A deep learning-ensemble framework for Pre-Meiyu season definition and synoptic-local weather classification

Abstract This study presents a deep learning framework to objectively classify Pre-Meiyu season synoptic weather types and their linkage to regional rainfall patterns. Using ERA5 reanalysis (1979–2023), we employ a suite of Convolutional Neural Network (CNN)-Autoencoders to extract latent features from 5-day moving mean atmospheric circulation fields. These features are utilized to objectively identify the duration of the Pre-Meiyu season via cosine similarity. Within this season, a set of CNN-Autoencoders is trained to extract latent features from daily mean atmospheric variables (ERA5) and daily rainfall (GPM IMERGE, TCCIP; 2001–2023). Unsupervised k-means clustering of these features identifies eight distinct synoptic weather types. Four types are associated with significant rainfall over Taiwan and are directly influenced by frontal systems and/or southwesterly flow. The other four correspond to drier conditions governed by the relative positions of the monsoon trough and subtropical high. Notably, a distinct type characterized by a southward-shifted southwesterly flow is linked to tropical cyclone activity. Subsequently, these latent features serve to train six machine learning classifiers to categorize daily weather types. The Ensemble Voting Classifier (EVC) demonstrates robust performance, with its accuracy, F 1-score, and precision all exceeding 0.9. Compared to individual models, the EVC improves overall mean scores while providing more consistent classifications, effectively reducing performance variability even for weather types where single models exhibited instability. These results demonstrate that combining deep learning feature extraction with an ensemble classifier provides a robust and accurate method for objectively characterizing Pre-Meiyu season weather systems.