List of hyperparameters used for model training.

This study develops the LGBM-3SC-CF product, a machine learning framework that integrates multi-source data (Himawari-8, ERA5, ASTERDEM, and rain gauge) to significantly improve rainfall classification and estimation accuracy for four coastal provinces in Central Vietnam. Methodologically, our core contribution lies in the proposed 3-stage classification architecture combined with a novel cloud masking technique. To address the severe class imbalance inherent in rainfall datasets, we further employed an effective data balancing technique based on rainfall intensity distribution alongside feature selection based on cloud- and rain-forming factors. The performance of the proposed rainfall product was compared with four existing regional rainfall products: IMERG Final Run V07, IMERG Early Run V06, GSMaP_MVK_Gauge V07, and PERSIANN-CCS. The LGBM-3SC-CF achieved the highest performance, with a Critical Success Index (CSI) of 0.55 and a Probability of Detection (POD) of 0.74, and achieved the Correlation Coefficient (CC) of 0.47 and the Modified Kling-Gupta Efficiency (mKGE) of 0.47. Furthermore, it obtained the lowest error metrics, with a Mean Absolute Error (MAE) of 2.66 mm/h and a Root Mean Squared Error (RMSE) of 5.48 mm/h. This study not only establishes a robust machine learning framework but also provides the essential methodological foundation for developing near real-time rainfall estimation models through the seamless substitution of the atmospheric reanalysis features with near real-time meteorological features.

本研究研发了LGBM-3SC-CF产品，这是一款整合多源数据（Himawari-8、ERA5、ASTERDEM及雨量计数据）的机器学习框架，可显著提升越南中部四个沿海省份的降雨分类与估测精度。在方法论层面，本研究的核心贡献在于提出了结合新型云掩膜技术的三阶段分类架构。针对降雨数据集固有的严重类别不平衡问题，本研究进一步采用了基于降雨强度分布的高效数据平衡技术，以及基于成云与成雨因子的特征选择方法。本研究将所提出的降雨产品与四款现有区域降雨产品开展性能对比：IMERG Final Run V07、IMERG Early Run V06、GSMaP_MVK_Gauge V07及PERSIANN-CCS。结果显示，LGBM-3SC-CF取得了最优性能，临界成功指数（Critical Success Index, CSI）达0.55，探测概率（Probability of Detection, POD）达0.74，相关系数（Correlation Coefficient, CC）为0.47，修正克林-古普塔效率（Modified Kling-Gupta Efficiency, mKGE）为0.47。此外，其误差指标亦为最低，平均绝对误差（Mean Absolute Error, MAE）为2.66 mm/h，均方根误差（Root Mean Squared Error, RMSE）为5.48 mm/h。本研究不仅构建了一套稳健的机器学习框架，还通过将大气再分析特征无缝替换为近实时气象特征，为开发近实时降雨估测模型提供了关键的方法论基础。