Evaluation of evapotranspiration models integrating convolutional neural network-predicted leaf area for Pak Choi (Brassica campestris ssp. chinensis) in greenhouse environments

Abstract

This study evaluates how predicted leaf area index (LAI) affects evapotranspiration (ET) model performance and uncertainty in greenhouse Pak Choi cultivation. Five ET models (Penman-Monteith, Stanghellini, Fynn, Shin, and Baille) were compared using both measured and Convolutional Neural Network-Predicted LAI data. Greenhouse environment experiments from June to August 2021 provided validation data under controlled conditions. LAI was estimated using image analysis with high accuracy (R² = 0.9986, RMSE = 0.0547 m²·m⁻²). Sensitivity analysis revealed that ET models were most responsive to radiation and LAI variations, with lower sensitivity to air temperature and relative humidity. Among physical models, the Fynn model demonstrated superior performance based on ET prediction accuracy (R² > 0.87), while the Shin model excelled among simplified approaches (R² > 0.92). Uncertainty propagation analysis revealed that the Stanghellini model exhibited the highest sensitivity to LAI estimation errors (12.55 W·m⁻² error when LAI error = 1.0 m²·m⁻²), whereas the Penman–Monteith model showed minimal sensitivity. Model performance remained consistent when using predicted versus measured LAI (R² > 0.99 for all models), indicating the robustness of image-based LAI estimation for ET modelling. This research provides quantitative insights into model selection and uncertainty assessment for precision irrigation management in protected cultivation systems, with particular applicability to leafy vegetable crops under controlled greenhouse conditions.