![FIG. 3. Scatterplots of monthly anomalies of PDSI vs observed soil moisture content within top 1-m [except (c), which is for top 0.5 m] depth within China during 1981–91. Here r is the correlation coefficient of all the data points. (a) Northeast China (14 soil moisture stations), (b) east China (7 stations), (c) north-central China (5 stations), and (d) south China (3 stations). Legend: symbols 1, C, 3, and ● are, respectively, for months 6, 7–8, 9, and 10 in (a), 3–4, 5–6, 7–9, and 10–12 in (b), and 5, 6, 7, and 8 in (c); symbols 1, C, 3, *, and ● are, respectively, for months 1–3, 4–5, 6–8, 9–10, and 11–12 in (d).](/figures/fig-3-scatterplots-of-monthly-anomalies-of-pdsi-vs-observed-2epwm95l.webp)
![FIG. 8. (top) The 1998–2002 mean difference of the PDSI calculated with and without temperature changes. (bottom) Linear trends of observed surface air temperature [8C (50 yr)21] during 1950–2002. Dashed contours have negative values.](/figures/fig-8-top-the-1998-2002-mean-difference-of-the-pdsi-mqey1vvj.webp)
![FIG. 7. Maps of linear trends of PDSI [change (50 yr)21, calculated with both precipitation and temperature changes] during (top) 1900–49 and (middle) 1950–2002. (bottom) The trends of PDSI calculated without temperature changes. Red (blue) areas indicate drying (wetting).](/figures/fig-7-maps-of-linear-trends-of-pdsi-change-50-yr-21-4x2aj46j.webp)
D
ECEMBER
2004 1117DAI ET AL.
q 2004 American Meteorological Society
A Global Dataset of Palmer Drought Severity Index for 1870–2002: Relationship with
Soil Moisture and Effects of Surface Warming
A
IGUO
D
AI
,K
EVIN
E. T
RENBERTH
,
AND
T
AOTAO
Q
IAN
National Center for Atmospheric Research,* Boulder, Colorado
(Manuscript received 24 February 2004, in final form 26 May 2004)
ABSTRACT
A monthly dataset of Palmer Drought Severity Index (PDSI) from 1870 to 2002 is derived using historical
precipitation and temperature data for global land areas on a 2.58 grid. Over Illinois, Mongolia, and parts of
China and the former Soviet Union, where soil moisture data are available, the PDSI is significantly correlated
(r 5 0.5 to 0.7) with observed soil moisture content within the top 1-m depth during warm-season months. The
strongest correlation is in late summer and autumn, and the weakest correlation is in spring, when snowmelt
plays an important role. Basin-averaged annual PDSI covary closely (r 5 0.6 to 0.8) with streamflow for seven
of world’s largest rivers and several smaller rivers examined. The results suggest that the PDSI is a good proxy
of both surface moisture conditions and streamflow. An empirical orthogonal function (EOF) analysis of the
PDSI reveals a fairly linear trend resulting from trends in precipitation and surface temperature and an El Nin˜o–
Southern Oscillation (ENSO)-induced mode of mostly interannual variations as the two leading patterns. The
global very dry areas, defined as PDSI ,23.0, have more than doubled since the 1970s, with a large jump in
the early 1980s due to an ENSO-induced precipitation decrease and a subsequent expansion primarily due to
surface warming, while global very wet areas (PDSI .13.0) declined slightly during the 1980s. Together, the
global land areas in either very dry or very wet conditions have increased from ;20% to 38% since 1972, with
surface warming as the primary cause after the mid-1980s. These results provide observational evidence for the
increasing risk of droughts as anthropogenic global warming progresses and produces both increased temperatures
and increased drying.
1. Introduction
Droughts and floods are extreme climate events that
percentage-wise are likely to change more rapidly than
the mean climate (Trenberth et al. 2003). Because they
are among the world’s costliest natural disasters and
affect a very large number of people each year (Wilhite
2000), it is important to monitor them and understand
and perhaps predict their variability. The potential for
large increases in these extreme climate events under
global warming is of particular concern (Trenberth et
al. 2004). However, the precise quantification of
droughts and wet spells is difficult because there are
many different definitions for these extreme events (e.g.,
meteorological, hydrological, and agricultural droughts;
see Wilhite 2000 and Keyantash and Dracup 2002) and
the criteria for determining the start and end of a drought
or wet spell also vary. Furthermore, historical records
of direct measurements of the dryness and wetness of
* The National Center for Atmospheric Research is sponsored by
the National Science Foundation.
Corresponding author address: A. Dai, National Center for At-
mospheric Research, P.O. Box 3000, Boulder, CO 80307-3000.
E-mail: adai@ucar.edu
the ground, such as soil moisture content (Robock et al.
2000), are sparse. In order to monitor droughts and wet
spells and to study their variability, numerous special-
ized indices have been devised using readily available
data such as precipitation and temperature (Heim 2000;
Keyantash and Dracup 2002).
The Palmer Drought Severity Index (PDSI) is the
most prominent index of meteorological drought used
in the United States (Heim 2002). The PDSI was created
by Palmer (1965) with the intent to measure the cu-
mulative departure (relative to local mean conditions)
in atmospheric moisture supply and demand at the sur-
face. It incorporates antecedent precipitation, moisture
supply, and moisture demand [based on the classic work
of Thornthwaite (1948)] into a hydrological accounting
system. Palmer used a two-layer bucket-type model for
soil moisture computations and made certain assump-
tions relating to field water-holding capacity and transfer
of moisture to and from the layers based on limited data
from the central United States (Palmer 1965; Heim
2002). The Palmer model also computes, as an inter-
mediate term in the computation of the PDSI, the Palmer
moisture anomaly index (Z index), which is a measure
of surface moisture anomaly for the current month with-
out the consideration of the antecedent conditions that
1118 V
OLUME
5JOURNAL OF HYDROMETEOROLOGY
characterize the PDSI. The Z index can track agricultural
drought, as it responds quickly to changes in soil mois-
ture (Karl 1986). The Z index relates to the PDSI
through the following equation (Palmer 1965, p. 22):
PDSI (m) 5 PDSI {m 2 1 1 [Z(m)/3 2 0.103 PDSI
(m 2 1)]}, where m is a month index. A detailed de-
scription of the Palmer model is given in Palmer (1965),
Alley (1984), and Karl (1986).
Theoretically, the PDSI is a standardized measure,
ranging from about 210 (dry) to 110 (wet), of surface
moisture conditions that allows comparisons across re-
gions and time. Guttman et al. (1992) found, however,
that the normal climate conditions tend to yield more
severe PDSI in the Great Plains than in other U.S. re-
gions. The PDSI is also imprecise in its treatment of all
precipitation as immediately available rainfall (i.e., no
snow accumulation on the ground), the effects of veg-
etation on surface evapotranspiration, the calibrationco-
efficients (Karl 1986), and some other processes (Alley
1984). For example, Palmer assumed that evapotrans-
piration occurs at the potential rate (i.e., according to
Thornthwaite 1948) from the top soil layer until all the
available moisture in this layer has been depleted. Only
then can moisture be removed from the underlying layer
of soil. Although these assumptions are not unreason-
able, they are only crude approximations even for bare
soil surfaces (Philip 1957). Other factors such as chang-
es in surface solar radiation due to changes in cloudiness
or aerosol concentrations (Abakumova et al. 1996; Lie-
pert 2002) are not considered explicitly in the Palmer
model, although the effect of solar radiation is implicitly
considered through surface temperature. Also, the PDSI
cannot reflect soil moisture conditions when the soil is
frozen, or when snow accumulation and melt are a sig-
nificant factor, such as during the winter and spring
months at mid- and high latitudes. Nevertheless, the
PDSI is still an approximate measure of the cumulative
effect of atmospheric moisture supply and demand (i.e.,
meteorological droughts) in these situations. We em-
phasize that, by design, the PDSI is not always a good
measure of soil moisture and thus agricultural droughts,
although we show that the PDSI does correlate with soil
moisture content during warm seasons.
On the positive side, the PDSI uses both precipitation
and surface air temperature as input, in contrast to many
other drought indices that are based on precipitation
alone (Keyantash and Dracup 2002). This allows the
PDSI to account for the basic effect of surface warming,
as has occurred during the twentieth century, on
droughts and wet spells. The effect of surface temper-
ature, which accounts for 10%–30% of PDSI’s variance,
comes through the potential evapotranspiration, which
was computed from Thornthwaite’s (1948) formula in
the Palmer model and used as a measure of the atmo-
spheric demand for moisture. As precipitation and sur-
face air temperature are the only two climate variables
with long historical records, the PDSI makes full use
of these data and can be readily calculated for the last
hundred years or so for most land areas.
Besides its routine use for monitoring droughts in the
United States, the PDSI has been used to study drought
climatology and variability in the United States (e.g.,
Karl and Koscielny 1982; Karl 1986), Europe (Domon-
kos et al. 2001; Lloyd-Hughes and Saunders 2002), Af-
rica (Ntale and Gan 2003), Brazil (dos Santos and Pereira
1999), and other areas. The PDSI was also used in tree-
ring-based reconstructions of droughts in the United
States (e.g., Cole and Cook 1998; Cook et al. 1999; Fye
et al. 2003). Most of these studies are regional and focus
on a particular location or nation. One exception is Dai
et al. (1998) who calculated the PDSI for global land
areas for 1900–95 and analyzed the influence of El
Nin˜o–Southern Oscillation (ENSO) on dry and wet ar-
eas around the globe. This study updates the global
PDSI dataset of Dai et al. (1998), provides a detailed
evaluation of the PDSI against available soil moisture
and streamflow data, examines the trends and leading
modes of variability in the twentieth-century PDSI
fields, and investigates the impact of surface warming
of the latter half of the twentieth century on global
drought and wet areas. The global PDSI dataset (avail-
able from http://www.cgd.ucar.edu/cas/catalog/climind/
pdsi.html) has been used by a number of groups and
will be updated periodically in the future. We emphasize
that the PDSI is better used on annual time scales and
should not be used as a measure of soil moisture content
during cold seasons at high latitudes. In addition, quan-
titative interpretations of dryness or wetness for a given
PDSI value depend on local mean climate conditions.
For example, a PDSI value of 14 may imply floods in
the central United States, but only moderate rainfall (by
central U.S. standards) in northern Africa.
2. Datasets and procedures
Table 1 lists the datasets used in this study. To cal-
culate the monthly PDSI, we used the Climate Research
Unit (CRU) surface air temperature data (Jones and
Moberg 2003; regridded to 2.5832.58 grid). Precipi-
tation data for 1948–2003 were obtained from the Na-
tional Centers for Environmental Prediction (NCEP)
Climate Prediction Center (Chen et al. 2002) and were
created by gridding data from about 5000 to 16 500
gauges during 1948–97 and about 3500 gauges for more
recent years using the optimal interpolation scheme; for
the pre-1948 period, we used the precipitation data from
Dai et al. (1997). The monthly anomalies of Dai et al.
(1997) were adjusted to have zero mean values for
1950–79 and then added to the 1950–79 mean of Chen
et al. (2002) to obtain the total precipitation used for
the PDSI calculation. For field water-holding capacity
(awc), we used a soil texture–based water-holding-ca-
pacity map from Webb et al. (1993). If awc is no more
than 2.54 cm (or 1 in.), then awc is assigned to the top
soil layer, and the bottom layer has zero capacity, oth-
D
ECEMBER
2004 1119DAI ET AL.
T
ABLE
1. Datasets used in this study. All are monthly.
Variables Type and coverage Resolution Period Source and reference
P Rain gauge, land 2.5832.58 1850–2003 Dai et al. (1997); Chen et al. (2002)
T Surface obs, land 58358 1851–2003 CRUTEM2; Jones and Moberg (2003)
Streamflow Station, land 1–1001 yr NCAR; Dai and Trenberth 2002
Soil moisture Station, land
Illinois
China
Mongolia
Former USSR
19 stations
43 stations
42 stations
50 stations
10–21 yr
1981–2001
1981–91
1978–93
1972–85
Robock et al. (2000)
Hollinger and Isard (1994)
Robock et al. (2000)
Robock et al. (2000)
Vinnikov and Yeserkepova (1991)
Soil water-holding capacity Derived, land 18318 climatology Webb et al. (1993)
erwise, the top layer has 2.54-cm water-holding capacity
while the bottom layer has (awc 2 2.54)-cm capacity.
The temperature and precipitation data likely contain
some errors; however, various data quality controls were
done to minimize data inhomogeneities by previous
analyses (e.g., Jones and Moberg 2003; Dai et al. 1997;
Chen et al. 2002), and the time series are thought to be
reliable over most land areas. The relatively low reso-
lution (2.5832.58) used here does not resolve small-
scale variations such as those over mountains.
The PDSI has often been used without rigorous eval-
uation as it is not directly comparable to any measured
variables, such as soil moisture content and streamflow.
Soil moisture content reflects the amount of precipita-
tion retained locally after runoff. This depends on many
factors, including field water-holding capacity (a func-
tion of soil texture and depth), antecedent soil condi-
tions, and precipitation frequency and intensity (Tren-
berth et al. 2003). For example, a severe thunderstorm
can create a lot of surface runoff or even flash floods,
but may leave the subsurface soil still dry. In contrast,
hours of light, stratiform rain can moisten the soil thor-
oughly with little runoff. Furthermore, surface and sub-
surface runoff may take weeks to months to reach the
downstream part of the world’s major rivers. These sit-
uations illustrate the complex relationships among local
atmospheric moisture supply and demand (based on
which the PDSI is computed), soil moisture, and runoff
or streamflow. Nevertheless, on regional and river-basin
scales and averaged annually, the PDSI, soil moisture,
and streamflow should correlate with each other, asthese
all are measures of large-scale droughts and wet spells
that are driven by regional atmospheric moisture supply
(i.e., precipitation) and demand (i.e., evapotranspira-
tion). Dai et al. (1998) showed that area-averaged PDSI
is significantly correlated (r 5 0.63 2 0.75) with stream-
flow of the twentieth century over the United States,
midlatitude Canada, Europe, and southeast Australia.
Here, we compare the PDSI with available soil moisture
data obtained from the Global Soil Moisture Data Bank
(http://climate.envsci.rutgers.edu/soilpmoisture/) (Ro-
bock et al. 2000) over Illinois, Mongolia, and parts of
China and the former Soviet Union (fUSSR). The soil
moisture stations [see Robock et al. (2000) for their
locations] were grouped into regions defined by latitudes
and longitudes to facilitate comparisons with the gridded
PDSI values. Nearby stations were averaged first and
then combined with other stations to obtain the arith-
metical mean for the region. We also integrate the PDSI
over a number of large river basins for comparison with
observed river flow (from Dai and Trenberth 2002) dur-
ing the twentieth century. In this comparison, we focus
on the annual time series, ignoring the time lag between
local runoff and downstream river flows. Significant
correlations between the PDSI and the other drought
measures should provide further support of the useful-
ness of the PDSI dataset.
3. PDSI versus soil moisture
Figure 1 compares the observed and Palmer model–
calculated soil moisture content for Illinois. Forthe sum-
mer half year, the Palmer model captures both the mean
seasonal and year-to-year variations in the top 0.9-m
depth very well with only small biases. This is remark-
able considering that the Palmer model is very simple
compared to modern land surface models (e.g., Dai et
al. 2003) and was driven by only monthly temperature
and precipitation. The largest bias, which does not affect
correlation, is in September when the calculated mean
soil moisture is lower and interannual anomalies are
larger than observed. Correlations between the calcu-
lated total moisture content and the observed moisture
content at each depth are strongest and significant up
to ;1.25 m depth in September, insignificant during
March–April, and significant up to only ;0.5 m depth
during May–June (Fig. 1c). The low correlation during
spring is not surprising as snowmelt is not considered
in the simple Palmer model but has large effects on soil
moisture in Illinois. The snow effects gradually diminish
during May–June, and soil moisture is increasingly af-
fected by monthly rainfall so that the Palmer model
performs better. Furthermore, Fig. 1b suggests that soil
moisture anomalies in the Palmer model and the data
have different lower bounds.
Significant correlations are also found between the
observed soil moisture content and the PDSI for Illinois
(Fig. 2). The interannual correlation (Fig. 2a) is reduced
considerably by the three lower-right points, and the
correlation at depths is lower than that shown in Fig.
1c except for May when the PDSI actually correlates
strongest with the observed soil moisture. This May
1120 V
OLUME
5JOURNAL OF HYDROMETEOROLOGY
F
IG
. 1. (a) Mean annual cycle of Palmer model–calculated (solid
line) and observed (dots, for top 0.9-m depth, 17 stations) soil mois-
ture content for Illinois. (b) Scatterplot of monthly anomalies of soil
moisture from the Palmer model and observations for Illinois during
1981–2001. Here r is the correlation coefficient of all the data points.
Legend: symbols 1, C, 3, *, and ● are, respectively, for months 5,
6, 7, 8, and 9. (c) Distribution of the correlation coefficient between
the Palmer and observed soil moisture as a function of month and
soil depth. Values below ;0.4 are statistically insignificant. There
are insufficient soil moisture data for months Oct–Feb.
F
IG
. 2. (a) Scatterplot of monthly anomalies of the PDSI and ob-
served soil moisture within top 0.9-m depth for Illinois during 1981–
2001. Here r is the correlation coefficient of all the data points.
Legend: symbols 1, C, 3, *, and ● are, respectively, for months 5,
6, 7, 8, and 9. (b) Distribution of the correlation coefficient between
the PDSI and observed soil moisture as a function of month and soil
depth. Values below ;0.4 are statistically insignificant. There are
insufficient soil moisture data for months Oct–Feb.
maximum is a bit surprising given the snow effect dis-
cussed above.
The Palmer model–calculated soil moisture content
has large biases for many of the regions in China, Mon-
golia, and fUSSR, as the model was tuned to conditions
typical for the central United States. On the other hand,
the PDSI and Z index are standardized indices that have
near-zero biases. As such, we compare the PDSI and Z
with observed soil moisture content outside the United
States.
Table 2 shows the correlation coefficients between
the observed monthly mean soil moisture content (in
top 1-m depth, except Illinois where it is top 0.9 m)
and the Palmer model–calculated soil moisture content,
Z index, PDSI, and observed precipitation for regions
where soil moisture data are available (from Robock et
al. 2000). The scatterplots of the PDSI versus observed
soil moisture (SM) content for the Chinese and fUSSR
regions are shown in Figs. 3 and 4, respectively. We
D
ECEMBER
2004 1121DAI ET AL.
T
ABLE
2. Correlation coefficients of monthly anomalies of observed SM content (top ;1 m depth) vs Palmer model–calculated soil moisture
(PSM), moisture anomaly index (Z), PDSI, and observed precipitation (P) over regions where soil moisture data are available. Boldface
numbers are statistically significant at the 5% level. See Table 1 for data periods.
Region (No. of stations) SM vs PSM SM vs Z SM vs PDSI SM vs P Months included
Illinois (17) 0.72 0.72 0.58 0.64 May–Sep
Northeast China (14)
(408–47.58N, 122.58–132.58E) 0.50 0.24 0.50 0.16 May–Oct
East China (7)
(32.58–358N, 1108–1208E) 0.58 0.51 0.63 0.39 Mar–Dec
North-central China (5)
(358–42.58N, 1058–1108E) 0.44 0.33 0.61 0.31 May–Aug
South China (3)
(22.58–258N, 102.58–1108E) 0.45 0.36 0.55 0.23 Jan–Dec
West Mongolia (7)
(458–508N, 908–97.58E) 0.44 0.49 0.50 0.42 Jun–Oct
Central Mongolia (25)
(458–508N, 97.58–107.58E) 0.42 0.33 0.52 0.33 May–Sep
East Mongolia (6)
(458–508N, 1108–1158E) 0.29 0.40 0.48 0.38 May–Sep
fUSSR box 1 (11)
(508–558N, 708–1008E) 0.45 0.42 0.54 0.34 Apr–Oct
fUSSR box 2 (14)
(47.58–558N, 458–608E) 0.67 0.59 0.69 0.45 May–Nov
fUSSR box 3 (8)
(508–608N, 27.58–408E) 0.77 0.57 0.71 0.38 Apr–Nov
fUSSR box 4 (9)
(42.58–47.58N, 52.58–77.58E) 0.42 0.58 0.50 0.57 Apr–Jul
excluded most winter and spring months from the cor-
relation since there were either insufficient moisturedata
or because snow interferes with the relationships with
the PDSI and Z index during these months, except for
South China where it rarely snows. In these comparisons
of monthly data, we consider only simultaneous cor-
relations. Table 2 and Figs. 2 and 4 show that the PDSI
is more consistently correlated with the observed SM
than are the modeled SM, Z index, and observed pre-
cipitation. The observed SM versus PDSI correlation
coefficients range from ;0.5 to 0.7, whereas the cor-
relation with the Z index and precipitation are generally
lower, as the Z index and precipitation time series have
more high-frequency variations than the PDSI. The
modeled SM correlates with the SM data better than the
Z index, but it can have large mean biases outside the
central United States, as the Palmer model was tuned
to central U.S. conditions. Figures 2a, 3, and 4 show
that the quantitative relationship (i.e., the slope)between
the observed SM and PDSI does not vary significantly
with month over most of the regions, except for east
China where October–December PDSI varies more with
SM than in other months.
As stated in sections 1 and 2, the Palmer model does
not deal with snow and some other processes affecting
soil moisture content. The PDSI was not designed to be
and should not be considered a direct measure of soil
moisture content. Therefore, one should not expect a
perfect correlation between the two. In addition, the soil
moisture data, which have only two–three measure-
ments per month and contain many missing values, like-
ly have large temporal and spatial sampling errors, es-
pecially for regions with a small number of stations such
as South China and west Mongolia (Table 2). These
errors also contribute to the scatter in Figs. 1b, 2a, 3,
and 4. The records of the soil moisture data outside the
United States are relatively short so that the SM versus
PDSI correlation at each depth and for each month are
noisy (not shown); nevertheless, the PDSI was found
to correlate with the SM up to 1-m depth in most of the
regions.
4. PDSI versus river flow
The time series of annual streamflow rates and basin-
averaged PDSI for world’s largest 10 rivers (except No.
5 Brahmaputra and No. 10 Mekong, both in southern
Asia, whose streamflow records are too short), plus 4
smaller rivers that have long records are compared (Fig.
5). In deriving the basin-averaged PDSI, we required
that more than half of the basin areas had data. This
mostly affects the PDSI of the earlier years of the time
series. Figure 5 shows that the basin-averaged PDSI
covaries with streamflow rates for all of the 12 rivers
except for the Yenisey. For example, the PDSI over the
Amazon basin closely follows the flow rates at Obidos
during the last three decades, it suggests low flow rates
in the 1960s and near-normal flows in the 1950s when
there were no streamflow data, and it matches the mea-
sured flow rates in the 1940s. For the Orinoco, Missis-
sippi, and Parana´, both the PDSI and streamflow show
some increases after the late 1950s, or early 1960s for
Parana´. For the Congo basin, large increases occurred
around 1960 in both the PDSI and steamflow; thereafter,
they both decreased gradually. Even over smaller river
basins such as the Columbia in the western United States