Geophysical Research Letters

Distinct causes for two principal U.S. droughts of the 20th century



[1] Diagnosis of observational and climate model data reveals that the two major U.S. droughts of the 20th Century had distinct causes. Drought severity over the Southern Plains during 1946–1956 is very likely attributable to remote influences of global sea surface temperatures (SSTs). The Southern Plains and adjacent Southwest are regions particularly sensitive to SST variability, and strong La Niña events that occurred during 1946–1956 exposed that region's drought vulnerability. Drought severity over the Northern Plains during 1932–1939 was likely triggered instead by random atmospheric variability. The Northern Plains lies within a region of comparatively low sensitivity to SST variability, and that region's drought exhibited little sensitivity to SST conditions during the Dust Bowl period. Our results indicate that the southern portions of the Great Plains lie within an epicenter of potentially skillful drought predictions for which an ocean observing system is also a vital drought early warning system.

1. Introduction

[2] There is a broad view that major U.S. droughts are influenced by variations in global sea surface temperatures [e.g., Cole et al., 2002; Hoerling and Kumar, 2003; Schubert et al., 2004a, 2004b; McCabe et al., 2004; Seager et al., 2005, 2008]. The implication is that drought early warning is rooted in ocean observations, and that skillful sea surface temperature (SST) predictions would be tantamount to skillful U.S. drought predictions. What is unclear is the spatial expression and intensity of the SST-induced drought signal for the conterminous U.S., and as such the expected skill of regional drought predictions is not known. Clarity on the causes for different drought events will thus be important for developing drought early warning capabilities.

[3] What, for instance, is responsible for the marked differences in regional extent of the two principal U.S. droughts of the 20th Century (Figure 1), and did these events have different causes? The existing picture regarding the 1930s and 1950s events in particular is muddled by both a lack of consistency among atmospheric general circulation model (AGCM) simulations, and by the fact that model and observational studies have not been carefully reconciled. In particular, the model studies have not calculated drought severity indices that can be related to the same indices calculated from observations such as shown in Figure 1.

Figure 1.

(left) The observed Standardized Precipitation Index (SPI, 12-month period) and (right) self-calibrated Palmer Drought Severity Index (PDSI) averaged for the periods (top) 1932–1939 and (middle) 1946–1956. Negative indices denote abnormally low surface moisture conditions. (bottom) Sea surface temperature departures for these two drought periods.

[4] We find that the 1930s and 1950s droughts had different origins based on diagnosis of the SST sensitivity of drought indices derived from observational and multi-model AGCM data. Much of the 1950s drought severity is attributable to SST forcing, whereas the 1930s drought is largely unrelated to SST forcing and was likely instead triggered by random atmospheric variability. We propose that drought early warning based on a sea surface temperature observing system would thus be particularly effective for the Southern Plains and Southwest U.S.

2. Data and Methods

[5] Monthly U.S. land surface temperature and precipitation data for 1895–2007 are based on the National Climate Data Center's (NCDC's) climate division analysis. Monthly global SST data are based on 1° gridded analysis of the UK Meteorological Office's HadISST version 2 [Rayner et al., 2003].

[6] Two drought indices are computed from monthly data. One is the standardized precipitation index (SPI) [McKee et al., 1993] which denotes the number of standard deviation departures of precipitation from the local climatology. We use a 12-month averaging window to compute the SPI (SPI-12). Because meteorological drought often develops from a combined effect of temperature and precipitation, we also compute the self-calibrated Palmer Drought Severity Index (PDSI) [Wells et al., 2004]. This index ranges from −4 for extremely dry conditions to +4 for extremely wet conditions.

[7] Atmospheric climate model simulations are used to determine the SST effects on the variability of U.S. drought indices. Three different AGCMs are used, each subjected to observed monthly varying global SSTs from 1903–2004. (The 3 models used, and their ensemble size are the NCAR Community Climate Model (CCM3) [Kiehl et al., 1998; Seager et al., 2005], 16 member; the GFDL AM2.1 model [GFDL Global Atmospheric Model Development Team, 2004], 10 member; the NASA Seasonal-to-Interannual Prediction Project (NSIPP) model [Schubert et al., 2004a, 2004b], 14 member.) For each of the 40 simulations, SPI-12 and PDSI indices are computed from the model's monthly temperature and precipitation data that are spatially interpolated to the U.S. climate division grid.

3. Twentieth Century Drought Variability and Its Oceanic Links

[8] Observationally derived SPI-12 and PDSI time series were computed for 1895–2007 averaged over the Northern Plains (Figure 2, left) and the Southern Plains (Figure 2, right; area-averages were computed over the gray stippled regions shown in Figure 2, middle). The 1930s and 1950s events are clearly the most severe and prolonged U.S. droughts over the Great Plains since 1895. Yet, diagnosis of the observational data implies that different processes contribute to drought and pluvials in the two regions. For instance, the northern and southern Plains drought time series have only weakly covaried during 1895–2007—0.32 for the monthly PDSI and 0.39 for the monthly SPI-12. As such, drought indices spatially averaged over the entire Great Plains region between the Mexican and Canadian borders do not sample a domain of coherent variability. This lack of coherency is particularly evident during the 1930s and 1950s; their SPI and PDSI drought patterns are spatially uncorrelated over the U.S. (see Figure 1). Further, different SST conditions were associated with these two major droughts (Figure 1, bottom). North Pacific SST anomalies during 1932–39 were opposite in sign to those during 1946–56, and colder SSTs occurred over the tropical east Pacific during the 1950s. Their most common SST condition was North Atlantic warmth.

Figure 2.

Time series (top) of the observed Standardized Precipitation Index (SPI; black) and self-calibrated Palmer Drought Severity Index (PDSI; red) for 1895–2007 averaged over (left) the Northern Plains and (right) the Southern Plains. Yellow band indicates the periods of severe drought over the Northern Plains and Southern Plains. The temporal correlation between global SST and (middle) SPI and (bottom) PDSI monthly time series for (left) the Northern Plains and (right) the Southern Plains. The Northern Plains area (MT, WY, CO, ND, SD, NE and KS) and the Southern Plains area (AZ, NM, OK, TX) are indicated by hatching in Figure 2 (middle).

[9] Figure 2 also shows the temporal correlation of SST with SPI-12 (Figure 2, middle) and PDSI (Figure 2, bottom) monthly time series for 1895–2007. Values exceeding 0.2 corresponding roughly to a 90% local significance are highlighted. For Northern Plains drought, only Atlantic SST correlations are apparent, with dry conditions occurring in concert with warm states of the North Atlantic Ocean as also highlighted in other empirical studies [e.g., McCabe et al., 2004]. For Southern Plains drought, a coherent Indo-Pacific SST pattern reminiscent of the El Niño/Southern Oscillation (ENSO) is evident, with dry conditions occurring in concert with cold La Niña phases of the ENSO cycle.

[10] Diagnosis of climate model simulations reveals that the two key empirical associations highlighted above are interpretable as cause-effect relationships: 1) drought variability over the Southern Plains is strongly forced by SST variability, with ENSO-like conditions being the principal driver, 2) SST forcing was the primary factor causing the 1950s Southern Plains drought, but the 1930s drought over the Northern Plains was unexplained by SST forcing. Figure 3 presents analysis of the multi-model AGCM simulations of 20th Century U.S. drought variability. The standard deviation of the monthly drought index values, averaged across all 40 model runs, is given in Figure 3 (top). Because the SPI and PDSI are normalized by each run's local climatological temperature and precipitation variability, the total standard deviation of these two drought indices is spatially uniform by design; for the SPI (PDSI), the standard deviation is close to 1 unit (2 units). The variability of drought indices resulting solely from SST variability (Figure 3, middle) is derived by first forming the 40-run ensemble averaged SPI-12 and PDSI for each month, and then calculating the standard deviation of the resultant time series. The result yields a distinct Southern Plains maximum where the SST-forced signal is 50% of the total drought index variability. Further, a distinct ENSO-like SST pattern is correlated with the model's Southern Plains drought index variability (Figure 3) in agreement with observations (cf. Figure 2).

Figure 3.

(top) The total monthly standard deviation of (left) AGCM Standardized Precipitation Index (SPI) and (left) self-calibrated Palmer Drought Severity Index (PDSI) based on averaging 40 individual simulations for 1903–2000. (middle) The SST-forced monthly standard deviation of (left) SPI and (right) PDSI based on the 40-run averaged indices during 1903–2000. (bottom) Temporal correlation between global SST and the 40-run ensemble averaged (left) SPI and (right) PDSI 1903–2000 monthly time series averaged over the Southern Plains.

[11] Consistent with such regionally dependent SST influences on drought, the ensemble mean simulations confirm that the two principal U.S. droughts of the 20th Century had distinct causes (Figure 4). Neither the SPI-12 nor the PDSI responses to SST forcing during the 1930s (Figure 4, top) supports the notion of a causal SST relationship over the Northern Plains. We note that despite the significantly warm North Atlantic SSTs during the 1930s, the models' Northern Plains climate is insensitive to forcing from that region. There are at least two interpretations of this result. One is that model biases exist that inhibit sensitivity. The other is that the observed empirical correlation (Figure 1) is not an indicator of cause-effect. It is noteworthy that the decade 1999–2008 has exhibited regional ocean conditions very similar to those of the 1930s—the North Atlantic SSTs have undergone a secular warming with amplitudes greater than witnessed during the 1930s and the tropical east Pacific has been cool consistent with a preponderance of La Niña events. Yet, a decadal average precipitation anomaly for 1999–2008 is most remarkable for the high annual precipitation over the entire Northern Plains, and the drought conditions one might have anticipated based on Figure 1 did not materialize.

Figure 4.

(left) The Standardized Precipitation Index (SPI) and (right) self-calibrated Palmer Drought Severity Index (PDSI) averaged for the periods (top) 1932–1939 and (middle) 1946–1956 based on the 40-run average of AGCM simulations. Negative indices denote abnormally low surface moisture conditions. (bottom) Estimated probability distribution functions (PDFs) of the (left) SPI and (right) PDSI values summarizing the 40 AGCM runs. Red (blue) PDFs are based on model index values that have been spatially averaged over the Southern (Northern) Plains for 1946–56 (1932–39). The Southern and Northern Plains areas are as indicated by hatching in Figure 2.

[12] Our finding that the Dust Bowl-era drought was not SST-caused differs from Schubert et al. [2004a, 2004b] and Seager et al. [2005]. This is partly due to our diagnosis of standardized precipitation indices, rather than just precipitation anomalies as in their studies. Precipitation anomalies alone are not drought proxies: large departures in humid climates may not indicate drought while small departures in arid climates may indicate drought. The precipitation departures of Schubert et al. [2004a, 2004b, Figure 1] and Seager et al. [2005, Figure 7] differ considerably from the SPI departures of our paper's Figures 1 and 4, with the latter indicating that drought severity was largest over the Northern Plains. Further, the Dust Bowl era was a very warm epoch over the Great Plains, which heightened PDSI drought severity. Yet, the climate models fail to reproduce such 1930s warming (not shown) thereby further diminishing their PDSI severity. Two of the three models used in our study are those also used in the Schubert et al. and Seager et al. papers, and the third model used herein is consistent with these in showing no drought response over the Northern Plains during the 1930s. It should be noted, however, our model ensemble does yield a modest drought response over the Southern Plains, consistent with the meteorological drought conditions observed there.

[13] We argue that other factors were likely of greater importance in triggering the 1930s drought over the Northern Plains, and one candidate is the role of random atmospheric variability. The blue curves in Figure 4 (bottom) are the probability distribution functions (PDFs) of Northern Plains drought indices summarizing results for all 40 model runs. The near-zero mean value of the PDFs indicates the lack of an SST-signal, whereas the spread of these distributions reveals the intensity of drought variability that arises solely from atmospheric internal variability. Individual AGCM runs are clearly capable of generating decade-long drought conditions over the Northern Plains, and the 1930s observed drought may have so arisen. Nonetheless, no single run generates drought severity as intense as observed (see Figure S1 of the auxiliary material for a 1903–2004 time series analysis of drought signal-to-noise ratios for the northern and southern Plains).

[14] A different story pertains to the drought of the 1950s. Both the SPI and PDSI responses to SST forcing consist of widespread drought over the Southern Plains and Southwest (Figure 4, middle), a pattern that is spatially congruent with the region's observed 1950's drought. The red curves in Figure 4 (bottom) are the PDFs of Southern Plains drought indices, and the shift of the distributions toward dry conditions is clear. Indeed, each of the 40 model runs generates a negative SPI-12 and PDSI index value for the decadal average of 1946–56 attesting to the strength of SST forcing.

[15] There is a distinct separation between the PDFs of drought indices for the 1930s and the 1950s, a consequence of two primary factors. First, drought indices over the Northern Plains are less sensitive to SST variability than over the Southern Plains. Second, the period of 1946–56 witnessed relatively strong and frequent La Niñas, an effective driver of Southern Plains drought. During the 1930s, tropical Pacific SSTs were comparatively weaker than during the 1950s (cf. Figure 1), though they were also slightly colder than normal and that likely explains the Southern Plains drying during the 1930s. Indeed, the fact that cool tropical Pacific conditions existed in both decades explains why the models generate virtually identical patterns of atmospheric circulation anomalies over the Pacific-North American region during the two periods (see Figure S2). Whereas the models reproduce the observed circulation anomalies during the 1950s, the observed circulation anomalies during the 1930s were opposite in phase to those simulated, again arguing for a different cause for the Northern Plains drought.

4. Summary and Discussion

[16] This study focused on the relationship between global sea surface temperatures and drought indices derived from both historical observations and a multi-model suite of climate simulations. Drought variability over the southern U.S. was shown to be highly sensitive to SST forcing, with the key SST condition resembling ENSO. Drought variability over the northern U.S. was shown to be considerably less sensitive to SST forcing. Consistent with this regional dependency of SST impacts, distinct factors were found to be responsible for the 1930s and the 1950s droughts. The region of maximum drought severity during 1932–1939 stretched from Kansas northward thru the Dakotas and exhibited little sensitivity to SST conditions during that era. Individual model runs generated decadally averaged drought over the Northern Plains as a consequence of atmospheric variability alone, and we speculated that the Dust Bowl might have been triggered by such a process. By contrast, the region of maximum drought severity during 1946–1956 stretched eastward from Arizona thru Texas and exhibited strong sensitivity to SST conditions—the ensemble average drought indices calculated from the model runs were as severely negative as the observed drought indices.

[17] As the U.S. seeks to realize its vision for an effective drought early warning system as part of a National Integrated Drought Information System [NOAA, 2007], it is important to clarify the scientific capability to anticipate when a 1930s or a 1950s-type drought could unfold again. Indeed, credibility of prediction systems is enhanced when supported by knowledge of the causes and underlying mechanisms of drought variability. Our analysis of these two principal U.S. droughts clarifies several key issues regarding the potential for skillful drought prediction. One, consistent with Schubert et al. [2004a, 2004b] and Seager et al. [2005], precipitation in the Great Plains is found to be sensitive to global SSTs. Our results reveal, however, that southern and northern portions of the Great Plains should be viewed as distinct regions for purposes of drought predictability—-mainly the southern portions of the Great Plains rest within an epicenter of potentially skillful drought predictions based on ocean observing systems alone.

[18] How does one explain the 1930s Dust Bowl period? In so far as it was an iconic event both in the Nation's climate history and its cultural history, explaining its origins continues to be particularly relevant. Consistent with our synthesis of three different models, Cook et al. [2009] using yet another AGCM also discovered that SST forcing alone fails to produce drought conditions over the Northern Plains during the 1930s. An important new part of the mystery that their study addresses is how human-induced land surface changes and soil dust aerosols could have affected the drought's severity, factors not represented in our suite of model runs. We thus offer the following view of the principal factors causing the Northern Plains drought during the Dust Bowl-era: Drying was triggered mostly by serendipitous atmospheric variability as argued herein, which then likely induced low soil moisture conditions that could have initiated positive feedbacks with the land surface as proposed by Schubert et al. [2004a]. This drying intensified further and subsequently spread geographically across much of the Northern Plains perhaps due to dust aerosol feedbacks as argued by Cook et al. [2009].