Grapevine crop evapotranspiration and crop coefficient forecasting using linear and non-linear multiple regression models

Vineyard irrigation management relies on accurate assessment of crop evapotranspiration (ETc). ETc is affected by the by type of plant, its physiological properties, and meteorological parameters. Rapid measurement of these factors facilitates quantification of ETc and enables skilled decision-making for data-driven irrigation. Our main objective was to quantify the performance of different modeling approaches for forecasting seasonal ETc using meteorological and vegetative data (e.g., leaf area) from five consecutive growing seasons (2013–2017) of Vitis vinifera 'Cabernet Sauvignon' vines. Time series of ETc was acquired from water balance from vines grown in drainage lysimeters within the vineyard. ETc forecasts were generated for each season using twelve regression models: six linear and six non-linear multivariate adaptive regression spline (MARS) models. Each regression model constituted a unique combination of variables, some relying on crop coefficient (Kc) and others based on direct ETc forecasting. The models were trained using data from four growing seasons and compared via measures of coefficient of determination (R2), residual standard deviation, and coefficient of variation. Each model was then tested using ETc forecasts for a fifth growing season, and compared to the measured ETc values using correlation, root mean squared error (RMSE), and normalized RMSE. Finally, a mean-seasonal rolling RMSE with a window of 7 days was used to assess the accuracy of the different models. The results show a clear advantage to using non-linear modeling for ETc forecasting (average RMSE range of 0.81–1.05 vs. 0.64–0.71 mm day−1, respectively). Furthermore, direct forecasting and Kc-based methods yielded similar results, and all models benefited from the incorporation of leaf area data. Similar outcomes were found for the rolling RMSE analysis, with improved model accuracy credited to the inclusion of leaf area, especially early in the season. Our findings confirm that advanced algorithms promote site-specific and location-oriented irrigation management.