Mapping high-resolution surface shortwave radiation over East Asia with the new generation geostationary meteorological satellite Himawari-8

ABSTRACT Surface shortwave radiation (SSR) plays an important role in global energy systems. The new generation of geostationary meteorological satellite Himawari-8, with higher spatiotemporal and spectral resolution, offers a new opportunity to retrieve SSR with higher accuracy. In this study, an improved algorithm was applied to estimate instantaneous, hourly, and daily mean SSR using cloud products from the Advanced Himawari Imager (AHI) onboard the Himawari-8 satellite. The validation against Baseline Surface Radiation Network (BSRN) stations showed a root mean square error (RMSE) of 95.8 W m−2 for instantaneous SSR, 82.4 W m−2 for hourly SSR, and 22.8 W m−2 for daily SSR and mean bias error (MBE) of −15.8 W m−2, −14.1 W m−2, and −6.6 W m−2. The validation against China Meteorological Administration (CMA) stations showed a RMSE of 99.5 W m−2 and MBE of −8.2 W m−2 for hourly SSR and RMSE of 27.7 W m−2 and MBE of −3.9 W m−2 for daily SSR, which are generally better than the Himawari-8 SSR product. Overall, the improved algorithm performed well on the new-generation geostationary satellite, with high accuracy and efficiency, and would contribute to surface process research and photovoltaic engineering applications.


Introduction
The solar radiation that reaches the Earth's surface in the wavelength range of 300-4000 nm is called surface shortwave radiation (SSR) (Huang et al. 2019).SSR plays an important role in the Earth's ecosystem and energy disposition.As the main energy source of Earth's system, SSR can promote surface energy exchange and radiation balance through interactions with the Earth's atmosphere, hydrosphere, biosphere, and other spheres (Wild 2009;Besharat, Dehghan, and Faghih 2013).In addition, solar-power generation systems, solar powered irrigation systems, and building lighting designs in human life are all dependent on the magnitude of SSR (Duzen and Aydin 2012;Wang et al. 2013;Teke, Yildirim, and Celik 2015).Therefore, accurate estimation of SSR is of great importance for studies on global radiation balance, surface process changes, and photovoltaic generation systems.
There are many methods for obtaining SSR, which can be divided into four main categories: ground observation, estimation based on ground-based meteorological stations, numerical simulations, and satellite remote sensing (Liang et al. 2019).Compared to the other three methods, satellite remote sensing has the advantage of extensive and rapid observation of targets.Therefore, in recent years, satellite remote sensing has become a common method for acquiring SSR.In particular, it is difficult to obtain ground measurements in harsh areas.The absence of observations in such regions will impact the calculation of global and local radiation budgets.Therefore, satellite remote sensing is a key and effective means of studying the local radiation balance (Zhang et al. 2014;Vindel et al. 2016).Satellite remote sensing can directly construct an empirical relationship between radiation and acquired satellite data to calculate SSR.On the other hand, SSR can also be retrieved by combining satellite data with radiation transfer theory.Moreover, satellite remote sensing can monitor clouds and aerosols dynamically compared with numerical simulation, so the SSR simulation based on satellite remote sensing has higher accuracy than that based on numerical simulation (Zhang et al. 2015;Zhang et al. 2016;Huang et al. 2019).
With the establishment of the International Satellite Cloud Climate Program (ISCCP) and Earth Radiation Balance Experiment (ERBE) in the 1980s, studies on radiation inversion based on satellite remote sensing have developed rapidly (Schiffer and Rossow 1983;Schiffer and Rossow 1985).Many algorithms have been developed based on the ISCCP and ERBE tasks, and the most widely used of which are the look-up table (LUT) algorithm proposed by Pinker and Laszlo (1992) and parameterization algorithm proposed by Li et al. (1993a).In the LUT algorithm, the solar radiation flux reaching the surface is calculated by establishing the relationship between the atmospheric transmittance and the TOA planetary albedo acquired by the satellite.Li et al. (1993a) developed a parametric relationship between the TOA irradiance and net surface solar irradiance for different atmospheric conditions.Using the solar zenith angle as an explicit input parameter, the net surface solar irradiance, a function of SSR and surface albedo can be estimated as long as the precipitable water and the albedo at the TOA are known.Both algorithms have been applied to produce multiyear global SSR products (Ma and Pinker 2012;Trentmann and Kothe 2016).
In the twenty-first century, methods and means of estimating SSR based on satellite remote sensing have matured.Various retrieval models consider different assumptions and influencing factors.The input factors are gradually expanded to include a number of meteorological and geographical parameters that affect solar radiation, such as planetary albedo, precipitable water, latitude, and longitude.Input sources were gradually expanded.This is partly because of the increasing accuracy of modern satellite technologies and sensors.They can detect the world regardless of geographic restrictions.With the increasing pervasiveness of SSR models and their wider application, the radiation areas that can be simulated have gradually expanded.Many global SSR products have emerged, including the International Satellite Cloud Climatology Project-Flux Data (ISCCP-FD), ERBE, Clouds and the Earth's Radiant Energy System (CERES), Global Energy and Water Exchanges-Surface Radiation Budget (GEWEX-SRB), the University of Maryland-Shortwave Radiation Budget (UMD-SRB) and MCD18A1, etc. (Li and Leighton 1993b;Gupta et al. 1999;Ma, Pinker, and Zhang 2007;Jia et al. 2018;Wang et al. 2020).The spatial resolution of these radiation products was relatively coarse, mostly at 1°or 2.5°.
Since 2014, a new generation of geostationary satellites with high spatial and temporal resolution, including FengYun-4, Geostationary Operational Environmental Satellites-R (GOES-R), and Himawari-8, has been successfully launched (Bessho et al. 2016;Yang et al. 2017).As the world's first geostationary meteorological satellite that can acquire color observation images, Himawari-8 carries a new sensor, the Advanced Himawari Imager (AHI), with 16 observation channels, including three visible, three near-infrared, and ten infrared bands.Compared with the previous generation of geostationary satellites, Himawari-8's observation frequency increased from once every 30 min to once every 10 min.For some specific regions, the temporal resolution can be increased to 2.5 or 0.5 min.Based on these improvements, many researchers have evaluated the radiation products of new satellites and developed new algorithms to estimate SSR based on new satellite products.Yu et al. (2018) compared the Himawari-8 SSR product with four other SSR products, such as the Modern-Era Retrospective analysis for Research and Applications Version 2 (MERRA-2) SSR products, finding that Himawari-8 showed the highest level of accuracy among the four chosen datasets.Ma et al. (2020) also developed a hybrid method to estimate the SSR for the new-generation Himawari-8 geostationary satellite.Overall, the higher spatiotemporal and spectral resolution of the Himawari-8 satellite has led to a significant increase in its ability to monitor clouds and aerosols, providing an important opportunity to estimate more accurate SSR (Bessho et al. 2016;Letu et al. 2020).
In our previous study, we developed an improved algorithm based on a physically based model (Li et al. 2022).Qin et al. (2015) developed the original algorithm.Tang et al. (2016) applied this algorithm to estimate instantaneous SSR on a global scale using MODIS Level-2 atmospheric and land products as driving data and validated the estimates at 44 Baseline Surface Radiation Network (BSRN) stations.The results showed an underestimation of the SSR in the presence of thick clouds.Therefore, we replaced the original cloud parameterization with that of FARMS developed by Xie, Sengupta, and Dudhia (2016).The cloud parameterization of FARMS is derived by fitting an exponential function to the cloud transmittance and reflectance produced by radiative transfer simulations, which provides an opportunity to improve the performance of Qin et al.'s model in heavy clouds.Combined with the clear-sky model developed by Qin et al. (2015) and the cloud parameterized scheme of FARMS, the improved algorithm was preliminarily evaluated using MODIS atmospheric and land products (Li et al. 2022).It was found that the improved algorithm estimates SSR with higher accuracy than that of the original model and does not significantly underestimate SSR under heavy clouds, suggesting that the improved algorithm can be potentially applied to next-generation geostationary satellite data.
In this study, the improved algorithm was used to map SSR in East Asia using products from the Himawari-8 satellite.In addition to the instantaneous SSR, hourly and daily average SSR were also retrieved.The remainder of this paper is organized as follows.Section 2 describes the improved algorithm used to map high-resolution SSR in East Asia.Section 3 introduces the data used in this research, including input data, validation data, and the Himawari-8 official SSR product.The results are discussed in Section 4. Finally, the conclusions are presented in Section 5. and input data.

Algorithm
The extinction of shortwave radiation in the atmosphere under clear sky conditions is mainly due to the scattering and absorption of gas molecules and suspended particles.However, when clouds are present in the sky, the scattering and absorption of clouds must also be considered.Compared with the original model, the cloud algorithm of the improved model uses computation from more discrete ordinate streams than simple forward and backward scattering analysis to calculate the cloud transmittances and reflectance.This makes the simulation of the reduction of shortwave radiation by clouds more accurate.The improved algorithm performed a separate and complete analysis of each extinction process, including cloud scattering and absorption, aerosol scattering and absorption, ozone absorption, homogeneous mixed gas absorption, and Rayleigh scattering.
The SSR estimation process under all-sky conditions (R all ) is expressed as follows: where C denotes the cloud fraction; R clr and R cld represent the SSR under clear-sky and cloudy conditions, respectively; m and R 0 denote the cosine of the solar zenith angle and the solar radiation flux at the TOA, respectively; t clr b , t clr d , r clr a , and r g refer to broadband direct radiative transmittance, diffuse radiative transmittance, atmospheric spherical albedo under clear skies, and surface albedo, respectively; and t cld b , t cld d , and r cld c represent the cloud direct transmittance, cloud diffuse transmittance, and cloud diffuse reflectance, respectively.The diffuse transmittance of the transmitted diffuse radiation through the cloud layer in the atmosphere is expressed as t clr dd .The detailed calculation methods for these parameters can be found in Li et al. (2022).
The input parameters of the improved model included the aerosol Ångström turbidity coefficient, cloud optical thickness, cloud particle effective radius, cloud fraction, solar zenith angle, total ozone amount, precipitable water, surface albedo, and surface pressure.Notably, the aerosol Ångström turbidity coefficient (β) is mainly obtained from the aerosol optical thickness estimation.Therefore, the influence of the aerosol optical thickness on the estimation results is mainly reflected by the aerosol turbidity coefficient.

Input data
As mentioned previously, the level-2 Himawari-8 cloud product provided by the Japan Meteorological Agency (JMA) was applied for SSR estimation in this study.The cloud product had spatial and temporal resolutions of 5 km and 10 min, respectively.Cloud parameters, including cloud optical thickness, cloud-effective radius, cloud top temperature, and cloud type, are extracted from the product.A cloud-type value of zero represents a cloudless sky.When the cloud type value is greater than zero, clouds are present in the sky.If the cloud top temperature is less than 253.1 K, the cloud type can be further judged as ice clouds and vice versa as water clouds (Tang et al. 2019).
In addition to the Himawari-8 cloud product, the MODIS atmospheric profile product (MYD08_D3), albedo product (MCD43C3), and ERA5 reanalysis data were used as auxiliary data in the estimation process.The data information is presented in Table 1.The MYD08_D3 product is a daily product with a spatial resolution of 1°and is mainly used to obtain information on aerosol loading.The temporal and spatial resolutions of the surface albedo data were 16 d and 5 km, respectively.The ERA5 product has a spatial resolution of 25 km and temporal resolution of 1 h, which is higher than that of other commonly used reanalysis data.This means that the ERA5 reanalysis can better describe the atmospheric conditions.Therefore, the atmospheric parameters required for the model, including surface pressure, precipitable water, and total ozone amount, are mainly obtained from the ERA5 product.To match the spatial and temporal resolutions of the Himawari-8 product, the auxiliary data were resampled to the same resolution as the cloud product.In addition, all input data cover the entire year 2017.

Validation data
In this study, we chose the BSRN and China Meteorological Administration (CMA) ground measurements for comparison with the estimated SSR. BSRN observations were first selected due to their high quality and high temporal resolution.To correspond to Himawari-8's observation range (80°E-160°W, 60°N-60°S), we selected ten BSRN stations.Figure 1 shows the locations of these stations.Detailed geographic information for these stations is presented in Table 2. To enhance the temporal representativeness of each site, we averaged the observations for 60 min for instantaneous SSR validation.This can also help reduce the impact of satellite imaging time.SSR data from 120 CMA stations were also used to validate our estimates.The spatial distribution of the stations is shown in Figure 1.The CMA established its own solar radiation observation network since 1957.The observed radiation parameters included global radiation, direct radiation, diffuse radiation, net radiation, and reflected radiation.Global radiation measurements were used in this study.The observations were quality-corrected using the quality-control scheme of Tang et al. (2010) to remove obviously erroneous and abnormal values.The uncertainty of the CMA radiation measurements after quality control was within 5% (Tang et al. 2019).

Himawari-8 official SSR product
In this study, an official Himawari-8 SSR product was selected for comparison with the validation results of the improved model.The Level-2 SSR product chosen in this section was produced using the algorithm proposed by Frouin and Murakami for estimating PAR (Frouin and Murakami 2007).The model uses plane-parallel radiative transfer theory based on the assumption that the effects of a clean atmosphere and clouds are separated.Notably, the same model produces two sets of SSR products with different spatial resolutions and coverage.The SSR product used in this study had a spatial and temporal resolution of 5 km and 10 min, covering the longitude range of 80°E to 160°W and latitude coverage range of 60°S to 60°N.

Validation of the results at bSRN stations
As mentioned above, the instantaneous, hourly mean, and daily mean SSR were estimated using the Himawari-8 product.The estimated SSR were first compared with in-situ observations collected at BSRN stations.Figure 2 shows the validation results for the instantaneous SSR.It can be seen that the correlation coefficient (R) between the estimates and observations reaches 0.95.The mean bias error (MBE) and root mean square error (RMSE) are −15.8W m −2 and 95.8 W m −2 , respectively.As shown in Table 3, the RMSE was lower than the results of Ma et al. (2020), who also estimated instantaneous SSR using the Himawai-8 Level-2 cloud product by a hybrid method combining a deep neural network (DNN) and radiative transfer model (RTM).The RMSE of the SSR estimated using the hybrid method was 117 W m −2 .The coefficient of determination (R 2 ) was 0.85, while the R 2 of this study was 0.9.This reveals that the instantaneous SSR simulated by the improved algorithm is in close agreement with ground measurements.
To further investigate the source of the negative MBE, the validation results of the instantaneous SSR at each BSRN site are shown in Figure 3. Owing to the few observations at the kwa site, the error metrics for this site were not analyzed.From the figure, we can see that the RMSEs range from 73.3-141.7 W m −2 at different stations.The MBEs range from −60.8-2.3W m −2 .The stations with high MBE and RMSE are mainly from the coc site located in the Cocos Islands and the ish site near eastern Taiwan Island, most likely caused by the quality of the cloud and aerosol input products.Both stations were located on offshore islands and along the coast.The turbid water in the offshore region makes the subsurface extremely complex and poses difficulties in detecting aerosol optical properties (Anderson et al. 2013).In addition, the high cloud frequency and complex cloud distribution at these two sites, located in the tropics, may affect the retrieval accuracy of cloud microphysical parameters, such as cloud optical thickness and cloud particle effective radius.We also compared the validation results at the eight BSRN stations between Yu et al.'s study and our algorithm (Table 4).Yu et al. (2018) evaluated the Himawari-8 SSR product with a spatial resolution of 5 km using BSRN measurements.The comparison shows that the RMSEs and MBEs for our SSR retrievals are generally lower than those of the Himawari-8 SSR product, except for a few sites.At the dwn site, the RMSE and absolute MBE of our estimates were lower than those of Yu et al.'s study by 40 W m −2 .
For some hydrological and evapotranspiration models, hourly SSR and daily SSR are also important input parameters (Alexandrov and Hoogenboom 2000;Chen, Chen, and Ju 2007).Based on the estimated instantaneous SSR with a 10-minute resolution, the hourly mean and daily mean SSR can be derived by averaging.The validation results for the hourly and daily mean SSR are shown in Figure 4.For the hourly results, the overall RMSE, MBE, and R were 82.4 W m −2 , −14.1 W m −2 , and 0.96, respectively.The daily results had an overall RMSE of 22.8 W m −2 , an MBE of −6.6 W m −2 , and an R of 0.98.As shown in Table 3, our retrievals have a better performance than Ma et al.'s study, except for the hourly validation results, which have a slightly higher RMSE than Ma et al.'s results.This is mainly because the estimated SSR data greater than or equal to five were used to calculate the hourly SSR in Ma et al.'s study, whereas the number in this study was three.In addition, based on the high temporal resolution of the Himawari-8 satellite, we also made an observation of the annual mean diurnal variation between the estimated SSR and the observations at each BSRN site.As is shown in Figure 5, the trends of the retrievals and observations at each station are very close, especially in the morning and afternoon.The deviation between the observed and estimated SSR tends to increase and then decrease as the time of day varies.Significant negative deviations occur mainly around noon, still for the coc and ish sites, where they are more pronounced.However, in general, the estimated results of the improved algorithm and the observed values are in good agreement at most stations.

Validation of the results at CMA stations
Due to man-made air pollution, aerosol concentrations are high in the eastern region of China.The aerosol number concentration is closely related to the cloud droplet number concentration and has a direct impact on the microphysical characteristics and radiative effects of clouds (Norris and Wild 2007;Rosenfeld et al. 2014;Tang et al. 2017).Therefore, the estimates from the improved algorithm were also validated using ground observations at 120 CMA stations.To quantitively evaluate the performance of the improved model, the validation results of the improved model are compared with the official Himawari-8 SSR product.The comparison results are shown in Figure 6.The left panel represents the hourly and daily validation results of the improved algorithm, whereas the right panel shows the hourly and daily validation results of the Himawari-8 official SSR product.The Himawari-8 official SSR product shows a significant overestimation (the MBEs of hourly and daily results of the Himawari-8 official product are 26.2W m −2 and 12.3 W m −2 ).Comparatively, the results of the improved algorithm have lower MBE values (The MBEs are −8.2W m −2 and −3.9 W m −2 , respectively).The RMSE of our estimated hourly SSR is 99.5 W m −2 , slightly lower than that of the Himawari-8 official product.Likewise, the RMSE of our daily retrievals was 27.7 W m −2 , while that of the official product was 29.3 W m −2 .This proves that in CMA stations with high aerosol concentrations, the improved algorithm performs better than the official Himawari-8 SSR product.

Spatial distribution of seasonal and annual mean SSR in 2017
Based on the estimated SSR data in 2017, we calculated the seasonal and annual mean SSR to observe the SSR distribution over East Asia.As shown in Figure 7, the SSR in the Western Pacific between the equator and 30°N exceeds 260 W m −2 in summer, whereas the SSR in the region between 40°N and 60°N is below 40 W m −2 in winter.This discrepancy proves that the distribution of SSR is significantly influenced by season and latitude, mainly caused by variations in the solar zenith angle.As the season changes and latitude increases, the solar zenith angle becomes lower, resulting in a  corresponding decrease in radiation values.In addition, it can be seen that areas located in the subtropical high-pressure belt are rich in SSR throughout the year, especially the Western Pacific and Australia.In the western Pacific, the central values of the two high SSR centers were greater than 280 W m −2 .The Australian continent had the highest SSR value in excess of 300 W m −2 .
The annual mean SSR distribution in China was very uneven.The highest annual mean SSR in China is located on the Tibetan Plateau, with a value of approximately 260 W m −2 .This is mainly due to the high altitude and thin air of the Tibetan Plateau, causing less extinction of solar radiation from the TOA to the ground than in other areas.In the 30 °N region, China has a lower annual SSR than other areas.Influenced by topography, there is more water vapor in the atmosphere in this region, resulting in stronger scattering and absorption of solar radiation.

Conclusion
In this study, we extended the improved algorithm for SSR estimation to the new-generation geostationary satellite Himawari-8 to map SSR with high spatial and temporal resolution in East Asia.The algorithm optimized the calculation of cloud transmittance and reflectance by introducing a cloud parameterization scheme over an efficient physically based model (Li et al. 2022).Validation at ten BSRN stations showed an R value of 0.95, MBE of −15.8 W m −2 , and RMSE of 95.8 W m −2 for the instantaneous results; an R value of 0.96, MBE of −14.1 W m −2 , and RMSE of 82.4 W m −2 for the hourly mean SSR; an R value of 0.98, MBE of −6.6 W m −2 , and RMSE of 22.8 W m −2 for the daily mean SSR.The estimates and observations are in close agreement at most BSRN sites.And the estimates obtained using the improved algorithm are generally more accurate than those obtained using other retrievals in previous studies.Moreover, the hourly and daily mean SSR estimated by our algorithm were compared with the existing Himawari-8 product at 120 CMA stations, revealing that our algorithm generally performed better than the official Himawari-8 SSR product.Influenced by variations in the solar zenith angle, the spatial distribution of the estimated seasonal and annual mean SSR in East Asia shows a clear seasonal and latitudinal distribution of SSR.The radiation values were high in the Western Pacific and Australia throughout the year and were located in areas of subtropical high pressure.However, the radiation distribution in the Chinese region shows unevenness.
During the estimation, the high errors of SSR at coastal and island stations were presumed to be a quality problem in inputs of aerosol and cloud products.The aerosol data used in the algorithm are mainly MODIS retrievals, and a large number of missing values are filled with the monthly mean climatology in the GOCART model (Chin et al. 2002), which may affect the accuracy of the calculated SSR to some extent.Attempts will be made in future to improve the accuracy of estimates by introducing more continuous aerosol data with fewer missing measurements, such as the MERRA-2 reanalysis aerosol data (Gueymard and Yang 2019).More accurate aerosol merging products are also in consideration (Yang and Gueymard 2020).

Figure 1 .
Figure 1.Spatial distribution of 10 BSRN stations and 120 CMA stations used in this study.Red pentagrams represent BSRN station locations.Blue crosses indicate CMA station locations.

Figure 2 .
Figure 2. Validation of the estimated instantaneous SSR at 10 BSRN stations in 2017.Units of MBE and RMSE are W m −2 .

Figure 3 .
Figure 3.Comparison of SSR between the estimates and observations at each BSRN station.The units of MBE and RMSE are W m −2 .

Figure 4 .
Figure 4. Similar to Figure 2 but for hourly and daily SSR.

Figure 5 .
Figure5.The diurnal variations between the estimated SSR and BSRN measurements.

Figure 6 .
Figure 6.Validation of the estimated hourly and daily SSR against the observed SSR at all CMA radiation stations.(a) and (c) present validation of the hourly and daily SSR estimated by our algorithm, while (b) and (d) show the validation of hourly and daily SSR in the Himawari-8 product.The units of MBE and RMSE are W m −2 .

Figure 7 .
Figure 7. Spatial distribution of seasonal mean and annual mean SSR estimated based on Himawari-8 data over East Asia in 2017.Panel (a)-(d) denote spring mean, summer mean, autumn mean, and winter mean, respectively.Panel (e) represent annual mean.MAM represents March, April, and May.JJA includes June, July, and August.SON denotes September, October, and November while DJF is December, January, and February.The unit of SSR is W m −2 .

Table 1 .
Basic information for the Himawari-8 product and other auxiliary data used in this study.

Table 2 .
Geographic information for ten BSRN stations.