Enhancing the Explainability of Representation of Surface Downward Shortwave Fluxes from Atmospheric Reanalysis Data Products

Abstract Reanalysis data products are widely utilized across disciplines; however, their explainability is limited. This study aimed to enhance the explainability of surface shortwave fluxes derived from atmospheric reanalysis data products and to propose methods for interpreting model behaviors that data users can implement. Enhancing explainability will improve the reliability of application results and support the selection of appropriate data products for diverse applications. To this end, a series of interpretation methods for the black-box learning model were introduced. This research evaluated global, direct, and diffuse shortwave fluxes using data covering Japan, comparing two reanalysis data products, MERRA-2 and ERA5, against the CERES data product as a reference. Although ERA5 demonstrated higher accuracy in global shortwave flux than MERRA-2, the discrepancies in shortwave flux representation between the two were attributed to input variables influencing the radiative transfer model. These variables included cloud properties, surface albedo, and clear-sky flux. The direct and diffuse fluxes exhibited lower accuracy than the global flux for both data products. Additionally, the results suggested that certain errors in direct and diffuse fluxes were associated with the radiation scheme. The evaluation indicated that no significant differences in errors were present in the shortwave radiation schemes of MERRA-2 and ERA5. Several factors such as cloud fraction at medium and low levels and the liquid water path at low levels influenced the errors in the radiation scheme.