Fusion of satellite and gauge precipitation observations through coupling spatio-temporal properties with tree-based machine learning

Abstract

Merging satellite and gauge observations is a promising solution for obtaining accurate precipitation data. Although machine learning based merging methods have shown excellent potential, their insufficient consideration of the spatial–temporal properties of precipitation greatly limits the performance of merging models. To address this problem, a novel merging approach is proposed here that couples Spatio-Temporal Properties and the Tree-based Machine Learning model (STPTML), aiming to improve the accuracy of precipitation estimation. This method focuses on two important spatio-temporal properties of precipitation: spatial correlation and temporal heterogeneity. Leveraging the intrinsic characteristics of tree-based machine learning models, an adaptive spatio-temporal encoding strategy is designed to transform these spatio-temporal properties into features that can be fully utilized by the tree model to achieve their organic coupling. The features guide the tree model to explore the spatio-temporal distribution patterns of precipitation, thereby promoting the high-level integration of satellite and gauge observations. Taking Hai River Basin as an example, the effectiveness of STPTML was verified using four typical tree models: random forest, LightGBM, XGBboost, and Catboost. The results show that: (1) STPTML greatly improved the accuracy of original satellite precipitation products compared to the state-of-the-art merging methods. (2) The proposed adaptive spatio-temporal encoding strategy exhibited broad effectiveness for tree-based models (3) The merged results greatly enhanced the reliability of satellite precipitation products in estimating rainfall erosivity. Overall, STPTML is an effective approach for the accurate estimation of precipitation, which furnish a reliable data foundation for research in the fields of meteorology and environmental science.