Improving the rainfall rate and kinetic energy estimation in Taiwan using machine learning approaches

This study investigates the seasonal variability of raindrop size distribution (RSD) in northern Taiwan and examines its implications for improving rainfall rate (R) and rainfall kinetic energy (KE) estimation. Long-term observations (2008–2021) from three Joss–Waldvogel disdrometers were analyzed for the Winter, Mei-yu, and Summer seasons to characterize seasonal microphysical differences in precipitation. The results reveal clear seasonal contrasts in RSD characteristics. Winter rainfall is dominated by small drops associated with stratiform precipitation, whereas summer rainfall contains a higher proportion of mid-to-large drops produced by convective systems, with Mei-yu exhibiting intermediate characteristics due to mixed precipitation regimes. These microphysical differences lead to systematic seasonal variations in radar reflectivity–rainfall rate (Z–R) and kinetic energy–reflectivity (KE–Z) relationships, demonstrating the necessity of season-specific retrieval models. To improve estimation accuracy, several machine learning (ML) regression approaches, including Decision Tree (DT), Random Forest (RF), Gradient Boosting (GB), and Support Vector Machine (SVM), were developed using radar reflectivity (Z) and mass- weighted mean diameter (D_m) as predictors. All nonlinear ML models outperformed conventional linear regression across seasons. Among the evaluated algorithms, the SVM regression model achieved the highest predictive accuracy, particularly during the Mei-yu and Summer seasons when precipitation microphysics exhibit greater variability. These findings demonstrate that incorporating seasonal microphysical characteristics and additional predictors such as D_m significantly enhances rainfall rate and kinetic energy estimation, with important implications for quantitative precipitation estimation and erosion-related applications in Taiwan.