Unprecedented Warming Revealed from Multi-proxy Reconstruction of Temperature in Southern China for the Past 160 Years

Several independent datasets show that the global mean surface temperature (combined land and ocean) increased by 0.85^°C (0.65^°C-1.06^°C) over the instrumental period of 1880-2012, with a linear trend of 0.64^° 0.15^°C (100 yr)^-1, and the land surface air temperature (LSAT) increased more evidently [0.92^° 0.15^°C (100 yr)^-1]. This warming occurred mostly in two periods: around 1900 to 1940 and around 1970 onwards, with a relatively warm interval around the 1940s (IPCC, 2013). In China, meanwhile, the LSAT warming trend has been reported at a rate of 0.5^°-0.8^°C (100 yr)^-1 for 1880-2008, in the 2nd National Assessment Report on Climate Change (EC-SCNARCC, 2011), but at a rate of 0.9^°-1.5^°C (100 yr)^-1 for 1909-2011 in the 3rd Assessment Report (EC-TCNARCC, 2015; Ding and Wang, 2016). This difference in the warming trend in China between two National Assessment Reports arises mainly from the difference among different studies in the construction method for the annual nationwide temperature before the 1950s (e.g.,Wang et al., 1998; Tang and Ren, 2005; Cao et al., 2013). For example, using instrumental data and temperature grades from 50 stations in 10 regions (5 stations for each region), together with proxy data (historical documents, ice cores, tree rings) for early years, (Wang et al., 1998) reconstructed a national temperature series (WYG series) from 1880 and reported a pronounced warm interval within the 1920s-1940s comparable to the 1990s. (Tang and Ren, 2005) developed another temperature series (TD series) representing the whole of China from 1873 based on instrumental data of monthly maximum and minimum temperature derived from 291 stations, evenly distributed, since 1951, together with corrected and interpolated data before 1950, which also showed the warm interval, but around the 1940s only. However, (Cao et al., 2013) and (Zhao et al., 2014) concluded that there was no warm interval before 1950 based on their annual mean series (T₁₃ series), which was derived from continuous and homogeneity-adjusted monthly temperature series for 18 stations in eastern and central China since the late 19th century. This dispute indicates that it is necessary to reconstruct regional temperature series from other data sources (such as high-resolution temperature proxies, e.g., phenological data, tree rings etc.) to compensate for the deficiency of instrumental data before 1950, because regular meteorological observations in most of China started in the 1950s. Therefore, in this paper, we report a multi-proxy reconstruction of annual mean (from November to the following October) temperature in southern China for the past 160 years.

Southern China is located in the southeast margin of the Eurasian continent (Fig. 1), characterized by subtropical and tropical humid climate. Figure 1 shows the study area, which was originally divided by (Wang et al., 1998) according to Chinese climate regionalization and the coherence of temperature changes. The instrumental gridded (resolution: 1^°× 1^°) dataset of monthly temperature (starting from 1951) in China (http://data.cma.cn/data/detail/dataCode/SEVP_CLI_ CHN_TEM_MON_GRID) is averaged to regional mean temperature anomalies (reference period: 1971-2000) for calibration. The information of all proxies used for reconstruction is presented in Table 1.

Figure 1.  Study area and locations of proxy data used for the annual temperature reconstruction. Bottom right: sub-regions divided by Wang et al. (1998), and southern China is highlighted by the gray area.

Our previous studies (Ding et al., 2015; Zheng et al., 2016) demonstrated that the southern limit of snowfall in the area, which is the latitude of the southernmost snowy site in winter, could well indicate winter temperature variability. This is because cold surges from the north are usually blocked by the Nanling Mountains (see Fig. 1) and only a strong cold front can cross the mountains to lead to snow in the south. Historical records of the southern limit of snowfall are from two sources. One is local gazettes (Zhang, 2013), which have special chapters to record abnormal weather, including the occurrence date, dekad, or month. The other is the archives of Yu (rainfall)-Xue (snowfall)-Fen (Chinese length units; 0.32 cm approximately)-Cun (3.2 cm approximately) during the Qing Dynasty (Ding et al., 2015), which comprise memos to the emperor by local officers in which snowfall events are noted with accurate dates. Modern data of the southern limit of snowfall are extracted from the Report of Monthly Surface Meteorological Records in China (Central Weather Bureau, 1975, 1978) and daily surface climate data (http://data.cma.cn/data/detail/dataCode/SURF_CLI_CHN_MUL_DAY).

Several studies have found that tree-ring chronology in this area is also a robust indicator of the temperature variability. For example (Table 1), tree growth is closely and positively related with late winter to spring temperature in low-altitude areas (<500 m; Chen et al., 2012), with early summer temperature in mid-altitude areas (around 1500 m; Cao et al., 2012), and with summer to early autumn temperature in high-altitude area (>3000 m; Wang, 2013).

Therefore, these four proxies (marked as XJ, CT, YA and HHS, see Table 1 for detail) contain the temperature signal for all four seasons, which makes it possible to synthesize the annual temperature. Then, the mean temperature from November to the following October is set as the dependent variable for the target reconstruction. It is also worth noting that all the tree-ring sample sites are located in mountain areas; thus, the proxy-reconstructed regional temperature series should not be affected by urbanization.

The dating and measurement errors for all the tree-ring width chronologies used are checked by the computer program COFECHA, and all chronologies are developed using the program ARSTAN. The methods for removing the age-related trend for tree-ring width series are "best fitting by negative exponential curve" or "linear trend with negative slope" (Cao et al., 2012; Chen et al., 2012). Considering different detrending processes, the tree-ring chronologies' variance contribution in the low-frequency climate signal may be reduced by statistical function fitting or enlarged by regional curve standardization (Yang et al., 2011); thus, we use a signal decomposition and synthesis method to calibrate these deviations in frequency spectrum (Zheng et al., 2015).

The approach used for annual temperature reconstruction in this study is "high and low frequency signal decomposition and synthesis", which has been used for reconstructing temperature series in Xinjiang, Northwest China. This method is not only able to successfully capture the decadal temperature variation and long-term trends of change from multiple proxies, but it can also preserve the annual temperature variability (Zheng et al., 2015). Firstly, both proxy and instrumental series are decomposed into high and low parts by the frequency of 1/T (T is the number of years for the FFT filter), to build a good relationship having higher variance explanation between the temperature and the proxy data. Then, the correlation coefficients (Fig. 2) between each proxy and annual temperature for different frequencies (the original series, first-order difference and low-pass smoothing by 1/T and T varied from 10 yr to 32 yr, in 2-yr steps, where 32 yr is the longest complete cycle that the calibration series could detect) show that all proxies are significantly correlated with annual temperature changes. While the XJ and YL series contain both high- and low-frequency signals of temperature variations, CT contains low-frequency and HHS contains high-frequency signals only. Therefore, HHS, YL and XJ are selected as the independent variables to build the high-frequency regression; and CT, YL and XJ are selected as the independent variables to build the low-frequency regression. The sum of the predicted R² values for high- and low-frequency regressions reaches a maximum when T=28 (Table 2), and the phase of the low-frequency reconstruction is completely coherent with the calibration series, which proves the model is stable at this decomposition. Finally, we choose T=28 to build the high- and low-frequency calibration equations for temperature synthesis reconstruction, which captures 65.8% of the variance of observations during 1952-2009 (Figs. 3a and b). In addition, Fig. 3 also illustrates the other temperature series, which include the previous reconstruction of annual temperature in southern China by (Wang et al., 1998), the observations at four sites (Hong Kong, Macao, Guangzhou and Fuzhou) in southern China after homogenization (Cao et al., 2013), and the LSAT of China (EC-TCNARCC, 2015; CMA, 2016) and the Northern Hemisphere (NH) (Jones et al., 2012; Stocker et al., 2013), for the purpose of comparison.

Figure 2.  Correlation coefficients between the regional temperature anomalies and each proxy series for different frequencies ("F" for first-order difference; "O" for original series; 10-32 for low-pass smoothing by 1/T, with T from 10-32 yr). To test the significance level of the relationship of FFT smoothing between proxy and temperature, the effective number of degrees of freedom for the correlation is calculated using the Monte Carlo method, as suggested by (Yan et al., 2004).

The multi-proxy temperature reconstruction (Fig. 3b) shows a robust centennial warming in southern China, with a linear trend of 0.47^°C (100 yr)^-1 from 1871 to 2009, after a rapid cooling within the 1860s. This warming trend is similar to that in China and the NH, with a start year of 1880 and 1875, respectively (Table 3). Compared with the previous reconstruction by (Wang et al., 1998), whose result is a linear trend of 0.09^°C (100 yr)^-1 over 1880-1996, our reconstruction shows a more evident warming trend, with a rate of increase of 0.22^°C (100 yr)^-1 for the same period. This is mainly because our reconstruction combines multi-proxy-based temperature variability from winter to the growing season, which captures the decadal temperature variation and long-term trends by low- and high-frequency signal decomposition and synthesis. However, compared with the warming rate shown from temperature observations [e.g., 1.26^°C (100 yr)^-1 during 1885-2010 in Hong Kong——the longest observation in southern China], our reconstruction underestimates the centennial linear trend. This might partly result from the proxy-based reconstruction only explaining 65.8% of the variance of temperature observations, and partly because the trend for the regional mean is lower than that for individual stations in cities affected by urbanization. For example, the trend of regional mean temperature observations for calibration in 1952-2010 is 0.148^°C (10 yr)^-1, while it is 0.191^°C (10 yr)^-1 for the mean temperature at the four stations of Hong Kong, Macao, Guangzhou and Fuzhou.

Figure 3.  Reconstruction of annual temperature anomalies (reference period: 1971-2000; the same for the other series) with 95% confidence intervals in southern China during 1850-2009, and comparison with other datasets: (a) comparison of the reconstruction and observations for 1952-2009; (b) annual (November-October) temperature reconstructed from multi-proxy data; (c) temperature reconstruction by Wang et al. (1998); (d) temperature observations at Hong Kong, Macao, Guangzhou and Fuzhou (Cao et al., 2013); (e) China LSAT (EC-TCNARCC, 2015; CMA, 2016); (f) NH LSAT (Jones et al., 2012; IPCC, 2013). Top labels for (a), bottom labels for (b) to (f).

Figure 3b also shows two rapid warming intervals. The first one starts from the late 1870s, and the warming rate is 0.125^°C (10 yr)^-1 during 1877-1938 (Table 3), in which the most rapid warming occurs during 1892-1916 at a rate of 0.308^°C (10 yr)^-1. This warming rate is similar to that from the reconstruction [0.164^°C (10 yr)^-1 during 1884-1946] of Wang et al. (1998) and the observations [0.114^°C (10 yr)^-1 during 1885-1938] at Hong Kong. Meanwhile, the warming rate is 0.146^°C (10 yr)^-1 during 1880-1946 for the whole of China, and 0.077^°C (10 yr)^-1 during 1880-1946 for the NH (Table 3). It is worth noting that our reconstruction shows the warmest year in this interval occurring in 1938, which is consistent with that shown in the mean observations of the four stations in southern China, as well as the NH. This rapid warming causes a moderate warm interval in the 1910s-1930s that is only a little higher than the mean of 1971-2000 in southern China. The second rapid warming occurs around 1970 onwards, and the warming rate is 0.258^°C (10 yr)^-1 during 1968-2007 (Table 3), with the most rapid warming in 1983-2002 at a rate of 0.512^°C (10 yr)^-1. In the same 40-yr period, the warming rate is 0.359^°C (10 yr)^-1 for the mean of Hong Kong, Macao, Guangzhou and Fuzhou observations, 0.413^°C (10 yr)^-1 for the whole of China, and 0.339^°C (10 yr)^-1 for the NH. The temperature during the 1980s-2000s is about 0.3^°C higher than the mean of 1971-2000 in southern China, which is consistent with that in China and the NH (Figs. 3e and f). These results suggest that the warming since the 1980s is unprecedented for the past 160 years.

The interdecadal variation of the temperature reconstruction also shows there to be two cooling intervals in southern China since 1850. The first one starts from the 1860s and is then followed by a cold interval until the early 1890s. The two coldest years are 1892 and 1893, whose temperature anomalies both reach about -1.0^°C. Specifically, the interannual variation of the reconstruction is the same as that in Hong Kong's observations, and similar to those for the whole of China and the NH. The second cooling starts from the 1940s and ends around 1970, but with a very low cooling rate. Besides, the temperature reconstruction shows a warming hiatus in the 2000s.

In addition, Table 4 gives the variance contributions of different frequency domains during the calibration period for the temperature observations and reconstructions based on ensemble empirical mode decomposition (Wu and Huang, 2009). The results show that the variance contributions of different frequency domains for our multi-proxy-based temperature reconstruction is closer to those for temperature observations (Hong Kong and regional mean) than previous reconstructions from tree-ring proxies only. Consequently, our reconstruction through signal decomposition and synthesis provides a more reliable trend estimation.

Based on the southern limit of snowfall, chronologies of tree-ring width and tree-ring stable oxygen isotope (\(\updelta^{18}\)O), a new annual temperature anomaly series for southern China during 1850-2009 is reconstructed using the method of signal decomposition and synthesis. This multi-proxy-based temperature reconstruction shows robust centennial warming, with a linear trend of 0.47^°C (100 yr)^-1 during 1871-2009. Moreover, on the decadal scale, it shows the first rapid cooling as having started from the 1860s, followed by a cold interval until the early 1890s, with the coldest years being 1892 and 1893. The first significant warming is from 1877 to 1938 [0.125^°C (10 yr)^-1], with the most rapid rate of increase being 0.308^°C (10 yr)^-1 during 1892-1916, resulting in a moderate warm interval during the 1910s-1930s. Then, a slightly temperature decline is apparent from the 1940s to the late 1960s. Another significant increase in temperature is shown to start around 1970 [0.258^°C (10 yr)^-1 during 1968-2007], with the highest rate being 0.512^°C (10 yr)^-1 during 1983-2002, though a warming hiatus occurs in the 2000s. Compared with the warm interval in the 1910s-1930s, the temperature in the 1980s-2000s is much higher. These results reveal that both the level of warmth and the warming rate from the 1980s are unprecedented since 1850. In addition, this multi-proxy-based temperature reconstruction provides an independent case to validate the global warming for the past 160 years and its hiatus recently, because the proxy data are not affected by urbanization.