Accurate and reliable estimation of actual evapotranspiration (ETa) is fundamental for hydrological modeling and effective water management. However, this remains a significant challenge due to the complex interplay between climate, soil, and vegetation, especially in data-scarce regions. This study introduces a novel hybrid approach to estimate monthly ETa by integrating a modified Budyko framework with an optimized machine learning model (XGBoost among thirteen other models) using deep learning for a watershed under non steadystate conditions in California’s Central Valley. Remote sensing data from ERA5 and TerraClimate datasets were utilized as primary inputs, and Eddy Covariance Towers data as observed data. While the standalone modified Budyko model, particularly the Zhang equation, provided a suitable estimation of ETa, the hybrid model consistently and significantly outperformed all standalone models. This demonstrates the superior predictive capability of the hybrid approach, which successfully mitigates the inherent weaknesses of both conceptual/ physical and pure data-driven models. The methodological innovation of optimizing the Budyko parameter’s temporal scale also improved performance by accounting for non-steady-state conditions of the watershed. A feature importance analysis using SHAP values highlighted the climate, soil, and vegetation indices as primary drivers of ETa. This research presents a scalable and cost-effective solution that leverages globally available data, making it a pragmatic and universally applicable tool for sustainable water management in developing countries and any data-scarce regions worldwide

Keywords