Revising midlatitude summer temperatures back to A.D. 600 based on a wood density network
Abstract
Annually resolved and millennium-long reconstructions of large-scale temperature variability are primarily composed of tree ring width (TRW) chronologies. Changes in ring width, however, have recently been shown to bias the ratio between low- and high-frequency signals. To overcome limitations in capturing the full spectrum of past temperature variability, we present a network of 15 maximum latewood density (MXD) chronologies distributed across the Northern Hemisphere extratropics. Independent subsets of continental-scale records consistently reveal high MXD before 1580 and after 1910, with below average values between these periods. Reconstructed extratropical summer temperatures reflect not only these long-term trends but also distinct cooling pulses after large volcanic eruptions. In contrast to TRW-dominated reconstructions, this MXD-based record indicates a delayed onset of the Little Ice Age by almost two centuries. The reduced memory inherent in MXD is likely responsible for the rapid recovery from volcanic-induced cooling in the fourteenth century and the continuation of warmer temperatures until ~1600.
Key Points
- First millennial climate reconstruction based on extratropical tree ring density network
- Improved ratio between low- and high-frequency temperature variability
- New insights into postvolcanic cooling and the onset of the Little Ice Age
1 Introduction
Integration of climate proxies and development of continental-scale temperature reconstructions of the past millennium has greatly improved our understanding of the spatial patterns of long-term climate variability [PAGES 2k Consortium, 2013]. It has been shown [Bunde et al., 2013; Franke et al., 2013], however, that the spectral characteristics of temperature fluctuations in proxy-based reconstructions differ systematically from observations, with the proxies overestimating the ratio of low- to high-frequency variability. This “red bias” [Franke et al., 2013] inherent to reconstruction spectra is a fundamental constraint limiting the assessment of preinstrumental climatic extremes [Anchukaitis et al., 2012; D'Arrigo et al., 2013; Esper et al., 2013], which in turn has implications for the estimation of the Earth's climate sensitivity through comparisons of climate model simulations with proxy-based reconstructions of preindustrial temperature variability [Frank et al., 2010].
Almost all millennial-scale temperature reconstructions assessed in the Intergovernmental Panel on Climate Change Fifth Assessment Report (IPCC AR5) [Masson-Delmotte et al., 2013] are either solely based on tree ring data or a combination of these data with multiple additional proxies—most of which have lower temporal resolution. The interannual variance in these records is nevertheless dominated by tree ring width (TRW) measurements, a parameter known to contain substantial biological memory [Frank and Esper, 2005] due to the persistence of needle generations and storage of reserves throughout dormancy [Fritts, 1976]. These effects translate into an overestimation of the persistence of postvolcanic temperature cooling [Frank et al., 2007a; Masson-Delmotte et al., 2013] and biased spectral characterization of tree ring-based temperature reconstructions [Bunde et al., 2013; Franke et al., 2013]. Maximum latewood density (MXD), in contrast, has the capacity to more directly respond to and recover from pulse-like cooling [Briffa et al., 1998] and disturbance [Esper et al., 2007a] events. Less physiological noise results in a stronger common signal among trees and stronger coherence between observational target and proxy data [Esper et al., 2013]. Moreover, compared to TRW, the relative magnitude of biologic age trends in MXD is smaller with respect to the standard deviation of the samples, therefore reducing the distortion effects that arise during the necessary process of detrending, which again supports the fidelity of millennial-scale temperature variance estimated from this parameter [Esper et al., 2012]. Nevertheless, a large-scale compilation of the longest MXD chronologies has not yet been used to estimate hemispheric-scale high- to low-frequency variability in temperature over the last millennium.
We introduce herein a network of the 15 longest available MXD records (Table S1 in the supporting information) from Asia, Europe, and North America (4/7/4), assess the climate signal inherent in these data, and develop an alternative Northern Hemisphere (NH) reconstruction. The proxy network comprises in total less site chronologies than the most recent TRW-based reconstruction [D'Arrigo et al., 2006] but is similarly replicated at the onset of the last millennium. Excluding TRW from our evaluations allows us to assess the potential benefits of a parameter free of biological memory effects and its ability to theoretically overcome the spectral limitation associated with TRW-derived estimates.
2 Data and Methods
2.1 Proxy Network
Proxy data for this study are publicly available at the International Tree Ring Data Bank or contributed by the original authors [Anchukaitis et al., 2013; Briffa et al., 2013; Büntgen et al., 2006, 2008; Esper et al., 2007b, 2012; Luckman and Wilson, 2005; Melvin et al., 2013; Myglan et al., 2012; Schweingruber et al., 1988]. All data sets were corrected separately for biological and site-related noise using regional curve standardization (RCS) [Briffa et al., 1992; Esper et al., 2003]. Residuals between the power-transformed [Cook and Peters, 1997] MXD series and an expected age-related growth curve were calculated and averaged using the robust biweight mean to form site chronologies. Tree ring data from Athabasca and Mangazia comprised site- or species-specific offsets in the mean growth level. These sites were divided into subgroups, according to species composition and growth rate, detrended separately using RCS, and recombined to form mean site chronologies (Text S1) [Luckman and Wilson, 2005].
The varying sample replication of the single site chronologies necessitates variance stabilization, here achieved by fitting a 100 year spline through the absolute departure values and dividing the chronologies by this spline curve. Other plausible tree ring standardization methods (e.g., signal-free detrending or replication-based variance correction) were also tested, yielding similar results (Figure S1 and Table S2). Fifteen detrended site chronologies longer than 600 years and replicated with at least three samples served as the final predictor network (Figure 1 and Table S1). The shortest record starts in 1363 and the oldest terminates in 1976, the last year of the calibration period. All records passed a benchmarking experiment evaluating the strength of local June-July-August (JJA) temperature responses via a random process. A 95% confidence level is derived from the probability distribution of correlations between local temperature and 1000 red noise time series with first-order autocorrelation matching each MXD record. Local correlations between each MXD chronology and temperature served as the weights in the hemispheric composites used in the reconstruction procedure. Records from the Alps and Scandinavia were averaged to regional composites to mitigate the European sampling bias.
2.2 Instrumental Target
The instrumental target is monthly mean JJA temperature derived from the 5 × 5° CRUTEM4v network (30°–90°) [Jones et al., 2012]. The field was truncated in 1901, due to an increasing number of empty grid cells (53% at the onset of the twentieth century and 86% in 1850) and subsequent variance reduction in the NH mean due to the gap filling via the regularized expectation maximization algorithm using ridge regression [Schneider, 2001]. Regional and hemispheric temperature averages were weighted by area, using the cosine latitude of local grid cell centers. For analyses of growth responses, we chose CRUTEM4v temperatures from the grid cell containing the corresponding tree site. Cells with very short records were replaced by records from neighboring cells (e.g., in eastern Siberia; Figure 1). Monthly averaged diurnal temperature maxima for JJA, tested as an alternative instrumental target [Wilson et al., 2007], yielded neither consistent nor significant improvements in the calibration results.
2.3 Calibration/Validation
The regional MXD chronologies were combined into an ensemble estimate of hemispheric temperature variability by calculating weighted composites based on moving correlations with local temperature. The composite averages were scaled against NH extratropical temperatures over the 1901–1976 period [von Storch et al., 2009]. Calibration and validation statistics (R2 and RE metrics [Briffa et al., 1988]) were estimated using a 38 year holdout window incremented by 1 year between 1901 and 1976 to derive an ensemble of 76 plausible reconstruction members (Figure S2). The procedure enabled the estimation of a calibration error derived from minimum and maximum reconstructed temperatures for each year. The decreasing number of predictors back in time, typical for dendroclimatological studies, results in increasing uncertainty in earlier periods, which was accounted for by recalculating the reconstruction and associated errors each time a site record dropped out (Figure S3) [Cook et al., 2002]. Such a nested approach also allowed the variance to be adjusted to account for the diminishing predictor network, as well as validation metrics and sampling error estimates to vary in time. Artificial reduction of site-specific replication by randomly drawing from all MXD series in the twentieth century revealed only negligible differences in estimated errors (Figure S4). Moreover, the sensitivity of the reconstruction against a changing predictor network and different calibration choices was assessed (Figures S5 and S6). The recent end of the proxy network is characterized by a quickly declining number of sites. The 1976–2002 period was thus reconstructed with only one nest containing seven records extending into the 21st century.
3 Results
The MXD network has reasonable coverage over the last millennium with all data sets reaching back to A.D. 1363. Each chronology explains a significant fraction of regional summer temperature variability, verified in a benchmarking experiment, and the combined correlation fields (p < 0.05) from all records represent 90% of the NH extratropical landmass (Figure 1). In addition to validating the MXD data against local observations and regional temperature fields, the combined site chronologies display strong coherence at continental (North America, Europe, and Asia) scales (Figure 2). The explained variance among 30 year low pass-filtered continental means over the common A.D. 1363–1976 period is 
21363–1976 = 0.45 (R2America/Asia = 0.47, R2America/Europe = 0.40, and R2Asia/Europe = 0.49) indicating sensitivity to common forcing and/or modes of variability at multidecadal to centennial timescales. The correlation coefficients decline during the early, less replicated portion of the records. At interannual timescales there is no coherence between the continental composite ( 
21363–1976 = 0.03), corresponding to the reduced spatial coherency of surface temperature fields in the higher-frequency domain [Jones et al., 2012].
The large-scale reconstruction correlates (Pearson's r) at r1901–1976 = 0.60 with modern JJA temperatures and estimates preinstrumental temperatures to have spanned 0.54 ± 0.34°C between the coldest 30 year periods during the Little Ice Age (LIA) (1627–1656 and 1809–1838) and the warmest such period during late medieval times (1396–1425) (Figure 3). The same temperature range is reconstructed from the LIA to the warmest most recent 30 year period (1927–1956). The reconstruction indicates moderate summer temperatures from ~ A.D. 900 to 1600 at the hemispheric scale, with maxima occurring in the late ninth and midfifteenth centuries, followed by cooler summers from ~ 1600 to 1900 and subsequent twentieth century warming.
The new summer temperature reconstruction was validated using stepwise ensemble techniques, a sliding calibration/verification approach and pseudoproxy experiments (Text S1 and Figures S2, S3, and S7; see Smerdon [2012] for a review). Despite considerable empirical uncertainty, the reconstruction proved robust to a number of methodological choices during calibration (Figures S6a–S6h). Potential divergence issues [Esper et al., 2010; Wilson et al., 2007], which become apparent in some of the local records in the late twentieth century, were evaluated by compiling two independent records derived from eight “nondiverging” and five “diverging” MXD site chronologies (Figures S8–S10). Despite low site replication, both subsamples share a considerable amount of low-frequency variability in their common period (1363–1976) followed by a more significant offset in the late twentieth century. In the complete composite, decoupling between the instrumental and proxy-derived temperatures is noted in the most recent decades. The new JJA temperature reconstruction also coheres closely with a shorter reconstruction based on 300+ MXD sites distributed over the NH extratropics, the so-called “Schweingruber MXD network” [Briffa et al., 2001] (Figure 3b). These two almost independent data sets (estimated data overlap < 1%) correlate at r = 0.70 over the past ~300 years. An offset of ~0.2°C evolving before 1450 is likely related to the changing spatial coverage in the Schweingruber network, while sample and site replication remain stable during this period in our new network of long MXD records.
4 Discussion and Conclusion
Comparison of the new MXD-based reconstruction with the multiproxy and TRW-based records assessed by the IPCC [Masson-Delmotte et al., 2013] reveals substantial differences at low and high frequencies (Figure 4). Modern warming is poorly represented in the new record questioning the ability of MXD to capture very warm temperatures. The last several decades of the reconstruction are, however, represented by a smaller number of chronologies. Among these, only a couple appears to be impacted by divergence, suggesting deficiencies regarding spatial representation or chronology quality, rather than a whole-scale limitation of the proxy. While the overall sequence of warmer conditions during medieval times, subsequent LIA cooling, and twentieth century warming is in accordance with the IPCC collection of reconstructions, the new record deviates substantially regarding the initiation of the LIA and the millennial-scale temperature amplitude. The Medieval Warm Period is less pronounced in our reconstruction, and summer temperatures remain at a relatively warm level throughout the ~1300–1450 period, when the existing reconstructions indicate a prolonged cooling trend. Residuals are largest in the late fifteenth and sixteenth centuries during which the MXD-based estimates point to warmer conditions, thereby delaying the onset of the LIA until the early seventeenth century—a substantial change important to the discussion about potential drivers of NH cooling [Schurer et al., 2014] and the interpretation of potential societal responses [Büntgen et al., 2011].
Some of these differences possibly arise as a result of varying seasonal or spatial coverage. Although most of the reconstructions in the IPCC ensemble are likely biased toward a summer signal, many of them target annual temperatures. For TRW-based reconstructions, this assumption is ascribed to the longer season portrayed by ring width due to biological-driven persistence [D'Arrigo et al., 2006; Frank et al., 2007a]. Thus, the more confined seasonal fingerprint of MXD could explain deviations to some extent. Other studies, in contrast, state that multidecadal variability should be largely independent of the season chosen [Cook et al., 2004], and we find some evidence for that at least during the observational period (Figures S6i and S6j). By focusing strictly on long MXD records, this reconstruction has a reduced spatial coverage during the calibration period compared to the most recent tree ring-based reconstruction [D'Arrigo et al., 2006]. Coherency with observational data is nevertheless good in the middle- and low-frequency domains and does not imply discrimination based on the network extent (Figure S11). Despite these findings, it is also clear that spatially expanding the extent of the NH MXD network would greatly benefit future paleoclimatological research.
A remaining rationale for deviating long-term trends and the delayed LIA onset likely originates from the different response of MXD data, compared to TRW, to a cluster of major volcanic eruptions in the thirteenth and subsequent centuries (compared to the relative lack of such events during medieval times) as revealed in Greenland and Antarctic ice core data [Gao et al., 2008]. For the TRW and multiproxy reconstructions, spectral power is shifted from annual and multidecadal timescales to lower frequencies, compared to the model simulations and MXD reconstruction, which indicates prolonged responses to climate variations in the traditional data (Figure 4b and Text S1). At the large spatial scale addressed here, these differences become visible during the periods subsequent to the three largest volcanic eruptions of the past millennium in 1257, 1452, and 1815 (Figures 4c–4e). Whereas the TRW-dominated reconstructions show a weak and temporally extended postvolcanic response, the MXD-based reconstructions reveal distinct postvolcanic cooling and recover to mean climatology much faster. Of particular interest is the temperature drop after the 1257 event, recorded in ice cores, providing further evidence for the precise dating of tree ring records over the entire millennium [Anchukaitis et al., 2012]. The MXD-based reconstruction from Briffa et al. [2001] indicates less cooling after the 1452 event, most likely related to decreased spatial replication in the earliest part of this record. Cooling in response to Tambora (1815) is less distinct as it coincides with the generally cold Dalton Minimum (~1790–1830) [Wagner and Zorita, 2005].
Model-simulated temperatures support the fast recovery from volcanic-induced cooling [Wigley et al., 2005] as recorded in the MXD-based reconstruction, although some models reveal a much stronger—and potentially too strong—temperature response [Landrum et al., 2013]. The temperature trend from the Medieval Warm Period into the LIA also appears reduced in some model simulations [Fernández-Donado et al., 2013]. It remains unclear, however, whether this is due to less persistent volcanic cooling, an underestimation of internal variability, or the influence of solar variability. Considering the differing spectral characteristics identified herein (Figure 4b), and confirmed by Bunde et al. [2013] and Franke et al. [2013], it seems important that our new MXD-based reconstruction has similar spectral properties to the simulated temperatures at frequencies ranging from 0.5 to 0.003 cycles per year. Although millennium-long oscillations are poorly sampled, our calculated beta values indicate that the new temperature history is not red biased, as is typical for the TRW and multiproxy reconstructions. Consideration of the new reconstruction might thus strengthen efforts to estimate the climate sensitivity to volcanic eruptions and the persistence of subsequent lower tropospheric cooling. The late initiation of the LIA around 1580 revealed in our reconstruction might additionally help to further elucidate ecosystem responses and historical linkages throughout the past millennium.
Acknowledgments
We thank all of our colleagues for developing millennial-length MXD chronologies and making these data available. This study was supported by the Mainz Geocycles Research Centre. Supplementary information is available in the online version of the paper. Data are available via the NOAA online server. Lamont contribution 7902.
The Editor thanks two anonymous reviewers for their assistance in evaluating this paper.
