Seasonal Estimation of Net Surface Shortwave Radiation Using Multiple Machine Learning Algorithms, Remote Sensing Observation, and In-Situ Station

Net surface shortwave radiation (NSSR) is a key parameter in the Earth’s energy cycle, greatly affecting global water and heat balance. Currently, a comprehensive comparative analysis regarding the accuracy of different models remains severely lacking, and there is also a notable deficiency in the systematic exploration of seasonal radiative drivers. Therefore, we developed a machine learning-based seasonal NSSR estimation model. By integrating in-situ observational data with multi-source remote sensing datasets, we achieved precise quantification of radiative fluxes. This proposed model framework employed three cutting-edge algorithms, namely Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM), to capture the non-linear interactions among radiative drivers across the four seasons. Through mechanistic sensitivity analysis, we quantified the impacts of key variables on NSSR prediction. The results unequivocally demonstrated that the RF algorithm demonstrated the best performance. Its seasonal R2 were 0.95 (spring), 0.89 (summer), 0.95 (autumn), and 0.96 (winter). The Solar Zenith Angle (SZA) dominated in spring and winter; its absence reduced R2 by 0.23 and raised RMSE by 20.66–26.42 W/m2. Meteorological factors mattered most in summer; excluding them cut R2 by 0.17 and hiked RMSE by 23.82 W/m2. This study provides actionable insights for terrestrial radiation budget research.