Robust and Explainable Objective Classification of Extreme Precipitation Events Using Machine Learning

Extreme precipitation events (EPEs) threaten socioeconomic development and public safety. Accurate classification of EPEs is essential for identifying key predictors and improving forecast accuracy. Existing methods mostly rely on subjective case-by-case analyses or objective clustering using the spatial patterns of all cases as input, however, suffering from vague classification boundaries and poor explainability. Here, we use a method combining Empirical Orthogonal Function (EOF) with K-means (or hierarchical) clustering to address these issues. Specifically, by applying EOF to geopotential height fields at 500 hPa (H₅₀₀) and using the first two principal components as input for clustering, EPEs occurring from May to October between 2010 and 2024 in Zhejiang Province can be effectively divided into three types: typhoon core (TC-type), Mei-yu (MY-type), and typhoon trough (TT-type). Hierarchical clustering further refines the TT-type into three subtypes: far-distance typhoon trough with a 500-hPa trough (FTT-500T-type), near-distance typhoon trough (NTT-type), and near-distance typhoon trough with a northwestern Pacific subtropical high dam (NTT-NWPSH-type). The MY-type is mainly influenced by meteorological variables at 700 and 850 hPa, while the TC- and TT-types are primarily affected by variables at 950 hPa. Topographical effects on the precipitation are evident in the TC-, NTT-, and NTT-NWPSH-types, but not in the MY- and FTT-500T-types. Accordingly, our study provides significant advantages for the objective classification of EPEs and subsequent in-depth investigations.