Machine learning models to predict daily actual evapotranspiration of citrus orchards under regulated deficit irrigation

Precise estimations of actual evapotranspiration (ETa) are essential for various environmental issues, including those related to agricultural ecosystem sustainability and water management. Indeed, the increasing demands of agricultural production, coupled with increasingly frequent drought events in many parts of the world, necessitate a more careful evaluation of crop water requirements. Artificial Intelligence-based models represent a promising alternative to the most common measurement techniques, e.g. using expensive Eddy Covariance (EC) towers. In this context, the main challenges are choosing the best possible model and selecting the most representative features. The objective of this research is to evaluate two different machine learning algorithms, namely Multi-Layer Perceptron (MLP) and Random Forest (RF), to predict daily actual evapotranspiration (ETa) in a citrus orchard typical of the Mediterranean ecosystem using different feature combinations. With many features available coming from various infield sensors, a thorough analysis was performed to measure feature importance, scatter matrix observations, and Pearson's correlation coefficient calculation, which resulted in the selection of 12 promising feature combinations. The models were calibrated under regulated deficit irrigation (RDI) conditions to estimate ETa and save irrigation water. On average up to 38.5% water savings were obtained, compared to full irrigation. Moreover, among the different input variables adopted, the soil water content (SWC) feature appears to have a prominent role in the prediction of ETa. Indeed, the presented results show that by choosing the appropriate input features, the accuracy of the proposed machine learning models remains acceptable even when the number of features is reduced to only 4. The best performance was achieved by the Random Forest method, with seven input features, obtaining a root mean square error (RMSE) and a coefficient of determination (R2) of 0.39 mm/day and 0.84, respectively. Finally, the results show that the joint use of SWC, weather and satellite data significantly improves the performance of evapotranspiration forecasts compared to models using only meteorological variables.