Forced and unforced variability of twentieth century North American droughts and pluvials
- First Online:
- Cite this article as:
- Cook, B.I., Cook, E.R., Anchukaitis, K.J. et al. Clim Dyn (2011) 37: 1097. doi:10.1007/s00382-010-0897-9
- 220 Views
Research on the forcing of drought and pluvial events over North America is dominated by general circulation model experiments that often have operational limitations (e.g., computational expense, ability to simulate relevant processes, etc). We use a statistically based modeling approach to investigate sea surface temperature (SST) forcing of the twentieth century pluvial (1905–1917) and drought (1932–1939, 1948–1957, 1998–2002) events. A principal component (PC) analysis of Palmer Drought Severity Index (PDSI) from the North American Drought Atlas separates the drought variability into five leading modes accounting for 62% of the underlying variance. Over the full period spanning these events (1900–2005), the first three PCs significantly correlate with SSTs in the equatorial Pacific (PC 1), North Pacific (PC 2), and North Atlantic (PC 3), with spatial patterns (as defined by the empirical orthogonal functions) consistent with our understanding of North American drought responses to SST forcing. We use a large ensemble statistical modeling approach to determine how successfully we can reproduce these drought/pluvial events using these three modes of variability. Using Pacific forcing only (PCs 1–2), we are able to reproduce the 1948–1957 drought and 1905–1917 pluvial above a 95% random noise threshold in over 90% of the ensemble members; the addition of Atlantic forcing (PCs 1–2–3) provides only marginal improvement. For the 1998–2002 drought, Pacific forcing reproduces the drought above noise in over 65% of the ensemble members, with the addition of Atlantic forcing increasing the number passing to over 80%. The severity of the drought, however, is underestimated in the ensemble median, suggesting this drought intensity can only be achieved through internal variability or other processes. Pacific only forcing does a poor job of reproducing the 1932–1939 drought pattern in the ensemble median, and less than one third of ensemble members exceed the noise threshold (28%). Inclusion of Atlantic forcing improves the ensemble median drought pattern and nearly doubles the number of ensemble members passing the noise threshold (52%). Even with the inclusion of Atlantic forcing, the intensity of the simulated 1932–1939 drought is muted, and the drought itself extends too far into the southwest and southern Great Plains. To an even greater extent than the 1998–2002 drought, these results suggest much of the variance in the 1932–1939 drought is dependent on processes other than SST forcing. This study highlights the importance of internal noise and non SST processes for hydroclimatic variability over North America, complementing existing research using general circulation models.
KeywordsDroughtForcingPredictabilityNorth AmericaDust Bowl
Persistent, multi-year drought and pluvial events have been a recurrent feature of North American hydroclimate since at least the time of the Medieval Climate Anomaly (Cook et al. 1999, 2004, 2007; Fye et al. 2003). For the United States (US), the monetary impact of droughts far exceeds the costs of other natural disasters, including fires, earthquakes, and hurricanes (Cook et al. 2007; Herweijer et al. 2007). Pluvials (anomalous wet periods) can have deleterious impacts as well (Trenberth and Guillemot 1996), and have also been instrumental in shaping resource use policies in the western US (Fye et al. 2003; Woodhouse et al. 2005). The societal importance of hydroclimatic events, coupled with evidence for recent trends towards increased desiccation (e.g., Seager et al. 2007; Touchan et al. 2010), has motivated a high level of interest in understanding the forcing and predictability of droughts and pluvials over North America (NA) (Herweijer et al. 2006; Seager et al. 2005b; Seager 2007; Schubert et al. 2009).
Experiments using general circulation models (GCM) have played a critical role in advancing our understanding of the dynamical mechanisms governing hydroclimatic variability (Hoerling and Kumar 2003; Schubert et al. 2009; Seager et al. 2003, 2005a; Seager 2007). When forced with either observed (Herweijer et al. 2006; Seager et al. 2005b) or idealized (Schubert et al. 2009) sea surface temperature (SST) patterns, GCMs are generally able to reproduce observed or expected patterns of drought with high fidelity (Herweijer et al. 2006; Seager et al. 2005b), albeit with some notable exceptions over NA (Cook et al. 2008; 2009; Seager et al. 2008a,b). Specifically, these experiments have helped illuminate the important role played by SSTs in the Pacific (Herweijer et al. 2006, 2007; Seager et al. 2005b; Seager 2007) and North Atlantic (Enfield et al. 2001; Kushnir et al. 2010; Mo et al. 2009; Wang et al. 2006) ocean basins.
Using GCMs to study drought and pluvial dynamics provides several major advantages over empirically based approaches. Based upon first principals, results from GCMs can be independently compared and verified against observational data sets and empirical studies, providing direct and independent insight into the physical mechanisms. GCMs are also informative for historical studies during time periods when observations are sparse or unavailable, including droughts during the Medieval Climate Anomaly (Feng et al. 2008; Seager et al. 2008a) and the nineteenth century (Herweijer et al. 2006). In addition to being forced by realistic boundary conditions and forcings, these models can also be driven with idealized forcings that are often larger or more persistent than observed (Schubert et al. 2009), maximizing the signal to noise ratio and allowing for clear dissection of causal mechanisms. These advantages are balanced by a number of significant shortcomings. GCMs are computationally expensive, limiting the number of ensembles or experiments that can be conducted. The models themselves are imperfect, often missing important processes or simulating existing processes incorrectly. Finally, it can be difficult to satisfyingly diagnose causal mechanisms in the models themselves when dealing with competing sources of variance, especially in cases where the signal to noise ratio is low.
Most statistically based studies of drought and pluvial dynamics over NA have focused on describing simple statistical associations between drought variability and various forcing factors (e.g., Andreadis et al. 2005; Cook et al. 1997; Mo and Schemm 2008). Here we investigate the variability and forcing of drought and pluvial events over NA using a statistically based modeling approach that allows for large ensemble simulations and empirical separation of modes representing distinct SST related sources of drought variability. Our study offers a statistically based complement to existing research using GCM experiments, and we use previous insights into the dynamics and physics of drought forcing to guide the interpretation of results from our analyses. We focus on the four persistent events during the twentieth century, events that, for the most part, have been extensively investigated using GCM modeling experiments and are typical of events that have occurred in the more distant past. Our analysis is organized around two research questions: (1) what is the relative importance of various regions of SST forcing for simulating these drought and pluvial events? and (2) how much of the variability during these events can be ascribed to the SST forcing?
2 Data and methods
2.1 North American Drought Atlas
The North American Drought Atlas (NADA) (http://www.ncdc.noaa.gov/paleo/pdsi.html) version 2a (Cook et al. 2007) is a tree ring proxy based reconstruction of the Palmer Drought Severity Index (PDSI) (Palmer 1965). PDSI is an index of drought, incorporating supply (precipitation) and demand (evaporation as a function of temperature) information in it’s calculation. It is locally normalized around a mean of zero, so that the standardized anomalies between different regions are directly comparable. Positive values of PDSI indicate wetter than normal (pluvial) conditions and negative values indicate drier (drought) conditions, with a typical range in values from −5 to +5. Tree ring chronologies and PDSI are effective at integrating over large spatial and temporal scales, making them ideal for use in studies of regional drought variability at seasonal timescales. The NADA product reconstructs PDSI for the summer (June–July–August) season using 1,821 tree ring chronologies, over a maximum of 286 2.5° × 2.5° gridpoints. This product is well validated and versions of the NADA have been used extensively in the study of North American drought variability (Cook et al. 1999; 2004; 2007; Fye et al. 2003; Herweijer et al. 2007). To develop our statistical model, we use data from the drought atlas for years 1400 C.E.–2005 C.E., representing 251 continuous in time grid point locations. The time span chosen represents a tradeoff between the number of years available in the NADA and the spatial coverage of the dataset, which progressively decreases further back in time. While our resulting analysis is focused on the four twentieth century events, using over 600 years in our model development helps us maximize the representativeness of our modes of drought variability in our analysis and model (described next section). Our region of interest is western North America (125°W − 95°W, 25°N − 50°N), a region of high drought variability encompassing the primary features of all the major twentieth century events. This region is outlined (dashed box) in Figs. 6, 7, 8 and 9.
2.2 Principal component analysis and modeling
In the drought modeling literature, it is common to conduct GCM experiments using idealized SST patterns (e.g., Schubert et al. 2009; Seager et al. 2005b). In a typical experiment, observed SSTs might be prescribed over the tropical Pacific or North Atlantic, with the rest of the ocean held to climatological SSTs or modeled as a thermodynamic mixed layer. Idealized SST patterns representing phases of ENSO or other SST modes are also often used. The point of such experiments is not to necessarily generate the most realistic patterns of observed drought variability; rather they are used to inform the underlying dynamics and the importance of various sources of forcing for specific drought and pluvial events. Our modeling approach and philosophy is quite similar. But instead of using a GCM with idealized SST patterns, we use a statistical model with empirically based modes of drought variability that represent distinct regions of SST forcing.
For all our modeling scenarios, we either prescribe (i.e., use the original PC scores without modification) the three SST related PCs (1, 2, and 3), or model them by randomizing their phase information, preserving the spectral characteristics and serial autocorrelation of the underlying PC time series (Ebisuzaki 1997). PCs 4 and 5 are always phase randomized to preserve any underlying autocorrelation or frequency structure, and within our study are always considered to represent spatially coherent, but unforced, variability. For these five PC time series, the associated EOF loadings are used to constrain the spatial expression of the PC scores and are not modified. The residual variance, beyond the leading five PC scores, is modeled as spatially uncorrelated white noise, resampled using Gaussian distributions derived at each grid cell separately. Regardless of whether PCs 1, 2, or 3 are modeled or prescribed, all five leading PCs are represented and the total variance of the original PDSI dataset is preserved. For each experiment, we conduct a 2,000 member ensemble in order to sufficiently sample the noise space, although the results generally converge after about 100–200 resamples. To quantify the match between our modeled PDSI fields and the PDSI from the NADA, we use the anomaly correlation statistic (AC) (Wilks 2006), the product–moment coefficient of linear correlation between variables corresponding to the same location. The AC is commonly used in the verification of forecast fields and can be interpreted identically as a Pearson correlation.
For our null hypothesis (that the occurrence of any drought or pluvial event is random), we simulated a full random ensemble, modeling all five leading PC scores and residual PDSI. We then calculated the AC between the NADA PDSI and the modeled PDSI for each null case ensemble member and each drought and pluvial event. From the full histogram of ensemble ACs, we use the AC associated with the 95th percentile to establish our noise confidence limit. In other words, in the forced simulations (described next), any ensemble member with a higher AC than this null limit is considered to be skillfull at reproducing the drought/pluvial event of interest beyond simply random noise.
For the forced ensemble experiments, we alternately prescribe the first three PC scores (representing the forced component of PDSI variability). By examining the ensemble median and comparing these ensembles against the null case, we can determine to what extent different droughts and the pluvial are dominated by different forcing, and how well (and often) that forcing can reproduce the various drought and pluvial events above the random noise. Our forcing cases (described later) are PC 1 (tropical Pacific forcing), PCs 1–2 (tropical + North Pacific forcing), and PCs 1–2–3 (Pacific + Atlantic forcing).
3.1 Observed twentieth century drought and pluvial events
3.2 PCA results
These sources of drought forcing (tropical Pacific, North Pacific, and Atlantic) have been identified in previous analyses (McCabe et al. 2004, 2008), providing further support that our selected PC/EOF pairs are dynamically meaningful in the context of drought variability over NA. The remaining two PCs (4 and 5) show few teleconnections to SSTs in any ocean basin (not shown), and their associated EOF loadings are difficult to interpret within the context of drought dynamics over NA. For our modeling exercise, we use these leading five PC/EOF pairs to represent the spatially coherent variability in the PDSI data, with the potential SST forced component of drought variability contained within the first three.
3.3 Modeling: ensemble median patterns
Figures 6, 7, 8 and 9 show the median PDSI patterns for each drought and pluvial event from our forced ensemble simulations: prescribed PC 1 (tropical Pacific forcing), PCs 1–2 (tropical + North Pacific forcing) and PCs 1–2–3 (Pacific + Atlantic forcing). The AC statistics comparing the ensemble median PDSI pattern and the PDSI pattern from the NADA are in the lower left corners of the model panels, calculated over the region bounded by the dashed lines (125°W − 95°W, 25°N − 50°N). As mentioned previously, the separation of PCs 1 and 2 may be somewhat artificial because of the orthogonality constraints imposed by the PCA. We therefore focus primarily on comparing the tropical+North Pacific case and the Pacific + Atlantic case.
3.4 Modeling: ensemble statistics
Increasing the level of forcing from PCs 1–2 (tropical + North Pacific forcing, red lines) to PCs 1–2–3 (Pacific + Atlantic forcing, blue lines) results in both increased ensemble median anomaly correlations and also the percentage of individual ensemble members that exceed the 95% significance threshold. Tropical + North Pacific forcing alone results in over 90% of the individual ensemble members passing the noise threshold for the 1905–1917 pluvial (96.6%) and the 1948–1957 drought (94.5%). The addition of the Atlantic forcing increases the proportion of passing ensemble members only marginally (99.0 and 96.3% for this pluvial and drought, respectively). For the 1998–2002 drought, forcing from the Pacific dominates (66.7% passing in full Pacific forcing case), but the Atlantic plays an important secondary role (82.1% passing in the Pacific + Atlantic case). This improvement is not reflected in the ensemble median, which only increases from 0.70 to 0.73 for the full Pacific and Pacific + Atlantic cases, respectively.
Of all four events, Atlantic forcing appears to be most important during the 1932–1939 Dust Bowl drought. In the tropical + North Pacific simulation, the ensemble median AC = 0.45, and less than one third (28.3%) of ensemble members pass the 95% threshold. With the addition of Atlantic forcing, the median AC increases to 0.61 and the number of ensemble members passing the noise limit nearly doubles (53.8%). Despite these improvements, however, the fidelity of the Dust Bowl simulations is much weaker compared to the other events. This supports several recent studies (Brönnimann et al. 2009; Cook et al. 2008, 2009, 2010; Hoerling et al. 2009), highlighting the importance of non-SST related factors during the Dust Bowl.
3.5 Model uncertainties
We tested the sensitivity of our modeling framework to our PCA methodology. Conducting the PCA on the correlation matrix gave similar results, although with a lower fraction of the variance in the leading modes, diminished EOF loadings in the west, and higher EOF loadings in the eastern US. This arises because, in the original PDSI dataset, the variance is higher in the west and the southwest compared to the eastern portion of the domain. A covariance matrix based PCA therefore weights the results towards the west and southwest; using the correlation matrix normalizes the variance, and more evenly distributes the loadings. Because we were most interested in drought in the western US, we chose to use results from the covariance matrix based PCA. We also experimented with varimax rotation of our EOFs, but found the results highly sensitive to our truncation criteria. Results from our unrotated analysis show good consistency with other North American drought studies that have found similar EOF loading patterns (Herweijer et al. 2007; McCabe et al. 2004, 2008). Given this similarity to other studies, and because we could find no compelling physically based reason to rotate, we decided to use results from the unrotated analysis.
4 Discussion and conclusions
Forcing from the tropical + North Pacific is sufficient to reproduce both the early twentieth century pluvial (1905–1917) and the 1950s drought (1948–1957). The median drought pattern from the tropical +North Pacific forcing ensemble captures both the spatial pattern and intensity of these events. The addition of Atlantic forcing does little to improve either the ensemble median drought/pluvial patterns or the proportion of ensemble members passing the 95% noise threshold.
For the 1998–2002 drought, the ensemble median from the tropical + North Pacific case generally captures the spatial distribution of the drought. The addition of Atlantic forcing has little impact on the median drought pattern, but substantially increases the proportion of ensemble members that exceed the noise threshold (66.7–82.1% passing). Even with Atlantic forcing, however, the median simulated drought pattern does not reach the intensity of the observed drought, implying other factors may be responsible.
In the Dust Bowl (1932–1939), tropical + North Pacific forcing does a poor job of reproducing the drought, with a median drought pattern centered too far west and less than one third of all ensemble members passing the noise threshold. Addition of Atlantic forcing improves the median drought pattern, and nearly doubles the percent of ensemble members passing the 95% threshold (from 28.3 to 53.7%). Even with the Atlantic forcing, however, the fidelity of our simulation of the Dust Bowl is much weaker than for the other droughts, emphasizing the important role of non SST forcing or internal noise for shaping anomalies during the Dust Bowl drought.
Of all events, Atlantic forcing appears to have been most important during the Dust Bowl drought, resulting in the biggest improvement in both the ensemble median and ensemble spread compared to the tropical +North Pacific forcing only case. Even then, only about half of the ensemble members from the Pacific + Atlantic forcing case exceed the 95% noise threshold; this (as well as the ensemble median AC) is much lower than comparable simulations of the other droughts and pluvial. This suggests a nearly equally important role for non SST sources of variability during the Dust Bowl. Whether this variability comes from internal noise in the climate system or is forced from other sources (e.g., the land surface) is currently widely debated (Brönnimann et al. 2009; Hoerling et al. 2009). Recent GCM based studies, however, support the idea that land surface and dust aerosols feedbacks during this period were important (Cook et al. 2008, 2009, 2010). Land degradation and subsequent dust storm activity during the Dust Bowl drought was widespread, intensive, and unprecedented over at least the instrumental period (Hansen and Libecap 2004). In the dust feedback hypothesis, increased dust aerosol loadings in the atmosphere from these dust storms led to a ‘Charney’ type feedback, with increased shortwave reflection, reduced net radiation at the surface and top of the atmosphere, and compensatory subsidence that inhibited precipitation. To adequately address this within the current modeling framework, however, we would need a way to statistically diagnose this variability coming from the land surface and dust aerosols. This is beyond the capabilities of the relatively simple model we have developed here.
Finally, results from the 1998–2002 ensemble suggest that, in some cases, the ensemble mean or median may not be the best metric for determining which factors are important during a given drought event. For this drought, the addition of Atlantic forcing did little to improve the ensemble median drought pattern, but did substantially increase the chance of any give ensemble member exceeding the noise threshold. This, in particular, highlights the potential problem with limited ensemble sizes in GCM experiments, which may not have a large enough sample size to resolve this shift in the distribution of the ensemble.
This project received support from the Climate Dynamics Program of the National Science Foundation under ATM-06-20066. The authors also wish to thank anonymous reviewers who greatly improved the quality of this manuscript. Lamont contribution number 7387.