Accurate rainfall estimation in tropical regions is often hindered by non-linear atmospheric interactions and extreme data imbalance. This study develops a multi-stage precipitation estimation framework— comprising binary classification, multi-class classification, and regression—using an optimized Extreme Gradient Boosting (XGBoost) architecture. Applied to East Java, Indonesia, the model integrates Himawari-8/9 satellite brightness temperatures, global atmospheric indices, and high-resolution topography. To mitigate the dominance of non-rain events (91.6% of the dataset), Stratified Random Under-sampling (RUS) was employed. Hyperparameters were tuned using Bayesian Optimization and evaluated via 10-fold site-based cross-validation to prevent spatial data leakage. Results show that the optimized model significantly outperforms the baseline. In the regression stage, MAE and RMSE decreased by 21.5% and 23.0%, respectively, while the Pearson correlation coefficient improved by 43.1%. In classification, the Critical Success Index (CSI) rose by 16.2% for binary and 34.5% for multi-class stages, indicating an enhanced capability to detect rare rainfall events. Performance gains were most pronounced in mountainous regions, suggesting improved representation of orographic effects. The proposed hierarchical framework demonstrates potential as an effective approach for satellite-based rainfall estimation in topographically diverse tropical regions.
Wisnawa et al. (Thu,) studied this question.