This paper presents a method to homogenize China’s surface solar radiation (SSR) data and uses the resulting homogenized SSR data to assess the SSR trend over the period 1958–2016. Neighboring surface sunshine duration (SSD) data are used as reference data to assess the SSR data homogeneity. A principal component analysis is applied to build a reference series, which is proven to be less sensitive to occasional data issues than using the arithmetic mean of data from adjacent stations. A relative or absolute test is applied to detect changepoints, depending on whether or not a suitable reference series is available. A quantile-matching method is used to adjust the data to diminish the inhomogeneities. As a result, 60 out of the 119 SSR stations were found to have inhomogeneity issues. These were mainly caused by changes in instrument and observation schedule. The nonclimatic changes exaggerated the SSR change rates in 1991–93 and resulted in a sudden rise in the national average SSR series, causing an unrealistically drastic trend reversal in the 1990s. This was diminished by the data homogenization. The homogenized data show that the national average SSR has been declining significantly over the period 1958–90; this dimming trend mostly diminished over the period 1991–2005 and was replaced by a brightening trend in the recent decade. From the homogenized SSR data, the 1958–90 and 1958–2005 dimming rate is estimated to be −6.13 ± 0.47 and −5.08 ± 0.27 W m−2 decade−1, respectively, and the 2005–16 brightening rate is 6.13 ± 1.77 W m−2 decade−1.
Surface solar radiation (SSR) is an important part of the global energy balance. It is the energy source of the general circulation and the water cycle, influences the plant photosynthesis and carbon uptake, and determines, to a large extent, the climate condition on our planet (Wild 2012; IPCC 2013).
SSR has been recorded since the 1920s and became a routine meteorological observation in the 1950s. Researchers attempted to analyze the global energy balance and climate change based on the long-term in situ SSR records over the world. Early results showed that SSR has significant decadal change characteristics (Ohmura and Lang 1989; Gilgen et al. 1998; Stanhill and Cohen 2001). It has been declining at widespread locations since the 1950s and began to increase in the 1980s, phenomena which are well known as “global dimming” and “global brightening,” respectively (Pinker et al. 2005; Wild et al. 2005, 2009; Wild 2012).
China’s SSR records demonstrated similar characteristics: downward trend before 1990, followed by an upward trend (Che et al. 2005; Qian et al. 2007; Shi et al. 2008; Xia 2010; Tang et al. 2011; Yang et al. 2013a). However, it has been noted that SSR observations in China experienced a major instrument change in the early 1990s, which would cause inhomogeneity issues and lead to potentially misleading conclusions. The reversal of the SSR trend might be influenced by artificial changes (inhomogeneities) in the observational records (Wang et al. 2013; Tang et al. 2013; Wang and Wild 2016).
Some potential data quality and homogeneity issues in the SSR data records have been noted and efforts have been made to identify and adjust them. Shi et al. (2008) developed an SSR data quality assessment method that is proven to be good at identifying artifacts in SSR records. Hakuba et al. (2012) used the standard normal homogeneity test (SNHT), Pettitt test, Buishand test, and Von Neumann test to identify inhomogeneities in the SSR data for the period 2000–07 at 172 European stations taken from the Global Energy Balance Archive (GEBA; Wild et al. 2017). They found that relocation and instrument change caused significant changepoints in historical SSR data records at 20 stations, which resulted in Europe’s SSR trend being underestimated. Sanchez-Lorenzo et al. (2013) chose highly correlated nearby stations to construct reference series and used SNHT to detect potential changepoints in the Spanish historic SSR data. They found inhomogeneities at three stations in Spain. Sanchez-Lorenzo et al. (2016) applied the SNHT and Craddock tests to 56 SSR series over Europe taken from GEBA and did not find large biases caused by inhomogeneities. Manara et al. (2016) homogenized the Italian SSR data using a relative test developed by Brunetti et al. (2006) and found significant differences before 1980 between the raw and adjusted data. Wang (2014) compared the observed and derived (from sunshine duration) SSR and found that the decadal variability of observed SSR may have nonnegligible errors due to instrument sensitivity drifting and instrument replacement.
Compared with the absolute method (which only uses statistical tests), the relative method is more objective and efficient for detecting inhomogeneities. It helps to reduce misjudgments due to artificial changes through the comparison with a reference series. The construction of the reference series for the relative test is critical.
Surface sunshine duration (SSD) is the accumulated duration of the direct radiation beam exceeding a certain threshold (usually taken at 120 W m−2); it is considered to be a useful proxy of SSR (Angstrom 1924; IPCC 2013; Sanchez-Romero et al. 2015; Stanhill and Achiman 2017). There are only 130 SSR stations in China, but there are more than 2400 SSD observation stations in China and only 64 of the SSD stations experienced instrument changes before 2010. Therefore, SSD series could be well suited as reference series for the SSR data. Wang et al. (2015) made an effort to homogenize the observed SSR dataset using SSD-derived SSR as a reference dataset and found that the national averaged SSR (averaged over 105 stations in China) decreased from 1961 to 1990 at a rate of −2.9 W m−2 decade−1 and remained stable afterward.
The main objective of this study is to generate a long-term homogenized dataset of SSR over China and to study the temporal changes over the period from 1958 to 2016. The dataset used in this study is described in section 2, whereas the method and an example of the inhomogeneity test and adjustment are presented in section 3. The results of changepoints and their causes and the trend analysis, including the national average SSR series for China, are shown and discussed in section 4. Finally, conclusions of this study are presented in section 5.
Monthly SSR and SSD data and the related metadata at Chinese routine weather stations were released by the China Meteorological Administration (CMA) National Meteorological Information Center (NMIC) in 2013 (http://data.cma.cn/enl). A data quality control has been applied, including the spike value test, the stuck value test (a test to identify long runs of the same value in the data series), and the spatial consistency test. All data that pass all tests are classified as credible; the others are diagnosed further by experts at the stations and identified as doubtful or wrong (Ren and Liu 2013). Only credible data are used in the present study.
a. SSR data
Solar radiation was first measured in China in 1957. There are 130 SSR observation stations in total: 91 stations began SSR measurements before 1990 but 26 ceased SSR observations before 1995, and 39 new stations were added to the SSR observation network after 1990. There are 55 stations with long-term (more than 50 yr) SSR records. Figure 1 shows the geographical distribution of these SSR stations. The blue and red dots are meteorological stations that began SSR measurement before and after 1990, respectively. The blue square stations have no SSR observation after 1995. It can be seen that some of the red dots nearly overlap with the blue squares; the stations in red dots started to have SSR observations, replacing the measurements at the blue square sites.
We obtain more long-term series by joining nearby station data. Table 1 lists the geographic information of 11 joint stations. The left- and right-hand side of Table 1 show stations that consist of the former and latter part of the joint series, respectively; they are in a short distance (about 30 km on average) from each other. This means that, in general, they have similar climate conditions and SSR trends. Nevertheless, the joined series will be tested for homogeneity later, and significant discontinuities caused by joining of stations will be removed by the homogenization procedure. The joint series can represent the regional long-term SSR change characteristics. Also, all dates of joining are recorded as dates of relocation in order to assess the influence of joining, which might be significant changepoints. Eventually, data from 119 stations are analyzed in this study, including 64 stations with long-term SSR records (>50 yr).
It should be noted that SSR was manually recorded by an observer at all routine weather stations before 1990, and the observation time (sampling time) was mainly decided by the manager of the stations. This results in various SSR observation schedules over China, changing from three to six times per day or even more, and might induce inhomogeneities. In the early 1990s, SSR began to be automatically measured with advanced instruments, which have a much higher accuracy and sensitivity.
Figure 2 shows the observation history of the SSR stations in China. The blue, green, and orange bars present the frequency of observation time change, relocation, and instrument change, respectively, at all SSR stations. It is evident that instrument changes mainly occurred in the 1990s, especially in the early 1990s, accompanied by observation time changes due to the transition from manual to automatic observations. It is worth noting that several types of pyranometers have been used by the routine weather stations before the 1990s, such as pyranometer models ∏3x3, DFY-1, and DFY-2 produced in China. Few orange bars can be found before 1990 because of a lack of instrument information records at most stations (Fig. 2). At the beginning of SSR observations, measurement schedules changed frequently until 1965. Then, the SSR measurement time and frequency at each station became stable but still not necessarily the same. This situation lasted until the use of automatic measurements in the early 1990s. Station relocation is another cause of inhomogeneity. Although there are some SSR station relocations nearly every year, a remarkable peak with more than five stations relocated in the same year is not found. Thus, relocation over time seems not to be a network-wide issue.
All of the above indicates that there may be significant inhomogeneities in China’s historical SSR data records.
b. Reference data
In this study, we use SSD data from 2479 meteorological stations taken from the CMA NMIC as reference data. Most stations began SSD measurements in the 1950s. The Campbell–Stokes sunshine recorders were employed at 75 stations, and 64 stations replaced them with Jordan recorders gradually by 2011. The other 97% of stations have been using Jordan sunshine recorders. This indicates that the historical SSD data records are hardly affected by instrument change, as instrument change is considered a major cause of SSR data inhomogeneities. Nevertheless, all the SSD data were tested for homogeneity (without using a reference series). The results showed that the homogeneity of SSD data is good in general; about 70% of the stations did not show a break and the identified changepoints did not show significant characteristics in the appearance time and region. These SSD data are suitable for use to construct a reference series using the principal component analysis method (see section 3a).
Many studies have pointed out that there is a high correlation between SSR and SSD (Almorox and Hontoria 2004; Yang et al. 2006; Sanchez-Lorenzo and Wild 2012; Sanchez-Romero et al. 2014; Manara et al. 2016, 2017). SSD data can serve as the reference for testing SSR inhomogeneities. It should be noted that a station relocation typically affects both SSR and SSD data. For example, a relocation from downtown to the outskirts of the urban areas could cause a sudden increase in both SSR and SSD. It is therefore hard to identify this type of changepoint in the SSR data by using the SSD data from the same station. In general, the nearby stations have similar long-term trends, and the sudden increase caused by relocation should not exist in the neighboring station’s data record. Thus, the SSD data from nearby stations can be used to construct an arguably better reference to identify inhomogeneities in the SSR data and to estimate their influence.
To check the feasibility, the correlation coefficient between the clear-sky index of radiation stations and sunshine percentage of the meteorological stations surrounding the SSR stations in China is calculated [the clear-sky index is the ratio between the measured SSR and the insolation at the top of the atmosphere (TOA); sunshine percentage is the ratio of measured SSD and the SSD at the TOA]. Considering the uncertainty associated with sample size, monthly anomaly series, rather than annual series, are tested here, to make the sample size 12 times as large as the annual series.
Figure 3 shows the average correlation coefficients (vertical axis) between the clear-sky index at the 119 SSR stations and their nearby SSD stations at different distances (horizontal axis). Obviously, the consistency is high between SSR and SSD at short distances, with the correlation coefficient being ≥0.8 within 100 km. Although the correlation weakens linearly with distance, it is still above 0.7 within 250 km. Thus, the SSD data series surrounding an SSR site should provide a good reference for testing the homogeneity of SSR data.
Figure 4 shows the number of SSD stations surrounding an SSR station whose sunshine percentage data series has a high correlation (>0.7) with the clear-sky index series at the SSR station. These are used as reference stations for the SSR series. There are 108 SSR stations that have at least four reference stations (blue dots) and seven stations that have no more than four reference stations (red circles). There are four stations that do not have suitable reference stations (red dots), which are almost all located in areas close to the national boundaries or with scarce weather stations (i.e., Xin Jiang, Inner Mongolia, Tibet, and the South China Sea island). The only exception is station Mount Emei, which is a single high-altitude station (elevation 3137 m) in the Sichuan plain region (southwestern China); the large altitude difference causes the low correlation with its surrounding SSD station data.
In summary, SSD records from nearby meteorological stations can be used in this study as reference time series for detecting changepoints in the SSR data series.
a. Reference construction method
The construction of reference series is very important in a relative homogeneity test. Many algorithms have been developed to improve the reliability and representativeness of reference series. (Peterson and Easterling 1994) proposed to build a monthly–annual temperature reference series by interpolating 3–4 highly correlated nearby stations’ data and to use a single nearby station as a daily temperature reference series. Peterson et al. (1998) developed a first difference method (FDM) to reduce the uncertainties caused by inhomogeneous and incomplete records, as well as different series lengths, and to allow the reference series to contain the observation information as complete as possible. Jones (1994) used a climate anomaly method (CAM) to produce global gridded temperature data.
The data from adjacent stations could be used to construct a reference series if they are homogeneous. However, changes in the observation system, relocation, instrument update, and measurement times occur at each observatory and could likely result in changepoints. In this case, a reference series constructed from adjacent stations’ data series may contain a combination of these inhomogeneity issues.
The principal component analysis (PCA) is a mathematical–statistical procedure that uses an orthogonal transformation to convert a set of observations of possibly correlated variables into a set of values of linearly uncorrelated variables called principal components (PCs). Often, its operation can be thought of as revealing the internal structure of the data in a way that best explains the variance in the data.
As mentioned in section 2, most stations in China did not change their SSD instruments in the last 60 years, and the SSD data have been recorded once per day throughout this period. Thus, the observational conditions, including new buildings and growing trees around the site and relocation, remain the major factors that potentially induce inhomogeneities in the SSD data. Changes in the observation conditions occur every year at some of the Chinese sites and mainly depend on individual station conditions. They seem to be random events in general, rather than a network-wide change. Thus, it is practicable to apply the PCA to reduce the effect of these random events, in order to extract the regional features from a set of SSD data series for use as a reference series.
The SSD data series at the four stations listed in Table 2 have been found to be homogeneous by the penalized maximal F (PMF) test (Wang 2008a) at the 5% significance level and thus were chosen to compose a reference series for comparing two reference construction methods: the PCA method and the arithmetical–regional-average method. The correlation coefficient of monthly anomalies between each SSD data series and the target SSR data series is higher than 0.7.
We have compared the performance of the two methods under various unsatisfactory data conditions (viz., some of the chosen series are inhomogeneous) by doing the following. To simulate a situation where 50% of nearby SSD series suffer from inhomogeneity issues, we first randomly chose two out of the four SSD series, then we added a constant value δi to all values between time ti + 1 and time ti + τi in the ith series (i = 1 or 2, i.e., one of the two chosen series). Here, value δi is randomly selected from −3, −2, −1, 1, 2, and 3, while ti is from years 0, 10, 20, 30, 40, and 50 and τi is from years 5, 10, 15, 20, 25, and 30 of the data series. When ti = 0, there is only one changepoint in the resulting inhomogeneous series; otherwise, there are two changepoints. The two inhomogeneous data series have changepoints of different sizes and at different times. We then constructed a reference series (regional SSD series) from the four SSD series by the PCA method and arithmetic-mean method; subsequently, the biases (RMSE and cosine similarity) between those unsatisfactory reference series and the homogeneous reference (without changepoints) series are calculated and used to assess the performance of two methods for constructing a reference series.
The root-mean-square error (RMSE) and cosine similarity [cos(θ)] were calculated for each reference construction method. The RMSE is widely used in absolute error estimations. The cosine similarity is a measure of similarity between two vectors of an inner product space that measures the cosine of the angle θ between them (Singhal 2001). It is a judgment of orientation but not magnitude: two vectors with the same orientation have a cosine similarity of 1, two vectors at 90° have a similarity of 0, and two vectors diametrically opposed have a similarity of −1, independent of their magnitude.
Figure 5 illustrates the comparison of RMSE (left) and cosine similarity (right) between the PCA and the arithmetic–regional-mean methods under imperfect data conditions (changepoints were imposed on two of the four SSD data series). It is evident that almost all points lie above the 1: 1 (gray dashed) line, suggesting that the PCA method has smaller RMSE and cosine similarity values than the arithmetic–regional-mean method. A further PMF test at the 5% significance level shows that 40.5% of the reference series constructed by the PCA method are homogenous, while this percentage of the arithmetic-mean method is only 18.0%. These results indicate that, to some extent, the PCA method would relieve the biases caused by the changepoints in the composing series and obtain more reliable reference series relative to the arithmetic–regional-mean method. The PCA method is better than the arithmetic-mean method because it extracts the “large scale” information from the set of neighboring stations’ data in the leading principal components, leaving most of the inhomogeneities in the remaining principal components that are not used to construct the reference series. Thus, the PCA method is used in this study.
The PCA method was applied to the sunshine percentage data series from the nearby SSD stations (at least four stations) that have high correlations (>0.7) with a target SSR data series (all nearby stations are within 250-km radius). The variance-weighted-average series of the first n principal components that account for 85% of the total variance was taken as the reference series for the target SSR data series in question.
b. Changepoints detection and adjustment methods
In this study, the RHtestsV4 software package (Wang and Feng 2013) was used to identify changepoints. The software package has been widely used to alleviate historic data homogeneity issues (Vincent et al. 2002; Zhang et al. 2005; Wan et al. 2010; Dai et al. 2011; Kuglitsch et al. 2012; Wang et al. 2014; Xu et al. 2013). It includes a relative test and an absolute test. The relative test is based on the penalized maximal T (PMT) test (Wang et al. 2007) and employs reference series to identify artificial sudden changes (shifts) in the data records. The absolute test is based on the PMF test (Wang 2008a), which can be used without reference series. Both methods account for the effect of the lag-1 autocorrelated noise term empirically and deal with multiple changepoints using a recursive testing algorithm (Wang 2008b). The effects of unequal sample sizes of the data before and after a time point being tested on the detection power were also diminished using an empirical function in each algorithm (Wang 2008b).
The relative method was applied to the clear-sky index data series for all SSR stations that have four or more highly correlated (>0.7) SSD (sunshine percentage) data series available to construct a reference series, that is, stations shown as blue dots in Fig. 4. All nearby SSD stations that have a correlation of 0.7 or higher with the target SSR data series were used to construct the reference series (see section 3a for a description of the method). The absolute method is also used at each station.
All changepoints that were found to be significant at the 5% level were retained and verified with metadata. For some stations whose metadata are incomplete (especially before the 1990s), the changepoints were diagnosed further with expert knowledge.
For all changepoints that were determined to be significant artificial changepoints, the quantile-matching (QM) adjustment method (Wang et al. 2010, 2014) and the reference series if available were used to adjust the SSR data series to the latest segment. The objective of the QM adjustments is to adjust the base series so that the empirical probability distributions of all segments of the detrended base series match each other. The adjustment values were estimated from the difference (base minus reference) series, or from the detrended base series when not using a reference series. The adjustment value depends on the empirical frequency of the datum to be adjusted; namely, it varies from one datum to another in the same segment, depending on their corresponding empirical frequencies (Wang et al. 2010, 2014).
c. An example of homogenization analysis
As an example, Fig. 6 shows the annual time series of sunshine percentage and raw and homogenized clear-sky index for the station Na Qu, along with the constructed reference series used to test the homogeneity of the clear-sky index series. The reference series was constructed using the sunshine percentage data from four nearby stations (Ding Qing, Suo Xian, Jiang Zi, and Dang Xiong) during 1962–2014. It is evident that there are big differences between the candidate clear-sky index series (red line) and the reference series (green line). At the 5% significance level, three shifts (around 1977, 1988, and 1993) were identified in the raw clear-sky index time series (red line). The sunshine percentage series at this station (blue line) and the constructed reference series (green line) do not show a similar change in 1993. In addition, the metadata show that automated SSR instruments were installed at the Na Qu station in 1993, which confirms the detected shift in 1993. There is no doubt that the jump in 1993 resulted from observation automation (including the instrument update and the increase of measurement frequency from once per several hours to once per several minutes). However, the causes for the other shifts in the clear-sky index series are unknown because of the lack of metadata in the early period.
It is worth noting that the significant downward shifts in 1977 and 1988 were also found in the sunshine percentage series at Na Qu (blue line), which were not identified in the corresponding reference and neighboring series. The Na Qu station has been using Jordan SSD recorders and has not been relocated since 1965. Thus, it can be concluded that local changes occurred at that time, which could have affected both SSR and SSD. Changes in the observational conditions, probably overlapped with changes of the SSR instrument, became the only cause inducing such a dramatic change. Thus, three changepoints were kept and adjusted by the QM method. The orange line is the adjusted series, which no longer has dramatic jumps and becomes more consistent with the SSD data series.
a. Changepoints and their causes
A total of 159 changepoints (shifts) were identified. These occurred at 60 of the 119 SSR stations; 51 stations have at least 2 shifts. The stations of inhomogeneous records are fairly evenly distributed over China, as shown in Fig. 7; nearly all provinces suffer from inhomogeneity issues. Instrument change is the major cause of the inhomogeneities in the SSR data; it resulted in 42 shifts, while 34 and 20 shifts resulted from changes in the observational conditions (including new buildings and growing trees around the observing site and relocation) and changes in the observation schedule, respectively. The exact causes for the remaining 63 discontinuities are unknown because of incomplete metadata.
Figure 8 shows the distribution of SSR discontinuities (at the 119 stations) over time during the period from 1957 to 2016 and their possible causes. The color bars stand for shifts caused by different factors. It can be seen that shifts occurred in almost every year during this period. The automation of SSR observations in the early 1990s, including instrument update and observation schedule change, affected the SSR data significantly. However, changes in observing frequency in the 1960s (Fig. 2) did not result in remarkable inhomogeneities. Discontinuities without metadata support (of unknown causes) are mainly seen before 1990 because of the incomplete metadata information for that period.
b. Effects on the consistency between SSR and SSD
As described in section 2, SSD is highly correlated with SSR, showing good spatial consistency. Thus, we also assess the homogenization effects by comparing the SSR level change with the corresponding SSD level change averaged over the neighboring stations.
In Fig. 9, Rc is the SSR level change defined as the ratio of the SSR level (mean) for the 5 years before to that for the 5 years after a changepoint. Rn is similar to Rc but for the average result of the nearby SSD stations (all stations within 250-km radius were used to calculate Rn). Clearly, there are remarkable changes within a short time in the raw SSR data (blue circles), with Rc ranging from 0.66 to 1.84, while the corresponding SSD changes at nearby SSD stations are minimal, with Rn varying from 0.86 to 1.14. The RMSE and absolute relative error between Rc of the raw SSR data and Rn reach 0.17 and 12.2%. However, for the adjusted data, the Rc and Rn values are comparable, with all green circles distributing around the 1: 1 line and the Rc varying from 0.88 to 1.10. The RMSE and absolute relative error between Rc of the adjusted data and Rn is 0.05 and 3.5%, respectively, which are much lower than those of the raw data. All this indicates that there is a significant irregularity in the raw SSR data causing spatial inconsistency. The SSR homogenization alleviates these issues effectively.
c. Effects on the historical SSR trend
Figure 10 displays the national average SSR annual anomaly series (1958–2016) of raw (blue) and homogenized (green) data for China. This SSR series is derived as follows:
- 1) Obtain the gridbox average SSR value for each 5° × 5° grid box, using the FDM (Peterson et al. 1998), which has been proven to be suitable for an irregular observation network with diverse data lengths, as well as starting and ending times, and which is able to use the maximum station density for the calculation of long-term changes. The formulas for calculating the FDM average SSR series for each box are given as denotes the average of X over a grid box, subscript i denotes the ith year, and superscript j denotes the jth station falling in the grid box. Variables n and m stand for the length of series and the number of stations in the box, respectively.
2) Calculate the SSR anomaly series for each grid box by subtracting the mean of the reference period 1961–90.
3) Obtain the national average SSR anomaly series by area-weighted averaging of the anomalies at all grid boxes that contain data over China, as in the CAM method (Jones 1994).
As shown in Fig. 10, there is a big difference between the raw and adjusted SSR anomaly series for the period after 1990, which is also known as the global brightening period. Some smaller differences are also seen in the period before 1960. The 1958–90 decline rate estimated from the adjusted SSR data series is −6.13 ± 0.47 W m−2 decade−1, which is slightly lower than the rate estimated from the raw (unadjusted) data series (see Table 3). The early 1990s’ nonclimatic dramatic rise (10.52 W m−2 in 2 yr) was largely removed by the homogenization. The homogenization also reduces the 1990–2016 brightening rate from 1.98 ± 0.76 to 0.01 ± 0.61 W m−2 decade−1. Both raw and adjusted series show a brightening trend since 2005 (Fig. 10 and Table 3). The homogenized data show that the national average SSR has been declining significantly over the period 1958–90. This dimming trend mostly diminished over the period 1991–2005 and was replaced by a brightening trend in the last decade or so, with the level of the current SSR still being much lower than the level before 1980 (Fig. 10). From the homogenized SSR data, the 1958–2005 dimming rate is estimated to be −5.08 ± 0.27 W m−2 decade−1, and the 2005–16 brightening rate is 6.13 ± 1.77 W m−2 decade−1 (Table 3).
Further, SSR trends at each individual station before and after the trend reversal period (the 1990s) are analyzed. For the 1961–90 trend analysis (Fig. 11), 80 stations were chosen because their data integrity is higher than 50% during this period and they have measurements during the first and last 5 years of this period. Similarly, 97 stations were chosen for the 1990–2016 trend analysis. Figure 11 compares the 1961–90 SSR trends between the raw and homogenized data at the 80 stations. Obviously, both the raw and homogenized data show a dimming trend at most stations across China; but the dimming trend is much stronger in the raw data than in the homogenized data: dimming rates of ≥9% decade −1 are seen at seven stations in the raw data but only at three stations in the homogenized data. Zhong Shan is the only station showing a dimming rate of more than 12% decade−1, which is seen in both the raw and homogenized data.
Figure 12 is similar to Fig. 11 but for the comparison of the 1990–2016 trends at 97 stations. The significant dimming trend seen in the previous period disappeared at most stations; it is even replaced by a brightening trend for this later period (Figs. 11 and 12). This can be seen in both raw and homogenized data. However, the brightening trend is seen at 20 more stations in the raw data than in the homogenized data (61 vs 41 stations); it is also exaggerated at several stations in the raw data. At some stations in Tibet, the raw and homogenized data show trends of the opposite signs (Fig. 12). Data inhomogeneities affect the SSR data significantly and reverse the SSR trend for the Tibet region. The nonclimatic changes in the historical data have largely masked the climate change signal and could have led to misleading conclusions. Nonclimatic changes in the historical data records caused by observation system and condition changes exaggerated the SSR change rates and resulted in a more dramatic trend reversal in the 1990s, which is reflected as a sudden jump in the national average SSR series in 1991–93 (Fig. 10).
Since the national average series (Fig. 10) shows an SSR trend reversal around 2005, we also estimated the 1961–2005 and 2005–16 trends. As shown in Fig. 13, the raw and homogenized data show similar trend characteristics, with an overall downward trend for the period 1961–2005, but the homogenized data show a slightly better spatial consistency in the trend patterns. As shown in Fig. 14, a brightening trend started to appear at about two-thirds of the SSR stations since 2005, although the estimated rates of change vary more notably, which may result from the short record lengths (only 12 data points in this period).
In this study, we have developed and applied a procedure to homogenize China’s SSR data and used the homogenized data to assess the SSR trends over the period 1958–2016. We have developed a PCA method to construct reference series from the best-correlated (≥0.7) neighboring SSD data series (four or more). We have shown that this PCA method is less sensitive to random inhomogeneity issues than the arithmetic-mean method. A relative test or an absolute test was used to detect changepoints, depending on whether or not a reference series is available. All breakpoints that were confirmed using metadata and expert judgment were retained and adjusted using the quantile-matching method (Wang et al. 2010, 2014).
As a result, 60 out of the 119 SSR stations were found to have inhomogeneity issues, with 159 artificial changepoints being identified. These were mainly caused by instrument change (42 shifts), observation condition changes including relocation (34 shifts), and observation schedule change (20 shifts). The causes of the remaining 64 changepoints could not be identified mainly due to incomplete metadata before the 1990s. The nonclimatic changes exaggerated the SSR change rates (Table 3) and resulted in a remarkable artificial sudden rise in the national average SSR series in 1991–93, causing a more drastic trend reversal in the 1990s (Fig. 10). The artificial jump was diminished by the data homogenization (Fig. 10).
Showing better spatial consistency, the homogenized data confirm that China experienced “global dimming” and “global brightening” in the past six decades or so, but the SSR trend reversal time is around 2005, later than previously reported (Che et al. 2005; Wild et al. 2005; Qian et al. 2007; Shi et al. 2008; Yang et al. 2013a; Wang and Wild 2016). The national average SSR anomaly series derived from the homogenized data shows a significant dimming trend (−6.13 ± 0.47 W m−2 decade−1) over the period 1958–90; this dimming trend mostly diminished over the period 1991–2005 (Fig. 10) and was replaced by a brightening trend in the recent decade or so (the 2005–16 brightening rate is estimated to be 6.13 ± 1.77 W m−2 decade−1; see Table 3).
Some studies have pointed out that the changing aerosol load over China might be the major cause of the dimming and brightening (Qian et al. 2007; Tang et al. 2011; Wang et al. 2012; Wild 2012; Yang et al. 2013b; Wang and Wild 2016). However, because of a lack of detailed historical data, it is still difficult to assess the exact influence of aerosol and to confirm that other factors such as clouds and water vapor do not play a major role in the long-term change. More efforts are needed, including reexamination of the historical data, development of gridded data products by merging in situ and satellite data, and modeling efforts.
This work is supported by the National Innovation Project for Meteorological Science and Technology: Quality Control, Fusion, and Reanalysis of Meteorological Observations under Grant CMAGGTD003-5 and Special Fund for Public Welfare Industry of Meteorology (Grant GYHY201306061).