A machine learning-driven semi-mechanistic model for estimating actual evapotranspiration: Integrating photosynthetic indicators with vapor pressure deficit

Abstract

Accurate estimation of actual crop evapotranspiration (ET_{c act}) is essential for optimizing water resource management and irrigation strategies, particularly in arid and semi-arid agricultural regions. Traditional models rely on extensive meteorological data, limiting their applicability in data-scarce areas. This study used on-site ground observation data with a 30-minute temporal resolution from a winter wheat field at the Yangling Station on the Guanzhong Plain, China, to evaluate the performance of machine learning-driven semi-mechanistic models driven by three machine learning methods (Ridge regression, Random Forest, and Support Vector Machine) in estimating ET_{c act}. These machine learning-driven semi-mechanistic models integrate photosynthetic indicators (Gross Primary Production, GPP; solar-induced chlorophyll fluorescence, SIF; near-infrared reflectance of vegetation, NIRv) with the square root of vapor pressure deficit (VPD^0.5) to enhance ET_{c act} estimation accuracy. The results showed that among the photosynthetic indicators, GPP and SIF exhibited a strong correlation with ET_{c act}. When combined with VPD^0.5, their correlation with ET_{c act} further increased by 0.10 and 0.05, respectively, while their response time to ET_{c act} variations was reduced by 2 hours and 1 hour. Notably, NIRv exhibited the weakest correlation with ET_{c act}, with a Pearson correlation coefficient of only 0.31, significantly lower than SIF (0.78) and GPP (0.69), indicating its limited effectiveness as an independent predictor. Furthermore, machine learning-driven semi-mechanistic models driven by machine learning achieved higher accuracy in ET_{c act} estimation than single-factor machine learning models and the Penman-Monteith equation incorporating the single crop coefficient method. Among them, the RF model based on SIF × VPD^0.5 achieved the best performance, with an R² of 0.86 and an RMSE of 0.69 mm/day. This study demonstrates that machine learning-driven semi-mechanistic models can significantly improve ET_{c act} estimation accuracy while reducing dependence on meteorological data. The proposed approach provides a new theoretical framework for improving water resource management and irrigation efficiency in arid and semi-arid agricultural regions, while also offering a scientific basis for future ET_{c act} estimation methods integrating remote sensing data.