1 Introduction

Tree rings, due to their annual resolution and relatively simple computation of climate–growth models, are valuable and widely used tools for the reconstruction of climate variability in pre-instrumental periods (Hughes 2011). Tree rings may complement the information from instrumental records during the early instrumental period, where the available meteorological records are typically scarce and where the uncertainty of the instrumental data is often very large (Böhm et al. 2010).

The best tree species to use are those showing the highest sensitivity to climate variability. These species include three of the most widespread conifers growing at the high altitudes of the treeline in the European Alps: European larch (Larix decidua Mill.), Norway spruce (Picea abies Karst.) and Swiss stone pine (Pinus cembra L.) (Beniston et al. 1997; Körner 1998; Grace et al. 2002; Pauli et al. 2003), species that have been widely used for dendroclimatic reconstructions in the region. In the European Alps, climate parameters, primarily the summer temperature, have been reconstructed both at local and regional scales using trees from high-altitude sites (e.g., Frank et al. 2005; Büntgen et al. 2005, 2006a, 2011; Corona et al. 2010; Coppola et al. 2013). These reconstructions also benefit from a unique availability of instrumental data, as the Alps and the neighboring areas are among those with the most available long-term meteorological records, with many series starting in the 18th century (Jones and Bradley 1992). The summer temperature is usually reconstructed based on strong relationships between the climate and tree-ring widths; however, on a monthly timescale, the periods that most influence tree-ring growth over the growing season (and therefore most influence the recorded climatic signal in the tree rings) are different for the three species: larch growth is mainly driven by the early summer (June) temperature (e.g., Carrer and Urbinati 2006; Coppola et al. 2012), while the mid-summer (July) temperature affects the Norway spruce and the Stone pine growths (Carrer et al. 2007; Oberhuber 2004; Leonelli et al. 2009).

Tree-ring based reconstructions are also complicated by the temporally changing patterns in the radial growth response to climate that have been outlined in several dendroclimatological studies conducted in high-altitude Alpine forests (e.g., Carrer and Urbinati 2004, 2006; Büntgen et al. 2006b; Leonelli et al. 2009; Coppola et al. 2012). The lack of sensitivity to temperature over certain periods during the growing season can potentially bias climate parameter reconstruction based on tree-ring widths and, to some extent, can be related to the so called ‘divergence problem’: the reduced tree-ring growth response occurring at some temperature-limited sites during the recent period of globally increasing temperature (e.g., Briffa et al. 1998a; D’Arrigo et al. 2008; Loehle 2009).

For climate reconstructions, it is crucial to deeply understand how the species composition in a given dataset may influence the temperature reconstructions and to what extent the incidence of trees highly sensitive to temperature may influence the quality of these reconstructions. However, these factors are not always considered.

The main objective of this paper is to show that the multi-species reconstruction of summer temperatures (JJA, June to August) in a selected region of the European Alps can be improved by considering trees that are highly sensitive to temperature (HSTT). The present study was based on 42 tree-ring chronologies of L.decidua, P.abies and P.cembra from high-altitude forest sites in a region of the central European Alps. This region includes five highly glacierized mountain groups: the Silvretta Group (Switzerland), the Ötztaler-Venoste Alps (Austria, Italy), the Bernina Group (Switzerland, Italy), the Ortles-Cevedale Group (Italy) and the Adamello-Presanella Group (Italy).

2 Methods

2.1 Chronology constructions

Ring-width chronologies from high-altitude sites (>1800 m a.s.l.) were collected from a region of the European Alps centered over the 1° × 1° grid cell 46°N 10°E from the ITRDB (http://www.ncdc.noaa.gov/data-access/paleoclimatology-data/datasets/tree-ring; September 2014) and from internal databases available to University of Milan and University of Pisa (Online Resource 1 for all references and site codification). To produce a larger dataset, the area boundaries were enlarged 15′ E- and W-ward, finally including forest sites from Italy (23 sites), Switzerland (15) and Austria (4) (Online Resource 2). Only three conifer species for which a dendroclimatic potential is well recognized were considered: European larch (Larix decidua Mill.), found at 11 sites; Norway spruce (Picea abies Karst.), 5 sites; and Swiss stone pine (Pinus cembra L.), 26 sites.

All tree-ring series were checked for measurement quality and correct cross-dating by visual assessment and statistically using the COFECHA software (http://www.ldeo.columbia.edu/tree-ring-laboratory/resources/software); series were definitively discarded from each site if they showed anomalous growth patterns and Pearson’s correlation coefficient values r < 0.30 vs. the respective mean site chronology calculated after removing the raw series itself. This selection ensured that the common signal recorded by trees at each site would be preserved and has also ensured coherent growth patterns in each tree-ring series considered for further analyses.

2.1.1 Constructing the site chronologies and regional chronologies of the species

At each site we built a chronology by applying the Regional Curve Standardization – RCS approach (Briffa et al. 1992; Briffa and Melvin 2011; Esper et al. 2003) – using the ARSTAN software (ver. 44 h3, http://www.ldeo.columbia.edu/tree-ring-laboratory/resources/software). Because the pith offset information was not available for most sites (which does not substantially influence the trends in the final tree-ring width chronologies; Esper et al. 2009), the series alignment by cambial age for constructing the RC was performed by considering the oldest measurement available in each raw series as the first year. After alignment, the resulting RC was smoothed with a cubic spline 10 % of the series length (Büntgen et al. 2006a). Detrended indexed series were obtained from each raw series by computing the ratios of measurements for each year, and a biweight robust mean was applied to all the indexed series to obtain the site chronologies. The obtained site chronologies were then truncated, discarding the early periods showing an Expressed Population Signal (Wigley et al. 1984) EPS < 0.85 or <0.8 (if the chronology did not show an EPS > 0.85). Chronologies with a mean EPS < 0.75 over the common interval 1870–1970 (where most of the considered chronologies are present) were discarded, except for P. abies due to their scarcity. Then, the selected site chronologies were averaged by species if presenting an r > 0.30 (or >0.29 for P.abies) with the JJA temperature series; these regional chronologies were named LADE, PCAB and PICE for L. decidua, P. abies and P. cembra, respectively. These regional chronologies were considered over the periods having at least 3 site chronologies. The PCAB chronology was only considered over the period having at least 2 site chronologies, not considering the non-overlapping, old series from site 14.

Two other regional chronologies were constructed: avgALL, by averaging the three previously constructed regional chronologies of all species, and avgLADE-PICE, by averaging only the regional chronologies from L. decidua and P. cembra.

2.1.2 Constructing the HSTT chronology

Finally, another chronology was constructed starting from the entire available dataset, selecting only those series >100 yr. and showing highly significant Glk index values (Gleichläufigkeit, year-to-year agreement between interval trends; Schweingruber 1988) vs. the respective mean site chronology. An RCS standardization approach with same settings described above was applied at each site to obtain the indexed individual series that were used for the further analysis (details of the applied method in Online Resource 3). The chronology was named HSTT (highly sensitive to temperature) because, for its construction, we used only those series with r > 0.30 vs. the JJA temperature series over the period covered by instrumental data (1763–2008) and vs. the growing chronology from the previous periods. When more series from the same trees were available (which was rare in the ITRDB dataset), only the series with the highest correlation was preserved. The HSTT was then constructed by averaging the selected indexed series. The purpose of this chronology was to show differences in growth patterns in the most temperature-sensitive trees with respect to the other trees in the area and for evaluating differences in the climate reconstructions performed using more canonical approaches.

An evaluation of the site fitness (SF) in containing HSTT trees was performed by calculating the percentage of summer-temperature sensitive series with respect to all series available at each site.

2.2 Climate data

Temperature records were extracted from an updated and improved version of the dataset presented in Brunetti et al. (2006). The dataset consists of monthly anomalies, with respect to averages over the 1971–2000 period, interpolated on a regular grid at a 1-degree spatial resolution. The initial set of station data used for the interpolation consists of quality checked and homogenized series (i.e., corrected for non-climatic signals due to all changes that affected stations, instruments, and measurement rules). Improvements over the Brunetti et al. (2006) version consist of increased reliability of the early instrumental period (before 1865) and the inclusion, in addition to the secular records presented in that paper, of the records from the Italian Air Force network (Simolo et al. 2010). Specifically, we used the monthly series of June, July and August and the summer season series JJA. The series from the area of interest for the present work are available for the period 1763–2014.

2.3 Climate reconstruction

Different climate reconstructions were performed based on the avgALL, avgLADE-PICE and HSTT chronologies as predictors and the JJA temperature series as the predictand using both simple regression models and scaling (Esper et al. 2005). The models were calibrated and evaluated for stability over half of the available period covered with both dendrochronological and climatic data and were verified over the other half; this procedure was then repeated inverting the two periods (defined as 1763–1885 and 1886–2008). The quality of the reconstructions was assessed using various statistics widely used in dendroclimatology for evaluating the relationships between estimated and actual temperature values: explained variance (R2) for the calibration; product mean (PM; Fritts 1976) reduction of error (RE; Fritts 1976) and coefficient of efficiency (CE; Briffa et al. 1988) for the verification.

3 Results

The constructed site chronologies having an EPS > 0.8 over the interval 1870–1970 (Table 1) show different gradients of EPS, mean sensitivity (i.e., the relative change in ring widths between two consecutive tree rings) and first order autocorrelation over elevation (Fig. 1). L.decidua and P.cembra in particular show similar mean EPS values, with slightly higher values at higher altitudes (Fig. 1a). The two species are more separated with respect to the mean sensitivity values (also presenting a positive gradient towards higher altitudes; both resulting R2 were statistically significant at the p < 0.05 level), with L.decidua always showing higher values than P. cembra (Fig. 1b). First order autocorrelation shows a slightly negative gradient towards higher altitudes especially for L.decidua, whereas for P.cembra, higher values are recorded on average and a much less marked negative gradient is found (Fig. 1c). These two species are also separated by mean values of intertree correlation (when at least two ring-width series per tree were available), with L.decidua showing higher values than P.cembra (Table 1; not shown).

Table 1 Main statistics of all site chronologies with EPS > 0.8; these chronologies were considered in the subsequent analysis. Only for PCAB, all chronologies were considered even given a low EPS. C.I. is the common interval considered for certain statistics (1870–1970). All sites are ordered according to species name and decreasing elevation; the site numbers correspond to the codification given in Online Resource 1, where site information and bibliographic references are reported
Full size table
Fig. 1
figure 1

Linear regression of EPS (a), mean sensitivity (b) and autocorrelation (c) with elevation. In all cases, the regression equations refer to L. decidua or P. cembra chronologies showing an EPS > 0.8; for P. abies, only the chronologies with the two highest EPS value are shown. Triangles = L. decidua sites; squares = P. abies sites; circles = P. cembra sites; larger triangles and larger circles refer to the chronologies that also showed r > 0.3 vs. the JJA temperature series over the period 1763–2008, which were used for constructing the regional chronologies of the species (see methods).

Full size image

The constructed regional chronologies of LADE (1675–2008), PCAB (1773–2005) and PICE (1650–2004) derived from site chronologies of the respective species show similar growth patterns over the common period 1773–2004 (Fig. 2a), even if a higher interannual variability is evident for L.decidua than for P.cembra and P.abies; also, in the recent period, markedly higher tree-ring growth rates are evident for L.decidua than the other two species.

Fig. 2
figure 2

a Standard chronologies constructed for the grid cell considered and the respective number of contributing site chronologies over time: LADE and PICE were derived from site chronologies truncated at EPS >0.85, and only those with r > 0.3 vs. JJA temperature were considered. For PCAB, all chronologies with r > 0.29 vs. JJA temperature were considered. The HSTT chronology is presented only since 1560 to compare it with the other chronologies; for HSTT, the number of contributing indexed series is also reported. b Correlation coefficient and Glk calculated over the common period 1773–2004 for LADE, PCAB, PICE, avgALL, avgLADE-PICE and HSTT vs. JJA temperature; all values are highly significant at the p < 0.001 level

Full size image

By comparing all regional chronologies over the common period (Table 2), the Glk index shows values ranging from 0.67 (LADE vs. PICE) to 0.95 (avgALL vs. avgLADE-PICE) and the correlation coefficient from 0.59 (LADE vs. PCAB) to 0.98 (avgALL vs. avgLADE-PICE). Also, the HSTT (1469–2011) chronology shows consistent patterns with the constructed regional chronologies, with a Glk ranging from 0.75 (with PCAB) to 0.87 (with avgALL and avgLADE-PICE) and a correlation coefficient ranging from 0.70 (with LADE and PCAB) to 0.89 (with PICE).

Table 2 Correlation coefficient and Glk calculated over the common period 1773–2004 between all regional chronologies constructed. All values are significant at the p < 0.001 level
Full size table

By evaluating the dendroclimatic potential of the regional chronologies in reconstructing the summer (JJA) temperature over the considered region, we found that the regional chronologies avgALL, avgLADE-PICE and HSTT always show higher correlations than the LADE, PCAB and PICE chronologies (Fig. 2b). The highest values are found for avgALL (r = 0.74, p < <0.001) and for HSTT (r = 0.76, p < <0.001; the correlation increases to 0.78 when considering the whole period covered by the temperature record). These correlation values are mainly due to the ability of the dendrochronological records in capturing the long-term variability of summer temperature, in fact the correlation coefficients become much lower if Gaussian 20-yr. low-pass filtered series are removed from the original ones. In this case the correlation coefficients with summer temperature fall to 0.49 for avgALL and to 0.43 for the HSTT. The highest synchronicity in interval changes expressed by the Glk is found for the LADE and avgALL chronologies (Glk = 0.70).

The reconstruction of past summer temperature shows that the regression model performs better than the scaling approach for avgALL and HSTT, with RE and CE always showing positive values and PM presenting similar values in the sub-periods (Table 3). The avgALL and avgLADE-PICE models also perform well with the scaling approach, which, on the contrary, does not show satisfactory CE values for HSTT. By comparing the avgALL and the HSTT reconstructions performed using regression models, the common signal recorded is evident even if certain differences emerge over some periods (Fig. 3a). In particular, by considering 20-yr. low-pass filtered values (Fig. 3b), the HSTT show lower values than the avgALL from 1725 to 1800 and from 1845 to 1910, whereas higher values were recorded in the period 1935–2008. A shift in the reconstructed temperature patterns with respect to the instrumental record is visible in 1821 (after the local minimum of 1816), and in 1951–52 (after the local maxima of 1945, 1947 and 1950).

Table 3 Reconstruction statistics calculated for both regression and scaling approaches over the respective sub-periods of calibration and verification
Full size table
Fig. 3
figure 3

Reconstruction of JJA temperature anomalies based on the HSTT and avgALL chronologies (a). The Gaussian 20-yr. low-pass filtered reconstructions are also reported (b). The HSTT reconstruction and the contribution of the series from different species over time (c): the vertical line in 1566 delimits the older period where the number of series of P.abies exceeds the L.decidua series. (d) Scatter plot of actual JJA values on predicted JJA temperature anomalies obtained by the avgALL (left) and HSTT (right) models. Regression equations and their respective determination coefficients are also reported

Full size image

The reconstructed summer temperature over the period 1567–2008 based on the HSTT chronology reveals a strong temperature increase established since approximately 1820 AD (Fig. 3c). The HSTT is mainly based on samples of P.cembra (a mean of approximately 120 samples per year) and L.decidua (32) samples, whereas the P.abies are less represented (11). Overall, the number of samples per year is >300 from 1655 to 2007 with a maximum of 312 samples over the period 1902–1920.

The predicted JJA temperature anomalies (by means of the two regression models avgALL and HSTT) show a good correspondence with the actual JJA values, with the former ones explaining up to 54.9 % (avgALL) and 60.1 % (HSTT) of the variance in the actual values (Fig. 3d). The extreme value of 2003, which is an outlier in the actual temperature series, does not correspond to the highest predicted values.

By evaluating the presence of HSTT trees using the site fitness (SF), we find higher mean values in L.decidua and P.cembra sites, whereas we find the lowest values for P.abies (Fig. 4 and table therein). SF values higher than 60 % are found only at sites 10 and 5 (L.decidua), whereas most sites of the three considered species have maximum values at approximately 40 %. L.decidua also shows two sites with SF = 0, whereas the minimum values were found at approximately 10 % for P.abies and P.cembra (sites 12 and 26). SF is correlated neither with the total number of series (r = −0.072) nor with elevation (r = 0.019).

Fig. 4
figure 4

Site fitness (SF) expressed as the percentage of HSTT series with respect to the total of series available at each site (red line). Mean values, standard deviation, median, maximum and minimum values of SF are reported in the included table

Full size image

To further verify the reliability of the two JJA temperature reconstructions obtained through regressions with the HSTT and avgALL chronologies, these reconstructions were compared with two summer temperature reconstructions for Central Europe and the European Alps freely available on the NOOA paleoclimatology dataset (Online Resource 4): Büntgen11 (Büntgen et al. 2011; https://www.ncdc.noaa.gov/paleo/study/10394) and Trachsel12 (Trachsel et al. 2012; https://www.ncdc.noaa.gov/paleo/study/12996). Lower differences between reconstructions were obtained comparing the HSTT and avgALL with the tree ring-based Büntgen11 reconstruction than when comparing the HSTT and avgALL with the Trachsel12 multiproxy summer temperature reconstruction. Overall, both the Büntgen11 and the Trachsel12 reconstructions tend to overestimate the summer temperature of the 19th century (in our study area), and the Trachsel12 shows the highest differences with the considered tree-ring based reconstructions in the period 1800–1860, comprising the Little Ice Age (LIA) peaks of 1816–1821 and the last great glacier advance in the region. The avgALL reconstruction tends to slightly underestimate the last period of temperature increase, after 1975.

4 Discussion

Temperature reconstructions from high-altitude mountain sites may be affected by many factors influencing tree growth: external variables, such as the physical environment with its geomorphic dynamism, climate comprising also its extremes, and recurrent insect outbreaks; and by many internal variables, related to tree age, forest growth, species-specific responses, and competition for resources. For temperature reconstructions, only high-altitude forest sites are usually selected to capture the highest, climatically driven, variability in tree-ring growth (e.g., Schweingruber 1996). In this study, we selected only sites at >1800 m a.s.l. to minimize other possible influences on tree growth related to internal forest dynamics, but we only found more sensitive tree-ring chronologies at higher elevations for larch, whereas the stone pine sensitivity showed a much weaker association with elevation. In contrast with that presumed, no relationship between the frequency of HSTT trees and altitude was found, showing that also at the highest altitudes the temperature signal can be masked by other signals. All series were standardized to minimize the forest internal influences as well as to remove age-related trends in the ring width series. In particular, tree age influences the tree-ring sensitivity to climate (Carrer and Urbinati 2004; Rossi et al. 2008), thus modulating the climatic signal recorded along the life of each tree. To minimize this problem, we selected only tree-ring series from >100 yr. old trees.

However, trees may record other signals over time, thus masking or altering the climatic signals. Environmental factors also in the remote sites of the European Alps are mainly related to the presence of human-related activities, some of the most important factors being forest thinning and fertilization related to cattle pastures (Motta and Nola 2001; Dirnböck et al. 2003). Also, topographic variables (e.g. elevation, aspect and slope) may influence the climatic signal recorded in tree rings; Leonelli et al. 2009; Bunn et al. 2011; Salzer et al. 2014), as well as geomorphological processes (the non-disruptive ones, like soil creep or sheetfloods; e.g., Pelfini et al. 2006).

Based on our results, we can underscore that the species composition in the pool of chronologies used for performing the reconstruction has a significant effect on the resulting reconstructions. As we showed, different species produce different responses to ongoing climatic changes: e.g., in the recent period since 1980, the LADE chronology (based on chronologies from 8 sites) showed a marked increasing trend of mean ring widths that follows the ongoing temperature trends on the Alps very well, whereas PCAB (4 sites) and PICE (12 sites) show much less positive responses, thus potentially resulting in a divergence between the tree-ring records and the temperature records. Indeed, evidence of this divergence problem is widely reported in the literature for both high-latitude and high-altitude sites (e.g., Jacoby and D’Arrigo 1995; Briffa et al. 1998a, 1998b; D’Arrigo et al. 2008; Loehle 2009); however, as we have found, divergence occurs according to site-specific and species-specific characteristics. In the European Alps, e.g., no divergence was found in a work based on L.decidua and P. abies (Büntgen et al. 2008), and unprecedentedly wider tree rings in Pinus longaeva have been found in high-altitude sites in western North America since the second half of the 20th century compared with the previous 3700 yr. (Salzer et al. 2009), thus suggesting a positive response to the globally increasing air temperatures.

The temperature reconstruction proposed here with the avgALL chronology attempts to overcome the problem of different species’ sensitivity to climate by representing each species with one regional chronology (the LADE, PCAB and PICE), even though the original dataset is composed of 26.2 % larch, 11.9 % Norway spruce and 61.9 % Swiss stone pine. Comparisons with the Büntgen11 and Trachsel12 JJA summer temperature reconstructions reveal that, overall, the avgALL (and the HSTT) chronologies enable reconstructions of lower summer temperature values during the LIA period for the 46°N 10°E grid cell.

The approach proposed here was also motivated by evaluating the influence of trees highly sensitive to temperature on reconstructing the summer temperature, because this datum is crucial in this highly glacierized region of the European Alps, where the longest series of mass balance are available (Zemp et al. 2009; WGMS 2012; Carturan et al. 2013) and where the potential for tree-ring chronologies in reconstructing past glacier fluctuations and mass balances has been already demonstrated (Leonelli et al. 2008, 2011). As we found, the HSTT performs well especially in reconstructing the long-term variability, whereas it is less efficient than the avgALL at the interannual scale. The HSTT reconstruction follows the long-term temperature signal better than the avgALL reconstruction and the HSTT model predicts lower temperatures values in the period before 1800 than does the avgALL, and higher values in the recent period, already since approximately 1935.

By considering the HSTT and avgALL reconstructions, they show different performances in the selected sub-periods (Tab. 3) and the HSTT shows a lower R2 value in the 1763–1885 period than in the following period. This lower R2 is due to a rather low long-term signal, which causes the R2 to depend almost only on the interannual variability. The proposed reconstruction over the entire period of 246 yr. of instrumental data shows that HSTT better predict the long-term summer temperature variations than the avgALL. Reducing the two subset periods to an extent of 92 yrs. (1793–1885, 1886–1978), thus truncating the instrumental series by 30 yr. at both series ends, the models still evidence a better performance of the HSTT (see Online Resource 5 for details).

Trees highly sensitive to temperature (representing on average 33 % of total trees in L.decidua sites and 27.5 % in P.cembra sites) are growing with larger tree rings than the other trees from the same sites in the region in recent period. HSTT trees confirm that the lower temperature phases during the LIA occurred over 1580–1620 and 1810–1830. Minor phases occurred between 1490 and 1530 and in series of approximately twenty years between 1630 and 1800. These results agree with some reconstructions of Holocene glacier expansions, occurred in Ortles-Cevedale and Adamello groups (Carturan et al. 2014; Baroni et al. 2014; Pelfini et al. 2014).

5 Conclusions

To better address the climate history of an important glacierized area in the central European Alps, we have presented two new dendroclimatic reconstructions of summer mean temperatures for a grid cell at 46° N 10° E. The tree-ring width chronologies of the three species in this study showed different sensitivities to climate and to the ongoing trends of increasing air temperatures over the region; however, they perform well as climate proxies. The best reconstruction performance on the long-term scale was observed for the HSTT chronology composed of only highly sensitive trees, thus representing a valuable approach for providing, annually resolved summer temperature information to be used also in glaciological models. Overall, we intend to stress the importance of testing the presence of trees highly sensitive to climate for detecting important dendroclimatic sites within a tree-ring network, as the incidence of these trees may influence climate reconstructions as much as the other known factors related to, e.g., site characteristics, tree age and species sensitivity. When performing reliable climate reconstructions, highly sensitive trees will indeed enhance the climate signal of tree-ring chronologies and reduce potential biases resulting from non-climatic influences on tree growth. Our results can further help to avoid the divergence problem, particularly for those periods where the avgALL curve showed a tendency to smooth the long-term summer temperature signal.