From Semi-Mechanistic Model to Explainable Machine Learning: A New Approach to Evapotranspiration Estimation

Accurate estimation of evapotranspiration (ET) is critical for understanding the terrestrial water and energy cycles, especially under the context of climate change and its impact on vegetation dynamics. Traditional semi-mechanistic models often struggle to accurately capture ET variability due to uncertainties in parameterizations and assumptions about vegetation responses to environmental drivers. In this study, we leverage the potential of machine learning (ML) models to improve ET estimation by integrating key biophysical and environmental variables: solar-induced chlorophyll fluorescence (SIF), photochemical reflectance index (PRI), and vapor pressure deficit (VPD). These variables provide a direct and dynamic representation of vegetation activity and environmental stress. Our results show that ML models outperform semi-mechanistic models in capturing ET dynamics across diverse spatial and temporal scales. Using explainable machine learning techniques, we further interpret the ML model's performance by identifying the relative importance of input variables and their interactions. SIF emerges as a dominant predictor, providing direct insights into photosynthetic activity and stomatal conductance. VPD is also shown to play a critical role, highlighting its influence on atmospheric demand for water. PRI contributes by offering a proxy for photoprotective mechanisms, which are crucial under stress conditions. The comparative analysis underscores the limitations of semi-mechanistic models in capturing non-linear relationships and rapid responses. Explainable ML techniques reveal that the improved performance stems from the ML model's ability to account for complex, non-linear interactions between variables and dynamically adjust to changing conditions. The findings have significant implications for hydrological modeling, water resource management, and climate change impact assessments.