Machine learning-based reconstruction of missing rainfall extremes: a comparative analysis with classical models

Abstract

The limited availability of daily rainfall data remains a key challenge in rainfall data analysis. This study assesses the effectiveness of spatial interpolation and bias correction techniques using satellite-derived rainfall data to fill missing observations in the Banten and Jakarta regions. Three interpolation methods inverse distance weighting (IDW), kriging, and spline were compared. Nine statistical and machine learning-based bias correction methods were applied to climate hazards group infrared precipitation with station data (CHIRPS), multi-source weighted-ensemble precipitation (MSWEP), and global precipitation measurement-integrated multi-satellite retrievals for GPM (GPM IMERG). Performance was evaluated using root mean square error (RMSE), mean absolute error (MAE), bias, Pearson correlation (R), and Kling-Gupta efficiency (KGE) in the expert team on climate change detection and indices (ETCCDI) extreme index. The research findings indicate that CHIRPS with quantile mapping (QM) bias correction delivers the best performance, followed by random forest regression (RFR) as the most accurate machine learning method. In spatial interpolation, IDW stands out as the leading method. Testing the extreme index ETCCDI confirms that CHIRPS-QM consistently outperforms machine learning and interpolation methods. In general, CHIRPS-QM and IDW represent the most effective combination of techniques for reconstructing daily rainfall, particularly extreme events. This study uniquely integrates spatial interpolation and bias correction in a unified evaluation.