Introduction

The Arctic exerts a large influence on global climate and is currently undergoing profound environmental changes1,2,3. In order to put these into a larger context and to anticipate future climate and environmental changes, we must understand the degree to which and the reasons why, Arctic temperatures have varied in the past. There are limited options for paleotemperature reconstruction in the Arctic region. Tree ring-based records are generally shorter than a millennium and low-frequency (multi-decadal to centennial) changes in temperature are probably not fully captured by such reconstructions4,5. Ice cores are geographically restricted (high latitudes and altitudes) and provide temperature estimates atop ice sheets, which can differ significantly from temperatures at sea level6. Lake sediments are excellent archives of Arctic temperature because they comprise various materials that can be used for paleoenvironmental reconstruction and, as a dominant feature of the Arctic landscape, provide the potential for a dense network of paleoclimate reconstructions. Lake-based reconstructions are contingent upon observed relationships between the sedimentary component of interest and one or more environmental or climatic variable and quantitative records require calibration.

Here we introduce “proxy-to-proxy” (PTP) calibration, using visible range scanning reflectance spectroscopic (VIS-RS) data, obtained through rapid, non-destructive and inexpensive analysis of sediment cores at very high resolution7,8. We report a PTP calibration between the VIS-RS RABD660:670 index7 and alkenone unsaturation () data9,10, from two overlapping cores from lake Braya Sø, west Greenland. was calibrated in situ in Braya Sø during two field seasons and a 5,600 year long record from Braya Sø sediments was replicated in a second lake in the region9. This and additional work11,12,13 provide strong evidence that reflects past lake water temperature. The index RABD660;6707 represents chlorin (molecules deriving from chlorophyll, primarily from autochthonous primary productivity) content within the sediment and therefore the balance between autochthonous primary productivity in the photic zone and organic matter preservation in lake sediments. Braya Sø is currently a meromictic lake with permanently anoxic bottom waters. The preservation of fine (< 1 mm), uninterrupted laminations in the sediments indicates that the lake bottom has been suboxic to anoxic during the past 7,000 years. Therefore, preservation potential is unlikely to have varied substantially and we interpret RABD660;670 of Braya Sø sediments as a record of autochthonous primary productivity.

Autochthonous productivity in arctic lakes is positively correlated with summer temperature14,15,16,17,18 through the impact of temperature on 1) the duration of the ice-free period (i.e., the productive season) and, potentially, 2) the nutrient flux to lakes from the watershed19. It follows that the chlorin content in Braya Sø sediments (quantified with RABD660;670) would be closely related to summer temperature.

Results

Developing the proxy-to-proxy calibration

Because the RABD660;670 data were obtained at much higher resolution than the data, for calibration purposes we numerically re-sampled them to match the sampling interval of the data from 0–84 cm in the core (Fig. 1). Both the timing and the amplitude of variability are very similar for both measurements, with the only notable discrepancies occurring between 54–61 and 69–84 cm. The Pearson's r of 0.76 for the regression between the RABD660;670 and data is highly significant (Pc = 8.51 10−8). The RMSEP (root mean square error of prediction) for both the Ordinary Least Squares (OLS) and Standard Major Axis (SMA) regression methods was small (OLS: 0.0243 and SMA: 0.0264 units) and the verification statistics high (OLS: RE = 0.62, CoE = 0.55; SMA: RE = 0.58, CoE = 0.50). Both the OLS and SMA models yield very similar results (with SMA producing only slightly larger amplitudes of variation) implying that the model choice is not critical in this case. The measurements below 84 cm were not used for the calibration because they reach anomalously high values (shaded area in Fig. 1). These high values correspond to the first appearance of alkenones in Braya Sø and, unlike the rest of the record, cannot be validated with results from a nearby lake9. The new chlorin based reconstruction strongly supports the temperature record presented in D'Andrea et al. (2011)9 and demonstrates that Lake Braya Sø is well suited to reconstruct past temperatures.

Figure 1
figure 1

Proxy-to-proxy calibration.

Left panel: Proxy-to-proxy calibration between the resampled VIS-RS RABD660;670 (blue) and alkenone data (red). The data below 84 cm were not used for the calibration (shaded area). Right panel: The regression plot and calibration statistics. DF: degrees of freedom, P: P-Value. The subscript “c” denotes values corrected for autocorrelation and trends.

Full size image

Improvements over original data series

When the calibration equation between RABD660;670 and is applied to the full RABD660;670 data set, sampling resolution of the -based temperature record is improved five-fold (661 instead of 131 measurements) (Fig. 2). Due to the additional uncertainty introduced by the PTP calibration, the RMSEP is larger (OLS: 1.66°C and SMA: 1.71°C) than for the reconstruction alone (1.33°C). We note that our reported RMSEP is an overestimate because it does not take into account the cross-validation of the two temperature proxies provided by the PTP approach; however, further work is needed to develop a statistical quantification of this impact.

Figure 2
figure 2

PTP-derived temperature reconstructions.

Ordinary least squares regression (OLS) in red; Standard Major Axis regression (SMA) in green. The dashed lines show the respective RMSEP. The alkenone-based () reconstruction and its RMSEP (vertical lines) in blue.

Full size image

Spectral analysis performed on the higher-resolution (PTP-based) OLS temperature reconstruction allows detection of multi-decadal variability that occurred over shorter timescales than are detectable by spectral analysis of the original -based temperature reconstruction of D'Andrea et al. (2011)9 (Fig. 3). The reconstruction exhibited two prominent frequency peaks at 230 and 297 years but those were not significant at the critical false-alarm level identified at 98% (see Methods). In contrast, the PTP reconstruction had frequency peaks exceeding the 99.7% false alarm level at 24–25, 63–64 and 81–82 years. The latter two correspond to the periodicities of 63–64 and ~81 years associated with the Atlantic Multidecadal Oscillation (AMO)20,21,22.

Figure 3
figure 3

Spectral analysis of time-series'.

Bias-corrected spectra of (A) the original temperature reconstruction and (B) the new PTP temperature reconstruction. Solid line is the spectrum, dashed line is the 99% and dotted line is the 95% false-alarm level. Frequencies are reported as 1/years and spectral amplitudes are plotted on a logarithmic decibel [dB] scale.

Full size image

Comparison with an independent AMO record

Figure 4 shows that our record shares a significant common spectral signal in the range of 51–64 years with the tree ring-based sea surface temperature (SST) reconstruction of Gray et al. (2004)23 which was used to reconstruct changes in the AMO since 1567 CE. When both records are band-pass filtered in the 50–70 year range to isolate the AMO signal, the reconstructed variance in both reconstructions is very similar (Fig. 5). We therefore suggest that our record provides insight into the long-term behavior of the AMO. A cross spectral analysis of the band-pass filtered records shows that the Gray et al. (2004)23 leads our new AMO series by ca. 20 years. Although there could be a mechanistic explanation for this phase shift, a systematic offset due to dating uncertainties cannot be excluded and indeed, this may be the more likely explanation. The full 7,000 year long band-pass filtered temperature record (Fig. 6) indicates that AMO intensity (or its influence on west Greenland's climate) was significantly higher during the last 2,000 years than prior to that. There is an intriguing period of particularly low variance from 2500–3500 BCE in the RABD660;670 record that potentially reflects an interval of very weak AMO variability. A 1,000 year period of suppressed AMO should be captured by other geological archives and additional reconstructions are needed to validate this surprising result. Over the last 2,000 years, our record also indicates that variance in the AMO band was largest during the Mediaeval Climate Anomaly (MCA; ca. 950–1250 CE)24 and reduced during the time interval associated with the Little Ice Age (LIA; 1250–1700 CE)25. Suppressed variance in the AMO frequency band during the LIA relative to the MCA has also been suggested based on analyses of ice core proxy records from Greenland26,27.

Figure 4
figure 4

Cross-spectral analysis with SST reconstruction.

Blackman-Harris 3-Term Cross-spectral coherence plot of the PTP (this study) and the Gray et al. (2004)23 SST record. The dashed line indicates the 90.6% false alarm level.

Full size image
Figure 5
figure 5

Band-pass filtering at periods of the Atlantic Multidecadal Oscillation.

Comparison of the 50–70 yr band-pass filtering results of the standardized PTP record of this study (black) with a proxy-based North Atlantic SST reconstruction23 (red) for the past 500 yrs. The inset shows a lagged correlation curve. The highest correlation between the two series is achieved when the filtered PTP record is shifted by 20 years (dashed gray line).

Full size image
Figure 6
figure 6

Band-pass filtered record beginning −5000 years C.E.

50–70 year band-pass filtered data (corresponding to the period of the AMO) for the entire PTP-based reconstruction, from –5000 to 2005 CE.

Full size image

Recent model-based results suggest that the AMO is a 20th century phenomenon driven by anthropogenic aerosol emissions28. However, our results indicate that the AMO has varied in intensity but has been a persistent mode of variability throughout the past 7,000 years. Aerosol forcing in the 20th century may have modulated the underlying AMO signal, but our results contradict the assertion that the AMO is purely anthropogenic.

Discussion

There have been three principal approaches to proxy calibration. In the calibration-in-time (CIT) approach, proxy measurements spanning the monitoring interval are regressed against long-term climatic data. This approach has been used extensively in tree-ring research29,30and has been applied in varved31,32,33 (annually laminated) and non-varved lakes11,15,34,35,36, but requires excellent chronological control. For non-varved sediments, radionuclide dating (such as 210Pb or 137Cs) can be used, but this is often hampered by low concentrations of the species of interest, particularly in the Arctic. Without an extremely accurate chronology in the uppermost part of the sediment core, it is not possible to precisely compare proxy data to meteorological series. Furthermore, the accuracy and error of the age model influences the maximal temporal resolution and root square mean error of prediction (RSMEP) of the climate reconstruction. Thus it is not surprising that the majority of high-resolution and quantitative temperature reconstructions from Arctic lakes are based on varved sediments37,38,39,40. Furthermore, for CIT to be successful, the top of the core must 1) be unaffected by non-climatic factors such as eutrophication or sediment slumping (turbidites), 2) have a sufficiently high sedimentation rate to provide a large enough sample set for statistically significant correlation and 3) have an intact sediment-water interface to ensure the upper samples represent modern sedimentation35. The CIT approach is further limited by the availability and quality of meteorological data. The Arctic generally has a very low density of meteorological stations with records usually covering fewer than 50 years. This makes the CIT approach particularly challenging and helps explain the scarcity of highly resolved and quantitative lake sediment temperature records from the Arctic41.

Another proxy calibration approach is the space-for-time (SFT) method, in which modern biological samples from lakes spanning a spatial climatic gradient are used to develop a proxy calibration. This approach has been used to develop diatom and chironomid transfer functions in the Arctic42,43 and “global” TEX86 and branched tetraether temperature calibrations for lake sediments44,45,46.

A third approach is to quantify the response of modern populations of organisms in situ to environmental parameters over one or more growth seasons. A calibration equation thus developed can then be applied to down-core sedimentary components. This approach has been successfully employed to develop lacustrine water temperature calibrations for alkenones, using the index9,12.

PTP calibration enables higher resolution paleoclimate reconstructions than are possible or practical using most sedimentological proxies of interest, due to the sampling limitations of those proxies. Another clear advantage of the PTP approach is that it can potentially be used in any region and type of environment. Unlike the CIT approach, PTP calibration does not depend on age model accuracy, which is often a limitation in non-varved lakes (the majority of lakes worldwide). Furthermore, for PTP calibration, the top of the sediment core need not be undisturbed or even intact. Another advantage is that the PTP correlations can be conducted using observed (in situ) data from the water column of the lake and the associated sediment core (as in this study) and is therefore possible even in the absence of a long local meteorological data set. Finally, PTP can substantially reduce the laboratory workload and costs associated with generating paleotemperature reconstructions. A statistically significant PTP calibration could be developed with as few as 50 correlation points (analyzed samples). Rapid scanning methods such as VIS-RS allow a meter of sediment to be analyzed at mm scale in under an hour. As demonstrated here, the PTP method can provide useful information, increasing both the length and temporal resolution of quantitative climate reconstructions.

Methods

Site

Lake Braya Sø (66°59′15″N, 51°02′0″W, ca. 170 m a.s.l) is located near Kangerlussuaq, West Greenland. It is a meromictic lake of 82 ha with a maximum depth of 24 m and a stable anoxic hypolimnion that favors preservation of pigments47,48.

Alkenones

The details of lipid extraction are described in D'Andrea et al. (2011)9. Briefly, free lipids were extracted from freeze-dried and homogenized samples at 0.5 cm intervals and alkenones were quantified by gas-chromatography-flame-ionization-detection. The -temperature calibration was established by combining in situ water temperature measurements made during two seasons of alkenone production in Braya Sø with previously published calibration data9,49. = ([C37:2] − [C37:4])/([C37:4] + [C37:3] + [C37:2]), where [C37:x] = mass of the alkenone having 37 carbon atoms and x double bonds. was originally defined in ref. 10.

VIS-RS

Because the sediment core (BS01-01) analyzed by D'Andrea et al. (2011)9 was consumed for alkenone analysis, the VIS-RS measurements were conducted on an overlapping core (BS03-01). The cores were easily stratigraphically aligned at mm-scale using laminations in the sediment (compared using high-resolution photographs from a Geotek Geoscan IV linescan). We measured VIS-RS (visible spectrum: 380–730 nm; spectral resolution: 10 nm) on the fresh unoxidized split sediment core at 2 mm intervals with a Spectrolino reflectance spectrometer (Gretag McBeth, Switzerland). The characteristic patterns of the reflectance spectra provide information on the composition and relative concentration of the sedimentary components. This method has been successfully used to measure organic compounds7,34,35 and lithogenic content50,51,52 in both marine and lake sediments. Here we use the parameter, Relative Absorption Band Depth at 660–760 nm (RABD660;670)7, indicative of total early diagenetic chloropigments (chlorin), which have absorption peaks between 660–670 nm (I-band). RABD660;670 was normalized by the mean spectrum reflectance (RABD660;670/Rmean)7. Figure 7 shows the measured reflection spectra of the sediment core of Braya Sø, the chlorin curve and how it is calculated.

Figure 7
figure 7

Reflectance spectroscopy in the visible range for a sediment core from Braya Sø.

(A) VIS-RS spectra from each 2 mm increment down core (length 120 cm). The bold line is the mean spectrum Rmean. (B) RABD660–670 data (chlorin). RABD660;670 = {[(6*R590 + 7*R730)/13]/Rmin(660;670)}, where R590 = reflectance at 590 nm, R730 = reflectance at 730 nm, Rmin(660;670) = minimum reflectance at 660 or 670 nm. Spectrum normalization and calculation of relative absorption band depths is fully described in ref. 7.

Full size image

Age model

The age model covering the last 7,000 years for Braya Sø is based on 6 AMS 14C dates of bulk sediment, 3 of wood and 1 of a gastropod shell9. A reservoir effect correction of 360 years was applied to the 14C dates prior to calibration9. The sediment core had an intact sediment-water interphase that was attributed the year 2005 CE of the coring9.

Numerical methods

The statistical basis of the PTP approach is a double calibration. First, we use the relation between the VIS-RS and alkenone data to transform the RABD660;670 data into “inferred” alkenone values. These values are then transformed into temperatures using the - temperature calibration. Because VIS-RS (2 mm) and alkenones (5–10 mm) were measured at a different resolution, the RABD660;670 data were re-sampled (linear interpolation) to match the data resolution for this calibration.

We used the Pearson product-moment coefficient to determine the correlation between the RABD660;670 and series. The P-values were corrected for trends and autocorrelation (Pc)53.

To statistically evaluate the PTP calibration we used the split period calibration approach and calculated the reduction of error (RE) and coefficient of efficiency (CoE) values following the method of Cook et al. (1994)29. The modern half of the samples was used for calibration and the older half for verification. We calculated the ten-fold cross-validated RMSEP using all depths to estimate the reconstruction error.

We used both an ordinary least squares regression (OLS) and Standard Major Axis regression (SMA) to transform the raw RABD660;670 series into “inferred ” () values. These i values were then transformed into temperatures with the -temperature calibration (T = 40.8 + 31.8, RMSEP: 1.3°C) of D'Andrea et al. (2011)9. The final error of the reconstruction was calculated by summing (square root of the sum of the squares) the RMSEP (in °C) of the two calibrations.

The PTP VIS-RS temperature reconstruction presented here and the reconstruction of D'Andrea et al. (2011)9 have non-regular temporal sampling intervals. Therefore, for the spectral analysis, we used the software REDFIT by Schulz and Mudelsee (2002)54 allowing the input of unevenly spaced time series to estimate their red-noise spectrum and test if peaks in the spectrum of our temperature reconstructions are significant against a first-order autoregressive (AR1) model. False alarm levels were determined within the software package.

To determine whether the tree-ring based SST reconstruction23 and our temperature reconstruction from Greenland share a common spectral phase in the domain of AMO, we calculated a Blackman-Harris 3-Term cross spectral analysis with the software SPECTRUM55 (Fig. 5). The analysis was performed on the raw, unevenly spaced reconstruction series. A significant common spectral phase was identified in the range of 51–64 years.

We then band-pass filtered our reconstruction series and the SST data of Gray et al. (2004)23 in the frequency band of AMO (50–70 years) using the Parks-McClellam algorithm. Because our reconstruction has non-regular temporal intervals, we linearly interpolated it to annual values before filtering it. Although data interpolation does introduce artifacts in the data series, their effect should be small in the low frequency band of interest. Means were removed before filtering. The spectral phase shift of 20 years between our band-pass filtered reconstruction and the band-pass filtered record by Gray et al. (2004)23 was determined by cross-correlation analysis.