Water Resources Research

Paleoreconstruction of cool season precipitation and warm season streamflow in the Pacific Northwest with applications to climate change assessments


  • Eric R. Lutz,

    1. Department of Civil and Environmental Engineering,University of Washington, Seattle, Washington,USA
    2. Now at Department of Earth Sciences,Dartmouth College, Hanover, New Hampshire,USA
    Search for more papers by this author
  • Alan F. Hamlet,

    1. Department of Civil and Environmental Engineering,University of Washington, Seattle, Washington,USA
    2. Center for Science in the Earth System (CSES) Climate Impacts Group,University of Washington, Seattle, Washington,USA
    Search for more papers by this author
  • Jeremy S. Littell

    1. Center for Science in the Earth System (CSES) Climate Impacts Group,University of Washington, Seattle, Washington,USA
    Search for more papers by this author


[1] Long-term streamflow reconstructions help characterize climate variability and extreme climatic events, such as droughts, which play a crucial role in water resource planning. Dendrohydrological reconstructions generally build on indirect associations between streamflow and radial tree growth, both of which depend on cool season precipitation in the Pacific Northwest. We develop a new approach that integrates a tree ring reconstruction of cool season precipitation with historical meteorological data and a physically based hydrologic model to reconstruct warm season streamflow and streamflow uncertainty. The Upper Yakima Basin in Washington state is used as a test case. We applied objective screening, principal components analysis, and multiple linear regression to reconstruct 366 years of basin-average cool season precipitation. The reconstruction was integrated with five temporally and spatially distributed cool season precipitation patterns spanning the historical range of natural variability. These distributed meteorological reconstructions were used as inputs to the Variable Infiltration Capacity (VIC) hydrologic model over the Yakima basin to produce an ensemble of warm season streamflows for each reconstructed year. The resultant streamflow reconstruction retains dendroclimatic information and quantifies the inherent uncertainty in warm season streamflow associated with historical meteorology. Finally, the meteorological reconstructions were systematically perturbed and used to drive the VIC to examine the potential impacts of climate change on warm season streamflow over the 366-year record. Despite projected wetter conditions in the future, less precipitation is stored as snowpack (due to warmer winter temperatures) and consequently warm season streamflow will be systematically reduced. The combination of long records of variability and systematic changes related to climate change provides useful information about the combined effects of natural variability and projected systematic changes in climate to support 21st-century water planning.

1. Introduction

[2] Streamflow reconstructions derived from tree ring proxy records are valuable to water resource planners because they can extend instrumental records by hundreds or even thousands of years. They also enhance our understanding of decadal- to centennial-scale climate variability and the occurrence of extreme events such as droughts [Cook et al., 1999] or reduced snowpack [Pederson et al., 2011], which often exceed the range of variability observed in historical records [Stockton, 1990; Loaiciga et al., 1993]. Such streamflow reconstructions can provide extended baseline conditions for planning efforts in the face of population growth, changing water demand, and changes in land use and resource management [Woodhouse et al., 2006; Lukas and Woodhouse, 2006; Rice et al., 2009]. However, the utility of dendrohydrologic reconstructions in water resources planning has two limitations. First, although such reconstructions can quantify annual to centennial variability (which often exceeds that evident in the historical record), tree ring records are rarely capable of capturing estimates of subseasonal variations, which are of great interest to water managers. Second, the future temperature and precipitation projected in climate change scenarios are often quite different from those associated with the underlying hydrologic variability informing the reconstructions. In particular, the loss of snowpack and the resulting decline in warm season (April–September) water availability due to climate change is a fundamental impact pathway for many water supply systems in the Pacific Northwest [Hamlet and Lettenmaier, 1999; Elsner et al., 2010; Vano et al., 2010] and the western United States as a whole. Thus, a central goal of this study has been to provide methods that develop more useful multicentury seasonal streamflow information to water resource managers in the Pacific Northwest (PNW) in support of long-term planning.

[3] In this paper we develop a new approach that combines tree ring proxy reconstructions of cool season (October–March) precipitation over several centuries with a physically based macroscale hydrologic model to estimate warm season (April–September) streamflow for both historical and future climates. The technique quantifies the uncertain relationship between cool season precipitation and warm season streamflow, simulates warm season streamflow over a long retrospective period (366 years), and quantifies the effects of projected future climate conditions on the warm season streamflow reconstructions. The approach improves our ability to simulate the range of seasonal responses associated with annual climate reconstructions, and to combine the paleoproxy record with hydrologic modeling of future conditions to better understand the range of possible future conditions.

[4] To our knowledge, this is the first study that integrates a spatially explicit hydrologic model with tree ring reconstructions of climate to estimate seasonal streamflow (and its variability) using physically based methods. Saito et al. [2008] coupled tree-ring-derived annual precipitation with a simple water balance hydrologic model to reconstruct streamflow of the Walker River at Coleville, CA, for the period 1939–2001. Gray and McCabe [2010] integrated tree-ring-derived annual precipitation estimates with a water balance model to examine the potential effects of climate change on the Yellowstone drainage. As with streamflow reconstructions derived directly from tree ring chronologies, a general limitation of these approaches is that uncertainties related to the natural temporal and spatial variability of meteorological conditions at daily, monthly, or seasonal timescales cannot be physically accounted for. For example, 2 years with nearly identical regional cool season precipitation (or annual, as in the cited studies) might have very different spatial patterns of snowpack (depending on the temperature during the season and the frequency and spatial distribution of storms), which subsequently influence warm season streamflow and consequently summer water supply. In our approach, we seek to address these limitations by explicitly accounting for the uncertainty in the analysis using a fully distributed, physically based hydrology model.

2. Methods

2.1. Experimental Design

[5] As noted above, warm season streamflow is a critical variable in assessing the ability to meet societal water demands in the Pacific Northwest (PNW) [Hamlet and Lettenmaier, 1999; Hamlet et al., 2010; Elsner et al., 2010; Vano et al., 2010]. Because of the Mediterranean climate in the PNW (wet winters, dry summers) and the influence of snowpack on the hydrologic cycle in many watersheds, warm season streamflow is strongly related to cool season precipitation. Most of the warm season streamflow of the Columbia, Snake, and Yakima Rivers, for example, is supplied by cool season (October–March) precipitation that is stored as snowpack until the early summer melt. For the Yakima River, in eastern WA state (see section 2.2), simple linear regression indicates a significant (p < 0.001) association between cool season precipitation and naturalized warm season streamflow that accounts for most of the streamflow variance in the historical period (r2 = 0.65; without outlier [1934], r2 = 0.73) (Figure 1). For the Columbia River at The Dalles, OR, more than 80% of the variance in annual flow is explained by cool season precipitation.

Figure 1.

Linear correlations between basin-average cool season precipitation and naturalized streamflow near Parker WA (USGS 12505000) for the historical period (black line) and excluding the outlier of 1934 (dashed line).

[6] Our approach, as shown schematically in Figure 2, is outlined as follows:

Figure 2.

Schematic diagram of experimental design for reconstructing streamflow. Data and processing steps are represented as parallelograms and rectangles, respectively. Section numbers of main methodologies are indicated in parentheses.

[7] 1. Reconstruct the Yakima Basin-average cool season precipitation for the period 1614–1979 from proxy relationships between historical meteorological data and tree ring chronologies.

[8] 2. Create an ensemble of five meteorological reconstructions that combine the 366-year basin-average cool season precipitation reconstruction with five observed spatial and temporal meteorological patterns selected from the historical record.

[9] 3. Reconstruct warm season streamflow for the period 1614–1979 by driving a physically based hydrologic model with the ensemble of meteorological reconstructions.

[10] 4. Simulate warm season streamflow under future climate conditions projected by a climate change scenario using the reconstructed meteorological ensembles for the proxy years 1614–1979.

2.2. Case Study

[11] To provide a case study, we applied our experimental approach to the Yakima River Basin upstream of the USGS gage station near Parker, WA (USGS 12505000; Latitude: 46.4972°N, Longitude: 120.4417°W; drainage area: 6889 km2) (Figure 3). This watershed is situated on the eastern slope of the Cascade Mountains in Washington. Warm season water supply is a crucial factor determining the economic benefits from irrigated agriculture in the basin [Vano et al., 2010].

Figure 3.

Overview map of Yakima Basin study area (gray area) above Parker Gauge (black diamond) and locations of 38 tree ring chronologies possessing negative (gray) or positive (white) associations between growth and cool season precipitation for same-year (circle) or subsequent-year growth response (diamond), based on a correlation coefficient |r| ≥ 0.25. Two chronologies possessed both negative same-year and positive subsequent-year growth correlations, resulting in a total of 40 time series being shown on the map. Symbol size corresponds with correlation coefficients (see legend for scale). State (USA) and province (Canada) boundaries and abbreviations are depicted.

[12] The Yakima River can be characterized as mildly snowmelt dominant, as defined by Hamlet and Lettenmaier [2007], with much of the summer streamflow volume deriving from snowpack. Although a portion of early winter precipitation coincides with above freezing temperatures and produces runoff, most of the precipitation occurs in the cool season and is stored as snowpack until the spring/summer snowmelt. Because snowpack storage is sensitive to temperature, systems such as the Yakima are expected to experience decreases in spring snowpack and earlier melt seasons in the warmer climate projected for the 21st century, resulting in reductions in the summer water supply and impacts to water management [Elsner et al., 2010; Maurer et al., 2007; Vano et al., 2010].

2.3. Macroscale Hydrologic Model

[13] The fully distributed, physically based Variable Infiltration Capacity (VIC) hydrologic model [Liang et al., 1994, 1996; Nijssen et al., 1997] has been widely used for characterizing observed natural variability and climate change impacts on hydrologic systems [e.g., Arora and Boer, 2006; Christensen and Lettenmaier, 2007; Elsner et al., 2010; Guo et al., 2009; Hayhoe et al., 2007; Vicuna et al., 2007]. The sensitivity of the integrated snow model to changing climate in the western United States and the PNW has been validated [Hamlet et al., 2005; Mote et al., 2005, 2007].

[14] Using calibrated model parameters prepared in a previous study by Elsner et al. [2010], we implemented the VIC hydrologic model over the Yakima Basin at 1/16° spatial resolution (longitudinal resolution ∼4.5 km, latitudinal resolution ∼7 km). Daily natural (i.e., without water management effects) streamflow values for the Yakima River near Parker (USGS 12505000) were simulated from 1916–2006. These values were then averaged for the warm season. To assess the hydrologic model performance, a Nash-Sutcliffe efficiency [Nash and Sutcliffe, 1970] of 0.71 was calculated from monthly values of calibrated streamflow simulations and naturalized streamflow observations obtained from the U.S. Bureau of Reclamation (http://www.usbr.gov/).

2.4. Historical Meteorological Driving Data

[15] We describe the source and processing of the main daily meteorological forcings for the VIC model, including gridded precipitation, temperature, and wind data. Additional variables such as solar radiation and humidity were estimated by the VIC model using daily maximum and minimum temperatures using algorithms described by Thornton and Running [1999] and Kimball et al. [1997].

[16] Gridded historical meteorological data were generated at 1/16° spatial resolution at daily time steps for the period 1915–2006. The primary sources of meteorological data included the National Oceanic and Atmospheric Administration (NOAA) Cooperative Observer Program (COOP; http://www.nws.noaa.gov/om/coop/), the U.S. Historical Climatology Network (HCN; http://cdiac.ornl.gov/epubs/ndp/ushcn/ushcn.html), and the Historical Canadian Climate Database (HCCD; http://www.cccma.ec.gc.ca/hccd/). The procedures are described by J. Deems and A. Hamlet [unpublished], who extended the methods of Maurer et al. [2002] and Hamlet and Lettenmaier [2005] at 1/16° spatial resolution. These refined methods incorporated the 30-arcsecond PRISM data set, derived from the period 1971–2000 by Daly et al. [2008], to adjust the long-term monthly means of both daily precipitation and temperature to match the PRISM monthly climatology. This hybrid technique preserves the monthly variability present in HCN and HCCD data sets, while retaining both the daily fluctuations derived from COOP data and the spatial patterns of climatic monthly means quantified in the PRISM data set.

[17] To remove long-term temperature trends in the historical period, the minimum and maximum temperature values were detrended, by month, using linear regressions pivoted on 1915 (following Hamlet and Lettenmaier [2007]). This detrending step was implemented because temperatures in the paleoclimatic record are not expected to reflect the observed rapid increases in temperature in the late 20th century [Wahl and Ammann, 2007].

[18] In addition to precipitation and temperature, for the period 1949–2006, daily wind speed values were downscaled from the National Centers for Environmental Prediction-National Center for Atmospheric Research (NCEP-NCAR) reanalysis products [Kalnay et al., 1996]. For the period 1915–1949, the daily wind speed climatology was derived from the 1949–2006 reanalysis, as described by Hamlet and Lettenmaier [2005].

[19] As described in greater detail in section 2.6, the complete meteorological data set was used to drive VIC modeling in the preliminary analysis to select spatial and temporal meteorological patterns from five representative water years. The patterns from these five representative years were then integrated with the tree ring reconstruction of precipitation to drive the paleo- and climate change VIC simulations (sections 2.8 and 2.9).

[20] The complete historical precipitation data set also provided the historical basin-average cool season precipitation observations Pobs for the tree ring reconstruction of cool season precipitation (section 2.7),

display math

where Pt,x is the daily total precipitation in each cell (x), which is seasonally summed (z = 182 or 183 cool season days) to estimate the total cool season precipitation in each cell, and then averaged over all cells in the domain (n = 275).

2.5. Tree Ring Reconstruction of Basin-Average Cool Season Precipitation

[21] Tree ring reconstructions of streamflow rely on the indirect association between streamflow and tree growth, both of which are dependent on precipitation, evaporation, soil moisture, length of the warm season (temperature and snow cover), and other factors [Fritts, 1976; Meko et al., 2010]. Dendroclimatological reconstructions of precipitation at annual, seasonal, and in some instances monthly timescales, have been successfully demonstrated in a number of past studies [D'Arrigo and Jacoby, 1991; Case and MacDonald, 1995; Gray and McCabe, 2010; Hamilton et al., 2001; Pisaric et al., 2009; Stahle et al., 2009; Watson and Luckman, 2001, 2004]. Such efforts have, to date, been limited in the Pacific Northwest compared to the southwest [but see Graumlich, 1987; Garfin and Hughes, 1996;, Knapp et al., 2004].

[22] Before describing the statistical methods applied to reconstruct cool season precipitation over the Yakima Basin, we summarize the dendroclimatological relationships of western North America that may drive any such reconstruction. The first known growth relationship is characterized by snow limitation, in which interannual tree growth is at least partially limited by high snowpack [Graumlich et al., 1989; Gedalof and Smith, 2001; Peterson and Peterson, 2001; Case and Peterson, 2005; Littell et al., 2008; Pederson et al., 2011]. Such negative correlations with winter snowpack are observed in the PNW in high-elevation tree species such as mountain hemlock (Tsuga mertensiana), subalpine fir (Abies lasiocarpa), subalpine larch (Larix lyallii), and occasionally in other species in places that experience temporally coherent, anticorrelated low-water availability because of Pacific teleconnections (e.g., in Nevada). Hence, incorporating such relationships can improve the skill of proxy reconstructions, although some variability will remain unresolved due to variations in the relationship between local climate and teleconnections over time.

[23] The second growth relationship is characterized by annual or seasonal water facilitation in which tree growth is limited by low-water availability [Brubaker, 1980; Cook et al., 2004;Graumlich, 1987; Gray and McCabe, 2010; Littell et al., 2008], which can be driven, at least in part, by the fate of snowmelt [Pederson et al., 2011]. In the PNW, positive correlations with precipitation, including winter, are most common in Douglas-fir (Pseudotsuga menziesii), ponderosa pine (Pinus ponderosa), western Juniper (Juniperus occidentalis), and sometimes species such as limber pine (Pinus flexilis), whitebark pine (Pinus albicaulis), and Engelmann spruce (Picea engelmannii). Depending on the precipitation regime, some species can be either snow limited or water facilitated in different parts of their biogeographic range (e.g., mountain hemlock [Peterson and Peterson, 2001]).

[24] The winter precipitation reconstruction was composed of four steps, including screening, empirical orthogonal function/principal components (EOF/PC) analysis, multiple linear regression modeling, and variance restoration. We considered a total of 498 chronologies in the Columbia Basin or within roughly 150 km of the boundaries of the basin: 327 from the International Tree-Ring Data Bank (ITRDB) for which raw ring widths were available, 124 Douglas-fir chronologies [Littell et al., 2008], 41 non-ITRDB chronologies from collaborators (see Acknowledgments), and six newly developed chronologies. The newly developed chronologies consisted of: three Larix lyallii chronologies from the North Cascades of WA and the headwaters of the Columbia River in MT (included in the work of Pederson et al. [2011]), two Pseudotsuga menziesii chronologies from central ID, and one Picea engelmannii from central ID. For the three subalpine larch chronologies, high-elevation cirques where snowpack lingers late into the season were targeted. For the two Douglas-fir chronologies, slopes with low soil development in two locations in central ID were targeted. The Engelmann spruce chronology was from a dry high-elevation site in central ID. In all six cases, site chronologies were developed from between 13 and 36 sampled trees of various ages, two cores per tree. Subfossil wood was used to extend all three larch chronologies and one Douglas-fir chronology. Only one of the larch chronologies (OLL, see Table 1) had a significant relationship with streamflow in the Yakima Basin.

[25] We screened these chronologies for those spanning the period 1614–1979 (to incorporate 1977 low flows and to include as many nonupdated chronologies as possible in the PNW), resulting in 219 (164 ITRDB chronologies and 56 non-ITRDB chronologies). We evaluated the raw cross-dating in the remaining 164 ITRDB chronologies (using the program COFECHA [Holmes, 1999]) and detrended and standardized them (using ARSTAN [Cook and Holmes, 1999]). All tree ring series (ITRDB and non-ITRDB) were conservatively detrended using a negative exponential, straight line fit through the mean or, in a few cases, cubic spline two-thirds the length of the series. Each standard chronology was calculated as the biweight robust mean of all series [Cook et al., 1990]. All 219 chronologies were then further screened against historical basin-average cool season precipitation Pobs. The 38 chronologies that exhibited statistically significant correlations (|rsig| >= 0.25, n = 62, p = 0.05) were retained for empirical orthogonal function/principal components (EOF/PC) analysis and are depicted in Figure 3.

Table 1. Negatively Correlated/Snow Limited Driving Tree-Ring Chronologies, Sources, and Contribution to Principal Components (Neg PC1, PC2, and PC3)a
ChronologyLocationSpecies Name/AbbreviationrrPC1rPC2rPC3
  • a

    r is the correlation coefficient between tree-ring growth in an individual chronology and cool season precipitation, indicating type of sensitivity. rPC1-3 are the correlation coefficients between chronologies and principal components used in the regression analysis. Uncited sources are unpublished material.

  • b

    Contributors of the International Tree-Ring Data Bank [2011].

  • c

    Peterson and Peterson [2001].

  • d

    M. Colenutt and B. Luckman, personal communication.

or042bCrater Lake, ORMtn. hemlock/TSME−0.400.24−0.100.22
or087bLane Plateau, ORMtn. hemlock/TSME−0.400.22−0.070.26
or066cMount Hood, ORMtn. hemlock/TSME−0.370.35−0.120.11
or069cMount Jefferson, ORMtn. hemlock/TSME−0.370.25−0.040.01
nv514bSpruce Mountain, NVGreat Basin bristlecone pine/PILO−0.340.210.13−0.05
wa104cMount Adams, WAMtn. hemlock/TSME−0.330.27−0.040.10
or086bHusband Lake, ORMtn. hemlock/TSME−0.320.27−0.060.17
or065cMount Hood, ORMtn. hemlock/TSME−0.300.22−0.030.20
nv052bMoody Mountain, NVWestern white pine/PIMO−0.300.130.810.10
wa097cHoh Lake, WAMtn. hemlock/TSME−0.290.17−0.010.11
wa081bMt. St. HelensPacific silver fir/ABAM−0.290.13−0.090.12
nv060bJackson MountainsWestern juniper/JUOC−
wa064bSunrise Lake, WASubalpine larch/LALY−0.280.370.03−0.62
nv049bPony Express, NVWestern white pine/PIMO−
or078cCrater Lake, ORMtn. hemlock/TSME−0.270.28−0.080.16
RL92dRowe Lakes, ABSubalpine larch/LALY−0.270.10−0.08−0.25
or026bAbbott Creek, ORDouglas-fir/PSME−
OLLOval Lakes, WASubalpine larch/LALY−−0.47
wa066bWhite Pass, WAMtn. hemlock/TSME−0.250.25−0.050.11
wa048bFrying Pan Creek, WADouglas-fir/PSME−0.250.21−0.06−0.05

[26] Empirical orthogonal function/principal components (EOF/PC) analysis via singular-value decomposition of the centered data matrix [e.g., Preisendorfer, 1988] was used to extract common signals from the individual chronologies. PC analysis has been applied in multiple hydroclimatic reconstructions [Gedalof et al., 2004, Hidalgo et al., 2000; Meko et al., 2001, 2007; Woodhouse et al., 2006]. To account for predictive skill in both temporally coherent growth-climate relationships and lagged climate-growth relationships [e.g., Wettstein et al., 2011], four separate EOF/PC analyses were carried out for chronologies that were negatively and positively correlated with Pobs, and for chronologies that were negatively and positively correlated with lag 1 cool season precipitation (growth in the year following the cool season precipitation anomaly). Standard chronologies were used as input to the EOF/PC process because the coherent variation among chronologies at even multidecadal timescales is probably climatic in nature in the Pacific Northwest and therefore should be retained [Wettstein et al., 2011]; the reconstruction method would preferentially incorporate in-common variability at decadal timescales.

[27] We used the eigenspectrum plot to identify 10 lead PCs (of 40 total) as potential predictors for the multiple linear regression model of basin-average cool season precipitation (we assumed the remaining PCs to be inseparable from noise). We applied a forward selection approach [Woodhouse et al., 2006] calibrated with the historical basin-average cool season precipitation Pobs (section 2.4). The leading PCs were entered hierarchically according to initial r2 values, but terms were retained in the model when p < 0.1 and the reduction of the error statistic (RE) was maximized. Interactions between frequently significant but unretained PCs were also explored. To aid in interpretation of the sign of regression predictors, because the sign of the loadings of the chronologies on the first PC time series is arbitrary, the sign of the time series was multiplied by −1 if all contributing chronologies were negatively correlated with the first PC (the case with the leading PC of the positively correlated chronologies and the leading PC of the lead 1 negatively correlated chronologies). Through this selection process the reconstructed basin-average cool season precipitation Preg was defined as a function of six PCs in five regression terms,

display math

[28] The first term is the second PC time series derived from chronologies which are positively correlated with cool season precipitation (drought-sensitive) primarily in the year prior (lead 1 tree rings, 16% total EOF variance explained, regression t = 3.9). The second and third terms in the regression were the first and second PC time series of chronologies negatively correlated with cool season precipitation (snow limited) in the year of growth (28% and 18%, respectively, EOF variance explained, regression t = −3.6 and −2.7). The fourth term is the second lead 1 PC time series of trees primarily positively correlated with precipitation in the year of growth (25% total EOF variance explained, regression t = 2.4). The last term was an interaction term between the third principal components of the chronologies negatively and positively correlated with winter precipitation in the year of growth (PC3 positive and negative explain 14% and 13% of the variance in respective positive and negative chronologies). This term appears to capture the contrast in variability between snow limited (larch chronologies in the central WA Cascades, OLL and WA064, Table 1) and precipitation-facilitated chronologies (ID010 and ID046, Table 2). Chronologies that contributed to the selected principal components of the snow-sensitive, water-sensitive, and lead 1 water-facilitated networks are listed in Tables 13, respectively.

Table 2. Water Facilitated Chronologies (Positively Correlated With Cool Season Precipitation), Sources, and Contribution to Principal Components (Pos PC3 and Pos PC2 lead 1)a
ChronologyLocationSpecies Name/AbbreviationrrPC2rPC3
  • a

    r is the correlation coefficient between tree-ring growth in an individual chronology and cool season precipitation, indicating type of sensitivity. rPC2-3 are the correlation coefficients between chronologies and principal components retained in the regression analysis.

  • b

    Contributors of the International Tree-Ring Data Bank [2011].

or085bMill Creek RNA, ORPonderosa pine/PIPO0.310.54−0.39
id010bSand Pass, IDWhitebark pine/PIAL0.29−0.160.62
cana220bWhirlpool Pt., ABLimber pine/PIFL0.26−0.38−0.02
or046bLookout Mt., ORPonderosa pine/PIPO0.260.490.67
or029bCross Canyon, ORPonderosa pine/PIPO0.250.54−0.05
Table 3. Tree-Ring Chronologies With Positive Correlation With Lag 1 Cool Season Precipitation, Sources, and Contribution to Principal Components (Lead1 Pos PC2)a
ChronologyLocationSpecies Name/AbbreviationrrPC2
id013bSleeping Deer Road, IDEngelmann spruce/PCEN0.250.44
wa104cMount Adams, WAMtn. hemlock/TSME0.29−0.12
Or065cMount Hood, ORMtn. hemlock/TSME0.26−0.69
wa101cMount Rainier, WAMtn. hemlock/TSME0.360.56

[29] The resulting regression model (Table 4) provided an adjusted calibration R2 of 0.51. For the calibration period (1917–1979), the correlation between Pobs and Preg was 0.74. To maximize the use of the observed record we chose a leave-one-out cross-validation to validate the reconstruction. A root-mean-square error of validation (RMSEv) of 118.5 mm (for the calibration period mean precipitation = 777.3 mm) and a reduction of error (RE) = 0.45 indicated a reasonably robust model.

Table 4. Regression Model for Reconstruction of Cool Season Precipitation Over the Yakima River Basin and Other Calibration/Validation Informationa
Regression TermEstimateStd. ErrortPr(>t)
  1. a

    aRegression Performance:

    Residual SE: 112.4 on 58 degrees of freedom.

    Multiple R2: 0.5529, adjusted R2: 0.5143.

    F-statistic: 14.34 on 5 and 58 DF, p-value: <0.0001.

    Durbin-Watson statistic = 2.2774, p-value = 0.819 (no autocorrelation in residuals).

    Leave-one-out validation: Reduction of error statistic (RE, or 1-PRESS/SSEnull): 0.45 (moderate model skill).

    Split-same validation, 1917–1947 model: R2: 0.54; RE: 0.39; sign agreement: 25 of 32 (moderate model skill).

    Split-same validation, 1948–1979 model: R2: 0.47; RE: 0.36; sign agreement: 24 of 31 (moderate model skill).


[30] We also used a split sample verification scheme to further test the reconstruction. The same model variables used in the full calibration model (1917–1979) were used in two separate regression equations calibrated on the first half of the period 1917–1947 and the second half of the period 1948–1979. The 1948–1979 values of the PC predictors were then input into the 1917–1947 regression and the predicted values tested against the observed 1948–1979 values for winter precipitation. Similarly, the 1917–1947 values of the predictors were input into the 1948–1979 regression and the predicted values tested against the observed 1917–1947 winter precipitation values. To assess the predictive ability of the models, we calculated RE statistics and used a sign test to characterize the number of disagreements in sign relative to the mean values of the observations and reconstructed values. For the model calibrated on 1917–1947, R2 = 0.54, RE = 0.39, and 25 of 32 pairs had sign agreements among the 1948–1979 predicted and observed values (binomial test p < 0.01). For the model calibrated on 1948–1979, R2 = 0.47, RE = 0.36, and 24 of 31 pairs had sign agreements among the 1917–1947 predicted and observed values (binomial test p < 0.01). This validation suggests that models built on subsets of the full calibration period exhibit similar skill to the full model and that the PC variables entered into the full model are reasonably robust to changes in the data set.

[31] To account for the variability inherently lost in regression reconstruction, we restored the variance of the reconstruction by standardizing the reconstruction time series to the calibration mean and variance. This was accomplished by multiplying the residuals of the reconstructed mean inline image by the ratio of the historical and reconstructed standard deviations (s):

display math

2.6. Meteorological Ensemble Selection

[32] Under unlimited processing resources, we could estimate the uncertainty of warm season streamflow in each retrospective year by combining the seasonal and daily variability from all 91 years of historical meteorological data with the paleoreconstructed estimate of cool season precipitation. Because of processing limitations, we selected spatial and temporal patterns of historical meteorological data (section 2.4) from five water years that span the range of warm season streamflow in the historical VIC model simulations. Here we describes the ensemble selection process.

[33] To produce an ensemble of preprocessed historical meteorological data that would reproduce the full range of warm season streamflow values for a set amount of cool season precipitation, five 91-year VIC test simulations were run. For each simulation, the historical daily variations of precipitation were scaled so that their cool seasonal sum matched a constant value, equal to the 5th, 25th, 50th, 75th, or 95th quantile value [Cunnane, 1988; Maidment, 1993] of cool season precipitation from the historical record. Each 91-year simulation produced a unique cumulative distribution (CDF) quantifying warm season streamflow variability under the prescribed amount of cool season precipitation.

[34] From these CDFs, five specific water years that consistently spanned the range of streamflows in all five prescriptions were selected to form a 5-year ensemble of historical meteorological data (Table 5). Most historical years maintained consistent streamflow ranking under the range of dry and wet cool season precipitation prescriptions (Figure 4). This demonstrates that, in the Upper Yakima Basin, variations in climate from year to year (such as variations in temperature or in the seasonality of cool season precipitation) consistently influenced warm season flow, regardless of the amount of cool season precipitation. Some anomalous years, which possessed abnormal summer precipitation values, were specifically avoided, since their ranking varied significantly with cool season precipitation amount (outliers in Figure 4).

Figure 4.

Comparison of warm season streamflow quantile values (QSSF) produced under five prescribed precipitation conditions, spanning the wetness range of the historical period (QPCP = 0.05, 0.25, 0.50, 0.75, 0.95) versus the annually based average of all five QSSF. Each distribution contained 91 years of historical data (1916–2006). Gray points represent individual QSSF for all years. Black points show selected years used to represent the range of uncertainty in warm season flow.

Table 5. Quantile Stability of the Selected Ensemble Yearsa
Year0. From Mean
  • a

    Summer streamflow quantile values of the selected ensemble years were derived from VIC simulations spanning dry (QWPRCP = 0.05) and wet (QWPRCP = 0.95) winter precipitation conditions.

19370.0610.0610.0500.0500.0390.0520.022−0.009 – 0.013
19530.2700.2700.2590.2700.2590.2660.011−0.004 – 0.007
19830.4780.4780.5110.5220.5550.5090.077−0.046 – 0.031
19330.7300.7410.7520.7520.7520.7450.022−0.007 – 0.015
19260.9500.9500.9610.9610.9710.9590.021−0.012 – 0.009

2.7. Integration of Paleoreconstructed Precipitation With Ensemble of Historical Meteorological Patterns

[35] We integrated the paleoreconstructed basin-average cool season precipitation Pdendro with the 5-year ensemble of historical meteorological data that robustly captures the range of variation in warm season flow (described above). For each of the five resampled meteorological data sets, the spatially and temporally distributed daily time step cool season precipitation Pt,x was scaled to match the (tree-ring-derived) basin-average cool season precipitation Pdendro for each retrospective year by multiplying Pt,x by the ratio of the basin-averages of tree ring and observed seasonal precipitation,

display math

where Pt,x,Y and Pobs,Y are the spatially distributed daily time step series and the basin-average cool season precipitation, from the resampled historical year Y. Thus, the reconstructed series inline image is a daily time step series that possesses spatial and temporal precipitation patterns from the historic year Y scaled to match Pdendro. This procedure was applied separately for each ensemble member, resulting in five separate 366-year meteorological reconstructions (Y = 1–5). Note that both the warm season precipitation and the temperature data for the entire water year were retained from the five resampled water years.

2.8. Paleoreconstruction of Warm Season Streamflow

[36] Five 366-year VIC simulations were executed, each forced by one of the five meteorological reconstructions described above. This produced five simulations of warm season streamflow spanning the range of uncertainty in each retrospective year. For the calibration period (1916–1979), 78.1% of the historical naturalized summer streamflow values fell within the simulated bands for the corresponding year and 100% within the extremes of the simulation (Figure 5). This analysis confirms that the range of warm season streamflows from the five-member ensemble appropriately bounds the observed warm season streamflows. The Wilcoxon Rank sum test [Wilcoxon, 1945] found the difference in central tendency of the historical naturalized streamflow and the combined-ensemble distribution (20 cms) to be insignificant (p = 0.2035) (Figure 6). The predictive strength of an individual ensemble member reproducing streamflow was tested using linear regression. The central ensemble member (green line in Figure 5) was significantly (p < 0.0001) correlated with the naturalized streamflow (black line in Figure 5) and described 57% of its variance (r2 = 0.57).

Figure 5.

Observed (black line) versus simulated warm season streamflow during tree ring calibration period (1915–1979). Five streamflow reconstructions (colored lines) were generated from five meteorological reconstructions that integrate tree ring-derived precipitation with historical meteorological data to reproduce a broad range of streamflows for each reconstructed year (red: 1937, orange: 1953, green: 1983, blue: 1933, purple: 1926).

Figure 6.

Boxplot comparison of observed and simulated warm season streamflow for the historic period 1916–1979. Black line represents the median, whiskers represent the range, box limits delineate the interquartile range (25th–75th percentile), and dark gray triangles on boxes represent approximate 95% nonparametric confidence intervals of the mean [McGill et al., 1978]. The delta value is the difference in median values, shown here to be statistically insignificant (p = 0.2035) using the Wilcoxon Rank sum test. Bracketed values show sample size.

2.9. Climate Change Scenario

[37] We evaluated how future climate changes may impact warm season streamflow in the Upper Yakima Basin for a projected 2040s (2030–2059) climate by comparing the paleoreconstructed streamflow with VIC simulations forced with a perturbed version of the meteorological data set based on a climate change scenario. We applied a composite delta method [Elsner et al., 2010], perturbing the meteorological forcings with the average projected monthly changes in temperature and precipitation (Table 6) from multiple GCM simulations run for the period 2030–2059 [Mote and Salathé, 2010] under the A1B emissions scenario [Nakićenovićet al., 2000]. In contrast to previous studies constrained by historic meteorological records, these simulations apply systematic future changes in climate to the entire 366 years of reconstructed climate used to drive the VIC model.

Table 6. Projected Changes in Monthly Air Temperature and Precipitation for 2040a
ΔP (%)
ΔTmin (°C)1.961.631.781.911.842.382.862.952.811.971.621.97
ΔTmax (°C)1.961.631.781.911.842.382.862.952.811.971.621.97

3. Results

3.1. Cool Season Precipitation Reconstruction

[38] Figure 7 shows the paleoreconstruction of basin-average cool season precipitation Pdendro. Multidecadal variability is apparent in this record. The reconstruction exhibits a weak (r2 = 0.015) but significant (p = 0.016) decreasing linear trend. This long-term trend is largely driven by the temporal distribution of extreme events. Withholding the two wettest winters (occurring in the early 1700s) and the two driest winters (occurring after 1800) reduced the statistical significance considerably (p = 0.04). We discuss this trend in more detail in section 4 below.

Figure 7.

Reconstructed basin-average cool season (October–March) precipitation Pdendro, 1614–1979. Solid black line is the estimated mean and dashed black line represents long-term linear trend (p = 0.016, r2 = 0.015). Gray line represents historical values Pobs.

3.2. Warm Season Streamflow Reconstructions

[39] We simulated daily time step streamflow (aggregated to warm season values) for the Upper Yakima from 1614 to 1979, using the VIC hydrologic model driven by extended meteorological reconstructions. For each reconstructed year, five warm season streamflow values (colored lines in Figure 8) were derived from the five meteorological reconstructions. Warm season streamflow values for the unperturbed paleoreconstruction and the climate change scenario are listed in Table 7.

Figure 8.

Summer streamflow reconstruction, 1614–1979. Five reconstructions were generated using the basin-average reconstructed cool season precipitation and an ensemble of historical data, representing a range of possible streamflows (red: 1937, orange: 1953, green: 1983, blue: 1933, purple: 1926). Solid black line represents observed (naturalized) streamflow.

Table 7. Basic Statistics of Summer Streamflow Reconstructions
Ensemble EmberSSFa (cms) PaleoreconstructionSSFa (cms) Climate Change A1BΔSSFa (cms)
(Hist. Source Year)MeanMin.Max.RangeMeanMin.Max.RangeMeanMin.Max.Range
  • a

    SSF is the summer streamflow and ΔSSF is the difference in streamflow during paleo and projected climatic conditions in 2040 under A1B emissions scenario.

Group mean2004934229310536168132−94−14−174−161
Group range107341711371222621719115−84654

[40] Paleoreconstructed warm season streamflow ranged between 32 and 420 cms. Similar to the cool season precipitation reconstruction (Figure 7), individual ensemble members of the warm season streamflow reconstruction possessed a significant (p = 0.015) long-term decreasing trend with weak explanatory value (r2 = 0.015). However, when all five streamflow ensemble members are treated as a single data set, the variability introduced by the individual ensemble members obscures any linear trend (p = 0.96, r2 < 0.0001). Interdecadal variations are also apparent. Between ensemble members, the warm season streamflow varied by 107 cms on average, with the largest spread (171 cms) occurring in wet years and the smallest spread (34 cms) occurring in dry years.

[41] The long-term paleoreconstructions of cool season precipitation and warm season flow for the Yakima Basin case study show an expanded range of variability in comparison with shorter records for the 20th century alone. In particular the paleoreconstuctions show an extended wet period from ∼1700–1750, and an extended dry period from ∼1825–1855, that are longer and/or more intense than similar episodes in the observed 20th century record. A number of reconstructed single-year floods and droughts in the paleorecord are more severe than reconstructions of the most severe events in the 20th century, again suggesting greater variability in the past.

[42] The climate change simulation (Figure 9) incorporated monthly changes of both temperature and precipitation (Table 6) and produced warm season streamflow values that were 47.5% lower than historical values (from 200 cms to 105 cms, Table 7). Individual values ranged between 26 and 308 cms (Table 7). Historical streamflow observations were typically higher than the entire ensemble of simulated streamflow under perturbed climatic conditions (Figure 10). As expected, the simulations for the climate change scenario show a dramatic reduction in warm season flow in comparison with the unperturbed paleoreconstruction (Figure 11). The statistically significant decrease in summer streamflow (Figure 12a) is coupled with a significant increase in winter streamflow (Figure 12b). These simulations are consistent with analogous results shown by Elsner et al. [2010] and Vano et al. [2010].

Figure 9.

Simulation of warm season streamflow under perturbed temperature and precipitation conditions for the proxy period 1614–1979. The five reconstructions were generated from the perturbed meteorological ensemble (red: 1937, orange: 1953, green: 1983, blue: 1933, purple: 1926). All reconstructions exhibit a reduction in streamflow under the climate change scenario. Black line represents observed (naturalized) 20th century streamflow (1915–1979).

Figure 10.

Comparison of warm season streamflow simulation under climate change conditions (colored lines, for proxy period 1915–1979) and observed (naturalized) streamflow for the historical period 1915–1979 (bold black line).

Figure 11.

Comparisons of warm season streamflow under paleo- (blue) and climate change (orange) conditions for proxy period 1614–1979 (maroon band represents overlapping conditions). Black lines represent ensemble mean values. (Note that years correspond with the source year of the basin-average cool season precipitation reconstruction before perturbation.)

Figure 12.

Boxplot comparisons of (a) warm season and (b) cool season reconstructed paleo and future seasonal streamflow based on the climate change perturbation. Black square represents the median, whiskers represent the range, box limits delineate the interquartile range. Narrow dark gray triangles on boxes represent approximate 95% nonparametric confidence intervals of the mean [McGill et al., 1978]. The delta value is the difference in median values, shown here to be highly significant (p < 0.0001) using the Wilcoxon Rank sum test. Bracketed values represent sample size.

[43] Unlike the unperturbed paleoreconstruction (Figure 8), the ensemble members of the climate change scenario are not as evenly distributed over the range of possible streamflow values and the ensemble members change in their relative ranks (Figures 9 and 10). This indicates that the ensemble members vary in their sensitivity to the prescribed changes in monthly air temperatures and precipitation. For example because the prescribed precipitation changes are ratio based, relatively wet summers respond more strongly to percent reductions in precipitation than relatively dry ones. In addition, although streamflow variability decreased within each ensemble member, the difference between the ensemble members' mean flows, described through their group range, increased from 107 cms to 122 cms (Table 7).

[44] Graphical comparisons between the extended chronologies and cumulative distributions of warm season streamflow for baseline and climate change perturbed conditions are shown in Figure 13. For the calibration period (1916–1979), there is general agreement between observed and simulated baseline flows, while streamflow under perturbed climate conditions is systematically reduced (Figure 13a). The difference between baseline and climate change simulations is also evident in their supercumulative distributions, when derived from the entire reconstruction period (1614–1979, Figure 13b).

Figure 13.

Comparison of streamflow cumulative distributions for (a) the calibration period (1916–1979) and (b) the full reconstruction (1614–1979). Dotted line represents observed values, solid and dashed lines represent simulated baseline and climate change scenarios, respectively. Simulation distributions are derived from supercdfs (all ensemble members).

4. Interpretation/Discussion

[45] The reconstructed warm season streamflow values show a large range of uncertainty in each retrospective year as a result of the selected historical meteorological driving data. This ensemble reconstruction approach demonstrates that warm season streamflow can vary substantially for a given value of the independent proxy reconstruction of cool season precipitation due to variations in timing and spatial distribution of meteorological events controlling warm season streamflow (e.g., peak snow accumulation, melt timing, etc.).

[46] The subtle but statistically significant long-term trend in the reconstructed cool season precipitation must be interpreted with some caution. While this may indicate a real decrease in winter precipitation over centennial timescales, it could also result from proxy reconstruction techniques or from changes in the hydroclimatic-tree growth relationships over time. Although the selected “snow-sensitive” chronologies may have close inverse correlations with snowpack in the historical period, growth can also be positively affected by warm spring and summer temperatures, which can result in earlier loss of snow cover and an extended growing season. Collection of additional winter precipitation sensitive trees in the future may help to further evaluate the apparent long-term trend in winter precipitation in the Yakima reconstruction.

[47] In section 2.6 we demonstrated that a small group of historical water years could be selected to estimate the range of simulated warm season flows for a given cool season precipitation volume. This approach was robust for deriving the paleoreconstructed values, evident in the consistent spacing between values of the five ensemble members (Figure 8). However, the selected ensemble members reacted differently to the prescribed temperature changes in the climate change scenario, resulting in an uneven spacing between ensemble members for these runs (Figure 9). Thus, using a larger ensemble of retrospective years, or further incorporating a temperature sensitivity analysis into the ensemble selection process, would likely improve the representation of uncertainty in the climate change scenario.

[48] Consistent with previous climate change studies [Elsner et al., 2010], the climate change scenario demonstrates that rising air temperatures will result in dramatic reductions in warm season streamflow by 2040 in the Yakima Basin, despite wetter winter conditions. These extended scenarios, however, provide additional realizations of observed streamflow variability (e.g., the sequencing, intensity, and duration of droughts), which are useful for water resources studies, such as those by Vano et al. [2008, 2010]. Thus, by combining the wider range of variation evident in the paleoreconstructions with projected future climate, the performance of water management systems under the combined effects of climate variability and change can be more thoroughly and completely analyzed. We can also use the VIC scenarios to quantify the expected increases in cool season streamflow (Figure 10). Although this study used a meteorological ensemble selected to quantify warm season streamflow uncertainty, the same technique could be applied to more accurately quantify past and future winter streamflow, which may change as much or more than summer streamflow under climate change in the PNW [Elsner et al., 2010].

5. Summary and Conclusions

[49] In this study we have integrated tree-ring proxy reconstructions of cool season precipitation with observed daily time step meteorological data sets to produce an ensemble of extended meteorological reconstructions that were used as inputs to the physically based macroscale variable infiltration capacity (VIC) hydrologic model for the period 1614–1979. This approach allowed us to reconstruct warm season streamflow over multiple centuries, quantify the physical uncertainty in these seasonal reconstructions, and estimate the systematic effects of changes in temperature and precipitation associated with climate change scenarios over a multicentury time frame.

[50] The methods developed for this paper provide a means for combining the best qualities of proxy hydroclimatic reconstructions, hydrologic modeling, and climate change studies. For example, interannual reconstructions of streamflow, even over many hundreds of years, are less useful to water resources managers if their subseasonal variation and uncertainty cannot be quantified or resolved. The use of a physically based hydrologic model allows us to provide information on a finer temporal and spatial scale than might otherwise be possible. Similarly, future climate change impacts studies that only consider the narrower hydrologic variability of the historical/instrumental record probably underestimate the range of variability in future conditions.

[51] The long-term paleoreconstructions of cool season precipitation and warm season flow for the Yakima Basin case study show an expanded range of variability in comparison with shorter records for the 20th century alone, with an extended wet period in the first half of the 18th century and an extended dry period in the mid-19th century that are longer and/or more intense than the extremes of the 20th century. Although it should be interpreted with some caution, a statistically significant though weak decreasing trend in cool season precipitation and summer streamflow is also apparent in the paleoreconstructions over the last 350 years or so.

[52] The sensitivity study integrating a scenario of temperature and precipitation change for the 2040s with the paleoreconstructions shows dramatic reductions in simulated warm season streamflow in the Yakima Basin. These experiments also demonstrate that use of the reconstructed seasonal streamflow values without temperature corrections for future conditions would be highly misleading in the context of 21st century water resources planning.

[53] There is evidence from the climate change simulations that the choice of historical water years used to estimate the uncertainty in the summer streamflow may not fully represent the range of uncertainty under warming scenarios. Future studies should use a wider range of resampled years to better quantify the uncertainty.

[54] Using these methods as a foundation, the possibility exists to conduct a number of “experiments” that attempt to quantify seasonal hydrologic extremes within the context of the annual variability of the proxy reconstructions. For example, using the low range (i.e., the 5th percentile) of the warm season flow simulations over a retrospective period, a “worst case” single-year drought scenario could be constructed (e.g., compare the 1977 drought to the 5th percentile estimate for all years in Figures 5 and 8). Such scenarios may be valuable in characterizing the performance of water resources systems under extreme low-flow conditions.

[55] Streamflow uncertainties associated with the variability of other physical drivers, such as changes in land cover, could be incorporated into this method of paleoreconstruction. Finally, while we have focused on reconstructing and simulating seasonal streamflow, the approaches used here can be readily applied to improving our understanding of the variability of other important hydrologic variables simulated by the VIC model, such as snowpack, soil moisture, annual peak flows, extreme low flows, and water deficit.


[56] This study has been funded through NOAA CPO SARP (NA07OAR4310371, N. Mantua, J. Littell, A. Hamlet, and C. Woodhouse). This paper is partially funded by the Joint Institute for the Study of the Atmosphere and Ocean (JISAO) under NOAA Cooperative Agreement NA17RJ1232, contribution 1831. Significant contributions were made by J. Deems in meteorological data processing, M. McGuire Elsner in VIC implementation, and P. Carrasco with programming support. We thank C. Woodhouse for suggestions that helped us develop this project. We also thank G. Pederson and S. Gray for the use of chronologies in the Northern Rockies, M. Colenutt and B. Luckman for the use of the Rowe chronology, and Contributors of the International Tree-Ring Data Bank, IGBP PAGES/World Data Center for Paleoclimatology, NOAA/NCDC Paleoclimatology Program, Boulder, CO, USA.