Maps of monthly self-calibrating Palmer Drought Severity Index (SC-PDSI) have been calculated for the period of 1901–2002 for Europe (35°–70°N, 10°W–60°E) with a spatial resolution of 0.5° × 0.5°. The recently introduced SC-PDSI is a convenient means of describing the spatial and temporal variability of moisture availability and is based on the more common Palmer Drought Severity Index. The SC-PDSI improves upon the PDSI by maintaining consistent behavior of the index over diverse climatological regions. This makes spatial comparisons of SC-PDSI values on continental scales more meaningful.
Over the region as a whole, the mid-1940s to early 1950s stand out as a persistent and exceptionally dry period, whereas the mid-1910s and late 1970s to early 1980s were very wet. The driest and wettest summers on record, in terms of the amplitude of the index averaged over Europe, were 1947 and 1915, respectively, while the years 1921 and 1981 saw over 11% and over 7% of Europe suffering from extreme dry or wet conditions, respectively.
Trends in summer moisture availability over Europe for the 1901–2002 period fail to be statistically significant, both in terms of spatial means of the drought index and in the area affected by drought. Moreover, evidence for widespread and unusual drying in European regions over the last few decades is not supported by the current work.
The Palmer Drought Severity Index (PDSI) is a measure of regional moisture availability that has been used extensively to study droughts and wet spells in the contiguous United States, particularly as the primary indication of the severity and extent of recent droughts (Palmer 1965; Heim 2002), with applications in other parts of the world emerging in the past decade (Briffa et al. 1994, hereafter BJH; Dai et al. 1998; Dai et al. 2004; Lloyd-Hughes and Saunders 2002). The computation of the index involves a classification of relative moisture conditions within 11 categories as defined by Palmer (1965) (Table 1). The index is based on water supply and demand, which is calculated using a rather complex water budget system based on historic records of precipitation and temperature and the soil characteristics of the site being considered. The quantities involved in the calculation are potential evapotranspiration, computed using the Thornthwaite (1948) method; the amount of moisture required to bring the soil to field capacity; the amount of moisture that is lost from the soil to evapotranspiration; and runoff. Based on potential values for these four quantities, Palmer (1965) defined the “climatically appropriate for existing conditions” precipitation, and it is the difference between this value and the actual precipitation that is at the heart of the PDSI. The departure from normal precipitation is then multiplied by a weighting factor, termed the “climatic characteristic,” to produce a “moisture anomaly index.” The purpose of the weighting is to adjust the departures from normal precipitation such that they are comparable among different areas and different months (Alley 1984). Subsequently, Palmer (1965) related the drought severity for a given month to the weighted sum of the moisture anomaly index of that month and the drought severity of the previous month. These latter weighting factors, termed the “duration factors,” determine the balance in the PDSI between sensitivity to short-period moisture fluctuations and a more persistent character. All weighting factors in Palmer’s algorithm were empirically derived from a limited amount of data, largely from the U.S. Great Plains, but are frequently treated as fixed parameters regardless of the climate regime in which the index is computed (Wells et al. 2004).
In the years since its development, the PDSI has become a standard for measuring meteorological drought, particularly in the United States (e.g., see the Drought Monitor, information available online at http://www.drought.unl.edu/dm/). However, the PDSI has been criticized for a variety of reasons of which the most significant is perhaps that it is not comparable between diverse climatological regions, despite Palmer’s efforts to ensure that it would be. This is a particular drawback when a PDSI climatology is produced at a continental rather than regional scale (BJH; Dai et al. 1998; Dai et al. 2004; Lloyd-Hughes and Saunders 2002). Another deficiency of the PDSI as formulated by Palmer (1965) is that it produces values that are ≥4 or ≤ −4 up to 15% or more of the time (Wells et al. 2004), hardly corresponding to the classification “extreme.” Moreover, the PDSI tends to have a slightly bimodal distribution (Wells et al. 2004, their Fig. 8), with maxima in the distribution outside of the “near normal” category.
The self-calibrating PDSI (SC-PDSI) as put forward by Wells et al. (2004) is more appropriate for geographical comparison of climates of diverse regions. Wells et al. (2004) improve the performance of the PDSI by automating the calculations Palmer made when he derived the empirical constants used in the PDSI algorithm. They achieve this by determining, for each location, the climatic characteristic weighting factor using data from only that location (rather than by using data of a small number of stations from different climates as was done originally). This scales the departure from normal precipitation with a factor uniquely appropriate to that location, and affects the range of values of the self-calibrating PDSI. Similarly, the duration factors are determined using data from that location only, which influences the sensitivity of the index for changes in the moisture regime.
The improved performance of the self-calibrating PDSI over the “original” PDSI is demonstrated in a European context in section 3.
The aim of this study is to present an initial analysis of maps of the summer SC-PDSI for Europe for the period of 1901–2002. It serves as an update to the analysis of BJH by using a temperature and precipitation dataset with increased resolution (by an order of magnitude) and improved spatial coverage. Finally, the use of the self-calibrating PDSI, rather than the original PDSI, should make comparisons of the index values between diverse climatological regions more appropriate.
Note that the influence of large-scale changes in water usage resulting from reservoir development, urbanization, or changes in irrigation are ignored in the index. The calculation of the index does not account for precipitation in the form of snow, but assumes all precipitation to be in the liquid phase; nor does the calculation take account of changes in the potential Water-Holding Capacity (WHC) of the soils when the ground freezes. For these latter reasons we restrict this analysis to exploring changes in summer moisture PDSI only.
Detailed descriptions how the PDSI is computed can be found in Alley (1984) and Karl (1986). A detailed description of the modifications to this algorithm to obtain the self-calibrating PDSI is given by Wells et al. (2004) and, more concisely, in appendix A to this paper.
The current paper is organized as follows. Section 2 describes the sources of data used to calculate the European SC-PDSI. Section 3 compares the self-calibrating PDSI against the traditional PDSI using the same data sources. Section 4 describes its results as a preliminary analysis of the data. The self-calibrating PDSI dataset for Europe is then compared against a recent alternative (and more expansive) moisture analysis based on the traditional PDSI reconstruction in section 5. The study is summarized and results are discussed in section 6.
2. Temperature, precipitation, and soil moisture data
Gridded precipitation and temperature data (0.5° × 0.5° resolution) are taken from the monthly set compiled by the Climatic Research Unit (Mitchell and Jones 2005). At the time of the analysis, these data covered the period from January 1901 to May 2002. To bring the analysis up to the end of 2002, the data for the remaining 7 months (June–December) were taken to be the appropriate monthly climatological mean values. The SC-PDSI values for the summer of 2002 should therefore be regarded as an approximation.
We used the soil texture–based WHC data from Webb et al. (1993) with some small modifications, and assumed this to be constant over the period under consideration. A 0.5° × 0.5° resolution WHC dataset is obtained from the 1.0° × 1.0° resolution Webb et al. (1993) data by copying values from overlapping grid boxes into the finer grid.
Near the coasts, the temperature and precipitation datasets and the regridded WHC dataset do not completely overlap while WHC values for small islands are absent in the Webb et al. (1993) dataset. This has been overcome by extrapolating the WHC dataset seaward, using a regional average of WHC values, or, in the case of the islands, including WHC values for islands from Groenendijk (1989) in the Webb et al. (1993) dataset. The extrapolation of the WHC dataset does not account for a possible difference in soil types near the coasts.
Webb et al. (1993) remark that it is very unlikely that all horizons of a soil profile simultaneously reach saturation or the wilting point (when only hydroscopic water remains). Palmer’s soil moisture model, however, assumes runoff to occur if, and only if, both the surface and lower layers reach their water-holding capacity (Alley 1984). This motivated us to substitute WHC values in the Webb et al. (1993) dataset, which exceeded 2 m, by a value of 400 mm. Though somewhat ad hoc, this value is based on the potential water-holding capacity given by Groenendijk (1989) for the areas most affected (northwest Scotland and northern Scandinavia).
A map illustrating the WHC values used in this study is given in Fig. 1.
Any changes in SC-PDSI resulting from a different choice of water-holding capacity dataset are documented in appendix B.
3. PDSI versus self-calibrating PDSI
Monthly maps of the SC-PDSI were compared with maps of the original PDSI, calculated using the same temperature, precipitation, and water-holding capacity data as the current analysis. The program used to compute the PDSI is similar to the one used by BJH and was provided by the National Climatic Data Center. Both the self-calibrating and original PDSI calculations were started in 1890, using climatological monthly mean values for temperature and precipitation at each grid box for the period of January 1890–December 1900. This procedure removes any potential spinup problems associated with the PDSI calculation. The spinup period was excluded from the calibration period when calculating SC-PDSI or PDSI maps.
Figures 2a and 2b show the frequency with which the original and self-calibrating PDSI report an extreme dry spell, that is, the percentage of the months in 1901–2002 in which the index ≤ −4. Figures 2c and 2d compare extreme wet spells (index ≥ 4). Between 1901and 2002, the PDSI is in either an extremely dry or wet spell, averaged over Europe, 5.5% and 5.6% of the months. The self-calibrating PDSI is scaled to give a frequency of occurrence of extreme dry or wet spells of around 2%. And indeed, averaged over Europe, the percentages for the self-calibrating PDSI are more appropriate with 1.9% and 2.1% of the months experiencing an extreme drought or wet spell respectively. For some regions, the difference between the two indices can be much larger than this: the percentage of months the PDSI is at or below −4 averaged over the Iberian Peninsula is 7.2% (2.0% SC-PDSI). Over a region in Eastern Europe, lying between 50° and 60°N, 50° and 60°E, it is 9.7% (2.2% SC-PDSI). The large differences between PDSI and SC-PDSI on regional scales relate to the inappropriateness of parameters in the original PDSI calculation.
The unrealistically frequent occurrence of extremely dry or wet spells in the metric of the PDSI, and the improvement of this in the self-calibrating PDSI, is summarized in Fig. 3. A probability distribution of index values is constructed for each grid box and for both the PDSI and SC-PDSI datasets. This gives two ensembles of probability distributions, and the ensemble means (i.e., the spatial average) for the PDSI and SC-PDSI distributions are shown in Figs. 3a and 3b, respectively. The dashed lines give the upper and lower boundaries of the 90th percentile of the ensemble of probability distributions, indicating that for 5% of the grid boxes, the probability of a monthly PDSI that is ≤ −4 (≥4) is at least 12% (∼13%) For the self-calibrating PDSI, these probabilities are ∼5% and ∼4%.
a. European self-calibrating PDSI values
Figure 4 shows strong decadal-scale variability in the summer [June–July–August (JJA)] SC-PDSI averaged over Europe (35°–70°N, 10°W–60°E). The mid-1940s to early 1950s stand out as a persistent and exceptionally dry period, whereas the mid-1910s and late 1970s to early 1980s were very wet. These periods also encompass most of the driest and wettest years of the twentieth century—the three driest years on record are 1947, 1921, and 1950, and the three wettest are 1915, 1981, and 1926. Figure 5 shows that the dry and wet spells in these years were widespread phenomena. The year 1921 stands out in this respect, with 38.4% (11.2%) of Europe affected by SC-PDSI values below −2 (−4). The wettest years were slightly less expansive, with 35.4% (7.2%) of Europe affected by SC-PDSI values above 2 (4) for the year 1981. For 1915 these percentages are 36.6% and 5.4%.
The area of Europe experiencing extreme moisture conditions (either wet or dry) is shown in Fig. 6. The upper histogram is the percentage area with SC-PDSI values outside the range of ±2. The lower histogram shows the area outside the range of ±4. The long-term means are generally stationary between 30% and 35% for moderate or worse conditions and ∼4% for extreme conditions. The three years with the most widespread extreme moisture conditions are 1921, 1941, and 1990 (|SC-PDSI| ≥ 2) and 1921, 1990, and 2001 (|SC-PDSI| ≥ 4). The years with exceptionally small areas affected by extreme moisture conditions are 1974 (moderate conditions) and 1965 (extreme conditions).
Although the area affected by extreme moisture availability in the years 2000 and 2001 is not unprecedented, it is unusual for two consecutive years to have values that exceed two standard deviations above or below the mean. The percentage areas affected by extreme conditions for 2000 and 2001 are 8.5% and 12.8%, while the mean plus two standard deviations is 8.3%. The area affected by at least moderately wet or dry conditions, that is, |SC-PDSI| ≥ 2, is 42.4% for the year 2000 and 44.8% for 2001, both again exceeding the mean-plus-one standard deviation.
The upward trend in the first few decades of the twentieth century in the area affected by dry or wet spells (both |SC-PDSI| ≥ 2 and |SC-PDSI| ≥ 4) may be related to a decreased variance level in the very early precipitation data. For parts of Europe, instrumental data at high spatial resolution are not available for this period, and data have been interpolated from surrounding stations. This procedure tends to suppress extreme values in the datasets.
Trends in moisture availability during the last 50 yr of this record fail to pass the 95% confidence level. This holds for Europe as a whole and for individual grid boxes. These data indicate that a possible trend toward summer desiccation over the last 50 yr cannot be distinguished from the strong (multi-) decadal variability of the SC-PDSI. Following Lloyd-Hughes and Saunders (2002), significance testing in the presence of serial autocorrelation, a typical characteristic of the PDSI data, should be performed using a variant of the Student’s t test [von Storch and Zwiers 1999, their Eq. (6.6.7)]. The adjusted Student’s t test statistic is
where n̂′x and n̂′y are estimates of the effective sample size. They are defined as
(and analogous for n̂′y). Note the factor 2 in the denominator; it is missing in the formula as printed in von Storch and Zwiers (1999) (see Prof. Von Storch’s Web site for details, which is available online at http://w3g.gkss.de/G/mitarbeiter/storch/). The autocorrelation ρx(k) is defined as
where X is the time series.
b. Regional moisture variability
To explore regional patterns of European drought, rather than using the rotated EOF approach of BJH, this study uses the empirical orthogonal teleconnection (EOT) technique as proposed by van den Dool et al. (2000). With the EOT approach one searches for that point in the dataset, the base point, which, by linear regression, explains the most variance of all of the other points combined. The first spatial pattern is the regression coefficient between the base point and all other points, and the first series is taken to be the time series of the raw data at the base point. The original dataset is then modified by subtracting the variance at each point for which the first mode accounts. The procedure is then repeated for the second, third, and subsequent modes. This method can be used to objectively define the most significant regional patterns of drought in Europe in an intuitively clear way.
The first 6 EOT patterns, explaining a combined variance of 31%, are shown in Fig. 7. Unlike EOFs, EOT patterns are not required by construction to be orthogonal. Yet remarkably, the EOT patterns obtained in the analysis of the summer SC-PDSI data are close to an orthogonal set, with spatial correlations between nonidentical EOTs less than 0.25 and frequently ≤0.05 (in 47% of the cases). The exception is the correlation between the pair (2, 6) at 0.5. The orthogonality in the largest part of the EOT patterns has to be a result dictated by the data, because the method to calculate EOTs requires orthogonality in the temporal domain only.
We now describe some characteristics of the patterns of regional moisture availability and associated temporal variability. To describe regional moisture variability associated with these patterns, we averaged the summer scPDSI values only within those areas where the loadings were greater or equal to 0.6 on the relevant EOT (Fig. 7). This simplifies the interpretation of the results in that the time series represent the actual moisture conditions in unambiguously defined regions, rather than orthogonalized versions of these time series. The explained variance of each regional pattern, exceptionally dry and wet years and multiyear (three years or longer) dry and wet spells, are tabulated in Table 2. Here SC-PDSI values of −2 (2) are used as a threshold for a drought (wet spell).
The pattern explaining most of the combined variance (EOT 1) is depicted in Fig. 7a and has its largest amplitude in an extensive region north of the Caspian Sea. Being the first EOT, this pattern is simply a “teleconnectivity” map (Wallace and Gutzler 1981) between a point north of the Caspian Sea and all other points in Europe. Its loading outside this area is nearly zero, which emphasizes the isolated regional character of droughts affecting this part of Europe.
The second mode describes coherent moisture variability in the Balkans. The Balkan region saw persistently dry summers from 1983 to 1994 with six years in this interval in the slightly dry category (Fig. 8). This drought continues into the current drought, alleviated only by the short intermission during the years 1997, 1998, and 1999.
Mode three describes moisture variability over most of France, England, the Low Countries, and into southern Sweden. The correlation weakens toward Germany and Poland.
The fourth mode describes moisture variability in the Kazakhstan area and is very localized with a near-zero loading outside this area. The Kazakhstan area experienced three persistent droughts during the period analyzed.
EOT 5 describes moisture variability from the Baltic to northwestern Russia and has variability with an opposite sign in Scandinavia, Turkey, and areas northeast of the Black Sea.
Finally, the sixth mode describes variability centered around southwestern France, extending into Italy, Spain, and northward to Brittany. Variability with an opposite sign is found in Scotland, Ireland, and Norway. Exceptional are the years 1945, 1946, and 1949, which are all in the severely dry category.
c. Connection with modes of atmospheric variability
The patterns of EOT 2 and EOT 6 show a dipole structure, and EOT 5 a tripole structure, suggesting possible relationships with modes of atmospheric variability. To investigate this, the summer SC-PDSI maps are correlated with some indices representing the dominant modes of extended winter [December–March (DJFM)] atmospheric variability. The modes of atmospheric variability are most pronounced in winter and the accumulative nature of soil moisture, and of the (SC-) PDSI, means it has memory of the antecedent weather.
The North Atlantic Oscillation (NAO) is the most dominant mode of atmospheric variability in the Atlantic sector, with its familiar centers over or just west of Greenland and one of opposite sign over the subtropical Atlantic (Barnston and Livezey 1987). A correlation map between summer SC-PDSI and the extended winter (DJFM) NAO index (Jones et al. 1997) over the period of 1901–2002 is shown in Fig. 9a. The correlations vary between −0.60 and 0.55, and the most positive correlations are found in western Scotland and southern Norway. The most negative correlations are found in large areas over Spain.
The correlation pattern of the NAO index with summer SC-PDSI resembles the EOT 6 pattern (Fig. 7). However, the EOT analysis fails to show the impact of the preceding winter NAO on summer soil moisture availability in Spain in a coherent, pan-European pattern, but instead represents SC-PDSI variability in Spain as a separate, disconnected mode. This latter mode is ranked as EOT 10.
The East Atlantic (EA) pattern is the second most prominent mode of low-frequency variability over the North Atlantic and consists of a north–south dipole of anomaly centers that span the entire North Atlantic Ocean from east to west. The anomaly centers in the EA pattern are displaced southeastward to those the NAO pattern (Barnston and Livezey 1987). The temporal variability of the EA pattern and that of the Eurasian patterns (discussed below) are provided by the National Oceanic and Atmospheric Administration’s (NOAA’s) Climate Prediction Center (National Weather Service), and the common period between these data and the summer SC-PDSI maps is 1950–2002. Correlations between the EA pattern and the SC-PDSI maps vary between about −0.50 and 0.45, and the most positive correlations are found in central Norway/Sweden and the Baltic States (Fig. 9b). The most negative correlations are found in the Mediterranean and southeastern Europe.
The impact of the East Atlantic pattern on summer soil moisture availability resembles EOT 2 (Fig. 7).
The Eurasian type-1 (EUR1) and type-2 (EUR2) patterns are both three-center reflections of an approximate east–west wave train over the Eurasian continent (Barnston and Livezey 1987). The EUR1 pattern has a primary circulation center that spans Scandinavia into the Arctic Ocean. Centers with opposite signs and weaker amplitudes are located over Spain and the Mediterranean area and over Mongolia/western China. The EUR2 pattern has a western center near Denmark and England, a center in northeast China, and an anomaly center of opposite sign, located north of the Caspian Sea (Barnston and Livezey 1987).
The correlations between the temporal variability of EUR1 and summer SC-PDSI vary between about −0.50 and 0.60, with the most positive correlations along the coast of the western Mediterranean and south of the Caspian Sea (Fig. 9c). The most negative correlations are found around the Low Countries, in northern Fennoscandia and Scotland.
Variability in the EUR2 pattern is correlated with the SC-PDSI maps at −0.45 and 0.45 (Fig. 9d). Highest correlations are found in large areas northeast of the Caspian Sea, and to a lesser degree in southeast Turkey. Most negative correlations are seen around the Adriatic Sea, extending into Hungary and Slovakia.
When computing the first 10 EOTs, it was found that none of these modes of summer SC-PDSI variability could be identified with the impact of either the EUR1 or EUR2 patterns.
Only weak relationships were found between summer SC-PDSI and the Southern Oscillation index of the preceding winter or spring season or the simultaneous summer season.
5. Comparison with the Dai et al. (2004) PDSI dataset
Recently, Dai et al. (2004) presented a global PDSI dataset for 1870–2002, which serves as an update to their earlier PDSI dataset (Dai et al. 1998). The drought index used by Dai et al. (2004) is the original PDSI, not the self-calibrating PDSI. The Dai et al. (2004) PDSI dataset (referred to as the Dai dataset hereafter) is based on a combination of pre-1948 precipitation data from Dai et al. (1997) and National Centers for Environmental Prediction (NCEP) Climate Prediction Center data (Chen et al. 2002) for the recent part of the record, with the latter combined with a large number of gauges (Dai et al. 2004). The precipitation dataset is on a 2.5° × 2.5° grid. The temperature dataset used in the Dai dataset is the Climatic Research Unit CRUTEM2 dataset (Jones and Moberg 2003), regridded to the 2.5° × 2.5° grid, and the water-holding capacity dataset is similar to the one used in current study (barring the small modifications), presented in section 2.
Summer (JJA) PDSI values for Europe from the Dai et al. (2004) dataset are extracted and compared with the European summer SC-PDSI values presented in this study.
Dai et al. (2004) take the view that the PDSI metric ranges between −10 and +10 to cover the spectrum from extremely dry to extremely wet conditions, rather than the conventional view that the PDSI varies between ∼−4 and +4. No values in the Dai dataset exceed the [−15, +15] interval.
The extreme years in the current SC-PDSI dataset and the Dai dataset generally agree: years with either extensive wet or dry conditions in Europe in the SC-PDSI dataset are also years with widespread wet or dry conditions in the Dai dataset, albeit at a much larger spatial scale (cf. Fig. 5 and Fig. 10). In the Dai dataset, the percentage area affected by drought in 1921 is 75.6% (PDSI ≤ −2) or 44.4% for severe drought (PDSI ≤ −4), whereas in the SC-PDSI dataset these percentages are 38.4% and 11.2%. The area affected by wet conditions in 1915 according to the Dai dataset is 69.7% (31.4%) for PDSI values ≥2(≥4), while in the SC-PDSI dataset these percentages are 36.6% (5.4%). However, in the Dai dataset, many more years are indicated to reach similar extreme levels. This is especially the case for the extremely dry years, where the years 1990, 2000, 2001, 2002, and 2003 reached similar levels of the area affected by extreme drought (PDSI ≤ −4) as that of 1921 (percentages 42.7%, 45.5%, 51.6%, 49.6%, and 45.1% respectively), whereas SC-PDSI values for these years are at or below ∼8%.
Locally, and for less extreme years, the Dai dataset has a different character than the current SC-PDSI dataset, as demonstrated for areas in central France and northeastern Spain (Fig. 11). Although some correspondence between the SC-PDSI and the Dai dataset PDSI time series for these locations exist (correlations are 0.78 and 0.65 for the French and Spanish records, respectively), a sizeable difference exists between the two for most of the years. Extremely large differences are found during the summer of 1911 in the Spanish area (Dai dataset PDSI: 8.29, SC-PDSI dataset: −0.82) and during the summer of 1991 in the French area (Dai dataset PDSI: −12.18, SC-PDSI dataset: −1.97). Furthermore, the summer of 2000 is indicated as extremely dry in the Spanish area according to the Dai dataset (PDSI value −7.00), whereas the SC-PDSI value indicates a slightly dry period only (SC-PDSI value −1.19). The difference in characterizing the summer of 2000 for the French area is not very large, with the Dai dataset PDSI value indicating a slightly dry summer at −1.92 and the SC-PDSI value of opposite sign, but in the near-normal category (at 0.39); however, the lack of consensus is in the trends of these values. The Dai dataset PDSI indicates a strong trend toward drier conditions (culminating in PDSI values of −3.14 and −8.40 during the summers of 2001 and 2002, respectively), while the SC-PDSI dataset indicates a trend toward wetter conditions over the last 10 yr of the records.
Motivated by the obvious importance of estimating trends in summer moisture availability, we computed the trends over the 1950–2002 period for both the Dai and SC-PDSI datasets. These trends were not subjected to a statistical significance test because our purpose is to compare the actual changes that are indicated by the datasets, rather than to determine whether these changes are unusually large. Dai et al. (2004, their section 6) calculated trends in their dataset using all monthly data rather than only summers. Here we repeat that calculation but use only summer mean PDSI values. The trends per 50 yr are shown in Fig. 12. There is general agreement in the sign of the trend, with the exception of the Low Countries, France, most of the British Isles, and the Kazakhstan area. The Dai dataset gives the impression that Europe, on the whole, is getting drier, which is a conclusion also reached in earlier work by Dai et al. (2004). This is also reflected in the percentage area showing a trend of −2 (50 yr)−1 or stronger: 31.2% for the Dai dataset and only 9.8% for the SC-PDSI dataset. The values for a trend of −4 (50 yr)−1 or stronger are 6.9% and 0.1%, respectively. The area affected by a trend toward wetter conditions [2 (50 yr)−1 or stronger] in the Dai dataset is 6.6%, while it is 11.1% in the SC-PDSI dataset. Note that the trend in the European mean summer SC-PDSI is not statistically significant at the 95% level, when autocorrelations in the dataset are taken into account, as argued in section 4a.
Concluding, the evidence for unusually strong or widespread drying in European regions over the last 50 yr, as indicated by the Dai dataset, is not supported by the evidence of the current work. The SC-PDSI maps show that the area percentage of Europe experiencing a drying is more than 3 times lower than that in the Dai dataset, while the area percentage that has become wetter is nearly twice as large in the SC-PDSI maps. Although climate models predict summer desiccation in a warmer greenhouse world, no evidence that this drying has set in is present in the current analysis.
6. Summary and discussion
We have described a high-spatial-resolution analysis of summer moisture variability for Europe based on the self-calibrating Palmer Drought Severity Index (SC-PDSI), calculated on a monthly basis for the period of 1901–2002. The self-calibrating Palmer Drought Severity Index is a variant of the more common PDSI, but represents a more appropriate means of comparing spatial relationships between areas of differing moisture climates because the SC-PDSI provides a more realistic metric of relative periods of drought or excessive moisture supply.
This study and the resulting dataset serve as an update to those of BJH because the input data are of higher spatial resolution, the data coverage is much improved, and the SC-PDSI is more appropriate than the PDSI as a metric for dry and wet spells.
To compare PDSI and SC-PDSI values, a PDSI dataset is produced based on the same input data as those of the SC-PDSI dataset. This shows that the percentage of months when the SC-PDSI indicates an extreme drought or an extreme wet spell is much less compared to the percentage of months classed as extreme when using the PDSI data. The percentages are also spatially homogeneous in the SC-PDSI dataset and do not suffer from large regional differences in the occurrence frequency of extreme conditions.
The diagram for mean summer (JJA) SC-PDSI (Fig. 4) has similar features to that of BJH (their Fig. 2a). The driest 2 yr on record are 1947 and 1921, similar to those of BJH. The third driest record in BJH is 1949, with little difference in drought-affected area between 1949 and 1950. In the current study the third driest year is 1950. The wettest years on record in the current study are 1915, 1981, and 1926, while BJH rank 1987, 1916, and 1915 as the wettest. The mean European summer PDSI values of BJH have generally much stronger extremes than the equivalent SC-PDSI values in this study, which is consistent with the observation that the SC-PDSI values much less frequently rank a drought or wet spell as being “extreme.” Despite the general agreement, individual years between the BJH and the current study can be quite different; for example, 1934 is the fifth driest year in the BJH study and is much drier than 1937–38, while in the current study the reverse is the case.
A general similarity between the figures showing the area affected by moderate or extreme wet or dry conditions is less obvious (Fig. 5 and BJH’s Fig. 2b). Again, extremely wet or dry years in the BJH study remain extremely wet or dry, but the values of the area affected in the SC-PDSI metric are much less extreme. Interestingly, the extensive nature of summer drought in the years 1989, 1990, and 1991, which largely prompted the BJH study, appears less extreme in the current study, although the area affected by extreme drought is still notable. The area affected by moderate drought (index ≤ −2) in the BJH study was ∼50%, comparable with such extreme years as 1921 and 1947. Here the area affected is 23.0% for 1989 and 30.1% for 1990, while 1921 stands out at 38.4%. The area affected by extreme drought (index ≤ −4) in the BJH study is a little over 20% for 1990, the second most extreme year after 1921 in that respect. In the current study, 1990 ranks fourth at 7%, after 1921, 2000, and 2001.
The present study shows that the upward trend in the total area subject to extremes (BJH, their Fig. 2c) from the mid-1980s to the early 1990s, which led to unprecedentedly high values in the decadally smoothed values in that period, has not continued. Values for the total percentage area subject to extreme moisture conditions in the years 1996–99 returned to normal levels at ∼2% from a maximum of nearly 10% in 1990 (Fig. 6).
Trends in summer SC-PDSI are not statistically significant considering the strong autocorrelation in the SC-PDSI time series for each grid box. This confirms the conclusion of earlier work (Lloyd-Hughes and Saunders 2002) regarding the absence of trends in European summer PDSI.
Most of the regional patterns defined by the empirical orthogonal teleconnection (EOT) patterns (Fig. 7) can be identified with the regional patterns identified by BJH, despite the fact that different methods were used to identify spatial patterns. In the latter study, regional patterns are identified by first subjecting the data to a principal component analysis, with the first nine patterns being rotated according to the varimax criterion. The patterns defined in this way are referred to as rotated principal components (RPCs). However, the amount of variance explained by the RPC and the EOT patterns differs, but mostly because of the greater spatial resolution of the current data.
Comparison of the maps of summer self-calibrating PDSI against part of an alternative, recently published PDSI dataset (Dai et al. 2004) shows some agreement between these two metrics of dry and wet conditions for European-averaged quantities at a qualitative level. At regional scales, large differences, both quantitative and qualitative, exist between the datasets.
The evidence for a very strong increase in the area of combined severe dry and wet conditions in Europe over the last three decades observed by Dai et al. (2004) is largely absent in the current dataset. The slight upward trend in the combined area affected by extreme moisture conditions (both |SC-PDSI| ≥ 2 and |SC-PDSI| ≥ 4) from the early 1970s to the end of the record is not statistically significant. The absence of a trend toward summer desiccation has recently also been observed in soil moisture records in the Ukraine (Robock et al. 2005) and supports conclusions in the current study.
Monthly maps of the self-calibrating PDSI for Europe are made available through the Web site of the Climatic Research Unit (available online at http://www.cru.uea.ac.uk).
The referees are thanked for their constructive comments, which helped to improve the clarity of this paper. GvdS is funded by the U.K. Natural Environment Research Council (NERC) through the RAPID Climate Change programme. KRB and TJO also acknowledge support from the EU project SOAP.
Calculation of the SC-PDSI
Palmer’s method is based on a water balance calculation using a simple two-layer soil moisture model. The local hydrological normals for each month are computed in terms of a set of coefficients (α, β, γ, δ) that satisfy
where PE is the potential evapotranspiration calculated from the Thornthwaite formula, ET the actual evapotranspiration, R the soil water recharge, PR the potential soil–water recharge, RO the runoff, PRO the potential runoff, L the water loss from the soil, and PL the potential water loss from the soil. The overbars signify that these are average quantities, derived using the appropriate month’s data averaged over the calibration period.
Evapotranspiration losses from the soil occur when PE > P, where P stands for precipitation. Losses from the surface layer (Ls) are assumed to occur at the potential rate. Losses from the underlying layer (Lu) depend on the initial moisture content PE and the combined soil water content (SWC) in both layers. The available water capacity (AWC) of the underlying layer depends on local soil type, and an appropriate value must be specified. Moisture cannot be removed from, or replenish, the underlying layer until the available moisture has been removed completely from or replenished in the surface layer. Hence,
where Ss, Su are the initial moisture values in the surface and underlying layers at the beginning of the month, respectively. Runoff cannot occur unless both layers reach their combined capacity.
The potential recharge, runoff, and loss terms are calculated as
where PLs and PLu are defined by
Potential recharge (PR) is defined as the amount of moisture required to bring the soil water up to field capacity (i.e., AWC). Potential loss (PL) is the soil moisture that could be lost to evapotranspiration if precipitation was zero for that month. Potential runoff (PRO) is defined as the potential precipitation (assumed to be equal to AWC) minus potential recharge. Given α, β, γ, δ, the PDSI calculation for a given month then starts by computing a precipitation anomaly
where P̂ = αPE + βPR + γPRO − δPL. The precipitation anomaly d is weighted with the climate characteristic K to yield the moisture anomaly index
Scaling d with K in effect standardizes the moisture anomaly by scaling it to a degree determined by local conditions. Palmer (1965) determined K empirically using locations from nine different sites in the seven states in the United States (Texas, Kansas, Iowa, North Dakota, Ohio, Pennsylvania, and Tennessee), and K separates into two parts
where D is the mean of the absolute values of d for that month over all years in the calibration period. The constant K̃ can be viewed as the ratio between the expected and observed values of the average annual sum of the moisture anomaly (Wells et al. 2004). The self-calibrating PDSI calculation relates K̃ to the expected and observed tails of the PDSI distribution, rather than the central tendency of the moisture anomaly. For the self-calibrating PDSI we define the coefficient K as
where PDSI2nd and PDSI98th are the 2nd and 98th percentiles of the PDSI distribution as computed using the climatic characteristic K′ (rather than K) (Wells et al. 2004).
Having calculated Z for month i, the PDSI value Xi is related to the previous PDSI value Xi−1 and the current moisture anomaly Zi. The weights of these two components are determined by the duration factors p and q
Palmer (1965) chose p = 0.897 and q = 1/3 and assumed them to be constant over space, but the self-calibrating PDSI determines duration factors for wet and dry conditions separately, which are appropriate for a given location (Wells et al. 2004).
Evaluation of the duration factors starts by calculating the accumulation of the moisture anomaly index over 10 periods ranging in length from 3 to 48 months. The highest (lowest) accumulated Z for each period represents the extreme wet and dry spells by which Palmer calibrated his index. The duration factors are calculated giving the slope m and intercept b, such that
is an optimal fit, in a least squares sense, to the accumulated Z for the 10 periods. To make the least squares fit of the accumulated Z truly representative of the most extreme dry or wet spells, a new line is generated, parallel to the fit, but coinciding with the most extreme accumulated Z value. This is done for wet and dry conditions separately; for wet conditions the least squares fit has its intercept increased, while the least squares fit for dry conditions has its intercept decreased to more negative values. This is illustrated in Fig. A1 for data from a grid box in East Anglia, United Kingdom.
where C represents the value of the calibration index; C = −4 for dry conditions and C = 4 for wet conditions.
Palmer (1965) modified the procedure according to whether a dry (wet) spell was already established. A spell is considered to have become established once the absolute value of PDSI exceeds 1.0. The end of a dry spell does not necessarily indicate the start of a wet spell and vice versa. The beginning and ending of spells are determined by continually monitoring the three PDSI series X1, X2, and X3, which denote the index for a wet spell becoming established, a dry spell becoming established, and a wet or dry spell that is currently established, respectively. Droughts and wet spells end when X3 returns to a value in [−0.5, 0.5]. Rather than wait for a spell to end, Palmer calculated the “percentage probability” Pe that the spell had ended,
where j* is the number of months before month i that a spell became established.
which is derived from (A10) by solving for Zi and substituting −0.50 and 0.50, respectively, for Xi, which signifies the boundaries of the “near normal” category for the index (Alley 1984). In the case of an established drought, Palmer (1965) noted that a value of Z = −0.5(1 − p)/q will produce a constant index value of −0.50 (analogous for wet conditions). Consequently, if Z ≥ −0.5(1 − p)/q then dry conditions will tend to end, and Z ≤ 0.5(1 − p)/q tends to end an established wet spell. Palmer (1965) related the variable Ui to the moisture anomaly Zi by
for dry and wet conditions, respectively. A dry or wet spell is deemed to have ended when Pe exceeds 0% and is later shown to have remained above 0% until it finally reached 100% (Karl 1986; Alley 1984). The application of this criterion in the determination of whether a dry or wet spell has ended may lead to a revision of previously computed PDSI values. This retrospective element in the PDSI calculations is referred to as “backtracking” (Wells et al. 2004).
Effects of Different Water-Holding Capacities on SC-PDSI
The aim of this appendix is to test the generally held view (BJH; Dai et al. 1998) that changing the potential water-holding capacity (WHC) has little substantial influence on the resulting PDSI values.
One of the three different WHC datasets put forward by Webb et al. (1993), the (slightly modified) WHC estimate derived from soil texture, is used to calculate the SC-PDSI in this study. These SC-PDSI maps have been compared against SC-PDSI values based on the other two datasets—WHC derived from root zone thickness and WHC derived from the soil profile. For notational convenience, the three datasets are abbreviated as “TEX,” “ROO,” and “ALL,” respectively. The differences in potential storage of soil water between the datasets are shown in Fig. B1. From these figures it is clear that the WHC based on TEX is, averaged over Europe, the most conservative estimate. However, over large parts of Norway, the TEX estimate gives higher values than the ROO estimate by up to 600 mm.
When averaged over Europe, the differences in SC-PDSI values related to the three different WHC datasets are typically less than 0.1 PDSI units, only reaching 0.25 in occasional summers. The mean value of the difference in European-averaged summer SC-PDSI values based on the TEX dataset and both the ROO and ALL datasets is negative, indicating that the WHC datasets with larger water storage capacity tend to give higher (“wetter”) SC-PDSI values. We relate this to the increased buffering against moisture loss in dry spells. This mitigates the effects of drought and results in less negative SC-PDSI values.
Although the time mean of the summer SC-PDSI values are little affected by the choice of the WHC dataset, temporal variability in SC-PDSI values changes considerably. Figure B2 shows the standard deviation of the difference in summer SC-PDSI values between the three SC-PDSI datasets. The standard deviation is considerably smaller than 1 for the largest part of Europe, but there are some areas with sizeable differences in variability between the SC-PDSI datasets. Areas in southern Spain, southern France, southern Italy, southern Greece, western Turkey, and southeast Turkey show a standard deviation larger than 1 for the difference between summer SC-PDSI values based on the TEX and ROO estimates, as well as for the difference between summer SC-PDSI values based on the TEX and ALL estimates. A similar conclusion holds for the region defined as an arc from Poland to the Ukraine, stretching northeast into Russia, and for an area in northern Scandinavia. The TEX- and ROO-based SC-PDSI datasets show large differences in variability over the largest part of Norway, a feature that is not seen when comparing the TEX- and ALL-based SC-PDSI datasets.
We relate the changes in variability between the three SC-PDSI datasets to a combination of the change in water storage in the soils, and the backtracking aspect in the SC-PDSI calculation (appendix A). A description of Palmer’s operating rules for backtracking is given by Alley (1984). The existence of backtracking means that that a small change in how the index is computed, or a change in the input data for this computation, may initiate backtracking. This then has a substantial effect on the final values of the index (Wells et al. 2004).
Corresponding author address: G. van der Schrier, Climatic Research Unit, School of Environmental Sciences, University of East Anglia, Norwich NR47TJ, United Kingdom. Email: firstname.lastname@example.org