A comparative assessment of machine learning, statistical, and hybrid approaches in amending ECMWF rainfall bias in the Batanghari river basin

Abstract

In tropical basins, rainfall predictions using international numerical weather prediction (NWP) models, like the European Centre for Medium-Range Weather Forecast (ECMWF), exhibit significant persistent biases and insufficient ensemble dispersion. This study evaluates statistical, machine learning (ML)-based, and hybrid bias correction methods to improve daily rainfall estimates for Indonesia’s Batanghari River Basin. Findings indicate that the ECMWF model significantly overestimates wet-day frequency (WDF) (81.5% versus 43.4% observed) and underestimates extreme maximum magnitudes, exhibiting an annual maximum rainfall bias of -15.6 mm/day. Standalone ML-based methods, such as random forest (RF) and extreme gradient boosting (XGBoost), experience significant variance deflation, inadequately addressing these substantial underestimations. In contrast, quantile mapping (QM) significantly reduces the intrinsic drizzle bias and achieves the highest monthly temporal stability among the tested frameworks (Kling-Gupta efficiency (KGE) = 0.341) with low processing requirements. Moreover, hybrid methodologies, including RF-QM and XGBoost-QM, successfully restore the natural ensemble spread and robustly reconstruct the upper tails of extreme rainfall, albeit at the expense of long term volume conservation (resulting in negative monthly KGE). Ultimately, QM is favored as an effective baseline for continuous hydrological modeling, whereas hybrid methodologies are highly suited for evaluating localized flash flood hazards and short-term severe intensities.