A Multistage XGBoost Architecture with Class Weighting and SMOTE for Imbalance-Aware Rain Rate Forecasting

Abstract

Rainfall-induced signal attenuation, or rain fade, remains a major challenge for satellite and 5G communications in tropical climates. This study proposes a multistage XGBoost framework for hourly rain rate forecasting that explicitly addresses the issue of class imbalance in tropical rainfall data. Using 14 years of Hydro-Estimator satellite data from Peninsular Malaysia (2009–2022), the framework decomposes the forecasting task into
three stages: (i) binary classification (rain/no rain), (ii) multi-class rain intensity classification, and (iii) regression to predict the actual rain rate. To mitigate the dominance of ’no rain’ and ’light rain’ cases, class weighting and SMOTE oversampling were applied during training. The results demonstrate that imbalance handling significantly improves detection of minority rain events. Binary classification recall improved from 0.36 to 0.64,
while intensity classification macro-F1 increased from 0.32 to 0.40. In the regression stage, the combined architecture achieved near-zero MAE (0.0000), RMSE (0.0019), and R² =1.0, highlighting the effectiveness of incorporating classification outputs into the regression model. Forecasting for 2023 showed that the SMOTE-enhanced model better captured extreme rainfall peaks and seasonal variability compared to baseline models. These findings confirm the potential of structured, imbalance-aware learning frameworks for accurate rainfall forecasting. The proposed method provides valuable input for adaptive communication strategies, including link margin adjustment and frequency switching, thereby enhancing
the resilience of wireless networks in tropical regions.