Evaluating Evapotranspiration in a Commercial Greenhouse: A Comparative Study of Microclimatic Factors and Machine-Learning Algorithms

Abstract

AbstractThe FAO-56 Penman-Monteith equation (FPME) is commonly used to calculate evapotranspiration and apply necessary irrigation, based on environmental data usually taken from a single measuring station. In this study, we hypothesized that the accuracy of the FPME is affected by microclimatic changes over time and space within the target area. Therefore, we tested the impact of numerous spatial and temporal environmental measurement points in a commercial greenhouse on the accuracy of the FPME, by comparing its evapotranspiration evaluation to the actual evaporation measured by dozens of weighing lysimeters throughout the year. Additionally, we harnessed the capabilities of machine-learning algorithms to utilize the extensive data acquired for predicting evapotranspiration. Our results revealed that the daily FPME exhibited a -22% to +22% discrepancy in accuracy, as compared to the lysimeters, with overestimation in the winter and underestimation in the summer. Interestingly, using more data points per day led to less accurate FPME evaluation. The conflict between an increased number of data points and a reduction in accuracy was explained by daily hysteresis. Machine-learning algorithms (Decision Tree, Random Forest, XGBoost and Neural Network) showed impressive accuracy in predicting evapotranspiration, when the model dataset contained temporal parameters (R2> 0.918). Furthermore, we demonstrated that spatial sampling had a stronger effect on the accuracy of predictions than the amount of the data collected. Specifically, when we used 10% of the original dataset (3.01e5 entries) with high consideration of spatial measurements, the best-performing models (Random Forest and XGBoost) were highly accurate (R2= 0.913 andR2= 0.935, respectively). The top three most influential features of all models were light, day and hour, underscoring the importance of the temporal dimension. This approach allowed us to explore the potential of leveraging advanced computational methods to improve the estimation of water loss under various environmental conditions.