SPFANet: A Spatial Polarization Feature Aggregation Network for Radar-Based Hydrometeor Classification

Hydrometeor classification from dual-polarization weather radar is important for precipitation monitoring, cloud microphysical analysis, and severe weather warnings. However, operational performance remains limited by non-meteorological echoes, incomplete use of the relationships among polarimetric variables, and severe class imbalance. To address these issues, we propose a Spatial Polarization Feature Aggregation Network (SPFANet) for hydrometeor classification using dual-polarization weather radar. The proposed method combines a non-meteorological echo removal procedure and a multi-feature radar image constructed from reflectivity, differential reflectivity, and the correlation coefficient. SPFANet is designed as a semantic segmentation model that integrates multi-level semantic representations, aggregates multiscale spatial features, and enhances latent polarimetric features. Experiments using KTLX (USA) observations show a high level of agreement between SPFANet predictions and the NOAA reference labels across different precipitation scenarios. Across five independent runs, SPFANet achieves mean Recall values above 81% for all five categories. It also obtains the highest mean F1 and IoU for ice crystals, graupel, and hail, as well as the highest mean Precision for hail. The overall evaluation metrics further demonstrate the effectiveness of SPFANet in hydrometeor classification. These findings highlight the potential of the proposed model for radar-based precipitation analysis and severe weather monitoring.