The recent release of the International Satellite Cloud
Climatology Project (ISCCP) HXG cloud products and new ERA5 reanalysis data
enabled us to produce a global surface solar radiation (SSR) dataset: a
16-year (2000–2015) high-resolution (3 h, 10 km) global SSR dataset using an
improved physical parameterization scheme. The main inputs were cloud
optical depth from ISCCP-HXG cloud products; the water vapor, surface
pressure and ozone from ERA5 reanalysis data; and albedo and aerosol from
Moderate Resolution Imaging Spectroradiometer (MODIS) products. The
estimated SSR data were evaluated against surface observations measured at 42 stations of the Baseline Surface Radiation Network (BSRN) and 90 radiation
stations of the China Meteorological Administration (CMA). Validation
against the BSRN data indicated that the mean bias error (MBE), root mean
square error (RMSE) and correlation coefficient (
Surface solar radiation (SSR), which drives the energy, water and carbon cycles of Earth's system, is the driving input for simulations of hydrology, ecology, agriculture and land-surface processes (Wild, 2009; Wang et al., 2012). The accuracy of SSR data influences simulations of runoff, gross primary productivity, growth and yield of crops, and land data assimilation (Wild, 2012; Jia et al., 2013). SSR is also an important variable that affects the speed of glacier melting (Yang et al., 2011). Variations of SSR also affect the rate of global warming and the change of pan evaporation (Wild et al., 2007; Qian et al., 2006).
Information on the spatiotemporal distribution of SSR is fundamental for the selection of sites for solar power plants, decisions on energy policy, optimization of solar power systems and operations management (Mondol et al., 2008; Sengupta et al., 2018). To address issues such as these, historical SSR data have been obtained mainly through ground-based observations, station-based estimates and satellite-based retrievals (Pinker and Laszlo, 1992; Li and Leighton, 1993; Liang et al., 2006; Zhang et al., 2004; Wang et al., 2011; Huang et al., 2011; Kato et al., 2013; Ma and Pinker, 2012; Zhang et al., 2014; Wang et al., 2015; Niu and Pinker, 2015).
Measurement by the accurately calibrated and well-maintained radiometer of pyranometer is the most effective method to obtain reliable long-term SSR data. Although these data are valuable for simulations of land surface processes, solar power applications and evaluation of satellite retrievals (Sengupta et al., 2018), the high cost of maintaining radiation radiometers means that networks of radiation stations are too sparsely distributed. However, networks of routine meteorological stations are denser than those of radiation stations, and the variables observed at routine meteorological stations can be used to estimate SSR. For example, based on sunshine duration data, Tang et al. (2013, 2018) constructed long-term datasets of both daily global radiation and direct radiation over China at more than 2400 routine meteorological stations of the China Meteorological Administration (CMA). These datasets are generally more accurate than those derived from satellite retrievals (Yang et al., 2010). However, station-based estimates of SSR can be conducted only at routine weather stations, many of which are sparsely distributed, often in remote regions and harsh environments.
Alternatively, remote sensing retrievals based on satellites can provide reliable spatiotemporally continuous SSR data, either globally or regionally. The many methods that have been developed to retrieve SSR from satellite data can be roughly divided into two categories: statistical methods and methods based on radiative transfer processes (Huang et al., 2019). According to Sengupta et al. (2018), these methods can also be subdivided into three types: empirical, semi-empirical and physical.
Empirical methods build function relationships between SSR measured at limited numbers of stations and satellite data by applying regression or artificial intelligence technology (Lu et al., 2011; Wei et al., 2019). Empirical methods may work well at some locations, but the ability to expand their coverage to broader regions is limited.
Semi-empirical methods generally combine a physical model for clear-sky conditions with an empirical scheme for cloudy conditions. A well-known semi-empirical method is the Heliosat method of Cano et al. (1986), from which several improved versions have since been developed (Hammer et al., 2003; Mueller et al., 2009; Posselt et al., 2012; Wang et al., 2014).
Physical methods are generally well-suited to generalization because they take into account the physics processes of transfer of solar radiation from the top of the atmosphere to the Earth's surface. The look-up table (LUT) and physical parameterization methods (Pinker and Laszlo, 1992; Liang et al., 2006; Lu et al., 2010; Qin et al., 2015; Xie et al., 2016; Huang et al., 2018) are two typical physical methods that have been widely used to estimate SSR from satellite data.
Several well-known global SSR datasets have been produced by physical
methods. These include the global energy and water cycle experiment surface
radiation budget (GEWEX-SRB; Pinker and Laszlo, 1992), the International
Satellite Cloud Climatology Project flux dataset (ISCCP-FD; Zhang et al.,
2004) and the Earth's Radiant Energy System (CERES) radiation products (Kato
et al., 2013). Although each of these have been widely used in various
fields, the spatial resolutions (
The greatest uncertainty in satellite retrievals of SSR is the lack of a high-quality cloud product, which severely limits the development of high-resolution, long-term global satellite SSR products. However, the release in 2017 of new, global, long-term ISCCP H-series cloud products at a spatial resolution of about 10 km has provided an opportunity to develop a long-term high-resolution global-scale climate dataset of SSR.
We developed a global-scale 16-year dataset (2000–2015) of SSR data from the new ISCCP H-series cloud products and ERA5 reanalysis data, validated the accuracy of this dataset with surface observations, and compared its performance with other global satellite products. Section 2 introduces the method we used to estimate SSR. Section 3 describes the input data we used for SSR estimation and the observation data used for SSR validation. In Sect. 4, we presented our evaluation of the SSR product and compared it with other satellite products. Data availability is given in Sect. 5, and Sect. 6 presents some conclusions and explores future work to further improve SSR products.
The method we used to estimate SSR with ISCCP H-series cloud data is mainly
based on the SUNFLUX scheme, which was developed by Sun et al. (2012, 2014)
and first used by Tang et al. (2017) to retrieve SSR data from Moderate
Resolution Imaging Spectroradiometer (MODIS) atmospheric and land products.
Their validation of their results against measurements at Baseline Surface Radiation Network (BSRN) stations
indicated a mean root mean square error (RMSE) of
To produce the 16-year SSR products at global scale, we used three types of input data.
The first of these was the level 2 ISCCP H-series cloud product HXG
(H-series pixel-level global, here called ISCCP-HXG), which is a globally
merged product generated based on the HGS (H-series gridded by satellite)
product. The resolutions of HXG are 3 h and 10 km, and the HXG cloud
products are available for the period from July 1983 to December 2015. Note
that the ISCCP-HXG data are 0.1
The second data type we used was the new ERA5 reanalysis data. Three
variables of the ERA5 reanalysis data were used: surface pressure, total
column water vapor and total column ozone. The resolutions of the ERA5
reanalysis data are 1 h and 25 km. To derive the same spatial resolution as
the ISCCP-HXG cloud product, we re-sampled the three variables of ERA5
reanalysis data to a spatial resolution of 10 km.
The third data type comprised aerosol and albedo data. The MODIS aerosol
(MOD08_D3 or MYD08_D3) and albedo (MCD43A3,
Schaaf et al., 2002) daily products were used. The MODIS AOD product of the
combined Dark Target and Deep Blue AOD at 0.55
In this study, we used radiation observations made in 2009 to validate the
accuracy of the global-scale SSR estimate. These radiation observations were
collected at two networks. The first set was the radiation observations
(with temporal resolution of 1 min) measured at 42 BSRN stations (Ohmura et al., 1998), which were marked as
red crosses in Fig. 1. Radiation observations measured at BSRN stations
are regarded as the most reliable radiation data due to the instruments of
highest available accuracy and careful maintenance (see website:
Distribution of radiation measurement stations used to evaluate the performance of the estimated SSR. The blue circles mark the locations of the 90 CMA radiation stations, and the red crosses mark those of the 42 BSRN stations. Note that two stations (labeled as DAR and DWN) in Australia and two stations (labeled as BIL and E13) in North America are very close to each other.
The second set was the daily radiation observations measured at 90 CMA radiation stations, which are denoted by black circles in Fig. 1. Though the pyranometers used to measure global radiation at CMA radiation stations were calibrated by a series of standard procedures (Yang et al., 2008), the observed radiation data collected at CMA radiation stations frequently include questionable values, which may have been a result of improper operation of instruments and/or instrument defects (Shi et al., 2008). To reduce the uncertainty caused by the questionable radiation data, we used a quality-check procedure (Tang et al., 2010) to exclude the spurious and erroneous measurements. The quality-check procedure consists of two steps. One is the physical threshold test to eliminate the obvious errors, and the other is the statistical test using an artificial neural network method to eliminate the more insidious errors. More detailed information about the two-step procedure can be found in the article of Tang et al. (2010).
Firstly, the estimated SSR were validated against the observations measured at the 42 BSRN stations at both instantaneous and daily scales. To reduce the uncertainties induced by broken clouds, we validated the estimated instantaneous SSR against hourly mean observed ones centered on the time of satellite overpass, according to the suggestion of Wang and Pinker (2009). To examine the effect of different spatial resolutions on the accuracy of our SSR estimates, in addition to the 10 km spatial resolution, we also evaluated our estimated SSR at spatial resolutions of 30, 50, 70, 90 and 110 km, derived by averaging the SSR values observed at the original scale of 10 km.
Accuracy for instantaneous SSR at 90 km scale (RMSE
Comparisons of our estimated instantaneous SSR at spatial
resolutions of
Effect of spatial resolution on accuracy of our estimated instantaneous SSR compared to observations at the 42 BSRN stations. A comparison with instantaneous SSR of ISCCP-FD is also shown.
To further illustrate this issue, the performances of our instantaneous SSR with different spatial resolutions at the 42 BSRN stations were given in Table 1, which suggests that the accuracy was clearly improved when the data were upscaled to 30 km, with a further slight improvement at 70 km, but that accuracy started to decrease at 90 km. The performance of the ISCCP-FD was also presented in Table 1. Apparently, the accuracy of our estimated instantaneous SSR is significantly higher than that of the ISCCP-FD. A further advantage of our dataset is that its spatial resolution is far higher than that of the ISCCP-FD products.
Figure 3 shows the spatial distribution of RMSE for the estimated
instantaneous SSR (spatial resolution 90 km) at all individual BSRN
stations. The RMSE was
Spatial distribution of RMSE (W m
Figure 4 presents the validation results for our estimated daily SSR at 42
BSRN stations. The MBE values were
Comparisons of our estimated daily SSR at spatial resolutions of
Effect of spatial resolution on accuracy of our estimated daily SSR compared to observations at 42 BSRN stations. A comparison with daily SSR of ISCCP-FD is also shown.
The spatial distribution of RMSE for our estimated daily SSR at spatial
resolution of 90 km (Fig. 5) showed that RMSE at most of the 42 BSRN
stations were
Spatial distribution of RMSE (W m
GWEWX-SRB and CERES are two other well-known and widely used global
satellite radiation products. Zhang et al. (2013; Fig. 8) evaluated the
performance of GEWEX-SRB SSR products with the mean 3 h observed data from
the BSRN and found that RMSEs for the instantaneous and daily SSR of
GEWEX-SRB were 88.3 and 35.5 W m
Thus, our estimated SSR based on ISCCP-HXG cloud products provided a more accurate, higher spatial resolution dataset than those of ISCCP-FD, GEWEX-SRB and CERES products.
Our estimated SSR were further evaluated against the observations collected
at the 90 CMA radiation stations at both daily and monthly scales. Figure 7
presents the validation results for the estimated daily SSR at spatial
resolutions of 10 and 90 km. The MBE, RMSE and
Comparison of CERES SSR products with observed SSR at 42 BSRN
stations for both
Comparisons of our estimated daily SSR at spatial resolutions of
Table 3 shows that the accuracy of our estimates of daily SSR clearly
improved when upscaled to 30 km spatial resolution and were most accurate at
90 km spatial resolution. RMSE and
Effect of spatial resolution on accuracy of our estimated daily SSR compared to observations at 90 CMA radiation stations. A comparison with daily SSR of ISCCP-FD is also shown.
Effect of spatial resolution on accuracy of our estimated monthly SSR compared to observations at 90 CMA radiation stations. A comparison with monthly SSR of ISCCP-FD data is also shown.
Spatial distribution of RMSE (W m
Figure 9 presents the validation results for our estimated monthly SSR. The
MBE, RMSE and
Comparisons of our estimated monthly SSR at spatial resolutions of
Comparison of CERES
Spatial distribution of global annual mean SSR (W m
The performances for CERES daily and monthly SSR were evaluated against observations at the 90 CMA radiation stations (Fig. 10) and also compared with those of our estimates from ISCCP-HXG (Table 4). The MBEs for CERES daily and monthly SSR were greater than those of our estimates at all scales, and the RMSE for CERES daily SSR was slightly smaller than that of our estimates at 10 km spatial resolution, but obviously greater than our estimates at spatial resolutions from 30 to 110 km. The RMSE for CERES monthly SSR was greater than those of our estimates at all scales. Thus, the accuracy of our estimates is generally higher than that of CERES.
Figure 11 presents the comparison of the global distribution of the annual mean SSR in 2009 between our retrievals and the ISCCP-FD SSR product. From the figure, it can be seen that the global distribution for our SSR estimate based on the ISCCP-HXG cloud products is almost the same as that of the ISCCP-FD SSR product, but the spatial resolution of our estimate is far higher than that of ISCCP-FD. There is no doubt that we can get more details that the coarse resolution product ISCCP-FD can not capture. For example, the region of high SSR clearly identified over the Tibetan Plateau by our estimate (Fig. 11a) is barely discernible in the ISCCP-FD-derived data (Fig. 11b). The high values are mainly from around the Equator and the low latitudes, and the low values are mainly from over the high latitudes and the Arctic and Antarctic regions. This phenomenon is primarily determined by the solar elevation angle. In addition, the relatively high values are also found over the Bolivian Plateau, the Tibetan Plateau and other high-altitude regions due to less radiative extinction over high altitudes.
The 16-year dataset of global SSR is available at the National Tibetan
Plateau Data Center (
This study produced a 16-year (2000–2015) global dataset of SSR (with
resolutions of 3 h and 10 km) based on recently updated ISCCP H-series cloud
products, new ERA5 reanalysis data and MODIS albedo and aerosol products
with a physically based scheme. The retrieved SSR dataset was evaluated
globally with observations collected at BSRN and CMA radiation stations.
Validation against observations collected at BSRN showed that the MBE and
RMSE were
The spatial resolution and accuracy of the new dataset are both higher than those of the global satellite radiation products of GEWEX-SRB, ISCCP-FD and CERES and will contribute to photovoltaic applications and research related to simulation of land surface processes. When reliable global aerosol and albedo datasets become available, we intend to expand our dataset of SSR estimates back to mid-1983. We also plan to expand the dataset beyond 2015 by using SSR estimates from new-generation geostationary satellites.
All authors discussed the results and contributed to the paper. WT calculated the dataset, analyzed the results, and drafted the paper.
The authors declare that they have no conflict of interest.
The CMA radiation station data were obtained from the
National Meteorological Information Center (NMIC), and the ISCCP-HXG cloud
products were obtained from the NOAA's National Centers for Environmental
Information (NCEI). The ERA5 reanalysis data and MODIS albedo and aerosol
data were downloaded from official websites (
This work was supported by the National Key Research and Development Program of China (grant nos. 2018YFA0605400 and 2017YFA0603604), the National Natural Science Foundation of China (grant nos. 41671372), the Youth Innovation Promotion Association CAS (no. 2017100), the 13th Five-Year Informatization Project of the Chinese Academy of Sciences (grant no. XXH13505-06), and the Strategic Priority Research Program of the Chinese Academy of Sciences (grant no. XDA20100102).
This paper was edited by Yasuhiro Murayama and reviewed by Guanghui Huang and one anonymous referee.