Tree-Based Machine Learning Estimation of Near-Real-Time Diffuse Solar Irradiance From Himawari-8 Satellite Data

Diffuse solar radiation (Rd), one of the two components of solar radiance, has been shown to have a fertilizing effect on plants and, thus, has implications for terrestrial carbon sinks and the water cycle. Currently available Rd data are mostly at hourly or daily scales, but their production very often depends on additional meteorological data. Furthermore, very few works have estimated Rd in near-real time, as dynamically changing clouds and aerosols pose great challenges for instantaneous estimation. In this study, we adopt the Himawari-8 satellite data as the main input, to estimate the near-real-time Rd using four tree-based machine learning algorithms, namely gradient boosting regression tree (GBRT), random forest (RF), XGBoost, and LightGBM, similarly. Although all models performed well, with R2 around 0.80 and rRMSE at 32.97%–37.2% for random tenfold cross-validation (CV), LightGBM has the highest efficiency. Further ground validation on independent datasets showed that the results are highly consistent across different time scales. Additional analyses revealed that different models behave differently under clear and nonclear sky conditions, but all perform poorly under clear sky conditions. Taken together, we foresee that this efficient method, based on Himawari-8 geostationary satellite data, can be widely applied to estimate near-real-time Rd in the future.