Artificial Intelligence-Driven Analysis of Radar Polarization Parameters for Different Rainfall IntensitiesDong-In Lee, Seong-Jun Hwang, Jisun Lee (AERI, PKNU)Pukyong National University/Atmospheric Environmental Research InstituteEmeritus Professor/Distinguished Researcher |
|
This study systematically evaluates four distinct deep learning architectures (Multi-Layer Perceptron, a DenseNet with an attention mechanism, ResNet101 and Vision Transformer) to identify the most effective model for polarimetric radar-based QPE. The top-performing model, trained and evaluated on a comprehensive dataset from the Seoul metropolitan area and Gyeonggi Province in South Korea (2020-2024) combining S-band radar data and ground observations, was subjected to an in-depth analysis using explainable AI (XAI).
A SHAP-based XAI analysis of the trained model revealed a physically consistent and meteorologically sound strategy. The analysis showed a dynamic shift in the dominant predictor variable with rainfall intensity: prioritizing Differential Reflectivity (Z_DR) for drizzle, shifting to Vertical and Horizontal Reflectivity (Z_V, Z_H) for light and moderate rain, returning to Z_DR for heavy rain, and ultimately relying on the attenuation-robust Specific Differential Phase (K_DP) in severe and extreme conditions. By demonstrating that a top-performing deep learning model learns an interpretable, physics-aligned strategy, this study bridges the gap between high-performance AI and operational reliability. This provides a foundation for improved early warning systems and real-time flood prediction for extreme weather phenomena, such as localized heavy rainfall, Changma, and typhoons etc. " |