A New Long-Term Downward Surface Solar Radiation Dataset over China from 1958 to 2015

Downward surface solar radiation (Rs) plays a dominant role in determining the climate and environment on the Earth. However, the densely distributed ground observations of Rs are usually insufficient to meet the increasing demand of the climate diagnosis and analysis well, so it is essential to build a long-term accurate Rs dataset. The extremely randomized trees (ERT) algorithm was used to generate Rs using routine meteorological observations (2000–2015) from the Climate Data Center of the Chinese Meteorological Administration (CDC/CMA). The estimated Rs values were validated against ground measurements at the national scale with an overall correlation coefficient value of 0.97, a mean bias of 0.04 Wm−2, a root-mean-square-error value of 23.12 Wm−2, and a mean relative error of 9.81%. It indicates that the estimated Rs from the ERT-based model is reasonably accurate. Moreover, the ERT-based model was used to generate a new daily Rs dataset at 756 CDC/CMA stations from 1958 to 2015. The long-term variation trends of Rs at 454 stations covering 46 consecutive years (1970–2015) were also analyzed. The Rs in China showed a significant decline trend (−1.1 Wm−2 per decade) during 1970–2015. A decreasing trend (−2.8 Wm−2 per decade) in Rs during 1970–1992 was observed, followed by a recovery trend (0.23 Wm−2 per decade) during 1992–2015. The recovery trends at individual stations were found at 233 out of 454 stations during 1970–2015, which were mainly located in southern and northern China. The new Rs dataset would substantially provide basic data for the related studies in agriculture, ecology, and meteorology.


Introduction
The downward surface solar radiation (Rs) plays a dominant role in the global radiation budget as it is a basic element of energy source on the earth [1,2]. It is an important driving force of various biological, chemical, and physical processes of the Earth's system [3][4][5][6]. Therefore, understanding and determining the variability of Rs are crucial for practical applications such as environmental, hydrological, and ecological studies [7][8][9][10][11].
Rs is not directly measurable using the satellite sensors due to the atmospheric influences, the introduction of empirical or physical-based models for Rs estimation will induce possible uncertainties [12]. Direct ground measurements are essential for Rs quantification since it is one of the However, studies on the Rs estimation based on the ERT method is rare compared to that based on the RF method.
Therefore, the main objective of this study was to develop an Rs estimation model based on the ERT method using quality-controlled radiation measurements and meteorological measurements from CDC/CMA. The Rs estimates based on the ERT approach were compared with ground measurements. Additionally, the spatial distributions, seasonal variations, and long-term trends of the estimated Rs over China from 1970 to 2015 were also analyzed based on the new dataset. Sections 2 and 3 introduce the data and methods used in this study. Section 4 introduces the results of the ERT-based model and presents the spatiotemporal analysis of the Rs estimates over China. Section 5 discusses the probable cause of variations in Rs. A short summary is given at the end of this paper.

Data
There is a total of 756 meteorological stations in China and only 96 meteorological stations have records of solar radiation since 1994. The Rs measurements by CDC/CMA started in 1957, and it is noted that the radiometers equipped at CDC/CMA stations had been updated during 1990-1993. Before releasing the radiation data, a quality control process was conducted by CDC/CMA. Nevertheless, some studies [43,64] suggested that the radiation data of CDC/CMA need to be inspected more strictly for further application. The method proposed by Tang et al. [43] was performed to examine the quality of radiation data. In this study, the "complete records" of meteorological data, which was defined as data that contains more than 20 days in every month, and 12 months in a year, were used for model construction at 96 CDC/CMA stations and reconstructing Rs at 756 CDC/CMA stations. To analyze the temporal variations of Rs in China, the reconstructed Rs data at 454 CDC/CMA stations with complete records for consecutive 46 years (1970-2015) were used for analysis.
The routine daily meteorological measurements, which include air pressure, air temperature, wind speed, relative humidity, daily precipitation, water vapor pressure, sunshine duration, and the Rs data collected from CDC/CMA were used to reconstruct Rs. In addition, the temporal information was also used to reduce the influence of the seasonal cycle of Rs, since Rs have the clear temporal variations. For each station, the cosine of the radian difference (T) is obtained according to Equation (1) proposed by Wei et al. [65], which is capable of minimizing the influence of the seasonal cycle.
where d represents the Julian day of the year (DOY), D denotes the total number of days in a year. Figure 1a displays the spatial distributions of 454 stations from CDC/CMA, the 96 radiation stations were denoted by the star symbols. This study also discussed the regional Rs trends in different climatic regions of China. Various approaches for classifying climatic types in China [66][67][68][69] were presented in previous studies. In this study, six climatic regions in China were classified according to the classification in Wang et al. [68] and Zhou et al. [69]. Figure 1b

Extremely Randomized Trees (ERT)
The decision tree, a nonlinear and nonparametric method, is widely used in regression and classification problems. Decision trees predict the value of a target variable using a set of values of input variables. The decision trees are capable of dealing with problems of large-scale data using plenty of training samples and input variables. The decision tree method is easy and clear to understand. The relevant variables can be recognized during the growth of trees, which provides robustness for the decision trees model [63]. However, the high sensitivity for training samples is the main inadequacy of the single tree [70,71], the low accuracy and high randomness of the single tree result from this high sensitivity restrict the application of the single tree, particularly in handle numerous datasets [63]. The ensemble of trees, such as RF and ERT, is capable of conquering the problem of single tree models.
RF, a powerful ensemble-learning method, was proposed by Breiman in 2001, which is widely applied as a classification and regression tool [72]. The RF method employs the bootstrap technique (Ibrahim and Khatib 2017), the bootstrap samples are randomly created and replaced from the training data. Using bootstrap samples in the RF model generates around one-third unused subset, named as out-of-bag data (OBB) [72], the rest data are called in-bag data. According to the minimized Gini index, the best split is determined among the subgroup of the random selection at each node.
ERT, a tree-based machine learning method, was proposed by Geurts et al. [73]. ERT is considered as the further development of the RF method. ERT has been widely used to solve diverse sequence-based prediction problems [74,75]. ERT introduces a more powerful randomization method to efficiently reduce the variance of models and excavate more significant information than other tree-based methods. There are three main parameters in ERT for regression problems including K, nmin, N, and M. The K denotes the number of random splits, the nmin indicates the minimum sample to split a node [63,73], the N is size of samples, and the M represents the number of trees of ERT. The growth of trees in ERT is conducted through exactly defining values of the K on each node to achieve pure outputs in all subsamples [63], the M controls the degree of variance reduction in the ensemble model [73]. In addition, the variance of ERT generally further decreases with the increase of M [63], but the bias may increase slightly. The ERT offers added robustness about obvious errors due to the marginal affection of outliers on ERT prediction. The variable importance measure is also provided by the ERT method, which is defined as follows:

Extremely Randomized Trees (ERT)
The decision tree, a nonlinear and nonparametric method, is widely used in regression and classification problems. Decision trees predict the value of a target variable using a set of values of input variables. The decision trees are capable of dealing with problems of large-scale data using plenty of training samples and input variables. The decision tree method is easy and clear to understand. The relevant variables can be recognized during the growth of trees, which provides robustness for the decision trees model [63]. However, the high sensitivity for training samples is the main inadequacy of the single tree [70,71], the low accuracy and high randomness of the single tree result from this high sensitivity restrict the application of the single tree, particularly in handle numerous datasets [63]. The ensemble of trees, such as RF and ERT, is capable of conquering the problem of single tree models.
RF, a powerful ensemble-learning method, was proposed by Breiman in 2001, which is widely applied as a classification and regression tool [72]. The RF method employs the bootstrap technique (Ibrahim and Khatib 2017), the bootstrap samples are randomly created and replaced from the training data. Using bootstrap samples in the RF model generates around one-third unused subset, named as out-of-bag data (OBB) [72], the rest data are called in-bag data. According to the minimized Gini index, the best split is determined among the subgroup of the random selection at each node.
ERT, a tree-based machine learning method, was proposed by Geurts et al. [73]. ERT is considered as the further development of the RF method. ERT has been widely used to solve diverse sequence-based prediction problems [74,75]. ERT introduces a more powerful randomization method to efficiently reduce the variance of models and excavate more significant information than other tree-based methods. There are three main parameters in ERT for regression problems including K, n min, N, and M. The K denotes the number of random splits, the n min indicates the minimum sample to split a node [63,73], the N is size of samples, and the M represents the number of trees of ERT. The growth of trees in ERT is conducted through exactly defining values of the K on each node to achieve pure outputs in all subsamples [63], the M controls the degree of variance reduction in the ensemble model [73]. In addition, the variance of ERT generally further decreases with the increase of M [63], but the bias may increase slightly. The ERT offers added robustness about obvious errors due to the marginal Sensors 2020, 20, 6167 5 of 23 affection of outliers on ERT prediction. The variable importance measure is also provided by the ERT method, which is defined as follows: where B c(t) is the OBB sample for a tree, X a is sample value, t is the tree number, c The ERT and RF methods are similar but different in two aspects. The ERT employs all training samples instead of the bootstrap algorithm during the growth procedure; and the ERT performs node splitting by random selections of cut-points, instead of the best node split based on the Gini index in the RF [76], which is calculated according to Equation (3): The ERT method generate a set of independent decision trees based on the features space F. In this study, the features space X = {air pressure, air temperature, wind speed, relative humidity, water vapor pressure, daily precipitation, sunshine duration, elevation, and cosine of the radian difference}. The followed Algorithm 1 illustrates the procedure. At the begining of the training stage, the ensemble tree set is initialized as empty. Then, each decision tree is built with randomly selected features without replacement. During the testing stage, a predicted value y i is obtained by each decision tree. The final result is the average of all the decision trees.

Mann-Kendall (M-K) Test
The nonparametric Mann-Kendall (M-K) test [77,78] was employed for detecting the significant temporal variation of Rs. It is a rank-based method to identify tendencies of time-series dataset. The relative magnitudes of the samples are compared by the M-K test rather than the data values themselves. In addition, the M-K test does not require that samples conform to a certain type of distribution. The M-K test is an effective approach to identify tendencies of meteorological and other relevant variables without being affected by certain data distribution and outliers. The M-K statistics s is defined as: where i is the number of data. T m and T n (m>n) are observations in time series. f () is given as: The variance of statistics s is given as: The standardized test statistics Z is given as: A positive Z value represents the increasing trend and a negative value represents the decreasing trend. The significance levels (p = 0.05 and 0.01) are used to identify the significant trend, corresponding Z 1−α/2 = ±2.58 and Z 1−α/2 = ±1.96, respectively.

Results and Analysis
Previous studies [35][36][37] reported that the update of measuring equipment of CDC/CMA could result in uncertainty in the long-term trend analysis of Rs in China by introducing breaks. To eliminate this effect, the Rs data after 2000 were applied to build the model in our study. The daily ground-measured Rs data collected from 96 CDC/CMA stations during 2000-2015 were used as target variables. The daily meteorological data used as predictors including air pressure, air temperature, wind speed, relative humidity, water vapor pressure, daily precipitation, and sunshine duration. The elevation data and cosine of the radian difference data at each station were also used as predictor variables. The daily Rs dataset was randomly split into two subsets for model construction: 80% for training the model, hence the rest 20% for assessing the performance of the ERT-based model. The k-fold cross-validation method was performed during the training procedure to evaluate the overall performance of the ERT-based model. The correlation coefficient (R), root-mean-square-error (RMSE), mean bias error (MBE), and mean relative error (MRE) were employed for evaluating the accuracy of the ERT-based model.

Evaluation Using Ground Measurements
The comparisons between the estimated daily Rs and corresponding ground measurements were conducted on national, regional, and station scales, respectively. The results of the ERT-based model for estimating Rs in the training and test stages are summarized in Table 1. The results on the national scale are displayed in Figure 2. For the training dataset, the ERT-based model obtained great performance  Figures 3 and 4 show the estimation results using all datasets in both training and test stage to evaluate the overall performance of the ERT-based model. As shown in Figure 3, the test results suggest that the ERT-based model performed well in estimating the Rs in six climatic regions with reasonable accuracy. The ERT-based model in NE and TP provided slightly better accuracy than other regions, with R of 0.99 and 0.99, RMSE of 9.90 and 10.64 Wm −2 , MBE of 0.75 and 0.34 Wm −2 , and MRE of 3.06% and 2.49%, respectively. In addition to validations on national and regional scales, the validation of the ERT-based model was also conducted at individual CDC/CMA stations. As displayed in Figure 4, the daily estimated Rs agreed well with the Rs measurements at most CDC/CMA stations. For example, the great performance was found at station 52818 (Geermu) with R, RMSE, MBE, and MRE of 0.99, 7.23 Wm −2 , −0.14 Wm −2 and 1.73%, respectively. In contrast, the ERT-based model did not provide great Rs estimates at station 57874 (Changning) with R, RMSE, MBE and MRE of 0.98, 19.90 Wm −2 , −1.64 Wm −2 , and 7.39%, respectively. Overall, the R values were 0.99 at 91 out of 96 stations. The RMSE, and MRE, values were lower than 15 Wm −2 and 5% at 92 stations. The MAE were lower than 2 Wm −2 at 86 stations. It is obvious that the ERT-based model successfully estimated accurate daily Rs at most CDC/CMA stations.         Sensors 2020, 20, x FOR PEER REVIEW 9 of 23   To assess the performance of the ERT-based model on different temporal scales, the validations on the daily and seasonal timescales were implemented as well.  Table 1 indicate that the ERT-based model was promising to estimate the Rs accurately on both the daily and seasonal timescales. Figures 5 and 6 demonstrate the estimation results using both the training and test datasets. As displayed in Figure 5, the performance of the ERT-based model was shown as a function of DOY at 96 CDC/CMA stations during 2000-2015. It illustrates that the ERT-based model performed well on most days. For example, the day 365 had relatively higher accuracy with R, RMSE, MBE, and MRE of 0.99, 5.53 Wm −2 , 0.06 Wm −2 , and 3.02%, respectively. In contrast, the ERT-based model did not give satisfactory Rs estimates on day 237 with R, RMSE, MBE, and MRE of 0.98, 13.74 Wm −2 , 1.43 Wm −2 , and 33.28%, respectively. Overall, the R values were higher than 0.98 on more than 345 days. The RMSE and MBE values were lower than 14 and 1 Wm −2 on more than 329 days. The MRE values were lower than 3.5% on more than 330 days. It demonstrates that the ERT-based model was capable of estimating the Rs with reasonable accuracy on most days in a year. Figure 6 shows the test results for the daily estimated   Table 1 indicate that the ERT-based model was promising to estimate the Rs accurately on both the daily and seasonal timescales. Figures 5 and 6 demonstrate the estimation results using both the training and test datasets. As displayed in Figure 5, the performance of the ERT-based model was shown as a function of DOY at 96 CDC/CMA stations during 2000-2015. It illustrates that the ERT-based model performed well on most days. For example, the day 365 had relatively higher accuracy with R, RMSE, MBE, and MRE of 0.99, 5.53 Wm −2 , 0.06 Wm −2 , and 3.02%, respectively. In contrast, the ERT-based model did not give satisfactory Rs estimates on day 237 with R, RMSE, MBE, and MRE of 0.98, 13.74 Wm −2 , 1.43 Wm −2 , and 33.28%, respectively. Overall, the R values were higher than 0.98 on more than 345 days. The RMSE and MBE values were lower than 14 and 1 Wm −2 on more than 329 days. The MRE values were lower than 3.5% on more than 330 days. It demonstrates that the ERT-based model was capable of estimating the Rs with reasonable accuracy on most days in a year. Figure 6 shows the test results for the daily estimated   The validation against ground measurements using all datasets in both training and test stages was also conducted on the synthetic timescales including the monthly, seasonal, and annual timescales (   The validation against ground measurements using all datasets in both training and test stages was also conducted on the synthetic timescales including the monthly, seasonal, and annual timescales ( The validation against ground measurements using all datasets in both training and test stages was also conducted on the synthetic timescales including the monthly, seasonal, and annual timescales (Figure 7). On the monthly timescale, the Rs estimates had an R of 0.99, an RMSE of 4.86 Wm −2 , an MBE of 0.02 Wm −2 , and an MRE of 1.83%. Those values were 0.99, 4.20 Wm −2 , 0.01 Wm −2 , and 1.59% at the seasonal timescale, and 0.99, 3.43 Wm −2 , −0.03 Wm −2 , and 1.39% at the annual timescale. These results suggest that synthetic datasets can capture spatiotemporal variations of Rs in China more accurately.

Spatial Variations
The above validation results show that the Rs estimates based on the ERT method are reasonably accurate compared to the ground measurements. Therefore, the ERT-based model was applied to reconstruct Rs data at 756 CDC/CMA stations during 1958−2015. The reconstructed Rs data at 454 CDC/CMA stations covering 46 consecutive years  were selected for analyzing trends of Rs in China.
The spatial distribution of annual mean Rs from 1970 to 2015 is displayed in Figure 8. It shows that Rs was higher in TP, NC, and NE than that in SW, EC, and SC. Apart from Sichuan and Guizhou, the Rs was spatially decreasing from western to eastern China, and the Rs in western China was decreasing with the increase of latitude. The TP had higher Rs values than other climatic regions. A previous study shows that the Rs in TP was the second highest globally and the highest in China [79]. Strong Rs in TP was mainly due to the small amount of cloud [29], rainfall, water vapor content [68], and good air transparency [80]. In contrast, the relatively lower Rs values were found in the Sichuan Basin, where the multiple hazes and low atmospheric transparency always happened due to the interactions between the cold and warm currents [81,82].

Spatial Variations
The above validation results show that the Rs estimates based on the ERT method are reasonably accurate compared to the ground measurements. Therefore, the ERT-based model was applied to reconstruct Rs data at 756 CDC/CMA stations during 1958−2015. The reconstructed Rs data at 454 CDC/CMA stations covering 46 consecutive years  were selected for analyzing trends of Rs in China.
The spatial distribution of annual mean Rs from 1970 to 2015 is displayed in Figure 8. It shows that Rs was higher in TP, NC, and NE than that in SW, EC, and SC. Apart from Sichuan and Guizhou, the Rs was spatially decreasing from western to eastern China, and the Rs in western China was decreasing with the increase of latitude. The TP had higher Rs values than other climatic regions. A previous study shows that the Rs in TP was the second highest globally and the highest in China [79]. Strong Rs in TP was mainly due to the small amount of cloud [29], rainfall, water vapor content [68], and good air transparency [80]. In contrast, the relatively lower Rs values were found in the Sichuan Basin, where the multiple hazes and low atmospheric transparency always happened due to the interactions between the cold and warm currents [81,82].

Seasonal Variations
The annual and seasonal mean Rs were obtained from individual stations with complete records observations. Figure 9 shows that the monthly Rs increased slowly from January to June then it declined gradually from July to December. This may relate to the changes of the annual cycle of solar zenith angle and the maximum sunshine duration over China. The monthly Rs in July was the largest with the value of 216.52 Wm −2 , and the Rs in December was the smallest with the value of 92.73 Wm −2 . The spatial distribution of the seasonal mean Rs over China between 1970 and 2015 is shown Figure  10. The average Rs in spring, summer, autumn, and winter were 187.22, 213.28, 143.99 and 105.06 Wm −2 , respectively. The seasonal Rs trends detected by the M-K test in four seasons from 1970 to 2015 are shown in Table 2. The Rs showed an increasing trend in spring and decline trends in summer (p < 0.01), autumn (p < 0.01), and winter over China. In spring, the Rs in NE exhibited a significant decline (p < 0.01). Similarly, decline of Rs was also found in SW and TP. The seasonal Rs showed increasing trends in EC, NC, and SC. In summer, the significant decreasing trends were detected in Rs in EC (p < 0.01) and NE (p < 0.05). The decline tendencies of Rs were also found in NC and SC. The seasonal Rs showed increasing trends in SW and TP but were not significant. In autumn, the significant decline of Rs was found in EC (p < 0.01) and NC (p < 0.05). Similarly, the decline tendencies

Seasonal Variations
The annual and seasonal mean Rs were obtained from individual stations with complete records observations. Figure 9 shows that the monthly Rs increased slowly from January to June then it declined gradually from July to December. This may relate to the changes of the annual cycle of solar zenith angle and the maximum sunshine duration over China. The monthly Rs in July was the largest with the value of 216.52 Wm  Table 2. The Rs showed an increasing trend in spring and decline trends in summer (p < 0.01), autumn (p < 0.01), and winter over China. In spring, the Rs in NE exhibited a significant decline (p < 0.01). Similarly, decline of Rs was also found in SW and TP. The seasonal Rs showed increasing trends in EC, NC, and SC. In summer, the significant decreasing trends were detected in Rs in EC (p < 0.01) and NE (p < 0.05). The decline tendencies of Rs were also found in NC and SC. The seasonal Rs showed increasing trends in SW and TP but were not significant. In autumn, the significant decline of Rs was found in EC (p < 0.01) and NC (p < 0.05). Similarly, the decline tendencies of Rs were also found in NE and SC but not significant. The increasing trends of Rs were detected in SW and TP. In winter, the Rs showed decline in all six climatic regions.
Sensors 2020, 20, x FOR PEER REVIEW 13 of 23 of Rs were also found in NE and SC but not significant. The increasing trends of Rs were detected in SW and TP. In winter, the Rs showed decline in all six climatic regions.   of Rs were also found in NE and SC but not significant. The increasing trends of Rs were detected in SW and TP. In winter, the Rs showed decline in all six climatic regions.

Decadal Variations
Previous studies show that the Rs in China showed significant decline trends before the 1980s by 2 to 8 Wm −2 per decade [25][26][27][28][29][30][31][32], which did not persist into the 1990s. However, the magnitudes of variations in Rs were still controversial. The decadal variations of the Rs over China during 1970-2015 were analyzed based on the reconstructed Rs data at 454 CDC/CMA stations. As given in Figure    The decadal variations of Rs in six climatic regions during 1970-2015, 1970-1992, and 1992-2015 were also analyzed based on the reconstructed Rs dataset. As shown in Table 2 and Figure 12

Discussion
To investigate the influence of input variables of the ERT-based model on the estimation results, the variable importance measures provided by the ERT method was conducted. As shown in Table  3, the importance of the input variables of the ERT-based model was in the order of sunshine duration,

Discussion
To investigate the influence of input variables of the ERT-based model on the estimation results, the variable importance measures provided by the ERT method was conducted. As shown in Table 3, the importance of the input variables of the ERT-based model was in the order of sunshine duration, cosine of the radian difference, air pressure, water vapor pressure, relative humidity, elevation, air temperature, wind speed and daily precipitation. The sunshine duration had significant influence on estimating Rs, while the wind speed and daily precipitation had relative low influence on the Rs estimation.  [83,84], the variations of the Rs in China may relate to the variations of sunshine duration. The variable importance measures indicate that the sunshine duration had significant influence on the Rs estimation. To further investigate probable causes of changes in Rs, the comparison of the variations of sunshine duration and Rs was conducted. Figure 14 shows the anomalies of the annual mean Rs and sunshine duration in China during 1970China during -2015China during , 1970China during -1992China during , and 1992China during -2015. It is clear that annual variations of the anomaly of Rs were almost in line with variations of sunshine duration from 1970 to 2015, before 1992 the sunshine duration exhibited a decreasing trend (0.18 h per decade), and an increasing trend (0.18 h per decade) afterward. As shown in Figure 15, it illustrates that the trends of sunshine duration were consistent with the trends of the Rs in six climatic regions from 1970 to 2015. The consistency of variations in sunshine duration and Rs indicates that the variations of Rs were likely due to long-term variations of sunshine duration. In the most recent years, a slight difference between the sunshine duration and the Rs was observed. This probably contributes to the influence of the aerosol variations on Rs according to the previous studies [37,48,[85][86][87][88]. For instance, Jia et al. [87] reported that the variations in Rs is likely due to the variations in the aerosols, cloud, and water vapor in north and south China. Wang and Pinker [85] found that the global average aerosol showed an increasing trend during 1979-2006, particularly in east and south Asia (including China). The increase in global average aerosol could be associated with a continuous decrease of Rs in China from 1979 to 2006. Qin et al. [86] pointed out that Rs over China declined gradually between 1980 and 2015, which may relate to the increasing aerosol radiative forcing effects over China in recent decades. It is clear that the spatiotemporal variations in Rs over China were complicated due to the effects of natural factors and human activities. The influence of other factors (e.g., aerosols, cloud, and topography) on Rs estimation are still needed to be investigated in future research.  Figure 15, it illustrates that the trends of sunshine duration were consistent with the trends of the Rs in six climatic regions from 1970 to 2015. The consistency of variations in sunshine duration and Rs indicates that the variations of Rs were likely due to long-term variations of sunshine duration. In the most recent years, a slight difference between the sunshine duration and the Rs was observed. This probably contributes to the influence of the aerosol variations on Rs according to the previous studies [37,48,[85][86][87][88]. For instance, Jia et al. [87] reported that the variations in Rs is likely due to the variations in the aerosols, cloud, and water vapor in north and south China. Wang and Pinker [85]

Conclusions
In this study, an Rs estimation model was constructed employing the ERT algorithm based on the quality-controlled daily Rs data and routine meteorological data. The accuracy and applicability of the ERT-based model in estimating daily Rs over China were investigated. Overall, the reconstruction of new daily long-term Rs with high accuracy would be a useful data source for the related climate change studies. More attention needs to be paid to quantitative analysis of Rs and other climatic factors.

Conclusions
In this study, an Rs estimation model was constructed employing the ERT algorithm based on the quality-controlled daily Rs data and routine meteorological data. The accuracy and applicability of the ERT-based model in estimating daily Rs over China were investigated. Overall, the reconstruction of new daily long-term Rs with high accuracy would be a useful data source for the related climate change studies. More attention needs to be paid to quantitative analysis of Rs and other climatic factors.