Offline Correction of CMIP6 HighResMIP Simulated Surface Solar Irradiance With 3D Sub‐Grid Terrain Radiative Effects

The surface solar irradiance (SSI) is crucial for the land‐atmosphere processes and remarkably affected by the topography over the rugged areas. However, the Coupled Model Intercomparison Project Phase 6 (CMIP6) HighResMIP models adopting the parallel‐plane radiative scheme without considering the sub‐grid terrain solar radiative effects (3DSTSRE) overestimate the SSI in the rugged areas and the overestimation increases with the sub‐grid terrain complexity. To reduce the biases of the SSI simulations, this study offline corrects the SSI simulations of CMIP6 HighResMIP models by a 3DSTSRE scheme. Results show that the SSI biases produced by the HighResMIP models in the rugged regions can be significantly reduced by adopting the 3DSTSRE offline correction, and the improvements increase with the sub‐grid terrain complexity, indicating that considering the 3DSTSRE in the climate models to improve the SSI simulations over rugged areas is necessary.


Introduction
Surface solar irradiance (SSI) is decisive to the surface energy balance, crucial for the land-atmosphere processes, and the ultimate energy driving the atmospheric motion (Liang et al., 2019;Wild, 2017).The SSI is regulated by the atmospheric and land surface conditions, such as the cloud, aerosol, gas molecule, and topography (Chakraborty & Lee, 2021;Firozjaei et al., 2020;Yang et al., 2016).The solar radiative effects of the atmospheric molecule, cloud, and aerosol are relatively homogenous over the region with tens of kilometers in diameter (Matus & L'Ecuyer, 2017;Yu et al., 2021).However, the terrains considerably increase the heterogenization in the spatial-temporal distribution of the SSI at the local scale (Nunez, 1980;Olson et al., 2019;Proy et al., 1989).The SSI fluxes over the rugged areas are reduced by the shading effects of local and surrounding terrains and added by the reflected solar radiation from the adjacent terrain (Chu et al., 2021;Dozier & Frew, 1990;Hay, 1993;Li et al., 2002).The solar radiation reflected from the surrounding snow-covered terrain can exceed 180 W•m 2 and account 60% of the total SSI at the typical mid-latitude alpine valley (von Rütte et al., 2021).The late sunrise and early sunset generally occur in the valleys with the different configurations of the terrains, which shorten the daily sunshine duration by several minutes to hours (Adamu et al., 2019;Chen, 2020;Zhang et al., 2018).The shading and obstruction of the sky by the terrains also decreases the SSI value by the average magnitude of 10 1 W•m 2 and maximum magnitude of 10 2 W•m 2 (Hao et al., 2019;Wang et al., 2018;Yan et al., 2018).For example, the daily SSI over the lower part of the Arizona's Meteor Crater is 6% lower than that over the crater rim and the maximum decrease of instantaneous SSI is ∼300 W•m 2 (Hoch & Whiteman, 2015).
Due to the essential roles of the SSI in the earth system, to accurately describe the processes determining the SSI in the numerical models has long been an important task (Edwards, 2011;Giorgi et al., 1993).Previous efforts focused on improving the description of radiative forcing by the cloud, aerosol, and atmospheric molecules (Randles et al., 2013;Ritter & Geleyn, 1992;Yi et al., 2016;Zakey et al., 2006).With the continuous increase of model resolution, the sub-grid scale terrain solar radiative effects (STSRE) become significant and nonnegligible in the numerical models (Colette et al., 2003;Essery & Marks, 2007;Hauge & Hole, 2003;Ruiz-Arias et al., 2011).During the last two decades, numerous studies have attempted to consider the STSRE in the numerical models (Liou et al., 2007;Manners et al., 2012;Müller & Scherer, 2005;Zhang et al., 2006Zhang et al., , 2022)).Three kinds of STSRE schemes have been developed based on the 2-dimensional idealized terrain radiation theory, the 3-dimensional real terrain radiation theory, and the Monte Carlo photon tracking technology (Gu et al., 2020;He et al., 2019;Huang et al., 2022;Lai et al., 2010;Lee et al., 2011).Numerical studies show that those schemes decline the biases of simulated SSI by 10-100 of watts per square meter for the normal and extreme case over the rugged areas respectively, and the improvement of the simulated SSI increases along with the increase of the horizontal resolution and sub-grid terrain complexity (Arthur et al., 2018;Cai et al., 2023;Fan et al., 2019;Gu et al., 2012;Lee et al., 2013Lee et al., , 2019)).The improved simulation of SSI leads to the better performance of the numerical models in reproducing the surface energy balance and thereafter temperature, atmospheric circulation, precipitation, etcetera (Feng & Zhang, 2007;Gu et al., 2022;Hao et al., 2021).However, these attempts are mostly conducted in the regional weather/climate models or land surface models instead of the global climate models (GCM).Although the uncertainty of the SSI simulated by the GCMs over the rugged areas have not been clearly revealed due to the lack of in situ observation in the previous studies (Hu et al., 2019;Markovic et al., 2008;Niu et al., 2023;Wang et al., 2022;Wild, 2008;Xu et al., 2022), the GCMs with the parallel-plane radiative transfer scheme may share the problems in simulating the SSI over the rugged areas to the regional numerical models due to the absence of the STSRE.
The High Resolution Model Intercomparison Project (HighResMIP, Haarsma et al., 2016) for the World Climate Research Program Coupled Model Intercomparison Project Phase 6 (CMIP6) gathers the state-of-the-art GCMs.The CMIP6 SSI data can provide the forcing data for land surface models or hydrological models and have been widely applied in studies related to potential evapotranspiration, solar energy projection, grain yield projection, and so on (Dutta et al., 2022;He et al., 2022;Heavens, 2021;Kim et al., 2023;Liu et al., 2020;Zheng et al., 2018).The 3D sub-grid terrain solar radiative effects (3DSTSRE) are not considered in the CMIP6 HighResMIP models (Table 1).Since complex terrain covers ∼30% of the world's total land area, it is necessary to evaluate and correct the CMIP6 HighResMIP SSI over the rugged areas.In this study, we preliminarily use the 3DSTSRE scheme (Huang et al., 2022) to offline correct the SSI simulated by the CMIP6 HighResMIP models and investigate to what extent the offline correction of 3DSTSRE can reduce the uncertainty of the simulated SSI over the rugged areas.This study not only provides a method to correct the CMIP6 HighResMIP SSI for SSI-related studies over the areas with complex terrains but also lays a foundation for future studies on online simulations using the GCMs with the 3DSTSRE scheme.The rest of the paper is arranged as follows: Section 2 introduces the data, 3DSTSRE scheme, methodology, and metrics.Section 3 presents the results.Section 4 draws the conclusion and discusses the limitation of this study and planning of future work.

Data
The data used in this study are listed as follows: 1.  Note.The models with the star (*) provide both the monthly and 3-hourly surface solar irradiance (SSI) data while the other models only provide the monthly SSI data.

Geophysical Research Letters
10.1029/2023GL107737 4. SSI observations from the monthly CMSAF Cloud, Albedo and Surface Radiation data set (CLARA-A2, Karlsson et al., 2017) with a horizontal resolution of 0.25°, the daily Global LAnd Surface Satellite (GLASS) data set (Liang et al., 2013) with a horizontal resolution of ∼5 km, and the TerraClimate data set of monthly climate and climatic water balance (Abatzoglou et al., 2018) with a horizontal resolution of ∼4 km during 2001-2014.The production of CLARA-A2 SSI considers the terrain effects (Karlsson et al., 2013(Karlsson et al., , 2017)).The GLASS and TerraClimate data sets include the satellite observations with the high horizontal resolution of 500-1,000 m as source data, which present the SSI over the complex terrain to a certain extent.

3DSTSRE Scheme
As shown in Figure 1a, the SSI over the rugged areas (E t↓ ) is greatly impacted by the local and surrounding terrains.E t↓ includes three components, namely, the direct solar radiation (E dirt↓ ), diffuse solar radiation (E dift↓ ), and additional solar radiation reflected from the surrounding terrain (E reft↓ ) (Dozier & Frew, 1990).The E dirt↓ is influenced by the shadows cast by the surrounding topography and the slope and slope orientation of the local terrain (self-shielding).The E dift↓ is controlled by the fraction of visible sky in the entire sky which is indicated by the sky view factor (SVF) (Figure 1a, Dirksen et al., 2019).SVF is 1.0 for the plane surface and gradually decreases with the proportions of sky blocked by terrain increasing (Dozier & Frew, 1990).Lower SVF indicates more complex terrain and less SSI due to the obstruction of topography (Jiao et al., 2019).E reft↓ is determined by the surface albedo and the configuration of surrounding terrain.
Figure 1b shows that the sub-grid scale terrains with the horizontal resolution of 3″ fluctuate sharply within a model grid with the horizontal resolution of 0.5°.The sub-grid terrain can significantly affect the grid-scale SSI.However, the GCMs adopt the plane-parallel radiative transfer scheme which treats the sub-grid topography as plane surface and totally ignores the 3DSTSRE.As a model grid with a horizontal resolution of 0.5°contains ∼360,000 sub-grids with the horizontal resolution of 3″, to explicitly calculate the SSI considering the 3DSTSRE at the sub-grids would result in heavy computing burden.Therefore, Huang et al. ( 2022) developed a 3DSTSRE scheme to correct the SSI components calculated by the plane-parallel radiative transfer scheme based on the static grid-scale terrain correction factors derived from the SRTM DEM data with the horizontal resolution of 3″ and the grid-scale albedo derived from the MODIS data.Detailed introduction for the Huang 3DSTSRE scheme can be found in Text S1 and Figure S1 in Supporting Information S1.

Offline Correction
As shown in Figure 1c, the 3-hourly SSI simulations of the 4 HighResMIP models (BCC-CSM2-HR, EC-Earth3P-HR, FGOALS-f3-H, and HadGEM3-GC31-HH) are processed to monthly data.Then the monthly SSI simulations of the 11 HighResMIP models are remapped onto the Gaussian grids with a horizontal resolution of 0.5°(here after omitted as "Gaussian grids") using the first-order conservative remapping method (Jones, 1999).The monthly multi-model ensemble (MME) mean SSI simulations of the 11 and 4 HighResMIP models with the plane-parallel radiative transfer scheme are referred as the 11MME_PP and 4MME_PP data, respectively.The suffix "PP" stands for the plane-parallel radiative transfer scheme.
The 3-hourly SSI simulations of the 4 HighResMIP models are offline corrected by the 3DSTSRE scheme at their native grid spacing, then time averaged to the monthly data, and lastly remapped to the Gaussian grids of 0.5°.The monthly SSI fluxes derived from the 3DSTSRE offline corrected 3-hourly SSI simulations of the 4 HighResMIP models are referred as the 4MME_TOP data.The suffix "TOP" stands for including the radiative effects of the topography, indicating the data with the offline correction of 3DSTSRE scheme.The grid-scale direct solar radiation (E dir,p↓ ), diffuse solar radiation (E dif,p↓ ), and albedo (a p ) at the plane surface are necessary for the offline correction of 3DSTSRE.In this study, the E dir,p↓ and E dif,p↓ are derived from the 3-hourly SSI simulations of the four models following the method of the Community Land Model Version 5 (details in Text S2 Supporting Information S1).The grid-scale a p at the native grids of each of the four models is derived from the MODIS albedo.
Seven of the 11 models only provide the monthly SSI (Table 1), however, the Huang 3DSTSRE scheme cannot correct the monthly SSI without diurnal variation.To solve this issue, the ratios between the monthly SSI during 2001-2014 between the 4MME_TOP and 4MME_PP data are used to represent the monthly radiative effects of the sub-grid terrain.The ratio is calculated by: where SSI 11MME PP are the monthly SSI simulations of the 11MME_PP using the plane-parallel radiative transfer scheme.The SSI 11MME TOP with the 3DSTSRE is stored in the 11MME_TOP data.
The daily GLASS SSI are firstly processed to monthly data.Then the monthly SSI from the CLARA-A2, GLASS, and TerraClimate data sets are remapped to the Gaussian grids of 0.5°, finally, the averages of the monthly observed SSI from the three products at the Gaussian grids of 0.5°are referred as OBS data.

Metrics
The smaller grid-scale SVF indicate that the sub-grid terrain obstructs the sky over the grid to a greater extent and lead to much stronger impact on the SSI at the grid (Cai et al., 2023).The SSI over the grids with very flat sub-grid scale topographies (SVF ≥ 0.99) are not discussed because the offline 3DSTSRE correction has extremely small impacts on them.Due to the missing values in the observations over the high-latitude areas, the evaluations are conducted over the land of Reg1 (130°W-180°E, 50°S-50°N) indicated by the blue-dash rectangle of Figure 1d.
It is clear that most of the grids with strong 3DSTSRE are located in the study area.In addition, four sub-regions highlighted in the red-dashed rectangles are utilized to reveal how the offline correction of the 3DSTSRE scheme operates in various locations (Reg2: 70°E-110°E, 25°N-40°N; Reg3: 0°-55°E, 30°N-50°N; Reg4: 130°W-85°W , 10°N-50°N; Reg5: 85°W-60°W, 50°S-10°N).The relative root mean square error (RRMSE) and relative error (RE) are used to assess the SSI simulations.Smaller RRMSE and absolute RE indicate better simulation of the SSI in quantity.The RRMSE and RE are calculated by: Where O and S represent the observation and simulation, respectively, n is the sample size.

Results
From Figure 2, the CMIP6 HighResMIP models clearly overestimated the SSI fluxes at most of the grids with complex terrain compared to the OBS data, and the overestimation increases with the sub-grid terrain complexity increasing or the grid-scale SVF decreasing (Figures 1d,2f,and 2g).The spatial distribution and magnitude of the SSI biases in the 11MME_PP data remarkably resemble to those in the 4MME_PP (Figures 2f and 2g).The largest overestimation of the SSI is located over the Tibetan Plateau and the Andes with the value more than 45 W•m 2 .After the offline correction of 3DSTSRE scheme, the SSI fluxes over the rugged areas in the 4MME_TOP and 11MME_TOP data can be decreased by 5-45 W•m 2 compared to those in the 4MME_PP and 11MME_PP data with much larger decreased values over the grids with much more complex terrains (Figures 1d,2h,and 2i).The RRMSE of the SSI in the 4MME_TOP and 11MME_TOP data are considerably decreased compared to those in the 4MME_PP and 11MME_PP data, indicating that the offline correction of 3DSTSRE scheme can significantly improve the SSI fluxes simulated by the CMIP6 HighResMIP models (Figures 2j and  2k).From Figures 2a-2e, the SSI of the 4MME_PP, 11MME_PP, 4MME_TOP, 11MME_TOP and OBS data averaged over the grids with the SVF ≤ 0.99 are 209.6, 209.4, 200.9, and 200.7, and 187.8 W•m 2 , respectively.Compared to the 4MME_PP and 11MME_PP data, the RE of the SSI simulations in the 4MME_TOP and 11MME_TOP data at the grids with the SVF ≤ 0.99 can be reduced from 11.6% to 7.0% and from 11.5% to 6.9% with the reductions of 39.7% and 40.0%, respectively, and the RRMSE of the SSI simulations in the 4MME_TOP and 11MME_TOP data at the grids with the SVF ≤ 0.99 are reduced from 16.9% to 13.9% and from 16.7% to 13.6% with the reductions of 17.8% and 18.7%, respectively.
The RRMSE of the SSI simulations in the 4MME_PP and 11MME_PP in each sub-region all increases with the grid-scale SVF decreasing or sub-grid terrain complexity increasing (Figures 3a-3e), indicating the important impact of the 3DSTSRE and the necessity of the terrain correction.Adopting the offline correction of 3DSTSRE scheme can clearly reduce the RRMSE of the SSI simulations in the 4MME_PP and 11MME_PP over each subregion by 10%-50% with much larger reduction at the grids with much lower SVF (Figures 3f-3j).Meanwhile, the RRMSE of the SSI simulations in the 4MME_TOP and 11MME_TOP including the 3DSTSRE are not sensitive to the sub-grid terrain complexity (Figures 3a-3e), suggesting that the biases in the 4MME_PP and 11MME_PP data due to the absence of the 3DSTSRE can be eliminated by the terrain correction to some extent.However, the RRMSE ranging from 12% to 17% over the sub-regions have been retained in the 4MME_TOP and 11MME_TOP data, which are related to the insufficient description of clouds, aerosols, atmospheric components, etcetera (He et al., 2023;Zheng et al., 2023).
To further illustrate the distribution of the errors in SSI simulations, the probability density function (PDF) and the cumulative density function (CDF) of the absolute relative error (ARE) of the monthly SSI simulations at the grids with the SVF ≤ 0.99 in sub-region Reg1 are illustrated in Figure 4. Compared to the 4MME_PP and 11MME_PP, the PDF of the ARE of the SSI simulations in the 4MME_TOP and 11MME_TOP increases at the AREs less than ∼12% and decreases at the AREs more than ∼12% with the largest increases at the ARE near 0 (Figures 4a and  4b), and the CDF of the ARE of the SSI simulations in the 4MME_TOP and 11MME_TOP is increased by 22.0% and 22.1% at the AREs less than 12% respectively (Figures 4c and 4d), indicating that the offline correction of 3DSTSRE can largely reduce the SSI simulation errors related to the terrain.
Overall, the CMIP6 HighResMIP models overestimate the SSI over the grids with the complex topographies and the overestimation increases with the complexity of the sub-grid terrain.The offline correction of 3DSTSRE scheme can effectively reduce the biases of SSI simulated by the parallel-plane radiative transfer scheme at areas with the complex sub-grid terrains and the improvements increase with the sub-grid terrain complexity.

Conclusion and Discussion
The SSI is very important in the land-atmosphere processes and significantly regulated by the topography.However, due to lack of ground-based SSI observations over the rugged areas, the performances of the simulated SSI over the complex terrain by the climate models are not well evaluated.The prosperously developed SSI data, based on observations from high-resolution satellite sensors, provide an invaluable opportunity to study SSI in rugged areas.In addition, the STSRE are still not well concerned in most of numerical models.This study uses the 3DSTSRE scheme to offline correct the simulated SSI fluxes from the CMIP6 HighResMIP Models and take the satellite-derived SSI data containing the information of terrain solar radiative effects as the reference data to evaluate the CMIP6 HighResMIP SSI fluxes over the rugged areas without and with the offline correction of 3DSTSRE scheme.
Without the offline correction of 3DSTSRE, the CMIP6 HighResMIP models obviously overestimate the SSI fluxes over the rugged areas between 50°S and 50°N globally in 2001-2014.The overestimation of the modeled SSI fluxes from each CMIP6 HighResMIP model and from the ensemble of the CMIP6 HighResMIP models increases sharply with the sub-grid scale terrain complexity.The RRMSE of the SSI fluxes from the MME mean of the CMIP6 HighResMIP Models increases from ∼15% to ∼30% with the sub-grid scale terrain complexity increasing.The offline correction of the 3DSTSRE scheme can effectively reduce the biases of simulated SSI fluxes from the MME mean of CMIP6 HighResMIP Models.For the grids with the most complex sub-grid terrain, the RRMSE of the SSI simulated by the CMIP6 models can be reduced by ∼50% due to the offline correction of 3DSTSRE scheme.Therefore, it is necessary, urgent, and beneficial to describe the 3DSTSRE into the climate modeling.

Geophysical Research Letters
10.1029/2023GL107737However, we have to admit that this study has two main shortcomings.One is that the normalized standard deviation of three observed monthly SSI is more than 10% over the Himalayas and Andes (Figure S3 in Supporting Information S1), indicating that the CLARA-A2, GLASS, and TerraClimate data sets exhibit discrepancies in the rugged areas.The simulated SSI are further evaluated by taking the CLARA-A2, GLASS, or TerraClimate alone as the reference data (Figures S4-S9 in Supporting Information S1).Figures S4-S9 in Supporting Information S1 also demonstrate that the RRMSE of the simulated SSI increases with the SVF decreasing, and the offline correction of the 3DSTSRE can clearly reduce the biases of the simulated SSI in the rugged areas, thereby enhancing the confidence in the offline correction of 3DSTSRE.The other one is that this study conducts an offline correction instead of applying the 3DSTSRE scheme to a certain numerical model, which neglects the important feedback in the land-atmosphere processes (such as the cloud, soil, albedo, snow cover, etcetera).We have adopted the 3DSTSRE scheme into the BCC-CSM2-HR model recently and the systematical impacts of the 3DSTSRE scheme on the performance of the BCC-CSM2-HR model will be reported in the future.The remained RRMSE of the SSI over the grids with different sub-grid terrain complexity after the offline correction of 3DSTSRE (Figure 3a) are ∼14%, which might come from the insufficient description of the other processes such as the clouds and aerosol.In addition, the reasons for the significant uncertainty in the simulated SSI over the regions with relatively flat terrains (such as the Amazon basin and the Changtang Plateau) should be investigated in the future.

Data Availability Statement
Data: The SRTM DEM data are provided by the International Center for Tropical Agriculture (CIAT) at https:// developers.google.com/earth-engine/datasets/catalog/CGIAR_SRTM90_V4;The daily GLASS SSI data (Liang et al., 2013) are retrieved from http://glass.umd.edu/Download.html;The monthly CLARA-A2 SSI data (Karlsson et al., 2020); The monthly TerraClimate SSI data (Abatzoglou et al., 2017); The monthly and 3-hourly   (No. 124 in 2023).We appreciate the High Performance Computing Center of Nanjing University and the National Key Scientific and Technological Infrastructure project "Earth System Numerical Simulation Facility" (EarthLab) for providing us the computing resource.We are grateful to ESGF, NASA, CIAT, UMD, EUMETSAT, and Abatzoglou et al. (2018) for allowing us to use their data sets.We show our warm appreciation and deepest respect to the editor and two anonymous reviewers for their constructive suggestions to greatly improve the manuscript.

Figure 1 .
Figure 1.The components of the surface solar irradiance over the rugged terrain (a), the sample diagram of the grid with the horizontal resolution of 0.5°and the sub-grid scale elevation with a horizontal resolution of 3″ in each grid (b), the data processing flow diagram (c), and the grid-scale sky view factor at the Gaussian grids of 0.5°across the world (d).The dashed rectangles in (d) shows the study sub-regions.

Figure 2 .
Figure2.The surface solar irradiance (SSI) with a horizontal resolution of 0.5°from the 4M_PP data (a), the 11M_PP data (b), the 4MME_TOP data (c), the 11MME_TOP data (d), and the OBS data (e) averaged over 2001-2014.The differences of the 14-year mean SSI fluxes between the 4MME_PP data and the OBS data (f), between the 11MME_PP data and the OBS data (g), between the 4MME_TOP data and the 4MME_PP data (h), and between the 11MME_TOP data and the 11MME_PP data (i).The differences of the relative root mean square errors (RRMSE) of the SSI between the 4MME_TOP data and the 4MME_PP data (j) and between the 11MME_TOP data and the 11MME_PP data (k) during 2001-2014.The black numbers in Figures(a-e) are the SSI averaged at the grids with SVF ≤ 0.99 in the subregion Reg1 shown in Figure1d.The red and blue numbers in Figures a-d are the relative errors of the 14-year mean SSI simulations and relative root mean square errors of the SSI simulations during 2001-2014 at the grids with SVF ≤ 0.99 in the sub-region Reg1 shown in Figure1d, respectively.

Figure 3 .
Figure 3.The relative root mean square errors of the surface solar irradiance (SSI) simulations from the 4MME_PP, 11MME_PP, 4MME_TOP, and 11MME_TOP data during 2001-2014 (a-e), and the relative changes of the relative root mean square errors of the SSI simulations with the 3DSTSRE offline correction during 2001-2014 compared to those without the 3DSTSRE offline correction at the grids with different grid-scale SVF in different sub-regions (f-j).Each sub-region is shown in Figure 1d.

Figure 4 .
Figure 4.The probability density function (PDF, a, b) and the cumulative density function (CDF, c, d) of the absolute relative error of the monthly surface solar irradiance simulations in the 4MME_PP, 4MME_TOP, 11MME_PP, and 11MME_TOP data at the grids with SVF ≤ 0.99 in the sub-region Reg1 during 2001-2014.

Table 1
The Models of Coupled Model Intercomparison Project Phase 6 HighResMIP Providing the Data in This StudyModel Geophysical Research Letters 10.1029/2023GL107737This study is supported by the National Natural Science Foundation of China under Grant 42375157, the National Key R&D Program of China under Grant 2022YFC3080500, the CAS "Light of West China" Program (E129030101), Open Research Fund Program of Plateau Atmosphere and Environment Key Laboratory of Sichuan Province (PAEKL-2023-K01), the Research Funds for the Frontiers Science Center for Critical Earth Material Cycling Nanjing University, the Fundamental Research Funds for the Central Universities (020914380103), the Jiangsu University "Blue Project" outstanding young teachers training object, the Jiangsu Collaborative Innovation Center for Climate Change, and the Postgraduate Research and Practice Innovation Program of Jiangsu Province