Using the fraction of transpirable soil water to estimate grapevine leaf water potential: comparing the classical statistical regression approach to machine learning algorithms

Abstract

Context and purpose of the study – Weather uncertainty is forcing Mediterranean winegrowers to adopt new irrigation strategies to cope with water scarcity while ensuring a sustainable yield and improved berry and wine quality standards. Therefore, more accurateand high-resolution monitoring of soil water content and vine water status is a major concern. Leaf water potential measured at pre-dawn (Y_PD) is considered to be in equilibrium with soil water potential and is highly correlated with soil water content at the soil depth where roots extract water.
The aim of this study is to evaluate a dataset of eco-physiological data collected in a 3-year vineyard irrigation trial to assess the explanatory power of the fraction of transpirable soil water (FTSW) to predict Y_PD by comparing the classical statistical regression approach with a machine learning algorithm (MLA).

Material and methods – Deficit irrigation trials were conducted from 2013 to 2015 in a commercial vineyard in the Alentejo (southern Portugal). Trial plot was planted with Vitis vinifera (L.) cv. Aragonez (ARA)(syn. Tempranillo), grafted onto 1103 Paulsen rootstock and spaced 1.5 m within and 3.0 m between N-S oriented rows. The experimental layout was a randomized complete block design with two treatments: sustained deficit irrigation (SDI – control; ~30% Etc) and regulated deficit irrigation (RDI; ~15% Etc) and 4 replicates per treatment. The Y_PD and soil water content were measured the day before and the day after each irrigation event by using a capacitance probe down to a soil depth of 1 m and a Scholander pressure chamber. Models predicting Y_PD from FTSW were trained on 600 data cases and validated on an independent dataset (10% of all available data) using MATLAB R2022b (Mathworks, USA) and STATISTICA 13 (Tibco, USA). 

Results – Our results show that 87.6% of the observed Y_PD variability is explained by the FTSW using a linear regression model (LRM) with a linear-logarithmic transformation of the independent variables. The accuracy of the prediction model, as measured by root mean squared error (RMSE), in the independent validation dataset, was 0.08 MPa. These results were compared to the estimation accuracy of a set of MLAs. Two support vector machine (SVM) algorithms with a quadratic and a medium Gaussian kernel function, and three Gaussian process regression (GPR) algorithms with an exponential, a squared exponential and a rational quadratic kernel functions were tested. All trained MLAs showed an accuracy in explaining the variability of the Y_PD (86-87%) similar to the LRM. An increase in the model explained variability of the independent dataset from 89 to 91% was observed in all MLAs, with an accuracy of 0.087 to 0.096MPa as measured by the RMSE.
Both statistical methods indicate that Y_PD can be estimated with good accuracy using FTSW as an explanatory variable. Regarding the comparative performance of the two types of statistical models no differences were found in the ability of the tested models to estimate Y_PD.