Reconstructing Northeastern United States temperatures using Atlantic white cedar tree rings

Initial AWC collections at eight sites throughout the Northeast were made between 2002 and 2003 (Hopton and Pederson 2005). In 2015 we updated and re-sampled three of these locations: Appleton, Maine, Saco Heath, Maine, and Westminster, Massachusetts (figure 1). These sites are in the northern region of the species range, with the forest in Appleton representing the northernmost known stand of thespecies (Stockwell 1999) (figure 1) and are relatively undisturbed by anthropogenic influences (Hopton and Pederson 2005, Pederson et al 2004) (table 1).

At the re-collected sites, AWC dominates the canopy and is largely even-aged. The Appleton, ME, site is a bog in the headwaters of the St. George River with a ground layer of Sphagmnum moss and fern species. Saco Heath, ME, is the only known domed bog to contain AWC, and is possibly the southernmost coalesced bog in the eastern United States (Laderman 1989). Saco Heath contains scattered aggregations of trees throughout as well as other areas of dense shrub dominated by blueberry (Vaccinium ssp.L.). AWC collected in Westminster, MA, are in a topographically depressed swamp environment that borders a wetland and ponds system. The Westminster swamp is considered to be a high elevation inland AWC swamp (NHESP 2007) (table 1) with interspersed tamarack (Larix laricina), red maple, and red spruce (Picea rubens Sarg.). All three sites stay wet for most, if not all, of the growing season, supplied with water from rainfall as well as flow from nearby rivers or ponds.

We collected AWC samples following standard dendrochronological techniques (Fritts 1976, Stokes and Smiley 1968), taking two to three increment cores per tree, then drying, mounting, and sanding the increment cores. To ensure we assigned the correct year to each annual ring, the increment cores were graphically and visually crossdated (Stokes and Smiley 1968, Yamaguchi 1990). Ring widths were measured at 0.001 mm precision and crossdating was statistically confirmed using COFECHA (Holmes 1983).

To remove the geometric growth trend and isolate the climate signal in the tree-ring series, we detrended and standardized the width measurements into site chronologies at all eight sites using both a standard negative exponential or linear growth curve (NEGEX) as well as a signal free method (SF) (Melvin and Briffa 2008, Briffa and Melvin 2011). Our tree-ring series all originate from living, mature, and canopy dominant conifers, which typically and historically have been treated with a NEGEX curve to retain climate signals (Fritts 1976). More recently, signal-free standardization has been developed to attempt to avoid possible trend distortion or end effects related to the presence of common medium-frequency variability (Melvin and Briffa 2008, Briffa and Melvin 2011). To account for changes in the number of series back in time, the variance in the chronologies was stabilized based on the interseries correlation (Cook et al 1995) and a 67% spline (Osborn et al 1997, Cook and Peters 1981, 1997). We used the autoregressive (AR)-standardized version of the chronology for our climate analysis to preserve the common autoregressive structure of the tree-ring data due to variations in climate (Cook 1985). We assessed the common signals in the tree-ring chronologies using the interseries correlation and expressed population signal (EPS) statistics. Interseries correlation is the mean correlation between all the cores. EPS indicates how well the sample of trees estimates the signal of a hypothetical population and is a function of sample size and interseries correlation (Wigley 1984). The length of the chronologies were limited to years where EPS values are at least 0.85, the conventional although arbitrary threshold value (Wigley 1984). We created a Northeast regional chronology for large-scale spatial climate analysis by averaging the ring-width measurements of the three updated sites (Appleton, Saco, Westminster) to create a single series (the 'regional chronology'). To detect possible disturbance signals in our regional chronology, we analyzed the individual series that composed the region chronology using both the Nowacki and Abrams' criteria 1997, and that of Lorimer and Frelich 1989.

We analyzed the association between tree-ring chronologies and local climate using gridded monthly temperature anomalies data from the NASA GISTEMP combined 250 km product(Hansen et al 2010), gridded monthly precipitation data from version 7 of the Global Precipitation Climatology Center (GPCC Schneider et al 2016), gridded sea surface temperatures from the UK Met Office Hadley Centre sea surface temperature (HADISST Rayner et al 2003). We calculated an Atlantic Multidecadal Oscillation (AMO) index using a weighted average of the UK Met Office Hadley Centre sea surface temperature gridded product for the Atlantic from 0° N−70° N (Rayner et al 2003, Enfield et al 2001). We performed seasonal correlation analyses as described by Meko et al (2011) to calculate the both the Pearson correlation and partial correlation coefficients of the chronologies with monthly and seasonal temperature and precipitation. Statistical significance was evaluated using exact simulation (Percival and Constantine 2006, Meko et al 2011). We computed seasonal correlations for our regional chronology based on temperature and precipitation averages over an area spanning from 41° N−48° N and 74° W to 62° W. To understand the spatial extent of the temperature signal in our trees, we calculated the Pearson correlation of the regional chronology with the GISTEMP temperature field. Pointwise statistically significant correlations were determined using the approach described by Ebisuzaki (1997).

Based on our site-level climate analysis, we reconstructed regional mean January through August temperature for the region from 41° N−48° N and 74° W to 62° W using a nested composite-plus-scale (CPS) approach (Meko 1997, Esper et al 2002, Cook et al 2002, Esper 2005). This method scales the tree-ring series to the mean and standard deviation of the instrumental observations during the calibration period, and then evaluates the fit between tree-ring reconstructed temperatures and the instrumental data during the validation period. We tested the sensitivity of the reconstruction to both our choice of detrending method and to the inclusion or exclusion of data from each of the three individual updated sites. We built reconstruction models using three nests (all of the recollected sites) and two nests (only the Maine sites), as well as building a reconstruction model for both SF and NEGEX chronologies. We used a split calibration/verification (1900−1955 and 1956−2010) to train and evaluate our models (Michaelson 1987, Meko 1997). We estimated the uncertainty of our reconstruction from the root mean square error (RMSE) of validation. The variance explained by our reconstruction was quantified using the R² statistic. We also used the reduction of error (RE) and coefficient of efficiency (CE) statistics to estimate reconstruction skill, with positive RE and CE values indicating the reconstruction performed better than a naive estimate of the mean (Cook et al 1999, Wahl and Ammann 2007).

There is a clear trend of increasing temperature correlation with higher latitudes in our extended AWC network (figure 1). However, this trend is not monotonic; the AWC series from Saco Heath has the highest correlation with winter-through-summer temperature but is at a lower latitude than the northernmost Appleton site. The cluster of sites near 41° N all have correlations with temperature less than r = 0.35. Our seasonal correlation analyses (figure 2) confirm that local temperature has the strongest correlation with the northern three sites and our temperature-sensitive AWC series show a broad winter-through-summer sensitivity with a peak seasonal correlation with mean temperatures in January through August (Hopton and Pederson 2005). Whilering widths of trees growing in cold environments generally reflect growing season spring and summer temperatures, tree-ring width can also be influenced by temperatures in the prior winter months (Jacoby et al 1996, Pederson et al 2004).

Zoom In Zoom Out Reset image size

The Appleton site has significant correlations with January–September temperatures, with the strongest correlations in the early spring. Appleton shows no significant precipitation partial correlation (figure 2(a)). Saco Heath has the highest correlation values with temperature, with seasonal correlations reaching r = 0.66 (figure 2(b)). The site also has statistically significant secondary and partial correlations with precipitation, especially in summer. Westminster swamp shows positive and significant winter through summer temperature correlations, and a seasonally narrow but significant summer precipitation signal (figure 2(c)).

Our complete regional chronology consists of 232 cores from 116 trees at three sites, with a series mean intercorrelation value of r = 0.49 over a common period of 170 years, from 1845−2014, and EPS values over 0.95 back to 1860 (Wigley 1984). No significant endogenic disturbance was detected in the chronology. We found that both SF and NEGEX chronologies had high Pearson correlations to regional temperature, but the SF detrending procedure increased the amplitude of the early and late ring-widths beyond reasonable growth patterns for the species (Cook et al 1995). Consequently, we used the NEGEX detrended chronologies for the remainder of our analysis and reconstruction, which preserved low frequency signals without exaggerating the growth trend over recent decades.

Similar to the individual sites of which it is composed, the strongest seasonal temperature signal for the regional chronology spans from winter through the end of the growing season (figure 3). The regional chronology also shows broad and significant correlation with temperature spatially, with significant correlations over eastern New England and with particularly high correlations of r > 0.50 over the adjacent North Atlantic to 65° W (figure 4, figure 6). Correlations over land cease to be significant south of Long Island, NY, but continue north of Maine into Canada. Western New England still has field correlation values over r = 0.4, but these fail to be statistically significant when accounting for autocorrelation (Ebisuzaki 1997).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

A skillful reconstruction of January through August mean temperature is possible using the two Maine tree-ring sites back to 1872, when the Saco Heath chronology currently ends (figure 5). Our model has positive RE and CE scores of 0.33 and 0.08, respectively, and an r² value of 0.34 from 1872−2014. Results of our cross validation for a three site (including Westminster) reconstruction had positive RE and CE values with an early validation period and late calibration period, but slightly negative CE values with a reversed calibration and verification period. These low frequency metrics can be sensitive to the calibration/verification period—particularly CE (Wahl and Smerdon 2012, Wahl and Ammann 2007)—and we found that relatively small shifts in thecalibration/validation period affected these statistics. Our sensitivity tests using SF instead of NEGEX chronologiesall yielded negative cross-validation RE and CE scores, indicating that low frequency behavior and detrending choices in this chronology strongly influenced reconstruction skill.

Our skillful reconstruction of Northeast temperature tracks instrumental temperature within ± 1 RMSE for most of the calibration and validation period. The tree-ring reconstruction captures both the long-term century-scale trend as well as multidecadal variability. Our chronologies have somewhat smaller rings in the earliest 1920s, latest 1960s, and late 1990s, compared to observed temperatures. The reconstruction shows a decade and a half of relatively stable mean temperatures from the late 1870s to the 1900s, prior to our calibration and validation periods, but indicatecold winter-through-summer temperatures in 1866 and 1868. However, in this portion of our reconstruction the skill is reduced because the Saco Heath chronology does not cover this period.