Modelling leaf water potential from physiological and meteorological variables – A machine learning approach

Abstract

Viticulture is a key economic sector in the mediterranean region. However, climate change is affecting global viticulture, increasing the frequency of heatwaves and drought events. Precision irrigation, supported by robust water status indicators (WSIs), is inevitable in most of the Mediterranean basin, and imperative when finding strategies to adapt to climate change phenomena. One of the most reliable WSIs is the leaf water potential (Ψleaf), which is determined via an intrusive and time-consuming method. The aim of this work was to evaluate which of the two Ψ methods (predawn water potential [Ψpd] and stem water potential [Ψstem]) better reflect the plants’ water status and identify the variables with better performance when predicting Ψleaf. Finally, model the Ψ selected.

Five varieties grown in the Alentejo region (Reguengos de Monsaraz, Portugal) were selected according to their described aniso/isohydric behavior: Syrah (Sy), Touriga Franca (TF), Vinhão (Vi), Touriga Nacional (TN) and Castelão (Cs), subjected to three long-term irrigation treatments: full irrigation (FI, 100% crop evapotranspiration), deficit irrigation (DI, 50% FI), and no irrigation (NI). Plant monitoring was performed in the 2023 season, over two weeks. Measurements included stomatal conductance (gs), Ψpd, Ψstem, thermal imaging, from which we calculated the difference between air and canopy temperatures (ΔT), the Crop Water Stress Index (CWSI) and the Jones index (Ig), and meteorological data. UAV flights with a multispectral and a thermal camera were also performed.

Ψpd and gs responded differently according to the irrigation treatment. Ψstem presented lower potential to discern between treatments, when compared to Ψpd. Mid-morning (MM) measurements presented the best correlations between WSIs. gs showed the best correlations between the other WSIs and was considered the best WSI to estimate Ψpd. Machine learning regression models were trained on meteorological, thermal and gs data to predict Ψpd. The model with the best performance was the ExtraTreesRegressor, with a coefficient of determination (R2) of 0.833 and a mean absolute error (MAE) of 0.072. UAV data also showed significant differences between treatments. Regression models were trained on the UAV data, and the best model was LinearRegression with a R2 of 0.751 and a MAE of 0.114.

We concluded that gs is an excellent WSI for estimating the Ψleaf, and the UAV data showed great potential for estimating Ψleaf.