Integrating Load-Cell Lysimetry and Machine Learning for Prediction of Daily Plant Transpiration

We conducted research to predict daily transpiration in crops by utilising a combination of machine learning (ML) models combined with extensive transpiration data from gravimetric load cells and ambient sensors. Our aim was to improve the accuracy of transpiration estimates. Data were collected from hundreds of plant specimens growing in two semi-controlled greenhouses over 7 years, automatically measuring key physiological traits (serving as our ground truth data) and meteorological variables with high temporal resolution and accuracy. We trained Decision Tree, Random Forest, XGBoost and Neural Network models on this data set to predict daily transpiration. The Random Forest and XGBoost models demonstrated high accuracy in predicting the whole plant transpiration, with R² values of 0.89 on the test set (cross-validation) and R² = 0.82 on holdout experiments. Ambient temperature was identified as the most influential environmental factor affecting transpiration. Our results emphasise the potential of ML for precise water management in agriculture, and simplify some of the complex and dynamic environmental forces that shape transpiration.