Categories of Machine Learning–Based Multisource Precipitation Estimation

Abstract Machine learning has been widely applied to precipitation estimation, but the lack of a formal definition for the machine learning–based precipitation estimation (MLPE) problem hampers consistent benchmarking, making it difficult to compare different machine learning (ML) approaches effectively. In this study, we defined MLPE problems into nine categories based on spatiotemporal estimation strategies and input and conducted experiments to explore their differences and commonalities using interpretability techniques. We found that spatial estimation models differ significantly in structure from temporal and spatiotemporal estimation models, while the latter two are relatively similar. Although spatial estimation models achieve the best performance, their extensibility is limited. In contrast, spatiotemporal estimation models offer greater extensibility, while they face a high risk of overfitting. In spatiotemporal estimation and temporal estimation, merging models that incorporate multiple satellite precipitation products slightly outperform calibration models that use a single product and significantly outperform interpolation models that rely on no satellite precipitation input. However, in spatial estimation, the differences were not statistically significant, with no substantial gaps discerned among the three approaches. By defining MLPE, relevant research can select appropriate benchmarks to compare algorithm performance.