A new transfer learning framework enables China's Fengyun-4A (FY-4A) geostationary satellite to estimate surface solar radiation (SSR) and its global, direct, and diffuse components with high accuracy. The method, reported in the Journal of Remote Sensing on April 29, 2026, adapts knowledge from the Himawari-8-based Cloud, Atmospheric Radiation and Renewal Energy Application (CARE) product, reducing dependence on auxiliary meteorological datasets. This advance provides a stronger data foundation for solar power forecasting, climate research, land-surface modeling, and sustainable energy planning.
Surface solar radiation controls Earth's energy balance, hydrological cycles, ecosystem processes, and the performance of solar photovoltaic (PV) and concentrating solar power systems. Ground-based radiometric networks offer the most reliable observations, but their stations are sparse and unevenly distributed, especially across oceans and developing regions. Reanalysis products provide broad coverage but may lose accuracy because of coarse resolution and simplified cloud–aerosol–radiation interactions. Satellite observations can fill this gap, yet many existing algorithms are sensor-specific, and most products focus mainly on global radiation rather than separately estimating direct and diffuse components. Based on these challenges, researchers from the Aerospace Information Research Institute, Chinese Academy of Sciences; Sichuan University of Science and Engineering; and the Institute of Atmospheric Physics, Chinese Academy of Sciences, conducted a deeper investigation into transferable, high-resolution solar radiation retrieval from Chinese geostationary satellites.
The study's key advance is a transfer learning strategy that carries radiative knowledge from Himawari-8 to FY-4A. The team first developed a deep neural network (DNN) model using Himawari-8 Level 1 (L1) observations and the CARE radiation product, then fine-tuned the pretrained model with FY-4A L1 data. The model uses top-of-atmosphere (TOA) reflectance and solar–satellite geometry as dynamic inputs, while Bayesian optimization automatically selects key hyperparameters to improve generalization and efficiency. Validation was performed using 33 ground stations from the Baseline Surface Radiation Network (BSRN), Bureau of Meteorology (BOM), and Global Tropical Moored Buoy Array (GTMBA) during 2018–2020. At representative BSRN sites, FY-4A achieved instantaneous root mean square errors (RMSEs) of 102.2, 117.5, and 83.1 W m⁻² for global, direct, and diffuse radiation, respectively. At the daily mean scale, the RMSEs dropped to 28.5, 30.1, and 22.6 W m⁻², showing strong performance across different temporal scales.
The authors said the study shows how knowledge from a mature satellite product can be transferred to another platform to build new operational capability. They said the framework allows FY-4A to estimate not only total sunlight but also the direct and diffuse components that determine how solar energy systems perform under clear, cloudy, and hazy conditions. They also emphasized that reducing reliance on auxiliary meteorological data makes the method more practical for near-real-time monitoring. In their view, the approach turns China's geostationary satellite observations into a more powerful resource for energy and climate applications.
The new FY-4A radiation product could help improve PV site assessment, power forecasting, grid management, climate modeling, and land-surface simulations. Direct radiation is especially important for concentrating solar power, while diffuse radiation affects PV output under cloudy or aerosol-rich skies. By resolving these components separately, the framework offers more actionable information than global radiation alone. The study also demonstrates that transfer learning can help overcome sensor differences and limited ground training data. Looking ahead, the same strategy could be extended to other Chinese geostationary satellites, including Fengyun-4B (FY-4B), supporting more reliable solar-energy monitoring across East Asia and beyond.
For further details, see the original study at DOI: 10.34133/remotesensing.1044 and related information at Chuanlink Innovations.
