Utilizing Multi-source Data and Machine Learning Models for Multi-class Rainfall Estimation in Central Vietnam

Abstract

This study proposes a multi-layer machine learning architecture for multi-class rainfall estimation in Central Vietnam. The input data includes Himawari-8 satellite imagery, ERA5 reanalysis data, ASTER DEM, and rain gauge observations. Four regional satellite-based rainfall products, including IMERG Final Run V07, IMERG Early Run V07, GSMaP_MVK_Gauge V08, and PERSIANN_CCS, were used as comparative datasets. Three machine learning algorithms, including Light Gradient Boosting Machine (LGBM), Extreme Gradient Boosting (XGB), and Random Forest (RF), were employed within the proposed architecture. Performance evaluation based on rain gauge observations showed that the LGBM-based rainfall product achieved the highest classification performance among the three surveyed products, with a Probability of Detection (POD) of 0.80, a Critical Success Index (CSI) of 0.54, a Matthews Correlation Coefficient (MCC) of 0.59, and a Symmetric Extremal Dependence Index (SEDI) of 0.58. Compared to the best-performing rainfall product (GSMaP_MVK_Gauge V08), the LGBM-based product demonstrated significant improvements in classification performance, with increases of 6.67% in POD, 8.00% in CSI, 11.32% in MCC, and 20.83% in SEDI. In terms of rainfall regression performance, the LGBM-based product also outperformed the other evaluated products, exhibiting the lowest errors, with a Mean Absolute Error (MAE) of 2.91 mm/h, Root Mean Square Error (RMSE) of 5.81 mm/h, and Mean Logarithmic Squared Error (MLSE) of 0.47.