Toby R. Ault [email protected] http://ecrl.eas.cornell.edu/

Relative Impacts of Mitigation, Temperature, and Precipitation on 21st- Century Megadrought Risk in the American Southwest

Part I What are we talking about when we talk about megadrought?

Part II Megadrought risk in a changing climate.

Part III Connection to seasonal prediction. “A megadrought (or mega-) is a prolonged drought lasting two decades or longer… They are suspected of playing a primary role in the collapse of several pre-industrial , including the Anasazi of the North American Southwest, the Khmer Empire of Cambodia, the Mayan of Mesoamerica, the Tiwanaku of Bolivia, and the Yuan Dynasty of China. As such they constitute one of the greatest threats to human .”

http://en.wikipedia.org/wiki/Megadrought “Decadal drought” aai h 1ya en(etclga bars). gray (vertical mean year 11 the in data etr rcptto aafrom: data precipitation century indt r rmteUiest fEs nlasCiaeRese as Climate Anglia’s decadal identify East We of 2005). University Jones the and (Mitchell from are data tion Fig. b) a) .Eee-errnigmaso omlzdplolmt re paleoclimate normalized of means running Eleven-year 2.

mm/year mm/year GreatPlainsprecipitation Southwestprecipitation 150 200 250 300 300 350 400 450

1910 1920 1930 a) 1940 h SSuhet and Southwest, US the 1950 40 Year 1960 -0.5 σ 1970 in in decadal mean(11yr) − b) 0 . 1980 5s t a n d a r dd e v i a t i o n so fr a w h ra lis Precipita- Plains. Great the rhUi’ S. dataset TS3.1 Unit’s arch Drought threshold Smooth precip. Raw precip. Decadal drought osrcin o 20th for constructions 1990 2000

Photo credit: Scott St. George Megadroughts – defined by duration

North American Drought Atlas (Cook et al., 2004)

Multidecadal megadroughts -0.5σ in multi-decadal (35yr) mean R EPORTS fered from severe drought, as well as ex- structions that extend back to AD 800. These United States (8, 9), but now with a much tensive areas of Mexico, particularly in the reconstructions were produced with a well- denser network of 602 centuries- to millennia- northern and western parts of the country. In tested principal components regression pro- long tree-ring chronologies used as predic- many of these areas, the 2002 drought was cedure developed previously to robustly tors of PDSI (7). As before (8), a split cal- actually part of an ongoing drought that started reconstruct drought across the coterminous ibration and verification scheme was used to in late 1999 or before, with widespread drought conditions already persisting for È3 Fig. 1. The North years. Drought abated in many areas by late American summer 2002 to early 2003, but severe drought drought reconstruction conditions have continued to affect the in- grid and the (mostly) western U.S. region, terior western United States throughout the circumscribed by the 2004 summer (2). thick black irregular This drought highlights both the extreme polygon. Out of the vulnerability of the semi-arid western United 286 total grid points, States to precipitation deficits and the need 103 are contained with- to better understand long-term drought var- in the Western region. Each of the 103 grid iability and its causes in North America. To points has an annu- this end, we have used centuries-long, an- ally resolved drought nually resolved tree-ring records to recon- reconstruction that struct annual changes in both drought and extends back to at wetness over large portions of North Amer- least AD 1380, with ica. The reconstructed drought metric is the 68 extending back to AD 800. These form summer-season Palmer Drought Severity In- the basis for our recon- dex (PDSI) (3), a widely used measure of struction of area af- relative drought and wetness over the United fected by drought in States (4) and other global land areas (5, 6). the West (7). New PDSI reconstructions have been pro- on March 25, 2011 duced on a 286-point 2.5- by 2.5- regular grid Fig. 2. (A) Smoothed LONG-TERM CHANGES IN DROUGHT AREA IN THE WEST (Fig. 1) (7). This grid covers most of North DAI reconstruction 100 America and is a substantial expansion of an (solid black curve) for A THE CENTRAL DATES OF THE the West, showing 80 1150 earlier 155-point 2- by 3- drought reconstruc- 936 1034 1253 SIGNIFICANT (p<0.05) EPOCHS tion grid that covered only the coterminous two-tailed 95% boot- ARE INDICATED WITH ARROWS strap confidence inter- 60 United States (8). In addition, our drought vals (dashed black reconstructions are 600 to 1200 years long curves) and the long- over much of the U.S. portion of the North term mean (thin hor- 40 American grid (particularly in the western izontal black line). Sixty- DRIER www.sciencemag.org United States), a substantial increase over the year smoothing was % DROUGHT AREA 20 applied to highlight È300 years available from the previously the multidecadal to 0 1613 1829 published reconstructions, which all began in WETTER 1321 1915 centennial changes in

1700. Finally, the variance restoration we aridity. The four driest 800 900 1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 2000 apply to the grid point reconstructions (7) epochs (P G 0.05, those YEAR allows for updates of those records to AD with confidence limits 2003 with instrumental PDSI data. Together, above the long-term mean in Fig. 2A) are be- DROUGHT TREND: 1900 - 2003 these attributes enable us to compare the 100 fore AD 1300, whereas Downloaded from current western U.S. drought to those that the four wettest (P G B are reconstructed to have occurred as far back 80 CURRENT 0.05) epochs occur af- DROUGHT as AD 800, a time period that includes the so- ter that date. The dif- called (MWP). ference between the 60 means of the AD 900 The region of interest here is contained 40 within an irregular black polygon on the to 1300 period (red line, 42.4%) and AD 1900 to

North American grid (Fig. 1), an area that 2003 period (blue line, % DROUGHT AREA 20 we henceforth refer to simply as the West. 30%) are also appar- Each of the 103 grid points in the West has a ent. The 12.4% differ- 0 1900 1920 1940 1960 1980 2000 summer PDSI reconstruction that covers the ence between the two YEAR common interval AD 1380 to 1978, with a periods translates into minimum of 68 grid points having recon- an average drought area (PDSI G –1) increase of 41.3% in the West during the earlier period. This difference is statistically significant (P G 0.001) given an equality-of-means t test with degrees of freedom corrected for first-order autocorrelation. Even so, some of the AD 900 to 1300 period 1Lamont-Doherty Earth Observatory, Palisades, NY PDSI estimates are extrapolations, because they fall outside the range of the instrumental PDSI 10964, USA. 2National Oceanic and Atmospheric data in the AD 1928 to 1978 calibration period (7). As regression-based estimates, these Administration (NOAA), National Climatic Data extrapolations have greater uncertainty compared to those that fall within the range of the Center, Boulder, CO 80305, USA. 3Laboratory of calibration period. However, they are still based on the actual growth histories of highly drought- Tree-Ring Research, University of Arizona, Tucson, 4 sensitive trees. Therefore, we argue that our DAI reconstruction is indicative of what really AZ 85721, USA. Department of Geosciences, Uni- happened in the West, even during the AD 900 to 1300 period of elevated aridity (7). (B) The versity of Arkansas, Fayetteville, AR 72701, USA. annually resolved AD 1900 to 2003 portion, which more clearly reveals the severity of the *To whom correspondence should be addressed. current drought relative to others in the 20th century and an irregular trend (red smoothed E-mail: [email protected] curve) toward increasing aridity since 1900.

1016 5 NOVEMBER 2004 VOL 306 SCIENCE www.sciencemag.org Reconstructed Colorado River Streamflow

L10705 MEKO ET AL.: MEDIEVAL DROUGHT IN UPPER COLORADO RIVER BASIN L10705

Figure 2. Time series plot of 25-year running mean of reconstructed flows. Flows are plotted as percentage of the 1906– 2004 mean of observed natural flows (18.53 billion cubic meters, or 15.03 million acre-ft). Confidence interval derived from 0.10 and 0.90 probability points of ensemble of 1000 noise-added reconstructions. Horizontal dashed line is lowest 25-year running mean of observed flows (1953–1977).

[8] A 25-year running mean of the reconstructed flows at least for this level of smoothing – conditions in the mid- illustrates the overall importance of variations at multi- 1100s in the UCRB were even drier than during the decadal time scales, and identifies intervals of amplified extremely widespread late-1500s North American mega- low-frequency variance (Figure 2). The most recent such drought [e.g., Stahle et al., 2000]. If ‘‘normal’’ is defined interval began in the mid-1800s and has continued to as the observed mean annual flow for 1906–2004, the present. The beginning of the tree-ring record is also anomalous flow for A.D. 1130–1154 was less than 84% characterized by amplified low-frequency variance, as of normal. By comparison, the lowest 25-year mean of evidenced by the swing from wet conditions to dry con- observed flows (1953–1977) was 87% of normal. Because ditions beginning in the early A.D. 800s. Running means regression biases the reconstructed flows toward the cali- are persistently below normal for most of the 9th century. bration-period mean, flows in the mid-1100s were quite The most prominent feature of the smoothed long-term possibly lower than indicated by the reconstruction. For reconstruction is the major period of low flow in the mid- example, the 80% confidence band plotted in Figure 2 1100s. The lowest reconstructed 25-year running mean suggests a greater than 10% probability that the true mean occurred in A.D 1130–1154. These results suggest that – for A.D. 1130–1154 was as low as 79% of normal.

Figure 3. Runs properties of 1100s drought. (a) Time series of reconstructed flow in units of billion cubic meters (BCM) for segment A.D. 1098–1202. Horizontal line at 18.53 BCM is observed mean for 1906–2004. (b) Time series of runs below the observed mean flow. Bars mark runs of two-or-more years. Run-length annotated below bar. Run-sum (cumulative departure from mean) given by length of bar.

3of5 Reconstructed Colorado River Streamflow

L10705 MEKO ET AL.: MEDIEVAL DROUGHT IN UPPER COLORADO RIVER BASIN L10705

Figure 2. Time series plot of 25-year running mean of reconstructed flows. Flows are plotted as percentage of the 1906– 2004 mean of observed natural flows (18.53 billion cubic meters, or 15.03 million acre-ft). Confidence interval derived from 0.10 and 0.90 probability points of ensemble of 1000 noise-added reconstructions. Horizontal dashed line is lowest 25-year running mean of observed flows (1953–1977).

[8] A 25-year running mean of the reconstructed flows at least for this level of smoothing – conditions in the mid- illustrates the overall importance of variations at multi- 1100s in the UCRB were even drier than during the decadal time scales, and identifies intervals of amplified extremely widespread late-1500s North American mega- low-frequency variance (Figure 2). The most recent such drought [e.g., Stahle et al., 2000]. If ‘‘normal’’ is defined interval began in the mid-1800s and has continued to as the observed mean annual flow for 1906–2004, the present. The beginning of the tree-ring record is also anomalous flow for A.D. 1130–1154 was less than 84% characterized by amplified low-frequency variance, as of normal. By comparison, the lowest 25-year mean of evidenced by the swing from wet conditions to dry con- observed flows (1953–1977) was 87% of normal. Because ditions beginning in the early A.D. 800s. Running means regression biases the reconstructed flows toward the cali- are persistently below normal for most of the 9th century. bration-period mean, flows in the mid-1100s were quite The most prominent feature of the smoothed long-term possibly lower than indicated by the reconstruction. For reconstruction is the major period of low flow in the mid- example, the 80% confidence band plotted in Figure 2 1100s. The lowest reconstructed 25-year running mean suggests a greater than 10% probability that the true mean occurred in A.D 1130–1154. These results suggest that – for A.D. 1130–1154 was as low as 79% of normal.

Figure 3. Runs properties of 1100s drought. (a) Time series of reconstructed flow in units of billion cubic meters (BCM) for segment A.D. 1098–1202. Horizontal line at 18.53 BCM is observed mean for 1906–2004. (b) Time series of runs below the observed mean flow. Bars mark runs of two-or-more years. Run-length annotated below bar. Run-sum (cumulative departure from mean) given by length of bar.

3of5 “Worst” megadrought from N. Amer. Drought Atlas (NADA)

Data from Cook et al., 2004 L05703 STAHLE ET AL.: MESOAMERICAN DROUGHTS L05703

Figure 2. (a) Tree‐ring reconstructed June PDSI for Mesoamerica (annual and smoothed estimates highlighting 30‐year departures), (b) EPS statistics for 100‐year segments, (c) sample size, and 50‐year subperiods illustrating the (d) Terminal Classic, (e) Toltec, (f) Aztec, and (g) Conquest‐era droughts. estimates from Mesoamerica based on sedimentary and [9] The June PDSI reconstruction also provides valuable, speleothem records, even though these other proxies are exactly dated information on early growing season climate dated with less accuracy and precision than is possible conditions during the rise and fall of pre‐Hispanic civiliza- with . The titanium record from lami- tions in Mesoamerica. For example, the reconstruction nated sediments at Laguna de Juanacatlan, Jalisco, is indicates that the Terminal Classic drought extended into the believed to represent a proxy for warm season rainfall central Mexican altiplano from AD 897–922, where it was over west‐central Mexico and covers the last 2000 years one of the worst megadroughts of the past 1200‐years based on an age‐depth model constrained by 26 AMS (Figures 2a and 2d). This may represent at least part of the radiocarbon dates [Metcalfe et al., 2010]. Within the same episode of extended drought reported from lake sedi- dating uncertainties of the titanium record (not shown), ments and speleothems in the Yucatan and Jalisco, within there are indications of drier conditions in Jalisco during the constraints of the various age estimates [Hodell et al., some of the megadroughts reconstructed from Queretaro, 1995; Medina‐Elizalde et al., 2010; Metcalfe et al., 2010]. especially during the early 10th, early 12th, and late The reconstruction also identifies Late Classic droughts 14th centuries (Figure 2a) [Metcalfe et al., 2010]. The centered at AD 810 and 860, which were previously early 16th century drought (1514–1539) is only partially reconstructed for the Caribbean sector from laminated synchronous with drier conditions in Jalisco, and histori- sediments in the Cariaco Basin [Haug et al., 2003]. Whether cal information indicates rising lake levels in the Valley adverse climate conditions were a key factor in the decline of Mexico from 1517–1524 and declining levels thereafter of Mayan city states during the Terminal Classic period is [O’Hara and Metcalfe, 1997]. The degree to which these contentious and alone cannot explain the complex chro- comparisons might be biased by dating uncertainties, nology of regional cultural changes identified in the differences in the seasonal climate response among the archaeological record during the late Classic/early Post‐ proxies, or real spatial differences in pre‐Hispanic or Classic periods [Demarest et al., 2004]. But the new Colonial climate over central Mexico will have to be reconstruction from Queretaro confirms the Terminal Clas- resolved with further research. But tree‐ring dating should sic drought, documents its penetration into the highlands of play an increasingly important role in the reconstruction central Mexico, and narrows the timing of this central of late Holocene climate over Mesoamerica because 900‐ Mexican megadrought to AD 897–922. to 1500‐year old Montezuma baldcypress have been [10] The Toltec state was the dominant imperial civiliza- sampled at other sites in central and southern Mexico and tion of central Mexico during the early Post‐Classic era and efforts to build exact chronologies are underway. archaeological, chronometric, and historical data indicate that the collapse of Tula occurred ca. 1150 [Diehl, 1983], a

3 of 4 L05703 STAHLE ET AL.: MESOAMERICAN DROUGHTS L05703

Decadal Drought (e.g., 30s, 50s): 1-2 per century Multidecadal Megadrought (>35 years): 1-2 per millennium

Figure 2. (a) Tree‐ring reconstructed June PDSI for Mesoamerica (annual and smoothed estimates highlighting 30‐year departures), (b) EPS statistics for 100‐year segments, (c) sample size, and 50‐year subperiods illustrating the (d) Terminal Classic, (e) Toltec, (f) Aztec, and (g) Conquest‐era droughts. estimates from Mesoamerica based on sedimentary and [9] The June PDSI reconstruction also provides valuable, speleothem records, even though these other proxies are exactly dated information on early growing season climate dated with less accuracy and precision than is possible conditions during the rise and fall of pre‐Hispanic civiliza- with dendrochronology. The titanium record from lami- tions in Mesoamerica. For example, the reconstruction nated sediments at Laguna de Juanacatlan, Jalisco, is indicates that the Terminal Classic drought extended into the believed to represent a proxy for warm season rainfall central Mexican altiplano from AD 897–922, where it was over west‐central Mexico and covers the last 2000 years one of the worst megadroughts of the past 1200‐years based on an age‐depth model constrained by 26 AMS (Figures 2a and 2d). This may represent at least part of the radiocarbon dates [Metcalfe et al., 2010]. Within the same episode of extended drought reported from lake sedi- dating uncertainties of the titanium record (not shown), ments and speleothems in the Yucatan and Jalisco, within there are indications of drier conditions in Jalisco during the constraints of the various age estimates [Hodell et al., some of the megadroughts reconstructed from Queretaro, 1995; Medina‐Elizalde et al., 2010; Metcalfe et al., 2010]. especially during the early 10th, early 12th, and late The reconstruction also identifies Late Classic droughts 14th centuries (Figure 2a) [Metcalfe et al., 2010]. The centered at AD 810 and 860, which were previously early 16th century drought (1514–1539) is only partially reconstructed for the Caribbean sector from laminated synchronous with drier conditions in Jalisco, and histori- sediments in the Cariaco Basin [Haug et al., 2003]. Whether cal information indicates rising lake levels in the Valley adverse climate conditions were a key factor in the decline of Mexico from 1517–1524 and declining levels thereafter of Mayan city states during the Terminal Classic period is [O’Hara and Metcalfe, 1997]. The degree to which these contentious and alone cannot explain the complex chro- comparisons might be biased by dating uncertainties, nology of regional cultural changes identified in the differences in the seasonal climate response among the archaeological record during the late Classic/early Post‐ proxies, or real spatial differences in pre‐Hispanic or Classic periods [Demarest et al., 2004]. But the new Colonial climate over central Mexico will have to be reconstruction from Queretaro confirms the Terminal Clas- resolved with further research. But tree‐ring dating should sic drought, documents its penetration into the highlands of play an increasingly important role in the reconstruction central Mexico, and narrows the timing of this central of late Holocene climate over Mesoamerica because 900‐ Mexican megadrought to AD 897–922. to 1500‐year old Montezuma baldcypress have been [10] The Toltec state was the dominant imperial civiliza- sampled at other sites in central and southern Mexico and tion of central Mexico during the early Post‐Classic era and efforts to build exact chronologies are underway. archaeological, chronometric, and historical data indicate that the collapse of Tula occurred ca. 1150 [Diehl, 1983], a

3 of 4 L05703 STAHLE ET AL.: MESOAMERICAN DROUGHTS L05703

But what about ?

Figure 2. (a) Tree‐ring reconstructed June PDSI for Mesoamerica (annual and smoothed estimates highlighting 30‐year departures), (b) EPS statistics for 100‐year segments, (c) sample size, and 50‐year subperiods illustrating the (d) Terminal Classic, (e) Toltec, (f) Aztec, and (g) Conquest‐era droughts. estimates from Mesoamerica based on sedimentary and [9] The June PDSI reconstruction also provides valuable, speleothem records, even though these other proxies are exactly dated information on early growing season climate dated with less accuracy and precision than is possible conditions during the rise and fall of pre‐Hispanic civiliza- with dendrochronology. The titanium record from lami- tions in Mesoamerica. For example, the reconstruction nated sediments at Laguna de Juanacatlan, Jalisco, is indicates that the Terminal Classic drought extended into the believed to represent a proxy for warm season rainfall central Mexican altiplano from AD 897–922, where it was over west‐central Mexico and covers the last 2000 years one of the worst megadroughts of the past 1200‐years based on an age‐depth model constrained by 26 AMS (Figures 2a and 2d). This may represent at least part of the radiocarbon dates [Metcalfe et al., 2010]. Within the same episode of extended drought reported from lake sedi- dating uncertainties of the titanium record (not shown), ments and speleothems in the Yucatan and Jalisco, within there are indications of drier conditions in Jalisco during the constraints of the various age estimates [Hodell et al., some of the megadroughts reconstructed from Queretaro, 1995; Medina‐Elizalde et al., 2010; Metcalfe et al., 2010]. especially during the early 10th, early 12th, and late The reconstruction also identifies Late Classic droughts 14th centuries (Figure 2a) [Metcalfe et al., 2010]. The centered at AD 810 and 860, which were previously early 16th century drought (1514–1539) is only partially reconstructed for the Caribbean sector from laminated synchronous with drier conditions in Jalisco, and histori- sediments in the Cariaco Basin [Haug et al., 2003]. Whether cal information indicates rising lake levels in the Valley adverse climate conditions were a key factor in the decline of Mexico from 1517–1524 and declining levels thereafter of Mayan city states during the Terminal Classic period is [O’Hara and Metcalfe, 1997]. The degree to which these contentious and alone cannot explain the complex chro- comparisons might be biased by dating uncertainties, nology of regional cultural changes identified in the differences in the seasonal climate response among the archaeological record during the late Classic/early Post‐ proxies, or real spatial differences in pre‐Hispanic or Classic periods [Demarest et al., 2004]. But the new Colonial climate over central Mexico will have to be reconstruction from Queretaro confirms the Terminal Clas- resolved with further research. But tree‐ring dating should sic drought, documents its penetration into the highlands of play an increasingly important role in the reconstruction central Mexico, and narrows the timing of this central of late Holocene climate over Mesoamerica because 900‐ Mexican megadrought to AD 897–922. to 1500‐year old Montezuma baldcypress have been [10] The Toltec state was the dominant imperial civiliza- sampled at other sites in central and southern Mexico and tion of central Mexico during the early Post‐Classic era and efforts to build exact chronologies are underway. archaeological, chronometric, and historical data indicate that the collapse of Tula occurred ca. 1150 [Diehl, 1983], a

3 of 4 7534 JOURNAL OF CLIMATE VOLUME 27

FIG. 4. Map of projected precipitation over the twenty-first-century (2005–2100) change in the RCP8.5 scenario shown as a percentage of twentieth-century precipitation change [as in the global maps of Diffenbaugh and Giorgi (2012)]. For each model, the number of available runs from each experiment is shown in parentheses in the following order: historical, piControl, RCP2.6, RCP4.5, and RCP8.5. The red box shows the greater southwestern United States to emphasize the focus of this study (308–408N, 1208–1038W). depend on their prior states (i.e., they have ‘‘memory’’). Wells et al. 2004). Similarly, the standardized pre- For example, PDSI models the surface water balance cipitation index (SPI) integrates anomalies over a num- using a simplified approximation of soil moisture, and ber of predefined lags to measure how aggregated rainfall has a built-in autocorrelation function (e.g., Alley 1984; anomalies deviate from their long-term averages. Small ensemble size…

1. Estimate forced response from (multi) model mean Pr {X=q}

“X” Small ensemble size…

1. Estimate forced response from (multi) model mean

2. Fill out PDF with Monte Carlo Emulators

“X” 7536 JOURNAL OF CLIMATE VOLUME 27

Fourier coefficients so that they are approximately power-law distributed:

1 2b/p2 if k 5 0 8 N ffiffi c 5 >   (4) k > > k 2b/p2 < otherwise. N ffiffi >   > Here the value:> of b is divided by p2 because it is being applied to the raw Fourier coefficients, which have am- ffiffiffi plitudes proportional to the square root of the power spectrum. The rescaled Fourier series X~p(k) is then used to generate power-law time series Xp(t) by taking the real part of the inverse Fourier transform of X~p(k):

N21 1 ~ 2i2pk[(t21)/N] Xp(t) 5 Re å Xp(k)e , t 5 1, ... , N . (N k50 ) (5) Small ensemble size… Finally, the mean and variance are restored to the values of the original white noise data (zero and unity, in this case). We used a value of 0.5 for b to rescale each realization

of Xw(t), which was suggested as an appropriate estimate 1. Estimate forced response from by Ault et al. (2013) from synthesis of tree-ring re- (multi) model mean constructions of precipitation, PDSI, and streamflow as well as non-tree-ring estimates of hydroclimate. As 2. Fill out PDF with a check, we calculated the power laws of the noises after Monte Carlo Emulators they had been rescaled. We found that the actual values of b varied from one realization to the next, but were generally between 0.4 and 0.6. This range agrees well 3. à Use paleo bracket low frequency with instrumental and paleoclimate estimates of this frequency characteristics parameter for the region, and is certainly within the observational uncertainty (Ault et al. 2013). Impor- FIG. 6. Statistics summarizing Monte Carlo simulations: (a) dis- tantly, time series with spectra scaled by power laws of tributions of years spent in decadal ($11 yr) drought conditions for ;0.5 will also appear to exhibit autocorrelation of about each type of noise (as a percentage of all realizations), (b) risk of at least one decadal ($11 yr) drought during any given 50-yr window 0.3, which in turn implies that the AR(1) and power-law in any realizations, and (c) risk of a multidecadal ($35 yr) drought realizations will behave very similarly on short time during any 50-yr window. Risk in (b) and (c) is expressed as scales, but not necessarily on longer ones (e.g., Pelletier a percentage of the total number of simulations. and Turcotte 1997; Ault et al. 2013). Finally, our use of power-law noises does not make any assumptions about the underlying climate dynamics governing the shape of awhitenoisetimeseriesXw(t), and filter it to conform to a predefined value of b: the power spectrum of hydroclimate: linear and non- linear processes alike may produce such spectral distri- N butions (Milotti 1995; Penland and Sardeshmukh 2012). ~ 2i2pk[(t21)/N] Xp(k) 5 ck å Xw(t)e , k 5 0, ... , N 2 1, Table 1 highlights a few key features of the two t51 models employed here. In particular, the noise models (3) used to estimate drought risk use parameters that do not vary across space, and all are scaled to the twentieth- where k are the standard Fourier frequencies and N is century mean and variance. The autocorrelation parameter

the length of the time series. The term ck rescales the of 0.3 is a middle-of-the-road value from the time series 7542 JOURNAL OF CLIMATE VOLUME 27

Ault et al., 2014 (JCLIM) FIG. 11. Multidecadal (.35 yr) megadrought risk estimates obtained from Monte Carlo simulations of projected pre- cipitation changes across all 27 CMIP5 models in three different climate change scenarios (columns) and for three different types of noises (rows). These maps express the average risk estimates obtained from Monte Carlo simulations of pre- cipitation in each model under three climate change scenarios. For each of the 27 individual CMIP5 models, risk is cal- culated as the percentage of the total number of Monte Carlo simulations (1000) that show a multidecadal megadrought. Here, those estimates of risk are averaged across the multimodel archive. 7542 JOURNAL OF CLIMATE VOLUME 27

Doesn’t factor in temperature…

FIG. 11. Multidecadal (.35 yr) megadrought risk estimates obtained from Monte Carlo simulations of projected pre- cipitation changes across all 27 CMIP5 models in three different climate change scenarios (columns) and for three different types of noises (rows). These maps express the average risk estimates obtained from Monte Carlo simulations of pre- cipitation in each model under three climate change scenarios. For each of the 27 individual CMIP5 models, risk is cal- culated as the percentage of the total number of Monte Carlo simulations (1000) that show a multidecadal megadrought. Here, those estimates of risk are averaged across the multimodel archive. Caveats (abbreviated list)

Models have systematic biases. à Examine multi-model ensemble

Ensemble sizes are too small. à Use Monte Carlo “Emulators”

PET estimates are uncertain à Don’t rely on just one.*

* In Cook et al., 2015 we used three; Cook et al., 2014 used eight; Smerdon et al, 2015 used four. 15 APRIL 2015 S M E15 R D A O NPRIL E T A2015 L . 3223 S M E R D O N E T A L . 3221 CanESM2

PET estimates are uncertain… PDSI – offline calculation (Thornthwaite, Penman-Monteith) CCSM4

Soil Moisture – model simulated

(Full Column Soil MoistureFIG. 8. A comparison & of30cm reconstructed PDSI_THSM) (black with gray shading) in the 4C region (see Fig. 1) with collocated model estimates of PDSI_TH (red), PDSI_PM (orange), normalized 30cmSM (black), and normalized FCSM (blue) during the historical and projection intervals. Results are shown for the first ensemble member of (top) the CanESM2 and (bottom) CCSM4 model simulations to allow direct visual comparisons between the reconstructed and model variance. All time series have been centered over the 1901–79 interval, the time period of common overlap. Reconstructions extend prior to 1800, but the 1800–2100 interval is chosen for ease of visual comparison.

models will not share consistently common features. particularly in the CCSM4 output). Larger differences These considerations therefore must inform attempts to are observed between the original time series, in which compare reconstructions and model simulations directly differences in twenty-first-century trends reduce the r2 over the interval of overlap and require more detailed values among the different soil-moisture metrics (in and specific analyses that are beyond the scope of this cases where trends are comparable among variables, the investigation. The bridging approach that we have out-FIGshared. 5. (left variances to in right) original Regional and detrended model time series estimates of PDSI_TH, PDSI_PM, normalized 30cmSM, and normalized FCSM during the lined herein nevertheless demonstrates the use ofhistorical the are similar). intervalFrom Secular (1901–2005) dryingSmerdon in PDSI_TH et in al., exceeds(top) 2015 thatthe (JCLIM) of CanESM2 and (bottom) CCSM4 models. The first ensemble member of each model is plotted twentieth century as a common interval for referencing all other modeled2 soil-moisture metrics over the twenty- reconstruction and model data in order to compare(interannualfirst centuryr inestimates all three regions, across a clear demonstration all five ensemble of members are shown in Figs. 6 and 7). In all cases, PDSI or soil-moisture normalizations earlier paleoclimatic intervals and model projectionsused of thethe1901–2005 tendency for PDSI_TH intervals to as overestimate a baseline, drying but time series are recentered from 1901 to 1979 to match the calibration/validation interval of the twenty-first century. during the projection interval. There are, however, re- the PDSIgional reconstructions. differences in the relative comparisons among b. Projection interval variables, as discussed below. Shared variance between the four soil-moisture vari- PDSI_TH includes secular trends in the 4C region that ables during the projection interval (2005–99) are given in some cases are larger in magnitude or opposite in sign in Figs. 6 and 7, whereas Figs. 8–10 provide comparisons relative to the three other metrics in both models (Fig. 8), between the four soil-moisture metrics for the firsta. en- Historicaldespite the fact thatinterval some of the other metrics compare the two PDSI variables and 30cmSM range between semble member from both models over the historical favorably to PSDI_TH over the full century or during about 0.5 and 0.75 across all regions. Comparisons be- interval through projection intervals. Shared variances specific intervals of time. PDSI_PM and 30cmSM compare for detrended time series over the projection interval, inThewell four throughout characterizations the projection interval in the of CCSM4 modeled soil-moisture tween the two PDSI variables and FCSM generally in- addition to the original time series, are also shown in model, whereas 30cmSM projects enhanced drying, rela- Figs. 6 and 7. The shared variances among all ofbalance the tive to PDSI_PM, in each in the region CanESM2 model. over FCSM the again historical interval dicate less shared variance than with 30cmSM because detrended soil-moisture metrics in the projection compares favorably to PDSI_PM and 30cmSM in the interval in all regions are generally comparable(1901–2005) to CCSM4 simulation, are whereas shown it suggests in aFig. wetting 5 trendfor in the CanESM2 and FCSM incorporates longer-scale variations and time those of the historical interval (with some exceptions,CCSM4the CanESM2 models. simulation Consistent that is not reflected with in any of the the observation-based lags that exceed the time scales that PDSI and near- estimates, these four measures of soil-moisture balance surface soil moisture more strongly sample. The depth yield internally consistent results over the twentieth cen- of sampling is not, however, the only factor, as indicated tury, and both formulations of PDSI reproduce modeled by the fact that shared variances between PDSI esti- soil moisture with high fidelity. Agreement between the mates and FCSM are larger than for 30cmSM in the SE two PDSI calculations is expected given that they are both region in the CanESM2 model and for some variable calibrated over the 1901–2012 period and therefore any pairings and ensemble members in both models in the unrealistic temperature-driven differences in the PDSI_TH NP. Moreover, FCSM spans a much greater depth in the calculation are minimized. PDSI_TH and PDSI_PM in fact CCSM4 model than in CanESM2, but the r2 values be- reveal a large amount of shared variance over the his- tween the PDSI variables and FCSM are comparable torical interval, matching or exceeding r2 values of 0.9 in or larger in CCSM4 than in CanESM2. Depth effects all regions over all ensemble members in both models are therefore not the only determining factor in the (Figs. 6 and 7). comparison. Comparisons between the two PDSI estimates and Comparisons between the two direct soil-moisture modeled soil moisture during the historical interval in- estimates (30cmSM and FCSM) are similar or worse dicate weaker but still high levels of shared variance within each model than the comparisons between the (Figs. 6 and 7). In most cases, PDSI_TH and PDSI_PM two PDSI estimates and either of the soil-moisture vari- compare best with 30cmSM. The shared variance is ables, with the one exception being in the CanESM2 strongest in CCSM4 for the 4C and SE regions, where r2 model in the SE. This highlights the fact that it is not values exceed 0.7 for comparisons between the two straightforward to determine which metric is most ap- PDSI variables and 30cmSM. These numbers reduce propriate as a measure of modeled soil-moisture vari- slightly for the NP where some ensemble members yield ability or even which soil-moisture target is the values between 0.5 and 0.6. Comparisons weaken in most appropriate analog to compare against PDSI. CanESM2, in which values of shared variance between Even within models, the agreement between direct 15 APRIL 2015 S M E R D O N E T A L . 3223

FIG. 8. A comparison of reconstructed PDSI_TH (black with gray shading) in the 4C region (see Fig. 1) with collocated model estimates of PDSI_TH (red), PDSI_PM (orange), normalized 30cmSM (black), and normalized FCSM (blue) during the historical and projection intervals. Results are shown for the first ensemble member of (top) the CanESM2 and (bottom) CCSM4 model simulations to allow direct visual comparisons between the reconstructed and model variance. All time series have been centered over the 1901–79 interval, the time period of common overlap. Reconstructions extend prior to 1800, but the 1800–2100 interval is chosenFrom Smerdon for ease of visualet al.,comparison. 2015 (JCLIM) models will not share consistently common features. particularly in the CCSM4 output). Larger differences These considerations therefore must inform attempts to are observed between the original time series, in which compare reconstructions and model simulations directly differences in twenty-first-century trends reduce the r2 over the interval of overlap and require more detailed values among the different soil-moisture metrics (in and specific analyses that are beyond the scope of this cases where trends are comparable among variables, the investigation. The bridging approach that we have out- shared variances in original and detrended time series lined herein nevertheless demonstrates the use of the are similar). Secular drying in PDSI_TH exceeds that of twentieth century as a common interval for referencing all other modeled soil-moisture metrics over the twenty- reconstruction and model data in order to compare first century in all three regions, a clear demonstration of earlier paleoclimatic intervals and model projections of the tendency for PDSI_TH to overestimate drying the twenty-first century. during the projection interval. There are, however, re- gional differences in the relative comparisons among b. Projection interval variables, as discussed below. Shared variance between the four soil-moisture vari- PDSI_TH includes secular trends in the 4C region that ables during the projection interval (2005–99) are given in some cases are larger in magnitude or opposite in sign in Figs. 6 and 7, whereas Figs. 8–10 provide comparisons relative to the three other metrics in both models (Fig. 8), between the four soil-moisture metrics for the first en- despite the fact that some of the other metrics compare semble member from both models over the historical favorably to PSDI_TH over the full century or during interval through projection intervals. Shared variances specific intervals of time. PDSI_PM and 30cmSM compare for detrended time series over the projection interval, in well throughout the projection interval in the CCSM4 addition to the original time series, are also shown in model, whereas 30cmSM projects enhanced drying, rela- Figs. 6 and 7. The shared variances among all of the tive to PDSI_PM, in the CanESM2 model. FCSM again detrended soil-moisture metrics in the projection compares favorably to PDSI_PM and 30cmSM in the interval in all regions are generally comparable to CCSM4 simulation, whereas it suggests a wetting trend in those of the historical interval (with some exceptions, the CanESM2 simulation that is not reflected in any of the Cook et al., 2015

Southwest

From Cook, Ault, & Smerdon, 2015 (Sci. Adv.) Cook et al., 2015

From Cook, Ault, & Smerdon, 2015 (Sci. Adv.) Cook et al., 2015

From Cook, Ault, & Smerdon, 2015 (Sci. Adv.) Cook et al., 2015

From Cook, Ault, & Smerdon, 2015 (Sci. Adv.) Cook et al., 2015

Does mitigation help?

From Cook, Ault, & Smerdon, 2015 (Sci. Adv.) Caveats (abbreviated list)

Models have systematic biases. à Examine multi-model ensemble

Ensemble sizes are too small. à Use Monte Carlo “Emulators”

PET estimates are uncertain à Don’t rely on just one.* Assess broader “plausible” Precip. uncertain à range of precip. outcomes Role of mitigation uncertain à Look at high and low emission outcomes

* In Cook et al., 2015 we used three; Cook et al., 2014 used eight; Smerdon et al, 2015 used four. 1. Calculate risk as function of changes in mean and variability (Δµ, δσ).

2. Overlay multiple metrics from different GCMs

4 3. Also evaluate different RCPs

Ault et al., 2016 (Today)

4 Figure 2: Joint (2D) PDF of megadrought risk for JJA PDSI given a range of changes in precipitation (%) and temperature shifts (T ). Overlaid on this 2D PDF are the estimates of changes in precipitation (again, as a percentage of the 1951-2000 base period) and changes in temperature (in degrees oC

6 Recap so far:

Part I Megadroughts in the past: 1-2 per millennium Linked to the demise of several pre-industrial civilizations

Part II Increased risk during climate change Temperature and Precip. both important (but T wins out) Mitigation helps

Part III Connection to seasonal prediction… Photo credit: Scott St. George Carrillo et al., (work in progress) Summary

Part I Megadroughts in the past: 1-2 per millennium Linked to the demise of several pre-industrial civilizations

Part II Increased risk during climate change Temperature and Precip. both important (but T wins out) Mitigation helps

Part III Timing is everything… Acknowledgements

Justin Mankin, Jason Smerdon, Ben Cook, Scott St. George, Carlos Carrillo, NSF