Data-driven hydrology: rainfall modeling in forest ecosystems using machine learning

Abstract

This study develops an Integrated Methodological Framework based on Machine Learning (ML) for the reconstruction, modeling, and transformation of rainfall in seasonal forest ecosystems, jointly addressing data availability, hydrological variability, and the fundamental ecohydrological processes controlling water input to the soil. First, a robust methodology is proposed for reconstructing continuous hourly and daily rainfall series in regions with incomplete monitoring networks. The approach combines automated quality control through unsupervised anomaly detection, bias correction of satellite and reanalysis products, and optimal model identification based on multi-objective criteria that balance statistical accuracy and hydrological consistency. Results show that ML models trained on quality-controlled data significantly outperform traditional statistical methods while preserving convective variability, which is essential for hydrological applications in tropical ecosystems. The study advances rainfall modeling by explicitly incorporating forest canopy interception processes. Using Machine Learning ensembles and hydrologically interpretable predictors, the models reveal strong spatiotemporal variability in interception, which cannot be adequately represented by constant coefficients. Overall, the study demonstrates that rainfall reconstruction and its transformation into effective precipitation are inseparable processes in the ecohydrology of seasonal forests, providing reproducible tools for hydrological cycle analysis, ecohydrological modeling, and water resources management in tropical forest ecosystems.