Instrumental temperature series in eastern and central China back to the nineteenth century

Authors


Abstract

[1] In this study, we bring together different source data sets and use quality control, interpolation, and homogeneity methods to construct a set of homogenized monthly mean surface air temperature (SAT) series for 18 stations in eastern and central China from the late nineteenth century. Missing values are statistically interpolated, and cross validation is used to assess the accuracy of the interpolation approaches. Results show that the errors of interpolation are small, and the interpolated values are statistically acceptable. Multiple homogeneity methods and all available metadata are used to assess the consistency of the time series and then to develop adjustments when necessary. Thirty-three homogeneity breakpoints are detected in the 18 stations, and the time series is adjusted to the latest segment of the data series. The adjusted annual mean SAT generally shows a range of trends of 1.0° to 4.2°C/100 years in northeastern and southeastern China and a range of trends of −0.3° to 1.0°C/100 years in central China near 30°N. Compared to the adjusted time series, the unadjusted time series underestimates the warming trend during the past 100 years. The regional and annual mean SAT over eastern and central China agrees well with estimates from a much denser station network over this region of China since 1951 and shows a warming trend of 1.52°C/100 years during 1909–2010.

1 Introduction

[2] The science of climate change has paid much attention to the inter-decadal variability and long-term trends of surface air temperature (SAT). Both long-term and homogeneous instrumental SAT data are essential for studying the characteristics of global and regional climate change. Over past decades, the establishment of long-time SAT series has attracted extensive attention worldwide and great progress has been achieved [Jones et al., 1999].

[3] Some well-known data sets of SAT have been constructed and extensively used to study climate changes. For example, Vose et al. [1992] constructed the Global Historical Climatology Network data set, which was further improved after rigorous quality control and homogenization [Peterson and Vose, 1997; Lawrimore et al., 2011]. Jones and Moberg [2003], Brohan et al. [2006], and Jones et al. [2012] constructed different versions of the global land surface air temperature data set, with the latest version being referred to as the Climatic Research Unit (CRU) version 4. In addition, Hansen et al. [2001, 2010] also established the Goddard Institute for Space Studies surface temperature data set. Based on these data sets, the Fourth Assessment Report of Intergovernmental Panel on Climate Change [Solomon et al., 2007] showed a warming trend of global mean SAT (land and oceans) since 1906, with a trend of 0.74°C over 1906–2005.

[4] The collection, compilation, and construction of long-term instrumental SAT data in China have been on-going processes over recent decades. Collections of the early observational temperature in China were first made in the 1980s and the 1990s [e.g., Wang, 1990; Tao et al., 1991; Tang and Lin, 1992; Lin et al., 1995; Wang et al., 1998]. These studies addressed the sparse early instrumental observations in China before 1950 including the number of missing records during the 1940s. One important fact they noticed was that no consistent observational schedules were followed at some stations. For example, Beijing time was used at most stations, while the local time or world time was also used at some stations during certain periods. Although a national surface meteorological observation network with over 2000 stations had been established in China from 1950, the sites of approximately 80% of these stations have been relocated at least once. One of the reasons for station relocations is that growth and expansion of cities has occurred, particularly across Eastern China over the past three decades, sometimes causing several moves. These changes in station positions and thermometers will probably cause significant inhomogeneities in the long-term SAT series and potentially add uncertainty to the estimation of trends of SAT in China [Yan et al., 2001]. Both the missing records and the inhomogeneous data can hinder the construction of homogeneous SAT time series for China, and thus China often appears as a data gap (e.g., when looking at trends in extreme temperatures [Alexander et al., 2006]).

[5] To obtain continuous series of China SAT during the past 100 years, some authors infilled the series and performed quality assessments of various kinds. For example, utilizing daily mean, maximum, and minimum temperatures, together with some documentary data in early years, Wang [1990] and Tang and Lin [1992] constructed several long-term series of SAT developing a regional mean SAT for China. Lin et al. [1995] similarly constructed a national time series of SAT for 1873–1995, referred to as the LYT series. Wang et al. [1998] further developed their earlier work and included some proxy data from historical documentary records, ice core records, and tree ring records principally over western China to construct a national average SAT time series for China during 1880–2008, referred to here as the WYG series. Meanwhile, linear regression, stepwise regression, and partial least squares regression have been applied to interpolate missing station values of monthly or annual mean SAT across China [e.g., Zhang and Sun, 1996; Jiang et al., 1999; Huang et al., 2004; Zhang et al., 2006; Li et al., 2008]. Using average values of the daily maximum and minimum temperatures to calculate a monthly mean SAT at more than 600 stations of China, Tang and Ren [2005] constructed a national mean SAT series for China during 1905–2001, referred to here as the TR series. Afterward, Tang et al. [2009] performed further quality control with some corrections to erroneous records before 1950 and chose 291 stations to reconstruct a national mean SAT series for China during 1873–2008, referred to as the TD series. In this series, however, the homogeneity of the time series was not discussed and only a limited number of series extended back before 1940. Using the homogeneity assessment technique of two-phase regression (TPR), Li et al. [2010] constructed time series of station SAT during 1873–2004, referred to as the LD series, in which the missing data before 1951 were not interpolated.

[6] Based on these various reconstructions of Chinese SAT, long-term changes in the past century have been examined. However, these studies indicate a wide divergence in linear trends of nationally-averaged SAT. For example, the linear trend of annual mean SAT for China during 1906–2005 (1908–2007) is 0.34°C/100 years (0.42°C/100 years) for the LYT series, 0.53°C/100 years (0.59°C/100 years) for the WYG series, 0.86°C/100 years (0.96°C/100 years) for the TD series, and 0.95°C/100 years (1.11°C/100 years) for the TR series [Tang et al., 2009]. The linear trend of the LD series is approximately 0.9°C/100 year during 1900–2006 [Li et al., 2010], referred to as the LQX series. This great range (from 0.30°C/100 years to 1.11°C/100 years) principally results from differences in the period before 1951 [Tang et al., 2009], where there is greater uncertainty in the series and potential flaws in some of the analysis methods used to develop the long-term SAT series. Also, all the previous studies did not provide much detail about the methods used to construct their long-term series for China SAT. Therefore, it is necessary to further reconstruct the time series of China SAT over the past 100 years.

[7] We conduct this study to establish in an open and objective way a new set of homogenized monthly mean SAT series in eastern and central China back to the nineteenth century by using quality control techniques, interpolation of missing records, and homogeneity assessment of the long-term series. Finally, the derived time series is used to analyze the characteristics of SAT climate changes over central and eastern China during 1909–2010.

2 Data

[8] In the present study, the originally observed SAT records and the associated metadata before 1951 come from two data sources. One is from the early work of “Two climatic data bases of long-term instrumental records of Chinese Academy of Sciences (CAS) in the People's Republic of China” [Tao et al., 1991]. Another is from the long-term series of SAT and metadata developed by the National Meteorological Information Center of the China Meteorological Administration (CMA) since 2002, including the newly digitalized long-term instrumental data set and metadata for 60 stations in the large cities of China. The metadata, archived at CMA, includes information on changes in observational instrumentation, times, and locations for each station. The originally observed records of SAT and the associated metadata during 1951–2010 come from 825 meteorological stations across China archived at CMA. Instrumental temperature records for Hong Kong and Macao come respectively from the homepages of the Hong Kong Observatory and the Macao Meteorological and Geophysical Bureau who collected the data.

[9] Additionally, the monthly mean SAT data of the CRU analysis with a horizontal resolution of 0.5° in both longitude and latitude (CRU TS3.10) during 1901–2011 [Harris et al., 2013] are interpolated to the locations of meteorological stations to compare with the constructed long-term SAT series from stations across China. A link to the stations used in CRU TS 3.10 is provided by Harris et al. [2013].

3 Compilation of Station Data and Interpolation of Missing Records

[10] Using the various observed data sets of SAT, we perform quality control to determine the reliability of each station SAT series and then consider each obviously erroneous value as a missing value. There are 2 months that have been set to missing values (as they are clearly erroneous) and these occurred at Hohehot in June 1924 and at Shenyang in September 1943. For some stations, the time series of monthly mean SAT from CAS and CMA is of different lengths. For example, at Tianjin, the time series from CAS is shorter than that of CMA (Figure 1a). However, at Fuzhou, the time series from CAS is longer than that from CMA. For these two cases, the data from the two sources before 1951 are joined. To avoid the inconsistency of calculating a daily mean SAT from different observation times each day, we follow Tang and Ren [2005] and define the daily mean temperature (TM) as the arithmetic mean of daily maximum and minimum temperatures (THL), rather than an arithmetic mean of multiple observations of temperature (TMT) during each day. However, the daily maximum and/or minimum SAT is occasionally missing at one station for certain periods before 1950 because the maximum and/or minimum SAT was not kept in the archive while the TMT was kept. In this case, we estimate the missing THL value from the available TMT series, using a linear regression to represent the relationship between the THL and TMT, and use an F test to determine the statistical significance of the relationship.

Figure 1.

Time series of monthly mean SAT (°C) at (a) Tianjin and (b) Fuzhou from the CAS (blue) and CMA (red) data sources.

[11] Finally, our study indicates that there are 468 meteorological stations across China (shown in Figure 2a) with some SAT records before 1951. These stations are mainly located to the east of 100°E. Of these 468, there are 152 stations with observational records with 10 or more years of data before 1951 (Figure 2b). After assessing the quality, reliability, and length of each SAT time series, we finally chose 18 stations in the eastern and central parts of China (shown in Figure 2c) with long records (more than 1000 months) and little missing data (less than 150 months) from the start of observations to 2010 to construct a continuous time series of monthly mean SAT. In the present study, we do not chose stations in western China because of large percentages of missing values.

Figure 2.

Distributions of (a) 468, (b) 152, and (c) 18 stations. In Figure 2c, the number corresponds to station names in Table 1.

Table 1. Missing Percentages of Records and the Number of Breakpoints at 18 Stations From Their Starting Dates to 2010
NumberStation IndexStation NameStarting Time (Month/Year)Length (Months)Missing MonthsMissing Percentage (%)Number of Homogeneity Breakpoints
Before 1950After 1951
145005Hong Kong01/18841524845.5100
245011Macao01/1901132000.0002
350527Hailar01/1909122412610.2900
450953Harbin01/19091224655.3103
553463Hohehot01/19151152554.7700
653772Taiyuan01/19161140554.8302
754342Shenyang05/19051268544.2604
854511Beijing09/189014441319.0704
954527Tianjin09/1890144400.0020
1054857Qingdao01/1900133213410.0600
1156778Kunming01/1921108000.0020
1257494Wuhan02/19051271776.0604
1357679Changsha01/19111200847.0001
1457816Guiyang10/1920108300.0001
1558238Nanjing01/19051272987.7002
1658367Shanghai01/1873165600.0001
1758847Fuzhou01/19051272322.5210
1859287Guangzhou03/19121186736.1604

[12] Table 1 shows the percentage of missing monthly mean SAT at the 18 stations. It is seen that there are five stations with continuous monthly mean SAT from their starting dates to 2010. These are Macao, Tianjin, Kunming, Guiyang, and Shanghai. There are 11 stations with missing percentages of 2.52% to 9.07% and there are two stations (Qingdao and Hailar) with missing percentages of 10.06% and 10.29%. Further analysis shows that these missing records mainly occur in the earlier period before 1950 except for a few missing values at Qingdao, Hohehot, and Taiyuan after 1951 (see Table 2). We note that it is often acceptable that a long-term series has a few missing months/years because some statistical calculation processes (e.g., national/regional mean) will likely reduce the influence of these missing records on the entire time series. However, in the present study, we select only 18 stations and there are a number of missing records at more than half stations during wartime in the 1940s, which will likely exert a greater influence on regional mean series. Thus, it is necessary to infill the missing values. To help with this, some short records at neighboring stations are available during the 1940s. In the following section, therefore, we shall first infill those missing records to have a near-complete number of stations during the study period.

Table 2. Error Estimation of Cross Validation of Interpolation Methods at 18 Stations From Their Starting Dates to 2010
Station No.Station NameMissing Time of Record (Month/Year)Missing Months in RecordInterpolation MethodError
MBE(°C)RMSE(°C)
45005Hong Kong01/1940–12/194684Integrated method0.080.555
50527Hailar09/1929–10/192913Integrated method0.000.951
01/1933–02/1933
04/1933–12/1933
11/1929–03/193072Gradient plus inverse distance square−0.281.134
03/1933
01/1934–12/1934
05/1941
05/1942–07/1945
01/1949–02/1950
08/1945–12/194841Relaxed integrated method0.001.414
50953Harbin10/1942–12/194218Integrated method0.120.843
02/1947–04/1948
08/194335Gradient plus inverse distance square0.450.879
12/1943–12/1945
01/1947
05/1948–12/1948
01/1946–12/194612Relaxed integrated method0.131.192
53463Hohehot07/1937–08/19374Integrated method0.581.226
01/1951
01/1953
06/192451Relaxed integrated method−0.171.554
09/1937–12/1938
01/1944–10/1946
53772Taiyuan10/1937–12/193851Integrated method−0.010.796
01/1944–10/1946
01/1947–02/1947
08/1955
12/1955
02/1949–05/19494Relaxed integrated method0.001.049
54342Shenyang05/194142Integrated method0.020.497
05/1942–12/1942
08/1943–09/1943
12/1943–12/1945
07/1948–12/1948
01/1946–12/194612Relaxed integrated method0.090.998
54511Beijing09/1890–12/1890131Integrated method−0.060.713
03/1900–04/1903
01/1904–12/1904
01/1909–12/1909
01/1912–12/1913
01/1915–03/1915
12/1926–06/1927
10/1927
01/1928–12/1928
02/1929–06/1929
09/1937–12/1938
54857Qingdao07/1914–03/1915134Integrated method0.110.555
09/1937–01/1938
01/1951–12/1960
57494Wuhan05/193849Integrated method−0.080.608
07/1938–10/1938
02/1941–04/1941
06/1941–12/1941
09/1943–10/1943
12/1943–07/1944
11/1944–01/1945
04/1946–12/1946
11/1938–12/193828Gradient plus inverse distance square0.140.691
01/1943–08/1943
11/1943
08/1944–10/1944
02/1945–03/1946
57679Changsha01/192362Integrated method0.000.524
11/1938–12/1939
01/1941–05/1944
04/1949–09/1949
06/1944–03/194622Gradient plus inverse distance square0.090.520
58238Nanjing11/191198Integrated method0.000.691
12/1937–12/1945
58847Fuzhou04/1941–11/194132Integrated method−0.020.933
07/1942–09/1942
04/1944–12/1945
59287Guangzhou07/192273Integrated method0.010.579
11/1938–12/1939
01/1943–12/1946
07/1947
10/1947–01/1948
06/1948–09/1948
10/1949–12/1949

[13] For the 13 stations with missing values, we develop an interpolation technique for the monthly mean SAT anomaly from the climatological mean by the following three steps that include different statistical methods. For the first step, an integrated method is designed by considering the neighboring reference stations and it includes three independent statistical approaches, that is, the standardized method [Steurer, 1985; Yu et al., 2012], partial least squares regression, and multivariate linear regression. Here, the reference stations are chosen mainly from those of 152 stations (shown in Figure 2b) which are within a 300 km distance from the candidate station and have data for more than 10 years before 1951. Meanwhile, the station elevation of the neighboring station (HR) must be close to that of the candidate station (H0), that is,

display math

[14] The temperature series at the reference stations is highly correlated with that at the candidate station. The three approaches are respectively applied to interpolate one missing value at a candidate station. Because the standardized method, the partial least squares regression, and the multivariate linear regression method may have different requirements for selecting reference stations, it is possible that one missing value at one station is interpolated by one, two, or all three methods. When this missing value can be interpolated by all three approaches, the median value of the three interpolated values is chosen as the final estimate; when the missing value is interpolated by two of the three approaches, the mean value of the two interpolated values is chosen as the final estimate; when the missing value is only interpolated by one of the three approaches, this interpolated value is chosen as the estimate.

[15] For those missing values which cannot be interpolated in the first-step interpolation, we additionally perform the second-step interpolation. For this step, we use a gradient plus inverse distance square method using neighboring reference stations, in which the weighting is inversely proportional to the square of the distance between the candidate station and its neighboring stations. Here, the neighboring stations are chosen from those of the 468 stations (Figure 2a) which are within a 300 km distance from the candidate station. The station elevation condition of the neighboring stations remains the same as the first step.

[16] Although some missing values are further interpolated using the second step, there were still some missing values before 1951 after this second step. For those remaining missing values, we perform the third step of interpolation. For this step, the interpolation method is the same as that of the first step except for a more relaxed distance/elevation threshold between the reference and candidate stations (Table 2). For example, at Taiyuan, Hohehot, Shenyang, Harbin, and Hailar stations, reference stations within a distance threshold of 400, 500, 700, 1000, and 1100 km away from the candidate station and an elevation difference threshold of 900, 1000, 100, 200, and 650 m are chosen in turn. The remaining missing monthly values after the second interpolation step are infilled after this third interpolation step.

[17] For all the interpolation methods, cross validation is used to assess the reliability of the interpolation methods. Here, cross validation is performed comparing interpolated estimates to the actual data values at the candidate station in months when the candidate station had data and the full period of available data could be used to estimate errors. Table 2 shows the mean bias error (MBE) and root mean square error (RMSE) of the cross validation for each interpolation step. It is seen from this table that MBE is generally small and falls between −0.28°C and 0.58°C, in which the absolute value of MBE beyond 0.2°C appears only at Hailar (on the second step), Harbin (on the second step), and Hohehot (on the first step). RMSE is smaller (below 1°C) at nine stations (Hong Kong, Shenyang, Beijing, Qingdao, Wuhan, Changsha, Nanjing, Fuzhou, and Guangzhou) for each interpolation step and is also generally below 1°C at Hailar, Harbin, and Taiyuan for the first and second interpolation steps. These results suggest that the interpolation errors are small. A relatively large RMSE (between 1.2°C and 1.6°C) mainly appears at three stations (Hailar, Harbin, and Hohehot) for 108 month interpolated values, accounting for 10.1% of the total interpolated values.

[18] Using the three interpolation steps, we have carried out an interpolation of 1068 monthly values at 13 stations, in which the first step interpolates 791 missing monthly values, the second step interpolates 157 missing monthly values, and the third step interpolates the remaining 120 missing monthly values. Finally, the continuous time series of monthly mean SAT from their start of observations to 2010 (generally at least or almost 100 years in length) is obtained for all 18 stations. Figure 3a shows the time series of monthly SAT at Nanjing that had long continuous missing records during 1938–1945. In the first interpolation step, the neighboring reference stations of Nanjing are Shanghai (that had no any missing values during this period), Xuzhou, and Dafeng. The results from cross validation show a very small MBE value and an RMSE value of 0.69°C, indicating a good interpolation. Figure 3b shows the time series of SAT at Hailar that has the largest RMSE for the third interpolation step from August 1945 to December 1948. At this station, 126 missing monthly values are interpolated using all three interpolation steps (see Table 2). Moreover, we also note that no extreme values of SAT are obtained from our interpolation, which is expected from any form of interpolation.

Figure 3.

Time series of monthly mean SAT (°C) at (a) Nanjing and (b) Hailar stations, in which the blue lines show the interpolated values.

4 Homogenization of SAT Time Series

[19] The inhomogeneity of a SAT time series may be a gradual trend and may also be a sudden change (breakpoint). A homogeneous time series is necessary when studying climate changes as this should exclude the influences of nonclimatic factors, such as changes in observational locations, times and instruments. A variety of homogenization methods [see Venema et al., 2012] have already been developed according to different climate elements or timescales assessed. It is useful to select specific homogeneity and adjustment methods for different elements. In the following section, we apply homogenization methods together with available metadata to perform the inhomogeneity detection and adjustment of the monthly temperature time series. However, the present work (actually most of current methods) principally deals only with breakpoints due to sudden changes in observing system such as relocation or “sudden” developments near the observing site (e.g., a new building) and so on. Consequently, gradual trend biases, possibly due to an enhancing urban heat island effect during recent decades for some studied sites here, might remain in the homogenized series, which will be discussed in a later paper.

4.1 Homogeneity Methods

[20] After comparatively analyzing the performance of several homogenization methods for time series of SAT across China [Cao and Yan, 2012], we use the RHtest version 3 software package [Wang and Feng, 2010] to detect and adjust inhomogeneities in this study. This package includes the PMTred algorithm [Wang, 2008a], which is based on the penalized maximum t (PMT) test [Wang et al., 2007], and the PMFred algorithm [Wang, 2008a], which is based on the penalized maximum F (PMF) test [Wang, 2008a, 2008b]. Both algorithms deal with multiple change points using a recursive testing algorithm and account for the first order autocorrelation, and thus have the word “red” in their names [Wang, 2008a]. The PMT test is a relative homogeneity test and thus must be used with a reference series. The PMF test can be used without a reference series. Cao and Yan [2012] and Li et al. [2013] showed that the PMTred algorithm is suitable for detecting multiple change points with a reference series representing a regional temperature series when the observational network is dense, while the PMFred algorithm is often applied to a sparse observational network where a reference series is difficult to develop. Therefore, in the present study, we use both PMFred and PMTred algorithms before and after 1950, respectively, with a statistical test conducted at the 1% significance level.

[21] For the time series of monthly mean SAT at the 18 stations before 1950, we use the PMFred algorithm without a reference series because the observational stations were sparse at that time. To increase the reliability of the PMFred method in detecting change points, we repeat the above work using the two-phase regression (TPR) method [Easterling and Peterson, 1995] and a running Student's t test. When one change point is detected by at least two of three methods (PMFred, TPR, and the running t test) and this discontinuity can also be detected for at least five continuous months or is supported by the metadata, it is accepted as a break point. When the change points detected by two methods occur close together (within 2 years), these two change points are considered as one change point and its occurrence time is determined by available metadata or is set to the first occurring year for no metadata.

[22] For the time series of monthly mean SAT at the 18 stations during 1951–2010, the PMTred algorithm considering a reference series is used to detect change points. Each reference series (for the 18 stations) is constructed on the basis of stations with continuous and homogeneous series (hereafter the reference stations), in which the reference stations are selected from 825 meteorological stations across China according to the following rules: (1) not further away from the tested station than 350 km, and (2) when HT < 2500 m, |HR-HT| ≤ 200 m; when HT ≥ 2500 m, |HR-HT| ≤ 500 m. Here, HT is the elevation of the tested station and HR is the elevation of the reference station. Additionally, the homogeneity of the time series at the reference station is first checked using the PMFred algorithm, and the reference series is highly correlated with the tested station. For each tested station when there are more than two reference stations, their arithmetic average is defined as a reference series.

4.2 Results of Homogeneous Detection and Adjustment

[23] Table 1 shows the detected break points at all 18 stations. It is seen from this table that there are no break points at four stations (Hong Kong, Hailar, Hohehot, and Qingdao), but there are 33 break points at the other 14 stations. Before 1951, there are five break points. Figure 4a further shows the time series for the monthly mean SAT anomaly at Kunming from January 1921 to December 1950. In this figure, the anomaly of SAT generally varied between 0°C and 4°C during 1930–1937, but rapidly decreased in May 1938, and generally varied in the 1°C to −3°C range in January 1945. The metadata information indicates that this sudden decrease is closely associated with a move from Yide station (with a station elevation of 1922 m) in Kunming to the Taihua Mountain station (with a station elevation of 2280 m) in Kunming in May 1938. This observation station then moved to Kunming airport (with a station elevation of 1906 m) in January 1946, which led to a rapid increase of the observed monthly mean SAT. Both of the significant change points are detected by all three methods (PMFred, TPR, and the running t test). Another break point, this time at Fuzhou, occurring in December 1923 (Figure 4b), is also detected by the three methods and is supported by the metadata information, which is associated with a move of the observation site. Figure 4c shows the time series of monthly mean SAT at Tianjin from September 1890 to December 1950. In the figure, SAT shows two break points, with a sudden increase of 1.6°C in March 1897 and a sudden decrease of 2.4°C in June 1907. These two break points are detected by the three methods and subsequently last for at least 5 months. However, no available metadata information supports these break points at Tianjin. In spite of this, we still accept them as change points.

Figure 4.

Time series of monthly mean SAT anomalies (°C) from the climatological mean at (a) Kunming, (b) Fuzhou, and (c) Tianjin. The red line is the linear trend for each period.

[24] From 1951, the PMTred method, conducted with a reference series, detects 28 break points at 11 stations (Macao, Harbin, Taiyuan, Shenyang, Beijing, Wuhan, Changsha, Guiyang, Nanjing, Shanghai, and Guangzhou) (see Table 1). According to the related metadata, 21 of these change points are due to changes in the station location, and three of them are related to instrument changes. The remaining four change points do not have metadata support. These four change points occur at Harbin in December 1995, at Shenyang in December 1975, and at Guangzhou in November 1987 and April 2001. It is possible that some changes of the local site conditions contribute to the four change points. They all are statistically significant change points, even though undocumented, and are thus retained for adjustments to homogenize the data time series.

[25] Each data time series that is identified to contain significant change point(s) is then adjusted to the latest segment of the data series, using the mean-adjustments in the output of RHtestsV3. Finally, we obtain the homogenized time series of monthly mean SAT at each of the 18 stations. In the following section, we analyze their long-term variations.

5 Climate Change of SAT in Central and Eastern China

[26] Figure 5a shows the linear trend of annual mean SAT without homogeneity adjustment at the 18 stations from 1901 or the start of observations to 2010. It is seen that the unadjusted SAT shows large increasing trends of 3.4°C/100 years, 2.3°C/100 years, 1.6°C/100 years, and 1.5°C/100 years at four stations (Hailar, Harbin, Shenyang, and Shanghai), a range of trends of 0.5° to 1.5°C/100 years at eight stations (Hohehot, Taiyuan, Beijing, Qingdao, Nanjing, Kunming, Guangzhou, and Hong Kong), and a range of smaller trends of ±0.5°C at six stations (Tianjin, Wuhan, Changsha, Guiyang, Fuzhou, and Macao). Compared to Figure 5a, the annual mean SAT after the homogeneity adjustment shows a larger trend at 11 stations (Harbin, Taiyuan, Shenyang, Beijing, Tianjin, Wuhan, Guiyang, Fuzhou, Kunming, Guangzhou, and Macao), with warming trends between 0.1° and 1.9°C/100 years (less than 1°C/100 year at seven stations), and negative trends at three stations (Changsha, Nanjing, and Shanghai), −0.7°C, −1°C, and −0.7°C/100 years, respectively (Figure 5b). The adjusted trends are larger at 3.4°C/100 years, 4.2°C/100 years, and 2.8°C/100 years at Hailar, Harbin, and Shenyang, showing more consistent linear trends. This result shows that the inhomogeneous time series appears to underestimate warming trends during the approximately 100 year period. A similar result has also been noted in many previous studies. These studies showed that when stations move from a city center locations to out-of-town locations, this normally makes SAT cooler [Li and Yan, 2009]. This phenomenon also occurred in central and eastern China. Figure 5a also presents a spatially less coherent pattern of the trends than Figure 5b. The adjusted annual mean SAT generally shows a range of trends between 1.5° and 4.2°C/100 years in northern China (Hailer, Harbin, Shenyang, Beijing, and Taiyuan), a range of trends between 1.0° and 3.3°C/100 years in southeastern China (Fuzhou, Kunming, Guangzhou, Hong Kong, and Macao), and a range of trends between −0.3° and 1.0°C/100 years in central and eastern China near 30°N (Guiyang, Changsha, Wuhan, Nanjing, and Shanghai).

Figure 5.

(a) Linear trends (°C/year) of unadjusted annual mean SAT at Hong Kong, Macao, Beijing, Tianjin, Qingdao, and Shanghai stations from 1901–2010 and from the start of observation to 2010 for the other 12 stations. (b) Same as Figure 5a but for the adjusted data. (c) Same as Figure 5a but for the CRU TS 3.10 data.

[27] Because Kunming and Guiyang stations are located in the southeastern part of the Tibetan Plateau, with elevations exceeding 1000 m, and their climate characteristics are different from those of other stations in eastern and central China, we analyze the regional temperature change by using only 16 stations without Kunming and Guiyang. It is evident from Figure 6a that only four stations had records before 1900. After this year, the number of station with records increased rapidly, reaching 12 stations in 1909 and 16 stations in 1916. Thus, the time series during 1909–2010 is used in the following analysis, giving a length of 102 years.

Figure 6.

(a) Temporal variations of station numbers during 1873–2010. (b) Same as in Figure 6a but for eastern China (green) and all of China (blue) during 1951–2010.

[28] To calculate the annual mean values of the adjusted and unadjusted SAT anomalies averaged over 16 stations, Beijing and Tianjin (Hong Kong, Macao and Guangzhou) are first averaged together because they are located close together, and then the large-scale average is calculated by a simple arithmetic average of the available 13 series. With this average, the adjusted and unadjusted SAT values for the 13 locations are referred to as TA13 and TUA13, respectively. Figure 7a shows time series of TA13 and TUA13. In this figure, TA13 and TUA13 time series show similar features and there is a high correlation of 0.99 between TA13 and TUA13 during 1909–2010. The series showed weak trends before the late 1960s, but after this rapid warming occurred. Trends for these series for various periods are given in Table 3. For the period 1951–2010, the linear trend of TA13 is 2.80°C/100 years, slightly smaller than that of TUA13 (2.97°C/100 years), with rates since 1909 less, both overall and for the earlier period of 1909–1950. These results show that the effects of the homogenization on linear trends of SAT are complicated in central and eastern China. Data homogenization reduces the linearly increasing trend during recent decades (1951–2010) while it appears to strengthen the linearly increasing trend during the earlier period (1909–1950) and increases the linear trends during the entire period (1909–2010).

Figure 7.

(a) Time series of TA13 (red) and TUA13 (blue) anomalies (°C) from 1971–2000 climatology during 1909–2010. (b) Same as in Figure 7a but for TA13 (red) and TCRU (blue) during 1909–2010 and TEast (green) and TChina (yellow) during 1951–2010.

Table 3. Linear Trends of Different Reconstructed SAT Series (°C/100a)
SAT Over China1909–19501909–20041909–20081909–20101951–2010
TA131.691.361.521.522.80
TUA131.231.091.271.292.97
TCRU2.601.201.311.302.60
TEast////2.44
TChina////2.26
LQX3.371.52///
TD3.35/1.00//
WYG2.46/0.58//

[29] To analyze the representativeness of TA13, we compare it with a regional and annual mean SAT over central and eastern China (to the east of 110°E; called the TEast) during 1951–2010. Because there are more than 400 stations to the east of 110°E, the distribution of these stations is dense and relative uniformly over central and eastern China in this period. Therefore, the time series of TEast is calculated as the arithmetic mean of SAT anomalies (from the climatological mean of 1971–2000). Figure 6b shows the change in the station number over eastern and central China. In this figure, the station number in eastern and central China exceeds 435 since 1960. The variability of the TEast anomaly approaches that of TA13 (Figure 7b), with a correlation coefficient of 0.99 during 1951–2010. Such a high correlation coefficient suggests that TA13 represents well the regional climate variation over eastern and central China. However, it is also noted that the linear trend of TA13 is slightly larger than that of TEast (2.44°C/100 years) (see Table 3).

[30] Moreover, the variability of annual mean TA13 is also highly consistent with that of the whole of China (from all the stations) (Figure 7b), with a correlation coefficient of 0.96 during 1951–2010. The number of stations used in the calculation of the China SAT value is shown in Figure 6b. For the period 1909–2008, the linear trend of the adjusted series is 1.52°C/100 years, which is larger than that (0.58°–1.0°C) of the 1909–2008 national mean series of China from Wang et al. [1998] and Tang and Ren. [2005], for which we have extended to 2008. These extensions were updated using the same stations used by the earlier studies. Compared to the linear trend of 1.52°C/100 years from LQX during 1909–2004 (available data only for 1900–2004), however, the linear trend of TA13 (1.36°C/100 years) during this period is smaller.

[31] Additionally, we also compare the 18 station SAT series with the CRU TS 3.10 data set [Harris et al., 2013]. Figure 5c shows the linear trend of the CRU annual mean SAT. In the figure, there are nine stations with a relatively large difference between the Chinese station data and the CRU data. This difference mainly appears near 40°N and to the south of this latitude, showing an underestimated warming trend of the CRU data at five stations for southeastern coastal regions of China and an overestimated warming trend in more northern parts. Figure 7b also shows the anomalies of the CRU annual mean SAT at the 16 stations (referred to as TCRU). Generally speaking, the anomalies of annual mean TCRU show consistent variability to TA13, with a correlation coefficient of 0.95 between TCRU and TA13 during 1909–2010, though the largest difference appears in 1946 and 1948, both years when station observations were sparsely located. The linear trend of TCRU is 1.30°C/100 years during 1909–2010, slightly smaller than that of TA13.

6 Conclusion and Discussion

[32] Construction of a long-term homogeneous time series is essential for research into climate change. Using quality control, interpolation, and homogeneity methods in this study, we objectively establish a set of homogenized monthly mean SAT series in eastern and central China back to the nineteenth century. Different source data sets are first compiled together, and the monthly mean SAT is defined as the arithmetic mean of the daily maximum and minimum temperatures. After assessing the length, quality, and reliability of the SAT time series, we chose 18 stations in eastern and central China with a small missing percentage of monthly records to construct the continuous time series.

[33] For those missing monthly values, we interpolate monthly mean SAT by using the standardized method, partial least squares regression, multivariate linear regression, and a gradient inverse distance weighting method in three steps, in which cross validation is used to assess the reliability of interpolation methods. The errors of these interpolations are small, with RMSE values below 1°C generally, and no extreme values are obtained from the interpolations. Thus, the interpolated values are statistically acceptable. Finally, the continuous time series from the start of observations to 2010 is obtained for each of the 18 stations. Several data homogenization methods and available metadata are used to perform the homogeneity detection and the adjustment of the series around the detected break points. For the time series before 1950, we use the PMFred, TPR methods, and the running Student's t test without a reference series to detect change points and use the RHtest to perform adjustment. For the time series during 1951–2010, the PMTred method with a reference series is used to detect change points and to adjust the series. Thirty-three break points are detected at the 14 stations from the start of observations to 2010, and the time series before the break point is adjusted to the latest segment of the data series. Finally, the homogenized time series is constructed at all 18 stations.

[34] Compared to the unadjusted series, the adjusted series of annual mean SAT generally shows a larger warming trend at most stations, with a range of trends between 1.0° and 4.2°C/100 years in northeastern China and southeastern China and a range of trends between −0.3° and 1.0°C/100 years in central and eastern China near 30°N, implying that the inhomogeneous series underestimates the warming trends during the past 100 years. Moreover, SAT from the CRU data also slightly underestimates the warming trend for the coasts of southeastern China but overestimates the warming trend in more northern parts of China.

[35] A regional annual mean SAT at 13 locations for eastern and central China agrees well with the variability of temperature at stations in the same region estimated from a much denser station network and is also highly consistent with that of the whole of China during 1951–2010. However, one should note that it is not enough to develop SAT series only for eastern and central stations of China to represent the variability of all of China. Thus, it will be necessary to construct long-term series of temperature over western China in the past century by combining limited observational data in western China and some observational data in neighboring countries before 1950. This will be addressed in future work.

Acknowledgments

[36] We thank Jingyun Zheng of the Chinese Academy of Sciences (CAS) for providing temperature data from CAS. We also thank both reviewers for their constructive comments. The long-term series of temperature data in Hong Kong and Macao is obtained from the homepages http://gb.weather.gov.hk/wxinfo/climat/specialday/html/select_c.shtml and http://www.smg.gov.mo/www/smg/centenary2/index_four.htm. This work is supported by the Strategic Priority Research Program-Climate Change: Carbon Budget and Relevant Issues of CAS (grant XDA05090101). PDJ has been supported by the US Dept. of Energy (grant DE-SC0005689).

Ancillary