Passive satellite hourly precipitation estimation over mainland China by combining cloud and meteorological parameters

Abstract

High-quality satellite quantitative precipitation estimation (QPE) is crucial for theoretical studies and disaster monitoring. However, it remains unclear which information is effective or relatively less valuable. Accurately eliminating ineffective variables and applying effective ones as predictors can further enhance the accuracy and computational efficiency for QPE. In this study, an hourly QPE algorithm was developed using three machine learning (ML) models, including Random Forest, XGBoost and LightGBM. We focused on obtaining high-precision precipitation estimations and further analyzing the contribution of different input variables. Sensitivity experiments revealed that satellite visible channels and cloud properties are key factors for accurate QPE. In contrast, information provided solely by infrared channels and meteorological variables is relatively limited. Among three ML models, LightGBM achieved the best QPE, and was comparable to, or even slightly better than GPM IMERG, which may be attributed to its incorporation of more effective variables and training with ground rain gauge. However, it sometimes underestimates heavy precipitation compared to GPM IMERG, probably due to few training samples and saturation of satellite spectral signals. The analysis of Shapley Additive Explanations (SHAP) indicates that QPE are more sensitive to cloud properties (e.g., cloud water path), but some meteorological factors, such as relative humidity at different pressure levels are becoming more important as the environment becomes drier. Additionally, the performance of ML model and GPM IMERG deeply relies on cloud type. These findings are expected to provide valuable references for the construction of future satellite QPE algorithms in terms of feature selection and data processing.