance of the correlations using Monte Carlo simulation (McCabe and Palecki 2006). A set of 10 000 pairs of time series (one for the number of droughts and one for the climate oscillation, AMO or PDO) were generated that possessed the mean, variability, and lag-1 correlation of the original data. The distribution of the correlations of each smoothed pair was then used to determine whether the observed correlation was statistically significant for a two-tailed test at the 95% level. The results show that the neither correlation is significant at this level, although the AMO correlation is significant at the 90% level.
5. Discussion and conclusions
a. Some caveats
Before summarizing the main results, we give some caveats concerning the simulated soil moisture data, including the forcing data, and the assumptions that impact their accuracy. We also highlight some of the uncertainties introduced by the clustering methodologies and the associated chosen thresholds. First, the soil moisture fields are taken from a single hydrologic model. Although we consider this to be currently our best estimate of soil moisture variability globally, the results may change if we used a different hydrologic model. Comparisons of drought for the United States for four land surface models forced by the same meteorology (Wang et al. 2009) show that, although major droughts are identified by all models, they can develop differently in terms of severity. This is, in part, due to intermodel differences in soil depth and water-holding capacities that impact soil moisture retention and persistence. There are also uncertainties and intermodel differences in the representation of vegetative controls on soil moisture.
Second, there are uncertainties in the meteorological forcings that were used to drive the simulation. The forcings are derived from observationally based gridded datasets with daily variability from reanalysis, as described in Sheffield et al. (2006). Although this is arguably the best estimate we have currently of continuous and consistent fields of meteorological forcings, there are nevertheless unknown biases and spurious trends in the data, caused by instrument errors and/or deficiencies in the methodologies, and these will impact the depiction of drought. This is especially likely in regions of sparse instrumentation, such as Africa, where, for example, the lack of rain gauges may have a large impact on the accuracy of simulated drought development.
Third, as mentioned in Sheffield and Wood (2007, 2008), the simulation makes a number of assumptions about land use change and anthropogenic influences. Specifically, we assume that land cover and land use does not change, although it varies spatially and seasonally, and that irrigation and water withdrawals do not play a major role in large-scale soil moisture variability and, thus, drought development.
Finally, the focus of the paper is on the identification and analysis of large-scale contiguous drought events that may give different results to, say, an areally averaged metric of drought, in terms of the number of events or the ranking of wet and dry years (this is noted for Europe in section 5b below). This is also dependent on the specific clustering method and related thresholds we have employed. It should be noted that the clustering algorithm does allow for droughts to traverse continents, either as a single cluster or from the merging of two or more subdroughts (such as the 1982–86 African drought). As subdroughts that are well separated in space are likely driven by different forcing mechanisms (e.g., SST patterns), this makes it somewhat difficult to interpret them when they are considered as a single event. The fact that the movement of the drought clusters is smooth and continuous does, however, hint at some common underlying mechanisms.
b. Consistency with other studies
Many other studies have analyzed large-scale drought, either as individual events or from a regional climatological perspective, and their impacts are generally well documented. In the following we qualitatively compare our results with some of these studies to gain insight into their robustness and consistency.
This study follows on from the continental U.S. analysis of A05, and we begin by looking at how well the results of that study are replicated here, given the obvious similarities but also the differences in their approaches. Both studies use VIC simulated soil moisture and the same cluster analysis for identifying droughts and their SAD characteristics. However, they differ in a several important ways, such as the spatial scale (0.5° continental United States versus 1.0° global for A05 and this study respectively), the temporal period (1920–2003 versus 1950–2000), the meteorological forcings (gauge-based precipitation and temperature versus bias-corrected reanalysis) and the minimum cluster threshold.
Excluding the 1930s Dust Bowl event, which we do not simulate, A05 found the 1950s drought (1950–57) to be the most severe for large areas and long durations. Although our chosen cluster threshold splits this drought into two events (1950–53 and 1954–57), we also find that these are the most severe for large extents and durations. Both studies identify the 1976–77 drought that covered much of the United States and southern Canada (1975–79 in A05) as one of the most severe for short durations. A05 also find that the 1998–2003 event is less severe but comparable to the 1950s (and 1930s) droughts, especially for smaller spatial extents. We also find this to be true for the equivalent 1999–2000 drought identified here, although the end of the simulation in 2000 hinders proper assessment of this event. For both studies, the important 1988 drought (1988–89 in this paper, 1987–93 in A05) is considered a severe event but does not contribute to the SAD envelope curves.
For Europe, Lloyd-Hughes and Saunders (2002) and van der Schrier et al. (2006) used PDSI and 3- and 12-month standardized precipitation index (SPI) and found the 1950s and 1990s to be the most drought prone. This is certainly reflected for the 1950s in the time series (Fig. 1) and the SAD envelope curves (Fig. 2) in which four out of the five drought events that form the European curves are from this period. Furthermore, the three events with the highest monthly severity-extent index are from the 1950s (Fig. 3). Although the 1995–96 drought ranks highly in terms of severity at 3-month duration, our analysis shows that the 1990s is actually the decade with the least events (6), with the 1980s having the most (11). At the monthly time scale, these two previous studies use essentially the same precipitation and temperature as the present study, yet differences in the modeling and analysis approaches can give quite different results in terms of identifying periods of drought. This is especially likely when comparing spatially averaged metrics [in the case of Lloyd-Hughes and Saunders (2002) and van der Schrier et al. (2006)] and spatially contiguous metrics as used here.
There has been considerable study of drought conditions in the Sahel (e.g., Hulme 1992; L’Hôte et al. 2002) because of the devastating impacts, and it is no surprise that the SAD curves are dominated by events that primarily affected this region (1973–74, 1982–84, and 1984–85). Elsewhere in Africa, Roualt and Richard (2005) identify the 1991–92 drought as the most spatially extensive event in southern Africa in terms of 6-month SPI, which is identified here as the most severe drought outside of the Sahel-centered events.
For Asia, the identified high-latitude droughts (e.g., 1984–88, 1978–80, 1981–83) are not well documented in the literature, given the sparse population and low economic activity in these regions, although indirect evidence from, for example, remotely sensed fire activity across the boreal forest (e.g., Balzter et al. 2007) is consistent with large-scale dry conditions during the 1980s. We found the drought events in central/southwest Asia (1950–52 and 1999–2000) to be very severe, especially at 3- and 12-month durations, and the latter is consistent with Barlow et al. (2002), who show this period to be unusual in terms of precipitation. The 1974–77 central Asian/European Russia drought was the most severe globally for 12- and 24-month durations at moderate extents (see Fig. 6). Meshcherskaya and Blazhevich (1997), in their assessment of dry and wet years in the Former Soviet Union (FSU), identified 1976 as the second driest year in the European part of the FSU (1981 was the driest), although this was in terms of the spatial extent of low values of a spring–summer moisture index. The 1997–98 (from eastern China to central Asia) drought is also reported by Zhou et al. (2005), who looked at PDSI over China from 1951 to 2003 and this event is typical of the reported increase in drought in northern China and its persistence since 1997.
In Oceania (Australia), the 1982–83 drought is considered to be one of the most severe and damaging events of the second half of the twentieth century mainly because it was intense (short and severe) causing over 3 billion Australian dollars ($AUS) in damages (Australian Bureau of Meteorology 2008). It does not appear on the 3-month SAD curve, however, and is only sixth on the list of most spatially extensive and lengthy events and so does not appear in Fig. 5, which is limited to the top five events. However, its severity is comparable to these events for 3-month durations, especially for relatively small spatial extents (not shown). The Australian Bureau of Meteorology also highlights the 1965 drought as particularly severe (Australian Bureau of Meteorology 2008) with an estimated 300–500 million $AUS in agricultural losses (Mpelasoka et al. 2007), and this forms most of the 3-month SAD envelope curve.
There are several references to drought in the Amazon in the literature, particularly those related to ENSO variability (e.g., the 1982–83 drought identified here as one of the longest events; Marengo et al. 1998). However, the drought of 1963–64 has received little attention (Marengo et al. 2008) despite being identified here as having the largest peak extent and the most severe up to 12-month duration.
c. Summary and conclusions
We have described the spatiotemporal characteristics of large-scale drought events globally over the second half of the twentieth century. Drought is defined in terms of anomalies of soil moisture derived from an observationally and reanalysis-driven simulation of the terrestrial water cycle. Cluster analysis is used to identify large-scale drought events (> 500 000 km2) as spatially connected regions of soil moisture below the 20th percentile. We have analyzed these events in terms of their severity, area, and duration and identified the severest droughts for each continent and globally for different spatial extents and durations.
Using this dataset and methodology, there have been 296 major droughts (> 500 000 km2 and lasting longer than 3 months) globally from 1950–2000. Half of these events lasted for 6 months or less and half were smaller than 3 million square kilometers. The longest drought event lasted for 49 months from 1984 to 1988 in Siberia. The most spatially extensive was in Africa during 1982–84, which covered over 11 million square kilometers in April 1983. On average there were about 4.5 unique droughts occurring globally in any month (or 55 yr−1), and this varies considerably from year to year. The mid-1950s showed the highest drought activity and the mid-1970s to mid-1980s the lowest activity. The number of very large droughts (> 5 000 000 km2) is significantly lower than the total number of droughts, partly a result of the size and shape of continental landmasses that limit the maximum size of a drought, but may also be related to the length scales of climate anomalies as driven by planetary waves (Herweijer and Seager 2008).
The global number of large-scale droughts of short duration (24 months or less) shows a weak relationship with ENSO variability in terms of correlations and composite analysis. There is a tendency for more (less) short-term drought events during El Niño (La Niña) episodes. The number of droughts tends to maximize or minimize (depending on the ENSO phase) toward the end of the ENSO year with considerably stronger composite values for stronger ENSO years as defined by larger SST anomalies. The smoothed number of longer-duration droughts (> 1 yr) varies on decadal time scales somewhat in line with the smoothed variation in the AMO and PDO, although only the correlation with the AMO is significant at the 90% level. Many studies (e.g., Kitzberger et al. 2007) have noted the regional influence of decadal variations in the northern Pacific and Atlantic Oceans, as characterized by the PDO and AMO. Other studies have shown that different combinations of modes of the PDO and AMO (along with ENSO) are associated with multiple regions (McCabe et al. 2004; Kitzberger et al. 2007) and may have a wide impact across the Northern Hemisphere (e.g., Sutton and Hodson 2005). Our results reflect this, with the majority (∼85%) of the identified long-term drought events occurring in the northern mid to high latitudes, indicating that these oceanic oscillations may have an appreciable influence on global drought occurrence.
In general, the identified droughts are well known and consistent with those reported in the scientific and popular literature. Notable exceptions are the less-well-known high latitude droughts in Siberia and Canada/Alaska whose long durations (up to 4 yr) are, nevertheless, likely to have impacted the local (albeit sparse) populations and fragile ecosystems (Hinzman et al. 2005). Other droughts are ranked highly in terms of severity and spatial extent yet are not well documented or analyzed, such as the 1965 Australian and 1963–64 South American droughts. As well as documenting these events, this study also provides incentive for further study into the occurrence and mechanisms of large-scale drought events, especially in sparsely populated and undermonitored regions. This further highlights the advantages of taking a physically based approach that considers the dynamics of the large-scale terrestrial water cycle in a consistent manner both spatially and temporally and therefore can identify and characterize droughts in a global context. The probabilistic basis of the soil-moisture-based drought index allows for intercomparison of drought events across regions and through time. It is therefore well suited to evaluation not only of changes in the preinstrumental relative to the observational period but also to evaluation of the implications of future climate for drought.
This work was supported by the U.S. Department of Energy (Grant DE-FG02-04ER63873 to the University of Washington and NA08OAR4310579 to Princeton University) and by the National Oceanic and Atmospheric Administration (Grant NA07OAR4310458 to the University of Washington). We thank the three anonymous reviewers for their helpful comments.
Corresponding author address: Justin Sheffield, Department of Civil and Environmental Engineering, Princeton University, Princeton, NJ 08544. Email: email@example.com