1 Introduction

The Tibetan Plateau (TP), with an average elevation greater than 4,000 m a.s.l., is considered as one of the Earth’s most prominent geomorphic features. It is also a climatically important region, due to its influences on large-scale atmospheric circulations over Asia (Li and Yanai 1996; Webster et al. 1998). The ecological ecosystems and hydrological cycles on the TP are sensitive to climate change (Jin et al. 2009; Huang et al. 2011; Wang et al. 2011a). Decades-long instrumental records have revealed an overall warming trend in surface air temperature over the TP during recent decades, especially during the winter season (Liu and Chen 2000) and at high elevations (Liu et al. 2009). However, meteorological stations on the TP are sparse and seldom have records prior to the 1950s. Therefore, our understanding of climatic variability from a long-term perspective is greatly dependent on a large number of high-resolution proxy climate records.

Tree-ring records have high temporal resolution and reliability, and can thus provide important information on long-term climate fluctuations. During the past decade, tree rings have been widely used to reconstruct the temperature change history over the TP (Bräuning and Mantwill 2004; Esper et al. 2003a; Fan et al. 2010; Gou et al. 2008; Liang et al. 2008; Liu et al. 2005; Yadav et al. 2011; Yang et al. 2010; Zhu et al. 2008). However, only a few tree ring-based temperature reconstructions have extended into the last millennium on the TP (Esper et al. 2003a; Liu et al. 2005; Yadav et al. 2011; Zhu et al. 2008), and most of these were derived from trees in the Qilian Mountains located in the northeastern part of the TP. To investigate historical temperature variability on the southeastern TP, Bräuning (1994) established a 1400-year-long tree-ring chronology in the Qamdo region; however, they failed to carry out appropriate calibration and verification testing between ring-width and climate data (Bräuning 1994, 2001; Bräuning and Grießinger 2006) due to the short period of overlap (1954–1991) between instrumental data and tree-ring records.

In this study, we developed new millennium-long ring-width chronologies for the Qamdo region using the signal-free regional curve standardization (RCS) and standard dendrochronological (STD) methodologies, and we reconstructed temperature variability during the past millennium for the southeastern TP. To reveal temperature variability over the TP and its relationship with the Northern Hemisphere (NH) climate change, we compared these chronologies with other temperature reconstructions from the TP and the NH. We also analyzed the relationship between TP temperature variability and Atlantic Ocean climate variability. In particular, we focused on decadal to multidecadal climate signals in our tree ring-based temperature reconstruction for the southeastern TP.

2 Materials and methods

2.1 Study area and climate

The study site of Tibetan juniper (Sabina tibetica Kom.) is located in the Zhujiao Mountains in the northwest part of Qamdo County, an area of sparsely distributed and well preserved old forests (Fig. 1). The forest has an open structure with a crown density of less than ca. 20 %. Dead tree trunks and snags cover an area of ca. 2 %. The tree-ring samples used in the current study were all sampled from the upper part of the forest belt (4,350–4,500 m a.s.l.).

Fig. 1
figure 1

Locations of tree-ring site, nearby meteorological stations and the closest PDSI grid point

Full size image

To better understand the regional climate characteristics, we selected three meteorological stations around the sampling site: Qamdo (31°09′ N, 97°10′ E, 3,306 m a.s.l., observation interval 1954–2010), Nangqian (32°12′ N, 96°29′ E, 3,644 m a.s.l., observation interval 1957–2010), and Yushu (31°01′ N, 97°01′ E, 3,681 m a.s.l., observation interval 1952–2010). The regional meteorological data series for the common period 1957–2010 were based on monthly data from the three stations using techniques outlined by Jones and Hulme (1996). Monthly values for each station were converted to z-scores relative to the period 1957–2010, and then averaged and converted back to dimensional precipitation and temperature values using the grand mean and grand standard deviations of the original monthly series (Wilson et al. 2005). For the common period 1957–2010, regional mean annual precipitation and temperature were 499.27 mm and 5.00 °C, respectively. July (14.10 °C) and January (−5.29 °C) were the warmest and coldest months, respectively. The growing season covered approximately May–September, during which more than 80 % of the annual total precipitation was received. The seasonality of temperature and rainfall in this region exhibits a monsoonal-type pattern.

2.2 Tree-ring data and detrending

During two field visits in September 2010 and April 2011, increment cores were extracted at breast height from each juniper tree using increment borers. The samples were air-dried, fine-sanded, and cross-dated using skeleton plots (Stokes and Smiley 1996). The ring-width of each tree core was measured with a LINTAB 6 measuring system at a resolution of 0.01 mm. The cross-dated tree-ring series were quality checked with COFECHA software (Holmes 1983). Cores having poor correlation with the master series, due to either age-related trends or different local habitats, were removed. Ultimately, 144 increment cores from 88 trees sampled in the upper part of the forest belt were included in construction of the ring-width chronology in the present study.

The traditional standard chronology (STD) was developed using ARSTAN software (Cook 1985). Before calculating the final chronology, the variance of each series was stabilized using a data-adaptive power transformation based on the local mean and standard deviation (Cook and Peters 1997). To remove age-related biological trends, a modified negative exponential curve (any k) was fitted to the raw series. A cubic smoothing spline function was then used to detrend the tree-ring sequences, with a 50 % frequency–response cutoff equal to 67 % of the series length. The tree-ring chronology was then obtained by calculating the residuals between the raw measurements and the fitted values, which can stabilize the variance and avoid systematical biases in the chronology, offering an improvement on the ratios method (Cook and Peters 1997). Furthermore, this procedure (data-adaptive power transformation and residuals) is particularly effective at avoiding the potential problem of index value inflation associated with the ratios method when annual measurement values are divided by a fitting curve whose values do not locally fit actual growth and approach zero (Frank et al. 2009). The expressed population signal (EPS; Wigley et al. 1984) with a threshold of 0.85 was used to assess the confidence of the chronology.

We also developed a second chronology using the regional curve standardization (RCS) detrending technique, a method that is often employed to retain low-frequency variations in mean tree-ring chronology (Briffa et al. 2002; Briffa and Melvin 2011; Esper et al. 2002a; Melvin and Briffa 2008). However, the RCS method is susceptible to the influence of non-climate-related factors that may bias RCS chronologies. For example, the “trend-in-signal” bias, which is most prevalent at the ends of chronologies, results from the removal of non-climate-related variance (Briffa and Melvin 2011; Melvin and Briffa 2008). The term ‘‘trend distortion’’ was used by Melvin and Briffa (2008) to describe this effect. Further, they described the “signal-free” standardization approach to mitigate the trend distortion problem (Briffa and Melvin 2011; Melvin and Briffa 2008). In the present study, the signal-free RCS method was used to develop the RCS chronology.

3 Results

3.1 Chronology characteristics

Figure 2 shows the age-aligned regional growth curve used for the signal-free RCS standardization, mean ring-width, and sample depth for each year. The regional growth curve shows an even distribution. The earliest 300-year section can be regarded as the fast, ‘grand period’ of growth, after which there was little downwards trend on a multicentury time scale. The final reliable chronology extends to A.D. 985–2010, based on 17+ trees and EPS > 0.85 (Wigley et al. 1984). The comparison between the RCS and STD chronologies (Fig. 3a) reveals close similarities in both the trend and amplitude of variations, with a correlation coefficient of 0.97 for the period 985–2010. As stated by Cook et al. (1995), the climate frequencies extracted from the STD chronology may have periods as long as one-third of the mean segment length of ring-width samples. Thus, the high coherence between RCS and STD chronologies in this study might be explained by the long mean segment length (662 years) of ring-width samples, which indicates the potential for preserving multi-centennial scale trends in the STD chronology.

Fig. 2
figure 2

Regional growth curve used for the “signal free” RCS standardization with mean ring-width, and its ±standard error (light grey shading). The lower graph shows the sample depth (dark grey shading)

Full size image
Fig. 3
figure 3

a Comparison between the RCS and STD chronologies; b Temporal variation in sample depth, mean segment length (MSL) and mean cambial age (MCA) for all samples and the subset of samples; c Comparison between two RCS chronologies calculated from all samples (144 samples) and the subset of samples (48 samples); d Comparison between the RCS chronology and original ring-width series. The original ring-width series was calculated from the 7 trees older than 800 years

Full size image

The mean sample segment length had a mean age of 662 years over the whole chronology period and was relatively even over time (Fig. 3b). The mean cambial age of the samples in each year is generally stable through time (Fig. 3b). Mean cambial age increasing since the 13th century may be caused by more old trees entering the mature stage over time. In addition, the mean sample tree age exhibited low values (lower than the mean age of 332 years in the chronology period) during the period 1100–1500 in response to the increasing number of young trees. The increase of mean tree age since the 13th century and the relatively large number of young trees in the period 1100–1500 might impart some bias in the RCS chronology slope (Esper et al. 2003b; Briffa and Melvin 2011; Büntgen et al. 2010; Cooper et al. 2012; Wilson et al. 2012).

To assess the bias associated with the post-13th century increase in mean tree age, we calculated a subset RCS chronology following the same method as that used to calculate the RCS chronology based on all samples. The subset of samples comprised 48 ring-width series younger than 600 years selected with an expectation of obtaining a stable mean tree age and sample depth during the chronology period. A relatively stable mean tree age was found for the subset RCS chronology during the chronology period (Fig. 3b). The consistent low-frequency variations in the two RCS chronologies (Fig. 3c) suggested that the RCS chronology developed in this study does not contain evidence for biases imparted by the increase of mean cambial age since the 13th century.

For the purpose of eliminating the possible bias imparted by young trees we calculated an original ring-width series using the 7 longest trees (>800 years), referring to the method used by Esper et al. (2002b). The specific process was as follows. We cut off the fastest-growing section (about 300 years of the earliest period), divided each ring-width series by its arithmetic mean, merged two ring-width series from the same tree into one series, converted all series to z-scores, and then calculated a mean ring-width anomaly chronology. The suitability of this method has been confirmed by Esper et al. (2002b) when applied to juniper trees, who adopted the rationale that uneven annual tree growth rates around the stem, and thus noticeable long-term age-trends are missing because of the existence of strip-bark growth forms in old junipers.

As shown in Fig. 3d, very similar fluctuations were seen in both the RCS chronology and original ring-width series. For the period 1100–1500, although there were potential biases in the RCS chronology, both tree-ring series showed coherent patterns of low-frequency variability on multidecadal time scales. These results suggest that the biases imparted by young trees played a minor role in influencing the trend of the RCS chronology in this study. Also, the consistent variation patterns among the RCS, STD, subset RCS chronologies and original ring-width series (Figs. 3a, b and d) confirm the robustness of the RCS chronology.

3.2 Tree growth–climate relationships

We assessed climate–tree growth relationships for the common period of ring-width and climate data. Because our STD and RCS chronologies had a very high correlation (r = 0.99) for the period 1957–2010, here we only show the analysis results for the relationship between the RCS chronology and climatic factors. Correlation and response function analyses for the RCS chronology with monthly temperature and precipitation were calculated from January of the previous year to September of the current year. We also compared the RCS chronology with the nearest Palmer Drought Severity Index (PDSI) grid-point data (2.5° × 2.5°; the center is 31.25°N, 96.25°E) from Dai et al. (2004). Because the earliest instrumental records from the nearest station (Qamdo station) did not begin until 1954 for this region, we extracted the most reliable portion (1954–2004) for correlation analysis. The climate–growth relationships were analyzed using DendroClim 2002 software (Biondi and Waikul 2004).

As can be seen in Fig. 4a, significant (p < 0.05) positive correlations were found between the RCS chronology and temperature in 10 months of the previous year. Among them, the strongest correlations occurred in November and December of the previous year, and January of the current year (r = 0.41, 0.50, and 0.47, respectively). The results of the response function analyses confirmed these relationships (Fig. 4b). Response coefficients between the chronology and previous November temperatures were significant at the 0.05 level. For the current year, the chronology was also positively correlated with January, February, April, July, August and September temperatures. Besides monthly temperature data, we also tested tree growth-climate relationships with different combinations of seasonal months. Since there were strong autocorrelations in our tree-ring chronologies, with respective AC1 values of 0.81 and 0.80 for the RCS and STD chronologies in the calibration period 1957–2010, the degrees of freedom were adjusted to account for the autocorrelation in two series when testing a correlation coefficient’s significance (Bretherton et al. 1999) in subsequent correlation analysis. As a result, the best relationship was found between the chronology and mean annual temperature in the previous year (January–December), with r = 0.71, adjusted degrees of freedom (adf) = 26 and p < 0.01. Moreover, the ring-width chronology was also significantly correlated with the temperature in July of the previous year to June of the current year (r = 0.60, adf = 20, p < 0.01), November of the previous year to January of the current year (r = 0.62, adf = 33, p < 0.01), and summer (June–August) of the previous year (r = 0.53, adf = 30, p < 0.01). These results suggest that temperature conditions, especially in the previous year, are an appropriate parameter to explain ring-width variance.

Fig. 4
figure 4

a Correlations of tree-ring index with monthly sums of precipitation, mean temperature and PDSI from January of the previous year to September of the current year. The horizontal red line indicates significance at the 0.05 level. b Same as (a) but for response functions. Asterisks indicate significance at the 0.05 level

Full size image

For precipitation, the correlations with ring-width chronology were lower than those with temperature. The RCS chronology was only significantly (p < 0.05) correlated with precipitation in May and October of the previous year, and March and May of the current year. We calculated correlations between the ring-width chronology and the totals of seasonal precipitation, and found that the totals of precipitation in spring (March–May) of the current year had a significant correlation with ring-width chronology (r = 0.48, adf = 28, p < 0.01). For PDSI, May of the current year as a single month had the strongest correlation (r = 0.48) with ring-width chronology. Although there were another six months of the current year in which monthly PDSI was significantly (p < 0.05) correlated with the ring-width chronology, each of their coefficients was less than 0.36. PDSI in spring (March–May) of the current year was correlated with the ring-width chronology (r = 0.42, adf = 31, p < 0.01). Response function analysis showed no significant response due to monthly precipitation or PDSI (Fig. 4b), suggesting the minor role of hydroclimate on tree growth at high altitude sites in this region. To further assess the relationship between PDSI and tree growth during the past centuries, we calculated the correlation between our chronology and the nearest PDSI grid-point data (31.25°N, 96.25°E) of the MADA gridded dataset (Cook et al. 2010). The correlation coefficient was relatively low and only reached 0.25 for their common period 1300–2005, indicating that the summer PDSI is not a primary factor, but rather a minor factor limiting tree growth at the upper portion of the forest belt in our study area during the past centuries.

The above results indicate that the previous year’s temperature may play the most important role in tree growth, although the May PDSI still has some effects. Climatic conditions of the previous year may precondition physiological processes within the tree and hence strongly influence tree growth of the current year (Bradley 1985). Consequently, the tree-ring record may contain a high autocorrelation. In this study, the strong influence of climate conditions in the previous year is corroborated by the wood anatomy which shows a clear earlywood structure and very little latewood. Additionally, analysis of the tree growth–climate relationship shows primary influences of temperature conditions in July–June, December–January, and January–December of the previous year. Due to it having the strongest relationship with tree growth, we selected annual mean temperature of the previous year (January–December) as the reconstructed variable in this study. Also, choosing the entire calendar year is of great practical meteorological benefit.

3.3 Temperature reconstruction

The correlation coefficients between instrumental temperature data and RCS (STD) chronology were 0.71 (0.72). These correlations are still significant at the 0.01 level after adjusting the degrees of freedom. In addition, significant correlations were found between the first differences of the reconstructions and instrumental data (r = 0.40 and 0.41, p < 0.01, respectively, for the RCS and STD reconstructions). These results indicate that the high first-order autocorrelations in both the STD and RCS chronologies did not play an important role in our identification of the previous year’s temperature as our target for reconstruction. We only show the RCS reconstruction in the following analysis because of coherent low- and high-frequency variations in the RCS and STD chronologies (Fig. 3a). The reconstruction interval began in A.D. 984, the period up to which the ring-width chronologies had an expressed population signal (EPS) > 0.85. A simple linear regression model (Y = 1.401X + 3.619) was obtained to reconstruct the annual mean temperature in the previous year. The RCS reconstruction explains 50 % of the observed instrumental temperature variance during the period 1957–2009.

The “leave-one-out” cross-validation method described by Michaelsen (1987) was employed to verify the reconstruction, because the available meteorological data were too few to carry out a robust split-sample calibration (Fritts 1976). The reduction of error (RE) had a positive value of 0.46 for the reconstruction (Table 1), confirming the reliability of the derived reconstruction (Fritts 1976). Statistically significant sign test (ST) and correlation (r) values between the recorded data and the “leave-one-out” estimates are all indications of the reconstruction’s validity (Table 1). Model residuals were reasonably normally distributed, showing low, insignificant (not significant at the 0.05 level) first autocorrelation values (serial correlation = 0.05), with Durbin-Watson (DW) values of 1.83 near the target of 2. To calibrate the first difference between the estimated value and the actual data, the results showed that actual records were significantly (p < 0.01) correlated with the “leave-one-out” estimated series, with a correlation coefficient of 0.31 for the reconstruction. In addition, the positive RE values indicated that the developed regression model provides a useful prediction (Fritts 1976; Cook et al. 1994). All these results support the robustness of these reconstruction models.

Table 1 Statistics of calibration and leave-one-out verification results for the common period 1957–2009
Full size table

The millennial length mean annual temperature reconstruction indicates centennial-scale cold conditions during A.D. 1000–1130 and A.D. 1600–1800 below the long-term mean of 4.98 °C (Fig. 5). A recent cool period in the 1960–1980s was associated with the advance of the Melang glacier in the Hengduan Mountains, southeastern TP (Zheng et al. 1999; He et al. 2003). With the exception of a cold interval in A.D. 1200–1220, a prolonged warm period occurred during A.D. 1140–1350. Eight of the ten warmest years occurred in the 2000s, with the warmest year in 2009, whereas seven of the ten coldest years occurred in the 17th century with the coldest year in 1726 (4.11 ± 0.44 °C). However, these warmest and coldest years are not statistically significant. The regression-based, reconstructed value-related uncertainty was calculated in the form of ± root mean square error (RMSE) (Cook et al. 2012).

Fig. 5
figure 5

Annually resolved reconstruction of Qamdo mean annual temperature for the period A.D. 984–2009 with 1 RMSE uncertainties. The annually resolved values are in red, the 30-year low-pass filtered values are in blue, and the uncertainties are in orange

Full size image

4 Discussion

4.1 Spatial representation of the reconstructed temperature

We chose 56 meteorological stations (90–103.5°E, 26.5–38.5°N) covering the eastern and central TP areas to investigate the spatial patterns of annual mean temperatures during their common period 1963–2007. An empirical orthogonal function (EOF) was used to reveal the spatial and temporal characteristics of temperature variability. The EOFs were calculated with the correlation matrix of the available 56 temperature time series over the period 1963–2007. The EOF1 mode explains more than 70 % of the total annual temperature variance. This pattern essentially represents the coherent temperature variability over the eastern and central TP, with the correlation coefficients >0.8 (43 stations) covering most of the area of interest (figure omitted). These results indicate that annual mean temperature variability has been highly consistent across the eastern and central TP during the past 50 years, which agrees well with the study of Xu et al. (2007).

We compared our reconstruction with the actual series and the first principal component (PC1) of 56 instrumental temperature records for the eastern and central TP. As can be seen in Fig. 6, the reconstruction shows a close similarity to PC1 and the actual series (in both the high- and low-frequency domains). PC1 is significantly correlated with the reconstruction (r = 0.75, adf = 18, p < 0.01) and the instrumental temperature record (r = 0.94, adf = 29, p < 0.01) during the period 1963–2007. Figure 7 showed the spatial correlation patterns between the reconstruction and temperature records of 56 stations. Most of the stations have correlation coefficients >0.60. To further highlight the spatial representation of climate change, we also calculated the spatial correlations of our reconstruction and the instrumental records with the CRU gridded dataset (TS3.1) (Mitchell and Jones 2005) during the period 1957–2009, using KNMI Climate Explorer (http://climexp.knmi.nl). As shown in Fig. 8b, the reconstruction correlates significantly with the gridded surface temperatures on a regional scale. The spatial correlation patterns are similar for the reconstructed and instrumental (Fig. 8a) temperatures. The values of the correlation coefficients were >0.50 for most areas of the TP, especially for the southern TP, where they were >0.60. The above results indicate that our temperature reconstruction is a reasonable proxy of temperature change for the whole TP.

Fig. 6
figure 6

Comparisons of the estimated mean annual temperature with the actual series and PC1 of 56 instrumental temperature records over the central and eastern TP

Full size image
Fig. 7
figure 7

Spatial correlations between the RCS reconstruction and temperature records from 56 meteorological stations during the period 1963–2007

Full size image
Fig. 8
figure 8

Spatial correlations of the instrumental (a) and RCS reconstructed (b) annual mean temperature for the southeastern TP with regional gridded annual mean temperatures during the period 1957–2009

Full size image

4.2 Comparisons with other temperature reconstructions

We compared the RCS reconstruction with the nearest temperature grid-point data (32.5°N, 97.5°E) from Mann et al. (2009) during the past millennium (figure omitted). In general, similar patterns of warm-cold variations were seen between our reconstruction and the subset of the temperature reconstruction of Mann et al. (2009) during the past millennium. The largest discrepancy was found in the 19th century, when our temperature reconstruction showed positive anomalies; in contrast, negative anomalies were seen in the subset of the temperature reconstruction of Mann et al. (2009). The discrepancy was partially caused by differences in the temperature sensitivities of climate proxies. It is obvious that more high-resolution temperature-sensitive records are needed to better understand temperature variability over the TP.

We compared our reconstruction with other temperature reconstructions from different parts of the TP and adjacent regions (Cook et al. 2003; Esper et al. 2003a; Liang et al. 2008; Zhu et al. 2008; Thompson et al. 2006). These tree ring-based temperature reconstructions were found to be closely related. The RCS reconstruction is most similar to the Nangqian summer minimum temperature reconstruction, during the period 1624–2002 (r = 0.45, adf = 214, p < 0.01). This strong similarity probably resulted from the relatively short distance between the two sampling sites. Our reconstruction also significantly correlates with the Wulan temperature reconstruction for the northeastern TP during the period 1000–2004 (r = 0.29, adf = 330, p < 0.01). Significant correlations were also found with the Nepal February–June temperature reconstruction by Cook et al. (2003), and with the west Tien Shan summer (June–September) temperature reconstruction by Esper et al. (2003a), implying a general similarity in interannual temperature variations across the eastern and western TP over the past millennium.

To compare low-frequency variations in temperature, a 30-year low-pass filter was applied to these tree ring-based temperature reconstructions. The reconstructions were also compared with the Puruogangri ice core δ18O record (raw data with a decadal resolution have already been interpolated to annual resolution) (Thompson et al. 2006). As could be expected, our temperature reconstructions generally show similar patterns of warm-cold variations to those of other temperature reconstructions. Discrepancies are partly due to the different seasonal responses of the tree-ring data used in different reconstructions. For example, some discrepancies were found in the second half of the 20th century when an extraordinary recent warming trend was seen in the Wulan September–April temperature record; this was less clearly evident in other series, which is possibly explained by a more pronounced warming trend in the winter half of the year than in the summer half of the year during the past decades (Liu and Chen 2000). However, before the 19th century, there were similar low-frequency fluctuations in the tree ring-based temperature reconstructions, indicating that temperature variations in different seasons were generally similar at decadal to centennial time scales, consistent with other studies based on tree-ring width (Liang et al. 2008; Yang et al. 2009, 2010). These results indicate generally synchronous temperature variations on decadal to centennial time scales across the TP during the past millennium.

The most prolonged warm period occurred during A.D. 1140–1350 (Fig. 9), which possibly corresponds to the Medieval Warm Period (MWP). It can be argued that our reconstructions do not cover the entire MWP because they only extend back to A.D. 984. However, the timing of the MWP in our reconstructions agrees well with that in the 2,000-year temperature record for the whole TP (Yang et al. 2003), indicating that our conclusions are reasonable. In addition, a recent reconstruction of summer temperature for temperature East Asia has also revealed a prolonged warm period during A.D. 1150–1350, with the exception of cold intervals around the 1,230 and 1,280 s (Cook et al. 2012). A common long-term cold episode occurred during the period 1600–1800, corresponding to the Little Ice Age (LIA).

Fig. 9
figure 9

Comparison of low-frequency temperature variations in different regions of the TP derived from tree rings and ice cores. All reconstructions were smoothed with a 30-year low-pass filter and normalized for each long-term average. The bars mark the Medieval Warm Period and Little Ice Age

Full size image

We also compared our reconstruction with large-scale temperature reconstructions for the Northern Hemisphere (NH) (Christiansen and Ljungqvist 2012; Crowley 2000; Mann et al. 2009; Moberg et al. 2005). As shown in Fig. 10, a large discrepancy between the TP and NH temperature records was found in the 11th century, when a cold climate characterised the TP while a warm climate was found in NH temperature records. The 17th and 18th centuries were the coldest period both over the TP and the NH during the last millennium (Christiansen and Ljungqvist 2012; Crowley 2000; Mann et al. 2009; Moberg et al. 2005). It has also been suggested that the 17th century was still the coldest century when extending further back through the last two millennia (Christiansen and Ljungqvist 2012). The 19th century was cold in the NH, but not obviously cold on the TP. However, on the whole, our temperature reconstruction is generally similar to these four NH temperature series during the last millennium, indicating that the southeastern TP experienced historical temperature variability on multidecadal timescales that was synchronous with variability in the NH. This agrees well with the study of Zhu et al. (2011), which suggested that the eastern TP and the NH have experienced similar temperature patterns at multidecadal time scale.

Fig. 10
figure 10

Comparison between the 30-year low-pass filtered Qamdo and Northern Hemisphere temperature reconstructions

Full size image

4.3 Temperature relationship with the Atlantic Multidecadal Oscillation (AMO)

To investigate the relationship of TP temperature variability with global sea surface temperature (SST) anomalies, we calculated spatial correlations between the RCS reconstruction and NCDC v2 SST data during the common period 1880–2009, using KNMI Climate Explorer (http://climexp.knmi.nl). The results suggested a close association between Atlantic SST anomalies and temperature changes on the southeastern TP (Fig. 11a). A closer relationship was found when the data were smoothed using a 21-year running mean (Fig. 11b), suggesting a significant influence of the AMO on multidecadal temperature variations over the southeastern TP. In addition, the existence of only a few significant correlations with Pacific SSTs (their annual data and 21-year running mean) suggested weak connections with the El Niño Southern Oscillation (ENSO) and the Pacific Decadal Oscillation (PDO). Associations between the TP climate changes and ENSO and PDO have been reported by many studies (e.g., Fang et al. 2009; Grigholm et al. 2009; Wang et al. 2003). Considering the stronger relationship with AMO, we focused on the influence of AMO on TP temperature variability.

Fig. 11
figure 11

Correlation patterns of the reconstructed temperature with the concurrent NCDC v2 SST data for the annual (a) and 21-year running mean data (b) during the period 1880–2009

Full size image

Significant correlations (r = 0.55, adf = 55, p < 0.01) were found between the RCS reconstruction and the instrument-recorded annual AMO index during the period 1856–2009 (the data were downloaded from the online data set: http://www.esrl.noaa.gov/psd/data/correlation/amon.us.long.mean.data). As shown in Fig. 12, the warm phases of the AMO were associated with positive temperature anomalies, and vice versa during the cool phases of the AMO; this agrees well with studies finding that Atlantic SST anomalies strongly affect the TP temperature and heat sources (Feng and Hu 2008). This has been supported by evidence from atmospheric general circulation models (AGCMs), demonstrating that during the warm AMO phases, temperatures are warmer in much of East Asia, especially in western China (Wang et al. 2009).

Fig. 12
figure 12

Comparisons of the AMO index with the TP and NH temperature series during the last 150 years

Full size image

Based on climate model simulations, Zhang et al. (2007) suggested that the AMO has played an important role in explaining the multidecadal variability in the NH mean temperature record in the 20th century. As shown in Fig. 12, we compared the instrument-recorded NH land surface mean temperature record (the data were downloaded from the web site http://cdiac.ornl.gov/trends/temp/jonescru/data.html) with the RCS reconstruction and the AMO index during the period 1850–2009. During the last 150 years, the NH and TP temperature variations have generally followed the AMO. Particularly worthy of note is that the close relationship between the AMO and NH temperatures has weakened in the 1850–1870s and after the 1960s, while a strong correlation still exists between the AMO and TP temperatures. This implies an important influence of the AMO on the surface heating over the TP, consistent with the study by Feng and Hu (2008). Observational evidence has indicated that the warm phases of the AMO are related to warm anomalies of temperature over Europe in all seasons (Sutton and Dong 2012). Such connection between AMO and temperature variability over Europe can also be obtained in climate model simulations (Knight et al. 2006). During the past five centuries, the AMO signals were detected by the six ring-width chronologies in northeastern Asia, suggesting the influence of AMO on climate variability in northeastern Asia (Wang et al. 2011b). Similar to these findings from the mid- and high-latitudes of the Eurasian continent, our results further suggest that the AMO can influence temperature variability in the southeastern TP. The evidence from both observations and model simulations indicated that the warm AMO phases are characterized by elevated temperatures over East Asia (Li and Bates 2007; Wang et al. 2009). Furthermore, this positive relationship between the warm (cool) AMO phases and above-average (below-average) temperatures over East Asia has been consistently stable and significant during the past millennium (for more details, see Wang et al. 2013). It has been suggested that the warm phases of the AMO are linked to enhanced mid-latitude westerlies (Grossmann and Klotzbach 2009), which induce negative surface air pressure anomalies over the North Atlantic extending to mid-latitude Eurasia (Knight et al. 2006; Li and Bates 2007). These factors in turn weaken the Siberian-Mongolian high-pressure system, and reduce the strength of the East Asian winter monsoon (Li and Bates 2007). As a result, the warm (cool) phases of the AMO are related to increasing (decreasing) temperatures over East Asia. Additionally, Li et al. (2008) conducted diagnostic simulations and showed a propagating Rossby wave train in the middle and upper troposphere extending from the Atlantic across Asia. This implies the possibility that warm Atlantic SSTs can directly warm up middle or upper troposphere of the Asian continent (Li et al. 2008). Although the mechanisms connecting Asian climate with the AMO are not fully understood, our findings provide good support for these reconstructed and model studies. The results presented here may contribute to a better understanding of climate variability and potential forcings of climate changes over East Asia.

5 Conclusions

Using the signal-free RCS and STD methodologies, we developed two millennium-long chronologies for the southeastern TP. The two chronologies are very similar in both their low- and high-frequency variations, with a correlation coefficient >0.97 for the last millennium. The RCS (STD) chronology exhibits strong and positive correlations with annual mean temperature in the previous year with r = 0.71 (0.72). Accordingly, based on the RCS chronology we reconstructed mean annual temperature variations over the period A.D. 984–2009, the period in which the ring-width chronology had an expressed population signal (EPS) >0.85. Our temperature reconstruction shows that two periods of prolonged cold conditions occurred during A.D. 1000–1130 and A.D. 1600–1800, and a prolonged warm episode occurred during A.D. 1140–1350. During the last decades, a rapid increase in tree growth was found in this study, which is a response to global warming.

Investigation of the spatial and temporal variations in annual mean temperature revealed coherent fluctuations in annual mean temperatures across the TP during the last 50 years. Spatial correlations of reconstructed temperatures with actual temperature records from 56 stations and the CRU gridded dataset indicated that our temperature reconstruction provides a reasonable proxy for temperature change over the whole TP. Comparisons of our reconstruction with other temperature records from the TP indicated that although there are some discrepancies between data series, in general, similar patterns of warm-cold variations suggested that decadal to centennial temperature variations have been generally synchronous across the TP during the past millennium. These reconstructions also indicate that the periods 1140–1350 and 1600–1800 were the common warm and cold periods over the TP. The comparison with NH temperature reconstructions revealed that generally similar patterns of warm-cold fluctuations were seen in the TP and NH on multidecadal time-scales during the last millennium. Significant positive correlations between our reconstruction and Atlantic SSTs suggest the influence of the AMO on surface heating of the TP. The mid-latitude westerly anomalies and propagation of Rossby waves related to the AMO may be the two channels linking North Atlantic SST anomalies and temperature variability over East Asia.

5.1 Data

Ring-width data will be made available online at the International Tree-Ring Data Bank (ITRDB). Climate data can be accessed from the China Meteorological Data Sharing Service System (http://data.cma.gov.cn/).