Tang et al. (2025) A novel hybrid approach for mapping global surface solar radiation with DSCOVR/EPIC: Combining deep learning with physical algorithm

Identification

• Journal: Remote Sensing of Environment
• Year: 2025
• Date: 2025-12-17
• Authors: Wenjun Tang, Jinwen Qi, Junmei He, Fuxin Zhu, C. Z. Yuan, Jin-yan Yang, Bing Hu
• DOI: 10.1016/j.rse.2025.115200

Research Groups

• National Tibetan Plateau Data Center (TPDC), State Key Laboratory of Tibetan Plateau Earth System, Environment and Resources (TPESER), Institute of Tibetan Plateau Research, Chinese Academy of Sciences, Beijing, China
• Institute of Urban Meteorology, China Meteorological Administration (CMA), Beijing, China
• University of Chinese Academy of Sciences, Beijing, China
• Faculty of Geography, Yunnan Normal University, Kunming, China
• College of Geography and Planning, Chengdu University of Technology, Chengdu, China

Short Summary

This study develops a novel hybrid approach integrating deep learning with physical algorithms, utilizing DSCOVR/EPIC observations to map global surface solar radiation (Rg) and its direct (Rdir) and diffuse (Rdif) components. The method demonstrates superior accuracy and spatial scalability compared to existing products, providing globally consistent high-resolution radiation data.

Objective

• To develop a novel method to map spatially homogeneous global products of Rg, Rdir, and Rdif at high resolution based on DSCOVR/EPIC observations.

Study Configuration

• Spatial Scale: Global coverage, 0.1° × 0.1° latitude-longitude grid, approximately 10 km at the sub-satellite point.
• Temporal Scale: 1 to 2 hours (10 to 22 times per day), hourly, daily, and monthly aggregates. Data period: January 2016 to December 2017.

Methodology and Data

• Models used:
  • DenseNet-based Convolutional Neural Network (CNN) for estimating cloud transmittance.
  • Physical parameterization scheme (from Tang et al., 2019) for calculating clear-sky Rg, Rdir, and Rdif.
  • Light Gradient Boosting Machine (LightGBM) model for estimating the diffuse ratio to separate Rdir and Rdif from Rg.
• Data sources:
  • Satellite: Deep Space Climate Observatory (DSCOVR) Earth Polychromatic Imaging Camera (EPIC) L1B raw spectral data (10 channels) and L2 cloud product (solar zenith and azimuth angles).
  • Reanalysis/Gridded: MERRA-2 aerosol data (hourly, 0.5° × 0.625°), MODIS albedo product (MCD43A3, 5 km, 16-day), ERA5 reanalysis (surface pressure, total column water vapor, total column ozone; hourly, 25 km).
  • In-situ (for training/validation):
    • Baseline Surface Radiation Network (BSRN): 42 stations (2017 for training/validation, 2016 for independent validation).
    • Solar Radiation (SOLRAD) network: 7 stations (2016 for independent validation).
    • China Meteorological Administration (CMA) radiation stations: 120 stations (2017 for independent validation).
    • Global Energy Balance Archive (GEBA) network: 408 stations (2016–2017 for independent validation).

Main Results

• The hybrid approach achieved high accuracy during the training phase for Rg (MBE: -0.1 W/m², RMSE: 21.6 W/m², R: 0.997), Rdir (MBE: -1.1 W/m², RMSE: 55.6 W/m², R: 0.98), and Rdif (MBE: 1.1 W/m², RMSE: 49.7 W/m², R: 0.91).
• Robust temporal scalability was demonstrated, with relative RMSE (rRMSE) for Rg, Rdir, and Rdif maintaining less than a 2% threshold elevation during independent time-based validation compared to the testing phase.
• The method showed superior accuracy compared to the CERES product at hourly, daily, and monthly scales, exhibiting lower RMSE and higher R values.
• Strong spatial scalability was validated across SOLRAD, CMA, and GEBA networks. For instantaneous Rg at SOLRAD stations, R was 0.96, MBE 5.4 W/m², and RMSE 79.7 W/m².
• At CMA stations, hourly Rg estimates achieved an MBE of 2.6 W/m², RMSE of 106.4 W/m², and R of 0.92, significantly outperforming other Fengyun-4A based retrievals (e.g., RMSE ~147.0 W/m² from Shi et al., 2023).
• Monthly Rg estimates validated against GEBA stations showed the best overall performance among compared products (Hao et al., 2020; BESS; CERES), with the lowest absolute MBE (2.5 W/m²), lowest RMSE (14.7 W/m²), and highest R (0.99).
• The hybrid method significantly outperformed the purely machine learning-based DSCOVR/EPIC product by Hao et al. (2020), reducing hourly RMSE for Rg, Rdir, and Rdif by 45.6 W/m², 55.2 W/m², and 21.7 W/m², respectively.
• Global spatial distribution mapping for 2016–2017 showed consistent patterns with other products, with high radiation in equatorial/low-latitude zones and elevated regions.

Contributions

• Introduces a novel hybrid approach that integrates deep learning (DenseNet-based CNN for cloud transmittance) with physical algorithms (for clear-sky Rg) using DSCOVR/EPIC observations.
• Addresses critical limitations in existing satellite-derived radiation products, specifically spatial inconsistencies from multi-source data fusion and high uncertainties in cloud optical property retrievals due to the absence of infrared bands in EPIC.
• Provides spatially homogeneous global products of Rg, Rdir, and Rdif at high temporal (1–2 hours) and spatial (0.1°) resolution, overcoming the coverage gaps of geostationary satellites and the low temporal resolution of polar-orbiting sensors.
• Demonstrates strong generalization capability and spatial scalability of the physics-constrained deep learning model by indirectly training cloud transmittance rather than directly training Rg.

Funding

• National Natural Science Foundation of China (Grant No. 42371370 and 42171360)
• National Key Research and Development Program of China (Grant No. 2024YFF0729102)

Citation

@article{Tang2025novel,
  author = {Tang, Wenjun and Qi, Jinwen and He, Junmei and Zhu, Fuxin and Yuan, C. Z. and Yang, Jin-yan and Hu, Bing},
  title = {A novel hybrid approach for mapping global surface solar radiation with DSCOVR/EPIC: Combining deep learning with physical algorithm},
  journal = {Remote Sensing of Environment},
  year = {2025},
  doi = {10.1016/j.rse.2025.115200},
  url = {https://doi.org/10.1016/j.rse.2025.115200}
}