Preparing for future hydroclimatic variability greatly benefits from long (i.e., multicentury) records at seasonal to annual time steps that have been gridded at kilometer-scale spatial intervals over a geographic region. Kriging is commonly used for optimal interpolation of environmental data, and space–time geostatistical models can improve kriging estimates when long temporal sequences of observations exist at relatively few points on the landscape. A network of 22 tree-ring chronologies from single-leaf pinyon (Pinus monophylla) in the central Great Basin of North America was used to extend hydroclimatic records both temporally and spatially. First, the line of organic correlation (LOC) method was used to reconstruct October–May total precipitation anomalies at each tree-ring site, as these ecotonal environments at the lower forest border are typically moisture-limited areas. Individual site reconstructions were then combined using a hierarchical model of spatiotemporal kriging that produced annual anomaly maps on a 12 km × 12 km grid during the period in common among all chronologies (1650–1976). Hydroclimatic episodes were numerically identified using their duration, magnitude, and peak. Precipitation anomalies were spatially more variable during wet years than during dry years, and the evolution of drought episodes over space and time could be visualized and quantified. The most remarkable episode in the entire reconstruction was the early 1900s pluvial, followed by the late 1800s drought. The 1930s Dust Bowl drought was among the top 10 hydroclimatic episodes in the past few centuries. These results directly address the needs of water and natural resource managers with respect to planning for worst-case scenarios of drought duration and magnitude at the watershed level.
Climate variability and change can influence multiple hydrologic characteristics, such as the proportion of snowfall versus rainfall (Knowles et al. 2006), the occurrence of droughts and floods (Cayan et al. 1999; Redmond et al. 2002), the timing of peak runoff (Stewart et al. 2004), and the amount of storage in reservoirs (Barnett and Pierce 2008; Barsugli et al. 2009). One strategy available to water managers for coping with the risk associated with these future impacts is to obtain a clear definition of past hydrological variability and extremes (Stakhiv 2011). In the western United States, as well as in other parts of the world, water resources are vulnerable to climate fluctuations (e.g., Rajagopalan et al. 2009; Seager et al. 2007), but instrumental records of precipitation, temperature, and surface-water flow are spatially and temporally limited, with length rarely exceeding the past few decades. At the same time, this region is rich with proxy records of hydroclimatic variables derived from tree rings (Loaiciga et al. 1993), given a unique combination of steep topography and semiarid conditions that limit wood formation, presence of old trees at both lower and upper tree lines, and a large body of ecological research on the climate sensitivity of these species (Fritts 1976; Speer 2010).
Blending instrumental and proxy records together with projected climate model simulations to inform water resource management is a topic that has received some attention (Gray and McCabe 2010; Prairie et al. 2008), but requires additional studies, as suggested in a joint report by the Bureau of Reclamation and the U.S. Army Corps of Engineers (Brekke 2011). In the Great Basin of North America, dendroclimatic reconstructions of moisture parameters have a long and productive history, starting with pioneering work in the 1930s (Antevs 1938; Hardman and Reil 1936), continuing with isolated efforts into the 1980s (Nichols 1989; Smith 1986), and arriving at the development of extremely long (>1000 yr) proxy records of precipitation in more recent times (Gray et al. 2004; Hughes and Graumlich 1996; Knight et al. 2010). Millennia-long tree-ring chronologies for the Great Basin were included in a spatially gridded dataset of drought reconstructions (Cook et al. 2004) at the relatively coarse scale of 2.5° of latitude and longitude (Cook and Krusic 2003).
Besides an overall arid-to-semiarid continental climate (Houghton et al. 1975), the Great Basin of North America is characterized by an upper and a lower tree line, with the latter being occupied by pinyon–juniper woodlands (West and Young 2000). These ecosystems, often called Nevada’s “pigmy forest,” are dominated by single-leaf pinyon (Pinus monophylla Torr. and Frém.) and Utah juniper (Juniperus osteosperma (Torr.) Little). Their climate is characterized by total annual precipitation usually between 300 and 500 mm, falling mostly during the cool season, and by a frost-free period ranging from ~90 to ~200 days. Single-leaf pinyon, the only one-needle pine in the world, can therefore survive both extreme drought and severely cold climate. It currently covers about 67 845 km2 (Cole et al. 2008), although its distribution has increased rapidly since Euro-American settlement (Romme et al. 2009), continuing a trend present throughout the Holocene (see Fig. 8-16 in Grayson 2011).
Among the many interpolation techniques available to produce gridded datasets from sparse point locations (Tabios and Salas 1985), geostatistical techniques such as kriging have various advantages (Biondi et al. 1994). In particular, kriging is the best linear unbiased estimator of the regionalized variable at unsampled locations, because it is constrained to produce residuals with zero mean and minimal variance (Isaaks and Srivastava 1989). In recent times space–time geostatistical models have been used to improve spatial estimation when long temporal sequences of observations exist at relatively few points on the landscape (Christakos 2000). Because the empirical estimation of space–time covariance models can become highly complex, many studies assume both spatiotemporal stationarity and the separability of spatial and temporal components (Kyriakidis and Journel 1999). These assumptions have been justified by arguing that no clear distance metric exists in space–time as there is in Euclidean space, given that time is ordered and unidirectional, while physical space is not. Recent developments have relaxed the separability assumption by introducing hierarchical models, estimated using Bayesian or empirical approaches (Cressie and Wikle 2011). In this framework, multivariate approaches can account for spatiotemporal covariates and nonstationarity (Fassó and Cameletti 2010). For instance, elevation can be entered as a covariate in order to compute a single kriging estimate for each two-dimensional gridpoint location rather than predict over all possible altitudes (Fassó and Cameletti 2009).
Increased spatial resolution is a necessity when simulating hydroclimatic processes in topographically complex terrain (Hijmans et al. 2005). With respect to water resources, Hoerling et al. (2009) found that using data spatially interpolated on 4- or 12-km grid cells improved agreement among models used for estimating future river runoff in the Colorado River basin, whose management is a vital issue for the western United States (National Research Council 2007). When long proxy records of climate exist at relatively sparse locations, the combination of temporal and spatial autocorrelation structures through space–time kriging provides an effective method for producing gridded proxy records at the same spacing as regional climate models and statistical downscaling techniques.
I present here an example of how proxy climate data could be interpolated by year at 12-km grid increments, which are easily comparable with downscaled regional climate simulations. Tree-ring chronologies were developed from single-leaf pinyon trees in the Great Basin of North America (Biondi et al. 2005) and calibrated against Parameter-Elevation Regressions on Independent Slopes Model (PRISM) data (Daly et al. 1994) to extend precipitation records. The study had mainly a hydrologic focus, with the objective to augment instrumental representations of severe droughts using much longer proxy records. In particular, I investigated the spatial and temporal features of climatic episodes, namely dry and wet spells, which have occurred over the past few centuries.
2. Materials and methods
A network of 22 tree-ring chronologies (Fig. 1) located in the eastern portion of Nevada (37.5°–39.8°N, 114.0°–116.0°W) was used to extend precipitation records. Data collection followed standard dendrochronological procedures, with 4.3-mm-wide increment cores taken from individual trees at or near breast height and from opposite sides of the tree along slope contours whenever possible (Grissino-Mayer 2003). Site locations were decided based on presence of old, healthy individuals of single-leaf pinyon. Wood cores were stored and dried inside paper straws in the field, then transported to the laboratory, where they were glued to grooved wood mounts. The cores were progressively sanded, first mechanically then by hand, until individual cells were clearly visible under a binocular stereo-zoom microscope with 10–50× magnification. All cross dating (Stokes and Smiley 1996) and locally absent ring assignment was done visually using a binocular microscope prior to ring-width measurement on a Velmex stage with 1-μm resolution. The ring-width measurement data were quality-controlled using the software program COFECHA (Grissino-Mayer 2001; Holmes 1983).
Most sites were sampled in 2008–09 by DendroLab (http://dendrolab.org/) personnel, and a few more chronologies for the study area were obtained from the public-domain International Tree-Ring Data Bank (ITRDB; Grissino-Mayer and Fritts 1997). Many ITRDB collections were completed decades ago, and their inclusion in the dataset is not a guarantee of quality. For instance, the oldest tree-ring collection for pinyon in Nevada (ID 043630) was not used because it listed Pinus edulis (the Colorado or southwestern pinyon) as the species. Another Pinus monophylla chronology (ID 461639) was excluded because it was less than 150 yr long, considerably shorter than expected given the longevity of the species.
The next step in the analysis combined all measurements for a site and species into a single “master” tree-ring chronology (Fritts 1976). This process requires standardization of the ring-width measurements because tree radial growth is a result of individual biological and ecological processes. So it is necessary to minimize age-related trends and other nonclimatic variability prior to combining multiple ring-width series together (Cook and Kairiukstis 1990). A cubic smoothing spline (Cook and Peters 1981) was fit to each ring-width series to avoid the drawbacks of other types of standardization functions, such as modified negative exponentials (Biondi and Qeadan 2008a). Ring indices were obtained as ratios between the ring-width measurements and the corresponding spline values. The median of all indices available for a year was used to produce the chronology value for that year, as follows:
where is the chronology value in year t, which is the median annual index; nt is the number of samples in year t, with nt ≥ 3; w is the cross-dated ring width (in millimeters, with 1000th digit resolution) of sample i in year t; y is the value of sample i in year t computed by fitting a cubic smoothing spline with 50% frequency response at a period of 100 yr to ring-width series i; and wt /yt is the dimensionless index value of sample i in year t. Ring-width measurements obtained from the ITRDB were also standardized using this formula. I performed numerical analyses using a combination of in-house scripts for SAS (Delwiche and Slaughter 2003) and the R numerical computing package (R Development Core Team 2012), together with task-specific software, either stand alone (Biondi and Waikul 2004) or modular (Bunn et al. 2012).
Calibration of dendroclimatic records with instrumental data was done using interpolated and spatially averaged records, namely, PRISM (Daly et al. 1994) and Climate Division (Guttman and Quayle 1996) data. PRISM monthly total precipitation and mean minimum and mean maximum air temperature are publicly available at 2.5-arc-min grid spacing. These data were downloaded from the DendroLab Arc Internet Map Server (ArcIMS) web server DendroNet GIS (http://www.dendrolab.org/GIS/) by clicking on the map grid cell that included the coordinates for a tree-ring site. The average annual cycle from 1895 to 2006 for these three instrumental variables was computed using all 22 PRISM cells. Monthly total precipitation and mean air temperature for Nevada Climate Divisions 2 and 3 during 1895–2006 were downloaded from the National Oceanic and Atmospheric Administration (NOAA)/National Climatic Data Center (NCDC) FTP site (ftp://ftp.ncdc.noaa.gov/pub/data/cirs/) for comparison with PRISM data. Time series patterns of average PRISM and Climate Division instrumental records were plotted by year using two seasonal combinations: cool and wet (October–May) and warm and dry (June–September).
Statistical relationships between each tree-ring chronology and PRISM monthly climate data were investigated for the water year (October–September) using bootstrapped correlation and response functions (Biondi 1997). Based on the average climate regime at the study sites and on the main climatic signals in the tree-ring chronologies, my target for record extension was the total October–May precipitation. Even though tree growth depends on water uptake by the roots, and not directly on precipitation, long-term instrumental observations were not available to calibrate tree-ring chronologies against soil-level processes. When the correlation between a tree-ring chronology and the local PRISM precipitation was less than 0.4, the reconstruction was not performed.
The line of organic correlation (LOC) method (Helsel and Hirsch 2002) was used for record extension to better represent uncertainty in the underlying processes (Biondi et al. 2010). LOC relies on minimizing not the sum of squared distances from the final regression line, but rather the area of all triangles formed by the horizontal and vertical distance to the regression line. It is therefore also known as least areas regression, standardized major axis regression, least products regression, and diagonal regression. In hydrology, the LOC method has been called the maintenance of variance extension (MOVE; Salas et al. 2008). Its main advantage is that the cumulative distribution function of the reconstructed values, including the variance and probabilities of extreme events, resembles the distribution of the instrumental data used for calibration. The LOC method was evaluated in detail for a reconstruction of streamflow variability in the Cleve Creek watershed (Strachan et al. 2012). The proxy record of October–May precipitation was produced to have an overall mean of zero, while deviations above and below this reference level, called anomalies, had physical units (millimeters).
I performed space–time estimation of tree-ring-derived precipitation for grid cell sizes comparable to those used for downscaled climate projections and model output from PRISM and regional climate models. Two different geostatistical approaches, neither one of them Bayesian in nature, were used to quantify relationships and to develop gridded records at 12-km intervals. One approach was based on a hierarchical model (Cressie and Wikle 2011) that combines a data model (i.e., the probability distribution given a hidden true process) with a process model (i.e., the probability distribution of the hidden process) to provide a conditional probability distribution of the hidden process given the data (the “posterior” of Bayesian statistics). Another approach was based on the linear combination of spatial and temporal autocorrelation structures, essentially considering time as an additional dimension with unidirectional features (Kyriakidis and Journel 1999).
The two methods were compared using the R software environment. Hierarchical space–time modeling was implemented using the Stem package (Fassó and Cameletti 2009), whereas the separable (product sum) exponential covariance model was applied using the gstat package (Pebesma 2013). The tree-ring reconstructions were stationary in time, and spatial structures were considered to be isotropic. Given the importance of orographic precipitation in the Great Basin (Houghton et al. 1975), elevation was entered as a covariate in the hierarchical model. This allowed the calculation of a single kriging estimate for each two-dimensional grid point, rather than over all possible altitudes, as would have been necessary in gstat. Since no missing values were allowed, the period 1650–1976 was selected to include 17 DendroLab chronologies and five ITRDB chronologies that effectively improved spatial coverage in the basins of interest.
All gridpoint estimates of October–May precipitation were averaged by year to obtain a single time series that could be analyzed in terms of episode duration, magnitude, and peak (Biondi et al. 2008). Duration is the number of time intervals the process remains continuously above (or below) a reference level. Magnitude is the sum of all series values for a given duration; hence, it is equivalent to the area above (or below) the reference level. Peak is the absolute maximum among all series values for a given duration. Episodes above (or below) a reference level are called positive (or negative). Analysis of episode parameters allows a less subjective identification of the “strongest,” “greatest,” or “most remarkable” periods, and although this approach is normally used for drought analysis, it can be applied to any cumulated deviations (Biondi et al. 2002). Each episode parameter was ranked separately, and the three ranks were added to obtain the final episode score: the higher the score, the stronger the episode.
The 22 tree-ring chronologies (Table 1) were distributed over 500 m of elevation [~(1930–2430) m] and a geographical area about 50% longer in the north–south direction (~230 km) than in the east–west direction (~155 km). Cross-dated tree ages reached a maximum of 784 yr, and mean segment length ranged from 192 to 443 yr with an average of 326 yr. Because of the remarkable tree longevity and large number of samples (32 on average) per site, this dendrochronological network comprised a total of 219 265 cross-dated ring-width measurements—roughly a mean of 10 000 measurements per site. Locally absent rings were numerous, ranging from 1.2% to 5.5% of the total number of measurements per site, with an average of 3%. Mean pairwise correlation between tree-ring chronologies during 1650–1976 was 0.7, and for that period, the first principal component accounted for 70% of the total variance.
In agreement with their numerical similarity, tree-ring chronologies showed highly coherent time series patters (Fig. 2). This was visually evident during major droughts, such as the late 1500s (Stahle et al. 2000) and the 1930s “Dust Bowl” (Fye et al. 2003). The chronologies also showed a high level of variability, as indicated by their standard deviation (from 0.292 to 0.467, with a mean of 0.359) and small lag-one autocorrelation (from 0.033 to 0.345, with a mean of 0.213). For comparison, Fritts and Shatz (1975) reported that 11 chronologies of southwestern pinyon (Pinus edulis) had higher standard deviation (mean of 0.395) and lag-one autocorrelation (mean of 0.361), although different standardization options and historical periods may have contributed to these dissimilarities. Rather than using mean sensitivity, which refers only to adjacent rings, the all-lag sensitivity of the chronologies was quantified by the Gini coefficient (Biondi and Qeadan 2008b). Its values (from 0.166 to 0.271, with a mean of 0.206) were higher than those previously reported for the same species and would have fallen right below those calculated for Pinus edulis during the 1880–1960 period (see Fig. 2 in Biondi and Qeadan 2008b), most likely because of the different standardization option used in this analysis.
Climate regime at the study areas was quantified by month (Fig. 3). The annual cycle of precipitation showed higher values and greater variability (especially from February through April) if PRISM data, rather than climate division ones, were used. Since PRISM data referred to local conditions and were restricted to mountain areas (mean elevation of tree-ring sites was 2280 m), temperatures were also cooler, especially from April to July. Despite these differences in absolute values, temporal changes were quite uniform (Fig. 3), as also suggested by extremely high correlations from 1895 to 2006 between PRISM and divisional precipitation (r = 0.9, p value < 0.0001 in both seasons) as well as temperature (r = 0.9, p value < 0.0001 for October–May; r = 0.8, p value < 0.0001 for June–September). In terms of overall trends, the divisional data showed a slight temperature increase over time (r = 0.4, p value < 0.0001), which was less evident in the PRISM data (r = 0.3, p value < 0.01). Finally, warm-season precipitation and temperature were inversely correlated in both datasets (r = −0.3 to −0.4), and a slight negative correlation (r = −0.2, p value < 0.05) also existed between precipitation in October–May and temperature during the following months (June–September). Since the PRISM data better captured the climatic regime at the study areas, and at the same time they were extremely similar to the divisional data in terms of time series patterns, they were then used for proxy calibration and reconstruction purposes.
A pooled spatiotemporal variogram was calculated as the average of 327 annual variograms for the 22-point reconstructions (Pebesma 2013). An appropriate model for the pooled variogram was given by a nugget plus exponential term with a range of about 100 km and a sill that was about 75% of the combined spatial variance (Fig. 4). Most of the spatial dependence was not temporally lagged, with very little spatial dependence for temporal lags greater than one (Fig. 5). A spatial range of 100 km and time range of 1 yr was therefore used for space–time kriging interpolation. The three driest years identified using an average of the gstat gridded values are the same as those calculated from an average of the Stem interpolation: 1934, 1782, and 1879. Similarly, the three wettest years were the same when results were averaged using one or the other method: 1726, 1914, and 1868. The spatial patterns, however, were not the same, as the nonhierarchical interpolation varied less over the study area in any given year, most likely because it did not incorporate elevation as an additional factor. Further analyses were therefore based on gridded values produced by the hierarchical method using the Stem package, which are shown year-by-year in Fig. 6.
As mentioned above, the three worst annual droughts in the 327-yr record were 1934 (mean anomaly of −181 mm, 3.1 standard deviations below the 0 reference level), 1782 (mean anomaly of −171 mm, 3.0 standard deviations below normal), and 1879 (mean anomaly of −150 mm, 2.6 standard deviations below normal). Although the entire study region experienced a drought in those years, the intensity of each event varied spatially (Fig. 7), and the availability of values on a 12-km grid would allow for detailed analyses of individual subwatersheds. Similarly, the three wettest years, 1726 (mean anomaly of 156 mm, 2.7 standard deviations above the 0 reference level), 1914 (mean anomaly of 140 mm, 2.4 standard deviations above normal), and 1868 (mean anomaly of 130 mm, 2.3 standard deviations above normal) occurred with different geographical features (Fig. 7). Wet years were usually more spatially variable than dry years: the three wettest years were characterized by three distinct patterns (Fig. 7, bottom, from left to right: U shaped, C shaped, and reverse-L shaped), while the three driest years displayed a more homogeneous spatial arrangement (roughly an east–west dipole; Fig. 7, top). The time series of the total October–May precipitation anomaly obtained by averaging the 315 space–time kriging values (Fig. 8) was almost perfectly correlated (r = 0.997) with the mean of the 22 proxy point reconstructions, but some of the rankings were different. For example, the third wettest year was 1838 (with 1868 immediately following) instead of 1868 (with 1838 immediately following). Because of this similarity, and of the higher spatial representation provided by the mean of the interpolated values, episode analysis was performed using the average of the 315 grid points.
A total of 150 wet and dry episodes were identified in the 327-yr record of proxy October–May total precipitation anomalies (Table 2a). Episode duration ranged from 1 to 11 yr, and the longest interval was 1905–15 (an 11-yr wet spell), followed by two 8-yr-long episodes, one dry (1876–83) and one wet (1848–55). The greatest magnitude also belonged to the early 1900s pluvial (1905–15), followed by another wet period (1723–28) and by a dry one (1706–10). The strongest peak occurred during the Dust Bowl drought (1933–36 in this record), followed by another dry episode (1780–83), and then by the wet spell that also had the second-largest magnitude (1723–28). According to a simple scoring rule (Table 2), the most remarkable episode was the early 1900s pluvial (1905–15), followed by the late 1800s drought (1876–83). The 1930s Dust Bowl drought (1933–36) was in eighth position, making it an important episode over the past few centuries, as was also found for this part of the Great Basin when reconstructing Cleve Creek streamflow over the past 550 yr (Strachan et al. 2012). To capture interannual variability mostly associated with the El Niño–Southern Oscillation (ENSO) phenomenon (Cayan et al. 1999), the annual precipitation anomalies were smoothed with a cubic spline having a 50% frequency response at a 7.5-yr period (Cook and Peters 1981). This new time series (Fig. 8) was composed of 55 episodes (Table 2b), among which the early 1900s pluvial (1904–24) remained the strongest one, and was then followed by five dry spells, starting with the 1930s Dust Bowl (1925–37) as the second-most remarkable episode and including the 1950s drought (1946–64) in fifth position.
Spatially interpolated climate data on regular grids are routinely employed for research and management purposes in a number of disciplines, including ecological (Guisan and Thuiller 2005), agricultural (Daly et al. 2008), and social (Wilby 2007) sciences. The widespread distribution of single-leaf pinyon in ecotonal and fire-protected conditions throughout the Great Basin allowed for the development of a moisture-sensitive tree-ring network spanning the past few centuries. Applying space–time kriging to obtain dendroclimatic records of precipitation interpolated by year at 12-km grid increments should then substantially enhance their management implications, as these “dendroclimatic surfaces” are one of the sources of information being studied by the Southern Nevada Water Authority to estimate the impacts of prolonged droughts in this region (Abatzoglou et al. 2012).
This initial effort was aimed at testing the applicability of space–time kriging for producing kilometer-scale proxy records of climate. In terms of year-by-year patterns, spatial variability was generally higher during wet years than during dry ones, and further research will be aimed at more formally quantifying such differences. Additional elaborations of the interpolated data could include a sum of all the annual values to give a composite map of the driest and wettest basins in the area and a map of time series standard deviations to identify the watersheds with higher precipitation variability, which could be useful to land planners. It is also expected that a larger area will be studied in the future, as space–time kriging is a promising alternative to the methods that are currently grouped under the name of “climate field reconstructions” (Dannenberg and Wise 2013) and that have been criticized for their statistical assumptions (McShane and Wyner 2011). For now it is clear that the mean of the 22 individual reconstructions and the mean of the 315 gridpoint reconstructions showed essentially the same variability; hence, spatial patterns themselves were the most important result derived from space–time kriging of site-specific records. In other words, the true value of kilometer-scale climate fields is the fine-grain spatial representation of dry and wet episodes, since their average temporal evolution over the landscape can be captured by a relatively limited set of points.
In fact, reconstructed time series patterns can often be compared in paleoclimatic studies across fairly distant regions because of their coarse spatial resolution. As an example, the area-averaged annual precipitation anomalies (Fig. 8) were compared to other dendroclimatic reconstructions that had been archived at the NOAA/NGDC Paleoclimatology Program. Total annual (from previous July to current June) precipitation for Nevada Climate Division 3 (Hughes and Graumlich 2000) as well as total annual (from previous June to current June) precipitation for Utah Climate Division 6 (Gray et al. 2005) provided two proxy records from areas located near the study region. The Hughes and Graumlich time series, coded H&G1996 because of the year reported on the NOAA/NGDC website, truly refers to a much larger section of the Great Basin, extending west all the way to the Sierra Nevada (Fig. 1). On the other hand, the Gray et al. reconstruction (coded G&a2004) pertains to an area outside the hydrographic boundary of the Great Basin (see Fig. 2-1 in Grayson 2011).
Other paleoclimatic records used for comparison had extended either a drought index or river runoff. Reconstructions of mean summer (current June to August) Palmer drought severity index (PDSI; Cook and Krusic 2004) for grid nodes 71 (40.0°N, 115.0°W) and 72 (37.5°N, 115.0°W) were averaged together, and their mean was coded C&a2004. The geographic region encompassed by each of the two grid nodes spanned a portion of my study area, plus several more Great Basin watersheds. Because of its biogeographical proximity, I also included the reconstructed Colorado River annual (from previous October to current September) streamflow at the Lees Ferry station (Meko et al. 2007). This river discharge time series (coded M&a2007) closely resembles other, shorter reconstructions of streamflow at the same location (Woodhouse and Lukas 2006). The five proxy time series showed pairwise linear correlations around 0.6 (Table 3), with the exception of the Hughes and Graumlich (2000) Great Basin reconstruction, which had the lowest values (0.2–0.3) when compared to the Utah (Gray et al. 2004) and Colorado River (Meko et al. 2007) ones. Overall, the October–May precipitation reconstruction presented here agreed with all other proxy records, as shown by correlations ranging from 0.5 with H&G1996 to 0.9 (the overall maximum) with C&a2004.
Information derived from the past, such as tree-ring reconstructions of hydroclimatic variables, requires careful consideration. On one hand, multicentury-long proxy time series complement and expand the perspective derived from analysis of much shorter instrumental records and from projections of future scenarios (Pederson et al. 2012). This is made particularly clear by considering that regional climate model predictions on seasonal time scales have shown little skill outside of the tropical Pacific in non-ENSO years (Goddard et al. 2001), and even less skill may occur in multidecadal regional climate projections (Pielke and Wilby 2012). On the other hand, the numerical development of tree-ring chronologies and the statistical tools employed in most reconstructions are based on assumptions of linearity and stationarity (National Research Council 2006). Although the line of organic correlation can generate more variable reconstructions than linear regression (Hirsch 1982), the best current solution to the announced “death of stationarity” (Milly et al. 2008) lies in both an understanding of its true statistical meaning (Matalas 2012) and in adopting a quantitative episode analysis to effectively place modern patterns into a long-term perspective. For instance, the early-1900s pluvial (Fye et al. 2004) was found repeatedly to be the most remarkable episode during the past few centuries as well as throughout the Common Era, and this information is vital to realize the extent of the “wet bias” that affects the instrumental record of hydroclimatic variables in the western United States.
Along the same lines, the 1930s Dust Bowl drought, which achieved top-10 status in the episode analysis presented here, was much less prominent (falling to the seventy-third position) in a 2300-yr precipitation reconstruction for the Walker River basin, on the east slope of the Sierra Nevada (Biondi et al. 2008). The current results confirm Strachan et al.’s (2012) finding that regional drought severity in the central Great Basin can differ from its realizations in the western portion of this region. In other words, the instrumental record better represents long-term hydroclimatic variability in the eastern portion of Nevada, near the Utah border, compared to its western portion on the eastern slopes of the Sierra Nevada at the border with California. Therefore, water management policies in eastern Nevada basins could reliably use the 1920–30s drought as a relevant scenario for extreme drought conditions. Because predicting changes in the statistics of regional climate over multiannual time periods is essential to understand climate forcings and feedbacks, a similar approach could prove highly effective for individual subwatersheds in other parts of the world.
Space–time geostatistics allow for the production of gridded proxy records of climate at kilometer-level resolution over multiple centuries. In this area of the Great Basin, spatial variability of October–May precipitation was reconstructed at 12-km grid intervals over more than three centuries. Whereas the early 1900s pluvial, a most remarkable episode in the last few centuries, can bias the instrumental record, water management policies in eastern Nevada basins could use the 1920–30s drought as an extreme case analog. These results directly address the needs of water and natural resource managers with respect to planning for worst-case scenarios of drought duration and magnitude at the watershed level. For instance, it is possible to analyze which geographical areas are more likely to be impacted by severe and sustained droughts at annual or multiannual time scales and at spatial resolutions commonly used by regional climate models. Multicentury long dendroclimatic records with kilometer-scale definition are therefore useful tools for designing management practices with the objective to achieve drought resiliency in individual watersheds.
The research was supported, in part, by Southern Nevada Water Authority; by the Office of the Vice President for Research at the University of Nevada, Reno; and by the U.S. National Science Foundation under Cooperative Agreement EPS-0814372 and Grant P2C2-0823480. Completion of the article was allowed by a Visiting Fellowship from the Cooperative Institute for Research in the Environmental Sciences (CIRES) of the University of Colorado Boulder, and by a Visiting Scientist Travel Grant to the WSL in Birmensdorf awarded by the Oeschger Centre for Climate Change Research, University of Bern, Switzerland. The views and conclusions contained in this document are those of the author and should not be interpreted as representing the opinions or policies of the funding agencies. I thank M. Cameletti, J. D. Salas, L. Saito, and J. Leising for helpful discussions of statistical and hydrological issues. I am also extremely grateful to the DendroLab personnel, especially Scotty Strachan, who contributed, either in the field or in the laboratory, to the development of the tree-ring network. The comments of two anonymous reviewers helped to finalize an earlier version of the manuscript.