The study area, which spans 36–42∘ N and 26–38∘ E, was based on the distribution of available tree-ring chronologies. This vast area covers much of western Anatolia and includes the western Black Sea, Marmara and western Mediterranean regions. Much of this area is characterized by a Mediterranean climate that is primarily controlled by polar and tropical air masses (Türkeş, 1996a; Deniz et al., 2011). In winter, polar fronts from the Balkan Peninsula bring cold air that is centered in the Mediterranean. Conversely, the dry, warm conditions in summer are dominated by weak frontal systems and maritime effects. Moreover, the Azores high-pressure system in summer and anticyclonic activity from the Siberian high-pressure system often cause below normal precipitation and dry subhumid conditions over the region (Türkeş, 1999; Deniz et al., 2011). In this Mediterranean climate, annual mean temperature and precipitation range from 3.6 to 20.1 ∘C and from 295 to 2220 mm, respectively, both of which are strongly controlled by elevation (Deniz et al., 2011).

To investigate past temperature conditions, we used a network of 23 tree-ring site chronologies (Fig. 1). Fifteen chronologies were produced by previous investigations (Mutlu et al., 2011; Akkemik et al., 2008; Köse et al., 2005, 2011, 2013) that focused on reconstructing precipitation in the study area. In addition, we sampled eight new study sites and developed tree-ring time series for these areas (Table 1). Increment cores were taken from living Pinus nigra Arnold and Pinus sylvestris L. trees and cross-sections were taken from Abies nordmanniana (Steven) Spach and Picea orientalis (L.) Link trunks.

Samples were processed using standard dendrochronological techniques (Stokes and Smiley, 1968; Orvis and Grissino-Mayer, 2002; Speer, 2010). Tree-ring widths were measured, then visually cross-dated using the list method (Yamaguchi, 1991). We used the computer program COFECHA, which uses segmented time-series correlation techniques, to statistically confirm our visual cross-dating (Holmes, 1983; Grissino-Mayer, 2001). Cross-dated tree-ring time series were then standardized by fitting a 67 % cubic smoothing spline with a 50 % cutoff frequency to remove non-climatic trends related to the age, size and the effects of stand dynamics using the ARSTAN program (Cook, 1985; Cook et al., 1990a). These detrended series were then pre-whitened with low-order autoregressive models to produce time series with a strong common signal and without biological persistence. These series may be more suitable to understand the effect of climate on tree growth, even if any persistence due to climate might be removed by pre-whitening. For each chronology, the individual series were averaged to a single chronology by computing the biweight robust means to reduce the influences of outliers (Cook et al., 1990b). In this research we used residual chronologies obtained from ARSTAN to reconstruct temperature.

The mean sensitivity, which is a metric representing the year-to-year variation in ring width (Fritts, 1976), was calculated for each chronology and compared. The minimum sample depth for each chronology was determined according to expressed population signal (EPS), which we used as a guide for assessing the likely loss of reconstruction accuracy. Although arbitrary, we required the commonly considered threshold of EPS > 0.85 (Wigley et al., 1984; Briffa and Jones, 1990).

We extracted high-resolution monthly temperature and precipitation records from the climate dataset CRU TS 3.23 gridded at 0.5∘ intervals (Jones and Harris, 2008) from KNMI Climate Explorer (http://climexp.knmi.nl) for 36–42∘ N, 26–38∘ E. The period 1930–2002 CE was chosen for the analysis because it maximized the number of station records within the study area.

First, the climate–growth relationships were investigated with response function analysis (RFA; Fritts, 1976) for biological year from the previous October to current October using the DENDROCLIM2002 program (Biondi and Waikul, 2004). This analysis is done to determine the months during which the tree growth is the most responsive to temperature. RFA results showed that precipitation from May to August and temperature in March and April have dominant control on tree-ring formation in the area. Second, we produced correlation maps showing correlation coefficients between tree-ring chronologies and the climate factors most important for tree growth, which are May–August precipitation and March–April temperature, to find the spatial structure of radial growth–climate relationship (St. George, 2014; St. George and Ault, 2014; Hellmann et al., 2016). For each site we used the closest gridded temperature and precipitation values.

The climate reconstruction is performed by regression based on the principal components (PCs) of the 23 chronologies within the study area. Principle component analysis (PCA) was done over the entire period in common with the tree-ring chronologies. The significant PCs were selected by stepwise regression. We combined forward selection with backward elimination setting p < 0.05 as entrance tolerance and p < 0.10 as exit tolerance. The final model obtained when the regression reaches a local minimum of the root-mean-squared error (RMSE). The order of entry of the PCs into the model was PC3, PC21, PC4, PC15, PC5, PC17, PC7, PC9, PC10. The regression equation is calibrated on the common period (1930–2002) between robust temperature time series and the selected tree-ring series. Third, the final reconstruction is based on bootstrap regression (Till and Guiot, 1990), a method designed to calculate appropriate confidence intervals for reconstructed values and explained variance even in cases of short time series. It consists in randomly resampling the calibration datasets to produce 1000 calibration equations based on a number of slightly different datasets.

The quality of the reconstruction is assessed by a number of standard statistics. The overall quality of fit of reconstruction is evaluated based on the determination coefficient (R2), which expresses the percentage of variance explained by the model and RMSE, which expresses the calibration error. This does not ensure the quality of the extrapolation which needs additional statistics based on independent observations, i.e., observations not used by the calibration (verification data). They are provided by the observations not resampled by the bootstrap process. The prediction RMSE (called RMSEP), the reduction of error (RE) and the coefficient of efficiency (CE) are calculated on the verification data and enable testing the predictive quality of the calibrated equations (Cook et al., 1994). Traditionally, a positive RE or CE value means a statistically significant reconstruction model, but bootstrap has the advantage of producing confidence intervals for such statistics without theoretical probability distribution; finally, we accept the RE and CE for which the lower confidence margin at 95 % are positive. This is more constraining than just accepting all positive REs and CEs. For additional verification, we also present traditional split-sample procedure results that divided the full period into two subsets of equal length (Meko and Graybill, 1995).

To identify the extreme March–April cold and warm events in the reconstruction, standard deviation (SD) values were used. Years 1 and 2 SD above and below the mean were identified as warm, very warm, cold and very cold years, respectively. As a way to assess the spatial representation of our temperature reconstruction, we conducted a spatial field correlation analysis between reconstructed values and the gridded CRU TS3.23 temperature field (Jones and Harris, 2008) for a broad region of the Mediterranean over the entire instrumental period (ca. 1930–2002). Finally, we compared our temperature reconstruction and also precipitation signal (PC1) against existing gridded temperature and hydroclimate reconstructions for Europe over the period 1800–2002. We performed spatial correlation analysis between (1) our temperature reconstruction and gridded temperature reconstructions for Europe (Xoplaki et al., 2005; Luterbacher et al., 2016) and OWDA (Cook et al., 2015); and (2) PC1 and summer precipitation reconstruction (Pauling et al., 2006) and OWDA (Cook et al., 2015). To assess the significance of the correlation between our reconstruction (y) and gridded reconstructions (xj, j=1…N), we have calculated significance thresholds based on a Monte Carlo technique. For each grid point j, we have calculated the correlation between xj and y, but with a random permutation of the values of our reconstruction. This is repeated 1000 times with a different permutation. The 1000 correlation coefficients thus obtained are expected to be zero as the correlation is established on non-corresponding years. The 95th quantile of these 1000 coefficients is assumed to be passed in less than 5 % of the cases. Then a correlation coefficient with a higher value is considered as positive with a 95 % confidence. These thresholds are obtained with a common permutation for all xj so that the spatial structure is conserved in the tests. The + sign is assigned to the xj with a correlation higher than an expected value under the noncorrelation hypothesis.

In addition to 15 chronologies developed by previous studies, we produced 6 P. nigra, 1 P. sylvestris and 1 A. nordmanniana/P. orientalis chronologies for this study (Table 2). The Çorum district produced two P. nigra chronologies: one being the longest (KAR; 627 years long) and the other the most sensitive to climate (SAH; mean sensitivity value of 0.25). Previous investigations of climate–tree growth relationships reported a mean sensitivity range of 0.13–0.25 for P. nigra in Turkey (Köse et al., 2011; Akkemik et al., 2008). The KAR, SAH and ERC chronologies (with mean sensitivity values from 0.22 to 0.25) were classified as very sensitive, and the SAV, HCR and PAY chronologies (mean sensitivity values range 0.17–0.18) contained values characteristic of being sensitive to climate. The lowest mean sensitivity value was obtained for the ART A. nordmanniana/P. orientalis chronology. Nonetheless, this chronology retained a statistically significant temperature signal (p < 0.05).

RFA coefficients of May to August precipitation are positively correlated with most of the tree-ring series (Fig. 2) and among them May and June coefficients are generally significant. The first principal component of the 23 chronologies, which explains 47 % of the tree-growth variance, is highly correlated with May–August total precipitation, statistically (r= 0.65, p < 0.001) and visually (Fig. 3). The high correlation was expected given that numerous studies also found similar results in Turkey (Akkemik, 2000a, b, 2003, Akkemik et al., 2005, 2008; Akkemik and Aras, 2005; Hughes et al., 2001; D'Arrigo and Cullen, 2001; Touchan et al., 2003, 2005a, b, 2007; Köse et al., 2011, 2012, 2013; Martin-Benitto et al., 2016). The influence of temperature was not as strong as May–August precipitation on radial growth, although generally positive in early spring (March and April; Fig. 2). Conversely, the ART chronology from northeastern Turkey contained a strong temperature signal, which was significantly positive in March.

Correlation maps representing influence of May–August precipitation (Fig. 4a) and March–April temperature (Fig. 4b) also showed that strength of the summer precipitation signal is higher and significant almost all over Turkey. Higher precipitation in summer has a positive effect on tree growth, because of long-lasting dry and warm conditions over Turkey (Türkeş, 1996b; Köse et al., 2012). Spring precipitation signals are generally positive and significant only for four tree-ring sites. The sites located at the upper distributions of the species generally showed higher correlations. The highest correlations obtained for Picea/Abies chronology (ART) from the Caucasus, and for Pinus nigra chronology (HCR) from the upper (about 1900 m) and southeastern distribution of the species. This black pine forest was still partly covered by snow from the previous year during the field work in fall. Higher temperatures in spring maybe cause snowmelt earlier and lead to produce larger annual rings. In addition to these chronologies, we also used the chronologies that revealed the influence of precipitation as well as temperature to reconstruct March–April temperature.

The higher-order PCs of the 23 chronologies are significantly correlated with the March–April temperature and, by nature, are independent of the precipitation signal (Table 3). The best selection for fit temperatures are obtained with the PC3, PC4, PC5, PC7, PC9, PC10, PC15, PC17, PC21, which explains together 25 % of the tree-ring chronologies. So, the temperature signal remains important in the tree-ring chronologies and can be reconstructed. The advantage to separate both signals through orthogonal PCs enable removal of unwanted noise for our temperature reconstruction. Thus, PC1 was not used as a potential predictor of temperature because it is largely dominated by precipitation (Table 3, Fig. 3). The last two PCs contain a too-small part of the total variance to be used in the regressions. However, even if Jolliffe (1982) and Hadi and Ling (1998) claimed that certain PCs with small eigenvalues (even the last one), which are commonly ignored by principal components regression methodology, may be related to the independent variable, we must be cautious with that because they may be much more dominated by noise than the first ones. So, the contribution of each PC to the regression sum of squares is also important for selection of PCs (Hadi and Ling, 1998). The findings of Jolliffe (1982) and Hadi and Ling (1998) provide a justification for using non-primary PCs, (e.g., of second and higher order) in our regression, given that correlations with temperature may be overpowered by affects from precipitation in our study area (E. R. Cook, personal communication, 2011).

Using this method, the calibration and verification statistics indicated a statistically significant reconstruction (Table 4, Fig. 5). For additional verification, we also present split-sample procedure results. Similarly, the bootstrap results, derived calibration and verification tests using this method indicated statistically significant RE and CE values (Table 5).

The regression model accounted for 67 % (Adj. R2= 0.64, p < 0.0001) of the actual temperature variance over the calibration period (1930–2002). Also, actual and reconstructed March–April temperature values had nearly identical trends during the period 1930–2002 (Fig. 5). Moreover, the tree-ring chronologies successfully simulated both high frequency and warming trends in the temperature data during this period. The reconstruction was more powerful at classifying warm events rather than cold events. Over the last 73 years, 8 of 10 warm events in the instrumental data were also observed in the reconstruction, while 5 of 9 cold events were captured. Similarly, previous tree-ring-based precipitation reconstructions for Turkey (Köse et al., 2011; Akkemik et al., 2008) were generally more successful in capturing dry years rather than wet years.

Our temperature reconstruction on the 1800–2002 period is obtained by bootstrap regression using 1000 iterations (Fig. 6). The confidence intervals are obtained from the range between the 2.5th and the 97.5th percentiles of the 1000 simulations. Low-frequency variability of our spring temperature reconstruction showed larger variability in the 19th century than the 20th century. For the pre-instrumental period (1800–1929), a total of 23 cold (1813, 1818, 1821, 1824, 1837, 1848, 1854, 1858, 1860, 1869, 1877–1878, 1880–1881, 1883, 1897–1898, 1905–1907, 1911–1912, 1923) and 13 warm (1801–1802, 1807, 1845, 1853, 1866, 1872–1873, 1879, 1885, 1890, 1901, 1926) events were determined. After comparing our results with event years obtained from May to June precipitation reconstructions from western Anatolia (Köse et al., 2011), the cold years 1818, 1848 and 1897 appeared to coincide with wet years and 1881 was a very wet year for the entire region. Furthermore, these years can be described as cold (in March–April) and wet (in May–June) for western Anatolia.

Among the warm periods in our reconstruction, conditions during the year 1879 were dry, 1895 wet, and 1901 very wet across the broad region of western Anatolia (Köse et al., 2011). Hence, we defined 1879 as a warm (in March–April) and dry year (in May–June), and 1895 and 1901 were warm and wet years. In the years 1895 and 1901 the combination of a warm early spring and a wet, late spring–summer caused enhanced radial growth in Turkey, interpreted as longer growing seasons without drought stress.

Of these event years, 1897 and 1898 were exceptionally cold and 1845, 1872 and 1873 were exceptionally warm. During the last 200 years, our reconstruction suggests that the coldest year was 1898 and the warmest year was 1873. The reconstructed extreme events also coincided with accounts from historical records. Server (2008) recounted the winter of 1898 as characterized by anomalously cold temperatures that persisted late into the spring season. A family that brought their livestock herds up into the plateau region in Kırşehir seeking food and water were suddenly covered in snow on 11 March 1898. This account of a late spring freeze supports the reconstruction record of spring temperatures across Turkey and offers corroboration to the quality of the reconstructed values.

Seyf (1985) reported that extreme summer temperature during the year 1873 resulted in widespread crop failure and famine. Historical documents recorded an infamous drought-derived famine that occurred in Anatolia from 1873 to 1874 (Quataert, 1996; Kuniholm, 1990), which claimed the lives of 250 000 people and a large number of cattle and sheep (Faroqhi, 2009). This drought caused widespread mortality of livestock and depopulation of rural areas through human mortality, and migration of people from rural to urban areas. Further, the German traveler Naumann (1893) reported a very dry and hot summer in Turkey during the year 1873 (Heinrich et al., 2013). Conditions worsened when the international stock exchanges crashed in 1873 (Zürcher, 2004). Our temperature record suggests that dry conditions during the early 1870s were possibly exacerbated by warm spring temperatures that likely carried into summer. A similar pattern of intensified drought by warm temperatures was demonstrated recently by Griffin and Anchukaitis (2014) for the current drought in California, USA.

Extreme cold and warm events were usually 1 year long, and the longest extreme cold and warm events were 2 and 3 years long, respectively. These results were similar with durations of extreme wet and dry events in Turkey (Touchan et al., 2003, 2005a, b, 2007; Akkemik and Aras, 2005; Akkemik et al., 2005, 2008, Köse et al., 2011; Güner et al., 2016). Moreover, seemingly innocuous short-term warm events, such as the 1807 event, were recorded across the Mediterranean and in high elevations of the European regions. Casty et al. (2005) reported the year 1807 as being one of the warmest alpine summers in the European Alps over the last 500 years. As such, a drought record from Nicault et al. (2008) echoes this finding, as a broad region of the Mediterranean Basin experienced drought conditions.

Heinrich et al. (2013) analyzed winter-to-spring (January–May) air temperature variability in Turkey since 1125 CE as revealed from a robust tree-ring carbon isotope record from Juniperus excelsa. Although they offered a long-term perspective of temperature over Turkey, the reconstruction model, which covered the period 1949–2006, explained 27 % of the variance in temperature since the year 1949. In this study, we provided a short-term perspective of temperature fluctuation based on a robust model (calibrated and verified 1930–2002; Adj. R2= 0.64; p < 0.0001). Yet, the Heinrich et al. (2013) temperature record did not capture the 20th century warming trend as found elsewhere (Wahl et al., 2010). However, their temperature trend does agree with trend analyses conducted on meteorological data from Turkey and other areas in the eastern Mediterranean region. The warming trend seen during our reconstruction calibration period (1930–2002) was similar to the data shown by Wahl et al. (2010) across the region and hemisphere. Further, the warming trends seen in our record agrees with data presented by Turkes and Sumer (2004), of which they attributed to increased urbanization in Turkey. Considering long-term changes in spring temperatures, the 19th century was characterized by more high-frequency fluctuations compared to the 20th century, which was defined by more gradual changes and includes the beginning of decreased DTRs in the region (Turkes and Sumer, 2004).

Spatial correlation analysis revealed that our network-based temperature reconstruction was representative of conditions across Turkey as well as the broader Mediterranean region (Fig. 7). During the period 1930–2002, estimated temperature values were highly significant (r range 0.5–0.6, p < 0.01) with instrumental conditions recorded from southern Ukraine to the west across Romania, and from northern areas of Libya and Egypt to the east across Iraq. The strength of the reconstruction model is evident in the broad spatial implications demonstrated by the temperature record. Thus, we interpret warm and cold periods and extreme events within the record with high confidence.

We compared our tree-ring-based temperature reconstruction with existing gridded temperature reconstructions for Europe (Xoplaki et al., 2005, Fig. 8a; Luterbacher et al., 2016, Fig. 8b) and the OWDA (Cook et al., 2015, Fig. 8c) for further validation of the reconstruction. Spatial correlations over the past 200 years were lower with reconstructed European summer temperature (May to July; Fig. 8b). Yet, we expected this result because of the paucity of Turkey derived proxies in the other reconstructions, as well as the differing seasons involved across the reconstructions. Similarly, our reconstruction showed weak correlations with summer drought index over Turkey. Besides comparing different seasons, perhaps this is because less precipitation begets drought conditions rather than high temperature in the region. The highest and significant (p < 0.05) correlations were found with European spring (March to May) temperature reconstruction over southeastern Europe, which are stronger over Turkey (Fig. 8a). We used the mean of corresponding grid points from European spring temperature reconstruction over the study area (36–42∘ N, 26–38∘ E) to show how the correlation changed over time (Fig. 9). The correlation coefficient was highly significant (r= 0.76, p < 0.001) during our calibration period (1930–2002). We found lower but still significant correlation (r= 0.35, p < 0.10) for the period of 1901–1929, which climatic records are very few over the region while available data has sufficient quality for most parts of Europe. These results give additional verification for our reconstruction. Moreover, our reconstruction has a weak, insignificant relationship (r= 0.13, p > 0.10) during the 19th century. This may be related to poor reconstructive skill of European spring temperature reconstruction over Turkey, which contains few proxies from the country (Xoplaki et al., 2005; Luterbacher et al., 2004). Nonetheless, these results demonstrate that tree-ring chronologies from Turkey can serve as useful temperature proxies for further spatial temperature reconstructions to fill the gaps in the area.

We also compared the precipitation signal (PC1) obtained from our tree-ring network with OWDA (Cook et al., 2015) and gridded European summer precipitation reconstruction (Pauling et al., 2006) to test the strength of the signal spatially (Fig. 8d and e, respectively). We calculated highly significant positive correlations with summer drought index over Turkey and neighboring European countries such as Greece, Bulgaria, Romania and Italy while significant correlations are lower for the northern Mediterranean countries (Fig. 8d). These results showed that summer precipitation signal represented by PC1 is very strong not only on instrumental period but also on pre-instrumental period and represents a large spatial coverage. We found low and insignificant correlations over Turkey and Mediterranean countries with European summer precipitation reconstruction (Fig. 8e). Pauling et al. (2006) stated that poor reconstructive skills determined over Turkey because of few instrumental record before the1930s.