Six temperature proxies of Scots pine from the interior of northern Fennoscandia combined in three frequency ranges

Research Article Six Temperature Proxies of Scots Pine from the Interior of Northern Fennoscandia Combined in Three Frequency Ranges Markus Lindholm,1Maxim G. Ogurtsov,2Risto Jalkanen,1 Björn E. Gunnarson,3and Tarmo Aalto1 1Metla, Rovaniemi Research Unit, P.O. Box 16, 96301 Rovaniemi, Finland 2A.F. Ioffe Physico-Technical Institute, St. Petersburg 194 021, Russia 3Bolin Centre for Climate Research, Department of Physical Geography and Quaternary Geology, Stockholm University, 106 91 Stockholm, Sweden Correspondence should be addressed to Markus Lindholm; [email protected] Received 20 December 2013; Revised 7 April 2014; Accepted 11 April 2014; Published 6 May 2014 Academic Editor: Silvio Gualdi Copyright © 2014 Markus Lindholm et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Six chronologies based on the growth of Scots pine from the inland of northern Fennoscandia were built to separately enhance low, medium, and higher frequencies in growth variability in 1000–2002. Several periodicities of growth were found in common in these data. Five of the low-frequency series have a significant oscillatory mode at 200–250 years of cycle length. Most series also have strong multidecadal scale variability and significant peaks at 33, 67, or 83–125 years. Reconstruction models for mean July and June–August as well as three longer period temperatures were built and compared using stringent verification statistics. We describe main differences in model performance (𝑅2= 0.53–0.62) between individual proxies as well as their various averages depending on provenance and proxy type, length of target period, and frequency range. A separate medium-frequency chronology (a proxy for June–August temperatures) is presented, which is closely similar in amplitude and duration to the last two cycles of the Atlantic multidecadal oscillation (AMO). The good synchrony between these two series is only hampered by a 10-year difference in timing. Recognizing a strong medium-frequency component in Fennoscandian climate proxies helps to explain part of the uncertainties in their 20th century trends. 1. Introduction Several recent studies have discussed the potential of highresolution proxies based on the growth of Scots pine from northern Fennoscandia for reconstruction of summer temperatures in particular at the low-frequency scale of variability [1–5]. The main concern has generally been the interesting temperature difference between medieval times, Little Ice Age, and the modern period viewing recent and projected warming within the context of natural variability. Less attentionisusuallypaidtothestrongmultidecadalcomponentof temperaturevariabilityintheobservationalaswellasproxy records, which may seriously hamper the identification of an amplified warming signal in the last century in the Arctic and surrounding regions [6]. Some internal climate controls mayhaveinfluencedregionalclimatesimultaneouslywith thecarbondioxideinducedwarming,andFennoscandian temperature proxies may have recorded both types of potentially coinciding, interacting, or even diverging signals in the decadal-to-centennial scales of variability. Multidecadal variability in Fennoscandian summertime climate may well be related to the AMO (sea surface temperatures (SST)), which has a period of about 40–80 years, suggested to arise from predictable internal variability of the ocean-atmosphere system [7,8]. In addition Arctic air temperature and pressure have been shown to display strong multidecadal variability on similar time scales [9]. The goal here is to study periodicity and trends in growth variability of six recently published and updated millennia length proxies of Scots pine from the northern timberline. If Hindawi Publishing Corporation Journal of Climatology Volume 2014, Article ID 578761, 13 pages http://dx.doi.org/10.1155/2014/578761 2Journal of Climatology the six series show consistent and synchronous interannualto-decadal, decadal-to-multidecadal, as well as centennial and longer types of variability, three versions of each series will be built in order to highlight the three frequency bands separately in these data as well as in various combinations to regional averages (not mixing frequency classes). Growth signals are considered more or less frequency dependent if they all show similar and coherent behavior in the suggested frequency ranges and may thus be usefully combined to regional high-frequency (h-f), medium-frequency (m-f), and low-frequency (l-f) chronologies. Current knowledge is relatively limited regarding spectral details of the l-f and mf trends and periodicities in these proxies (except for some older versions of individual proxies for relatively narrow frequency bands [10,11]). All the six series are known predominantly as summer temperature proxies [1,4,5,12,13]. This study will provide a synopsis of their rather complex potential as predictors of high-summer (July), standard summer season (June– August), and even some longer warm period temperatures in the three broad frequency ranges. The results will serve as practical guidelines for selecting proper proxy types and choice of indexing methods with respect to required response period as well as frequency band. Such knowledge is crucial in various multiproxy applications (see, e.g., [4]) where the amplitude, duration, and timing of changes are important. In particular, our goal is to reconstruct temperature variability at the multidecadal scale, which would be useful in interregional comparisons of meaningful periodicity in land surface temperatures in the subarctic region as well as between land and sea surface temperatures of larger fields in the search for periodic patterns (both intrinsic and external to the climate system). In order to gain insight into this topic we will here compare the well-known cycles of annual AMO with m-f periodicity in these data. Since Gray et al. [14] successfully reconstructed the AMO using a large network of tree-growth chronologies (including one early version of our data series), it will be interesting to test an AMO model now with these sixlongerandupdatedseries. Incalibrationmodelperformancewillbeevaluatedin each case based on known stringent verification tests. The aim is to provide feasible models in three scales from h-f to l-f with increasing proportion of low frequency variance (where h-f necessarily overlaps in m-f and m-f overlaps in l-f). The linear relationship between tree growth at the three frequency ranges and temperature over the whole range of target variations (unfiltered) is analyzed. It would be reasonable to expect model fit generally to increase if meaningful lower frequencies are added to the pool of predictors (from h-f to l-f). However the situation is generally more complicated as targets are relatively short series as compared to the proxies, and the longest trends in tree growth can only haveapartialmatchintemperature.Inanabout100-year temperature record (usual in calibrations in this region), the detection of verifiable trends is restricted to a maximum of about 50 years (interpolation). However, if a good partial fit of a longer trend is found, some extrapolation is usually reasonable. The results also aim to contribute to the largely missing debate on the uncertainties and even disagreement of regional reconstructions or proxies as compared to the considerable attention over the discrepancies in hemispheric reconstructions [15]. 2. Materials and Methods The main interest here is in the inland region, which is occupied by the Fennoscandian Shield and forms a relatively homogeneous peneplane at the modest altitudes of about 200to500ma.s.l.withsharpclimaticandgeobotanical boundaries in the east and west [16–20]. Climate in this interior region is arguably more homogeneous and continental without the more marine coastal regions. General location for all six data sets is the northern timberline, between the Swedish Scandes and the Khibiny Low Mountains region. The data include the recently updated and bias-corrected ring width (SWR) and maximum density (SXD) from Sweden [5,21,22], ring width (FRW), height increment (FHI), and maximum density (FXD) from Finland [4,23], as well as ring width (RRW) from the Kola peninsula, Russia [4, 24]. The ring width (SRW and FRW) data sets are large, including data from 650 and 536 trees, respectively. The others are smaller, consisting of samples from 167 (FHI) to 78 trees (FXD). Height data (FHI) are shifted by one year for making comparisons possible, because height growth reflects conditions in the previous year [25,26]. Four long, monthly climate records represent the regional temperatures over their common period from 1908 to 2002 [17]. Tornedalen (Sweden) composite record is available as continuous between 1816 and 2002 [27], Karasjok (Norway) record since 1876, Karesuando (Sweden) record since 1890, and Sodankyl¨ a (Finland) record since 1908. Tornedalen and Karesuando are the closest to the western (Swedish RW and XD) sampling sites, Karasjok and Sodankyl¨ atothecentral (FinnishRW,HI,andXD)sites,andSodankyl ¨ aistheclosest to the eastern (Russian RW) site. Karasjok is the northernmost and Sodankyl¨ a and Tornedalen the southernmost climate records. Additional verification is obtained from the Bottenviken compilation series [28], which is based on the data from six stations (Abisko, Karesuando, Kvikkjokk, Jokkmokk, Haparanda, and Pite˚ a, mainly from somewhat more western and southern locations than the ones used in calibrations here) in northern Sweden. Moreover, annual AMO anomalies [29] are used in multidecadal comparisons. Regional curve standardization (RCS; see [5,30–32]), 180-year and 30-year splines (see [33–35]) were used in indexing. They are well-known methods in dendroclimatology and frequently applied in targeting various frequencydependent responses in proxy-based reconstructions [36, 37]. The RCS method is expected to preserve any l-f signal up to wavelengths exceeding the lengths of the individual segments used in building the chronologies. On the other hand, the 180-year splines will highlight m-f signal by removing the lowermost frequencies, the most multicenturytimescale variance potentially present in the data. The 30year spline extracts the h-f signal and will remove in addition muchofthemultidecadalandalllongerscalevariances. The three frequency ranges correspond to interannual-todecadal, interannual-to-multidecadal, and centennial types Journal of Climatology 3 of variability in the time domain. Furthermore, digital lowpass filters were used to extract the chronology variance with afrequencylowerthan10years[38–40]. We used Fourier analysis to define the spectral content in the chronologies, that is, a spectral description in terms of cycles of varying length, the actual frequencies that generate the original series. The wavelet approach has advantages over the more traditional methods for analyzing potentially nonstationary signals, which have discontinuities and nonperiodic characteristics [41–43]. In the complex interactions in climate there are a number of components which tend to damp out the more rapid fluctuations. Thus climate time series with lower frequency/longer period cycles have to contain a greater proportion of the observed variance to achieve the same significance as higher frequency/shorter period features [44]. In the time domain confidence intervals (c.i.) were calculated separately for the l-f and m-f ranges of growth variability using nonparametric bootstrap method (sampling with replacement) [45,46]. Arithmetic averages (AA)andweightedaverages(WA)(see[47] and references therein) were applied in combining regional chronologies. Simple linear regression models (transfer functions) [12, 48–50]weredevelopedforeachofthethreeversionsofthe sixchronologiesandtheiraveragestobeusedinturnas climate predictors. Individual models were then tested in split period calibration verification, where the total calibration period (1908–2002) was divided into two equal 48-year halves: 1908–1955 and 1955–2002. These subperiods, used for calibration during one period and verification during the other, are referred to as early calibration-late verification and late calibration-early verification (EC-LV and LC-EV, resp.). Both subperiods should produce positive values of reduction of error (RE) [12,38,51] and coefficient of efficiency (CE) [12,51] statistics for the model in order to pass the verification tests. RE and CE values may vary from +1 to −∞,with0 indicating that the reconstruction model performs no better as a predictor than the calibration (RE) or verification (CE) period mean value. The models are also compared using coefficient of determination (𝑅2) and explained variance (𝑟2). 3. Results 3.1. Periodicity, Trends, and Interannual Shifts in Growth. Five of the six l-f series (RCS, no filtering) share significant oscillatorymodeat200–250yearsofcyclelength(Figure 1) and SXD shows even longer-term features (Figure 1(b)). Most series also have strong multidecadal scale variability, viz. significant peaks at 67 years in SRW and SXD, as well as at 83–125yearsinFRW,FXD,andRRW.Althoughallseries have cumulated some concentration of variance at m-f scale, the two western series (SRW and SXD) have 67-year peaks and both of the XD series have a (possibly related) peak at 33 years. In addition the two XD series differ evidently from the others with still distinctly higher frequency content at 0.1–0.3 cpa (Figure 1).Basedontheseresultsthecommon spectral content of these data is concentrated principally on the relatively narrow frequency ranges from 0.005 cpa to 0.5 cpa; that is, periodicities are from 200 to 2 years. Next, the six l-f time series were low-pass filtered, normalized, and then averaged (c.i. around the mean in Figure 2(a), in 1006–1996, since six years are lost at both ends due to filtering). Correlations between these l-f series are all positive, ranging from 0.81 (FRW-FHI) to 0.21 (FHI-RRW) with a mean value of 0.52. Correlation declines in the three RW series with increasing distance; the 𝑟-value (from west to east) is 0.55 between SRW and FRW, 0.34 between SRW and RRW, and 0.49 between FRW and RRW. The two density series (SXD and FXD) have high linear association (𝑟 = 0.66). However their correlation is lower than that between either the two western or central RW and XD series (SRW-SXD, 𝑟 = 0.73 and FRW-FXD, 𝑟 = 0.77). The mean regional l-f chronology (Figure 2(a))shows an overall increasing trend (regression of growth on time in 1006–1996; 𝑦 = 0.0006𝑥), reflecting the diverse longterm rates of change in the original series. Producing a distribution of the mean values and then locating the lower and upper bounds of this distribution, five significant (95% c.i.) periods of positive growth in the average series are discerned: 1083–1104, 1159–1178, 1428–1436, 1750–1769, and the longest from 1918 to 2002, separated by periods of poor growth of varying lengths and magnitudes between them. These periods are determined in the regional growth signal within the uncertainty limited by (sources of error) interregional differences (e.g., western, central, and eastern provenance) as well as differences due to proxy type (RW, XD. and HI). The highest values in the second millennium are clearly recorded in the last century (Figure 2(a)). The bootstrapped confidence intervals are nonsymmetric and particularly wide in the first and last centuries. The six m-f (filtered) series are more synchronous than the l-f series as their average has narrower c.i. and many more significant periods (Figures 2(a) and 2(b)) in addition to having a higher mean correlation (𝑟 = 0.58). Correlation isthehighestbetweenFRWandFXD(𝑟 = 0.74)aswellas between FRW and FHI (𝑟 = 0.74) and the lowest between SXD and FHI (𝑟 = 0.39) as well as between SXD and RRW (𝑟 = 0.39). Twelve distinct and significant (95% c.i.) periods of above average growth are dated (Figure 1(b)): 1083–1101, 1157–1183, 1283–1289, 1411–1447, 1489–1498, 1535–1547, 1559– 1572, 1625–1634, 1653–1665, 1750–1767, 1850–1862, and the longest in 1920–1956. The confidence limits indicate wider spread in the original six series in the 13th and 14th centuries than in the rest of the m-f series. The average of the six h-f (no filtering) series represents the h-f variability in this work (Figure 2(c)). Average correlation between all the six series is 0.41. It is the highest between SRW and FRW (𝑟 = 0.68) and the lowest between SXD and FHI (𝑟 = 0.22). In the mean series growth was the lowest in 1601 (−2.6 s.d. units below the mean) and the highestin1826(2.4s.d.unitsabovethemean).Thegreatest biennial shifts (difference between any two consecutive years) occurred between 1600 and 1601 (3.8 s.d. from 1.2 to −2.6) as well as between 1640 and 1641 (3.8 s.d. from 1.8 to −2.0). In the 20th century, 1903 was unusually low and 1937 was high in the growth index. It is worth noticing that even this h-f series has evident decadal fluctuations. 4Journal of Climatology 60 40 20 0 0 250 111 67 0.95 c.l. 0.1 0.2 Frequency (yr−1) Spectral power density (𝜔) (a) 20 10 111 67 0.95 c.l. 500 0.1 0.2 Frequency (yr−1) Spectral power density (𝜔) 0 0 (b) 40 20 0.95 c.l. 200–250 83 0.1 0.2 Spectral power density (𝜔) Frequency (yr−1) 0 0 (c) 120 80 40 0.95 c.l. 200 0.1 0.2 Frequency (yr−1) Spectral power density (𝜔) 0 0 (d) 30 20 10 0.95 c.l. 200 83–125 33 0.1 0.2 Frequency (yr−1) Spectral power density (𝜔) 0 0 (e) 30 15 45 0.95 c.l. 200 100 0.1 0.2 Frequency (yr−1) Spectral power density (𝜔) 0 0 (f) Figure 1: The Fourier spectra of the six growth-based chronologies (RCS indexing) of Scots pine. Smooth lines are 0.95 confidence levels, calculated for red noise with AR(1) coefficients 𝛼:(a)SRW(𝛼 = 0.7), (b) SXD (𝛼 = 0.20), (c) FRW (𝛼 = 0.77), (d) FHI (𝛼 = 0.75), (e) FXD (𝛼 = 0.43), and (f) RRW (𝛼 = 0.69). Cpa is cycles per annum. 1000 1200 1400 1600 1800 2000 0 1 2 3 −2 −1 (a) 1000 1200 1400 1600 1800 2000 0 1 2 3 −2 −1 (b) 1000 1200 1400 1600 1800 2000 −4 −2 0 2 4 (c) Figure 2: C.i. for the mean l-f series (RCS, low-pass filtering; (a)), for the m-f series (low-pass filtering of the six 180-year spline indexed series; (b)) and the average of h-f series (30-year splines, without low-pass filtering; (c)). Journal of Climatology 5 −2 0 Tornedalen 2 1910 1930 1950 1970 1990 (a) −2 0 2 Karasjok 1910 1930 1950 1970 1990 (b) −2 0 2 Karesuando 1910 1930 1950 1970 1990 (c) −2 0 2 Sodankylä 1910 1930 1950 1970 1990 (d) −2 0 2 Average 1910 1930 1950 1970 1990 (e) Figure 3: Mean June–August temperatures of the four records and their average (z-scores in 1908–2002). 3.2. Individual Proxies versus July and June–August Temperatures. Our climatic targets are arithmetic averages of normalized monthly data. July temperatures are first modeled using individual proxies and then June–August temperatures using individuals as well as their various combinations. The strength of targeted regional signal is indicated by correlation of June–August mean temperatures among the four stations (Figure 3), which vary between 0.88 (TornedalenKarasjok) and 0.94 (Karesuando-Sodankyl¨ a), while the mean correlationis0.92.ThisJune–Augustmeanhasevenhigher correlation (𝑟 = 0.96,𝑟2= 0.93)withtheJune–Augustmean of the Bottenviken regional temperature record exclusively from northern Sweden [28], used in calibrations in some previous studies [5,21]. All four records show, for example, cool conditions in the early 20th century until around 1915, the relative warmth of the 1930s, and a recent warming since the late 1980s. Summer temperature (June–August, normalized scale) varies between −1.32and1.88,whichhereformthe limits for interpolation (Figure 3). The mean temperature has a modest positive trend (𝑦 = −0.07+0.0014𝑥)in1908–2002. When the six l-f proxies were individually regressed on mean July temperature (Table 1(A)), only FXD and RRW produced positive verification statistics. Although SXD has positive RE values in both periods, the CEs are marginally negative. Among the m-f series (Table 1(B)), both of the western series (SRW and SXD) clearly pass the verification tests, while FRW and FHI do not. FXD and RRW series have improved (higher than in l-f) positive RE and CE values. In the higher frequencies (Table 1(C)), in addition to all three RW series, also FHI passes both tests in both periods. However,bothSXDandFXDfalljustbelowzero(−0.02) using the more searching CE statistic in the LC-EV period. The same proxies were then calibrated against mean June–August temperature. In the l-f range (Table 2(A)), SXD, FXD,andRRWpassthetests.NowSXDproducesthehighest RE and CE values in both periods. In the medium frequencies (as previously in the case of July, Table 1(B)), the four series (SRW, SXD, FXD, and RRW) show positive verification performance. FRW and FHI again have negative RE and CE values in the LC-EV period. In the higher frequencies (Table 2(C)), all six series pass the tests. The three RW series have generally lower explained variance (𝑟2≤0.21)aswellas lowerREandCEvalues(≥0.18) than the two other types of proxies (XD and HI), each of which has 𝑟2≥ 0.31 and both RE and CE ≥0.28. 3.3. June–August Temperature Signal in l-f, m-f, and h-f Averages. All six series (their l-f, m-f, and h-f variants) wereaveragedtoasimplemean(AA)andtoaweighted mean (WA) and tested as combined predictors of June– August temperature (Table 3(A)–(C)). Simple mean (AA) produced positive verification results (both RE and CE during both periods) only in the higher frequencies (Table 3(B), Figure 2(c)). The failure of l-f and m-f models here as well as in previous comparisons (Tables 1and 2)isduetoexcess growth variability as compared to temperature in the LCEV period, which is most pronounced in FRW and FHI. However,thetestvaluesforAAareonlyslightlyweaker than for WA. WA also had positive results in the lowermost frequencies (Table 3(A)). In the medium frequencies, neither method (AA or WA) resulted in an acceptable model (Table 3(B)). In the next step only those series which individually passed RE and CE tests (in both transfer models for June– August temperatures, Table 2) were weighted and combined, namely, the three l-f series (3RCS; 0.56 ∗SXD + 0.44 ∗FXD +0.14∗RRW) and four m-f series (4Spline; 0.22 ∗SRW + 0.54 ∗SXD + 0.41 ∗FXD + 0.16 ∗RRW). When these were calibrated against June–August temperature, both of them passed the verification tests (Table 3(C)). It should be noted 6Journal of Climatology Table 1: Six proxies versus mean July temperatures (average of the four records, no filtering) in 1908–2002. Reduction of error (RE), coefficient of efficiency (CE), and 𝑟2during EC-LV (early calibration in 1908–1955 and late verification in 1955–2002) and LC-EV (late calibration in 1955–2002 and early verification in 1908–1955). Only test results with RE and CE >0 are included. EC-LV LC-EV RE CE 𝑟2RE CE 𝑟2 (A) RCS indexing FXD 0.29 0.21 0.35 0.10 0.02 0.14 RRW 0.25 0.17 0.19 0.21 0.14 0.16 (B) 180-year spline indexing SRW 0.30 0.22 0.29 0.24 0.18 0.32 SXD 0.26 0.18 0.25 0.15 0.08 0.12 FXD 0.35 0.28 0.36 0.14 0.07 0.15 RRW 0.27 0.19 0.21 0.23 0.17 0.18 (C) 30-year spline indexing SRW 0.27 0.19 0.30 0.4 0.34 0.43 FRW 0.29 0.21 0.39 0.13 0.05 0.24 FHI 0.52 0.47 0.57 0.23 0.17 0.23 RRW 0.18 0.09 0.25 0.27 0.20 0.33 Table 2: Six proxies versus mean June–August temperatures in 1908–2002 (no filtering). RE, CE, and 𝑟2during EC-LV (1908–1955 and 1955– 2002) and LC-EV (1955–2002 and 1908–1955). Only test results with RE and CE >0 are included. EC-LV LC-EV RE CE 𝑟2RE CE 𝑟2 (A) RCS indexing SXD 0.53 0.53 0.56 0.53 0.53 0.56 FXD 0.43 0.42 0.44 0.42 0.42 0.44 RRW 0.1 0.1 0.11 0.15 0.15 0.17 (B) 180-year spline indexing SRW 0.17 0.17 0.21 0.07 0.07 0.28 SXD 0.47 0.47 0.54 0.45 0.45 0.57 FXD 0.35 0.35 0.44 0.27 0.26 0.43 RRW 0.12 0.12 0.13 0.15 0.15 0.18 (C) 30-year spline indexing SRW 0.18 0.17 0.21 0.14 0.13 0.20 SXD 0.49 0.49 0.50 0.41 0.41 0.42 FRW 0.14 0.14 0.16 0.11 0.10 0.15 FHI 0.46 0.46 0.48 0.28 0.28 0.31 FXD 0.39 0.39 0.40 0.37 0.37 0.38 RRW 0.16 0.16 0.17 0.14 0.14 0.15 that since sample replication is at least five in 1000–2002 in each of the six series it is thus >15 in 3RCS model, >20 in 4Spline model, >10 in the XD models, and >30 in models where all six series were included. Because of the persistence in the time series of tree growth theyareusuallyfilteredtoenhancethesignaltobeanalyzed. It is often recommendable to also reconstruct l-f and h-f variations separately smoothing the proxy and instrumental series prior to calibration [3,52]. In order to further study the correspondence between these two mean series (3RCS and 4Spline) and the target above decadal scales, the l-f and mf proxies as well as the June–August temperature were lowpass filtered (10-year smoothing; see [38–40]) and then tested again (using correspondingly shorter 42-year calibration andverificationperiodsin1914–1996duetofiltering).The verifications of these models produced positive results for both;however,herethel-fmodelwassuperiortothem-f model. The four successful multiproxy WA model combinations for June–August temperature (Table 3)wererecalibrated using the full 95-year calibration period (Figure 4). The mean of all six RCS-indexed series (with the highest weight onSXDandthelowestonFRWandRRW)produceda slightly inferior model as compared to the 3RCS series (𝑅2= 0.57–0.59,3RCSinFigure 4(c)). In the m-f scale the only successful combination (4Spline, Figure 4(b))has Journal of Climatology 7 Table 3: Simple mean (AA) and weighted mean (WA) of all six series versus mean June–August temperatures (no filtering). The WA of RCSbased (A) and 30-year spline indexed series (B) (see also Figure 4(a)). Only test results with RE and CE >0 are included. The WAs of those three RCS-indexed series (3RCS; 0.56 ∗SXD + 0.44 ∗FXD + 0.14 ∗RRW) and four 180-year spline indexed series (4Spline; 0.22 ∗SRW + 0.54 ∗SXD + 0.41 ∗FXD + 0.16 ∗RRW) which individually passed RE and CE tests (see Table 2) versus mean June–August temperatures (C; the full models shown in Figures 4(b) and 4(c)). EC-LV LC-EV RE CE 𝑟2RE CE 𝑟2 (A) RCS WA 0.55 0.55 0.67 0.21 0.21 0.53 (B) 30-year spline indexed AA 0.56 0.56 0.57 0.50 0.50 0.51 WA 0.65 0.65 0.65 0.58 0.57 0.58 (C) WAs of selected three RCS and four 180-year spline indexed series 3RCS 0.58 0.58 0.61 0.53 0.52 0.58 4Spline 0.46 0.46 0.60 0.28 0.28 0.60 intermediate values between l-f and h-f scales. The WA of all six h-f series (30-year spline indexed, Figure 4(a))shows the best fit with the target (𝑅2= 0.62). Here the modeled andobservedvalueshaveparticularlysynchronous(one-or two-year) peaks in, for example, 1923, 1928–1929, 1949, 1962, and 1979–1980 (Figure 4(c)). The l-f model reproduces, for example, the longer increasing trend from 1908 to 1937 as well as that of the last fifteen years more consistently than the h-f series (Figures 4(a)–4(c)). However, the reproduction of the twin peaks in 1969–1974 is noticeably poorer. The two density series (SXD and FXD) were further combined (WA) in both the l-f and m-f ranges and calibrated against longer warm season periods (Table 4). Three different periods were used, four months mean temperature from MaytoAugust(A),fivemonthsmeanfromApriltoAugust (B), and six months mean from April to September (C). All these six proxy combinations (both l-f and m-f) pass the verification tests (RE and CE in both periods: EC-LV andLC-EV)forallthreetemperatureperiods.Modelfit (𝑅2) recalibrated using all available data varies from 0.48 to 0.58.TheweightsinWAareratherequalforSXDandFXD, ranging from 0.51 to 0.34, but slightly favoring SXD. The best empiricalmodels(forMay–Augusttemperatures,Table 4)for the l-f and m-f bandwidths of the two XD-series were derived by recalibration in 1908–2002 with 𝑅2= 0.58 forthel-fand 𝑅2= 0.53 for the m-f model. 3.4. Building a Proxy for Subcentury Scale Temperature Variability. Acompositem-fsummertemperatureproxywas produced using the 4Spline variant (WA of SRW, SXD, FXD, and RRW) (Figure 5(a), calibration in Table 3(C) and Figure 4(b)). Several significant (above 0.99 c.l.) multidecadal characteristics from periods of 33 and 67 years up to a peak exceeding a century (111 years) are highly significant in the Fourier spectrum (Figure 5(c)).Thewaveletspectrumshows that the fluctuations spreading over these bandwidths are particularly apparent in the 12th and 17th centuries. On the other hand there is a noticeable lack of significant features residing in the decadal to bidecadal ranges (Figures 5(b) and 5(c)). In this multidecadal scale the 20th century does not stand out as unusual in the past millennium. In order to assess the correspondence of amplitude, duration, and timing of recent cycles between Fennoscandia and the North Atlantic the m-f proxy (4Spline based) and AMO were also visually compared (Figure 6). Correlation between the two series (in 1866–1992) is 0.27, rising to 0.49 using 10year moving averages (MA, Figure 6(a)) and to 0.56 using 20year MA. Presuming Fennoscandian proxy variability is preceding the AMO cycle (evident in Figure 6(a))andcorrecting for (removing) this 10-year time shift the correlation rises to 0.34, 0.73, and 0.82, respectively. The 20-year MA effectively removes all higher frequencies and the relationship seems in agreement particularly during early to mid-20th century. This illustrates that multidecadal variability in Fennoscandian and North Atlantic climates is closely similar in scale and length. To assess the potential of these data in modelling the AMO (e.g., time stability of such statistical relationship), all sixm-fproxieswerealsocomparedtoannualSSTanomalies. Screening the pool of candidate predictors it was also possible to build a transfer model for the AMO. The WA of three series (SRW, FXD, and RRW) was used with the same calibration (1922–1990) as well as verification (1856–1921) periods as in Gray et al. [14]. This model passed the verification trials (RE = 0.05 and CE = 0.05), but model performance is the lowest (𝑅2= 0.17)inthiswork. 4. Discussion Five of the six series from different parts of northern Fennoscandia show evidence of a dominant 200–250 years of periodicity. Similar periodicity was found significant in a harmonic decomposition of the average of six central European instrumental as well as stalagmite proxy series from theAustrianAlpsin500–1935[53]. However, the pronounced minimum during recent centuries appears in our proxies somewhat later (in the early 20th century) as compared to the 1880s noted by L¨ udecke et al. [53]. Previously Helama et 8Journal of Climatology 1910 1920 1930 1940 1950 1960 1970 1980 1990 2000 0 1 2 −1 (a) High-frequency model; all six 𝑅2=0.62 1910 1920 1930 1940 1950 1960 1970 1980 1990 2000 0 1 2 −1 (b) Medium-frequency model; 4Spline 𝑅2=0.56 1910 1920 1930 1940 1950 1960 1970 1980 1990 2000 0 1 2 −1 (c) Low-frequency model; 3RCS 𝑅2=0.59(all six model 𝑅2= 0.57) Figure 4: Final linear regression models on June–August mean as target (dotted line) with various combinations of proxies recalibrated in 1908–2002 (solid line): (a) WA of all six h-f (30-year spline indexed) series, (b) WA of SRW, SXD, FXD, and RRW (four 180-year spline indexed, 4Spline) m-f series, and (c) WA of the SXD, FXD, and RRW (3RCS series). Table 4: Weighted averages of the Swedish and Finnish density series (SXD and FXD) built using two types of indexing (RCS and 180-year splines, Sp) versus a four-month mean temperature (no filtering) from May to August (A), a five-month mean from April to August (B), and a six-month mean from April to September (C). Only test results with RE and CE >0 are included. EC-LV LC-EV RE CE 𝑟2RE CE 𝑟2 RCS A 0.55 0.54 0.56 0.55 0.55 0.57 Sp A 0.33 0.33 0.55 0.29 0.28 0.58 RCS B 0.49 0.48 0.50 0.49 0.48 0.51 Sp B 0.26 0.25 0.47 0.25 0.25 0.50 RCS C 0.44 0.43 0.45 0.53 0.52 0.53 Sp C 0.24 0.24 0.41 0.32 0.32 0.52 Weights used in averaging: 0.51 ∗SXD + 0.47 ∗FXD (RCS A); 0.47 ∗SXD + 0.42 ∗FXD (SP A); 0.47 ∗SXD + 0.41 ∗FXD (RCS B); 0.42 ∗SXD + 0.35 ∗ FXD (Sp B); 0.46 ∗SXD + 0.39 ∗FXD (RCS C); 0.41 ∗SXD + 0.34 ∗FXD (Sp C). al. [54] discussed centennial and multidecadal temperature variability over Northern Fennoscandia that bears a potential link to oceanic origins [8,55] and reviewed the paleoclimatic literature relevant to the topic. Several individual Fennoscandian tree growth-based proxies have indicated prominent spectral features at about 23, 30, and 90 years, 23–33 years, and 30.8–31.8 and 80.3–87.7 years in Finland [11,25,56] as well as relatively time-stable peaks at 32–33 years and at ∼55–100 years in Sweden [10]. In an analysis of seven northern hemisphere temperature reconstructions (including, e.g., [15,57–59]) Ogurtsov et al. [60] reported that they have an unambiguous 60–80-year multidecadal variability in common (AD 1000–1930), which is close to the range of a 67-year cycle indicated by our data. The 4Spline reconstruction without the more or less discrepant secular characteristics (see [17]) clearly records a bimodal structure—c.a. 110-year and 60–70-year variations— as well as a 33-year climatic cycle (similar to the Bruckner cycle). Based on these evidence there exists obvious potential for analyses in subcentury scales and a real possibility to gain insight into the extent of general natural variability in the subarctic region. In the lower frequencies of these data (low-pass filtering of the RCS and 180-year spline indexed series) several growth surges and troughs coincide in each group and they were also dated in the averages in the time domain (statistically significantfluctuations;fiveinl-fand12inm-frange).The latest, 37-year period (1920–1956) is the longest continuous multidecadal scale surge and the 85-year period (1918–2002) is the longest l-f surge. The combined m-f chronology is evidently more consistent than the l-f chronology as the more diverging l-f trends are left out. Overall signal strength (measured as mean correlation) between the six series is higher in m-f (𝑟 = 0.58)thaneitherinl-f(𝑟 = 0.52) or in h-f series (𝑟 = 0.41).Atleastpartofthel-fdisagreementispossiblydue to the (noisy) RCS method, which is usually recommended forlargedatasetsofvariousageclassesoftreesforeach year, ideally grown under a range of representative ecological conditions [10,21,61–64]. Even the ample data bases of SRW and FRW may not fully meet these requirements. Journal of Climatology 9 1000 1200 1400 1600 1800 2000 Years Index 4 3 2 1 0 −1 −2 −3 (a) 1000 1200 1400 1600 1800 2000 296,3 195,5 129 85,1 56,1 37 24,4 16,1 10,6 7 4,6 Time scale (year) Years 6,500 6,000 5,000 5,500 4,500 4,000 3,500 3,000 2,500 2,000 1,500 1,000 0,500 0,000 (b) 20 15 10 5 0 0 0,05 0,10 0,15 0,20 Frequency 𝜔(yr−1) 67 yr 33 yr 111 yr 0.999 c.l. 0.99 c.l. Spectral power density (c) Figure 5: The m-f proxy of summer temperatures since AD 1000 ((a), unfiltered). Wavelet spectrum (b) and Fourier spectrum (c) of this time series. 1860 1880 1900 1920 1940 1960 1980 2000 −1.5 −0.5 0.5 10-year running means (a) 1860 1880 1900 1920 1940 1960 1980 2000 −1.5 −0.5 0.5 20-year running means (b) Figure 6: Comparison of annual AMO cycle (dash) and the Fennoscandian multidecadal (m-f) summer temperature proxy (solid) during common period using 10-year (a) and 20-year (b) smoothing. The four instrumental temperature records used in calibrations show remarkable agreement justifying their averaging to a regional mean. July is usually one of the most important factors among monthly temperatures related to various growth parameters from different parts of the region [12,21,23,24,54]. When the six proxies were individually regressedonmeanJulytemperature,onlyFXDandRRW passed the verification trials in the l-f range, SRW and SXD in the m-f range, and all but the two XD series in the h-f range. The growth response of June–August temperatures has oftenbeenfoundtimestableandthisperiodhasrecentlybeen used in temperature reconstructions in the region [2–5,17]. A simple mean of all six proxy series (AA) works as a verifiable predictor of mean June–August temperature only in the hf range (30-year spline indexing, Figure 2(c))inthesedata. Althoughallseriespassthetests,theSXD,FHI,andFXDhave superior individual calibration and verification performance in this context. This is consistent with the results obtained