Estimating kiwifruit leaf stomatal conductance through artificial neural networks

There is considerable literature attempting to model leaf stomatal conductance (gs), both through mechanistic models and data-driven models. This study aimed at estimating gs of kiwifruit vines through the comparison of artificial neural network models trained with different variables. To do so, in 2023, kiwifruit vines were submitted to four irrigation treatments, which were established as percentages of crop evapotranspiration (ETc), being 100, 68, 57 and 40% ETc. Nine days during the 2023 fruit growing season gs measurements were performed throughout the day. Four vines per treatment were monitored with sap flow sensors. Additionally, to perform a further test of the performance of the models, in 2024 an experiment was conducted in a different orchard, on vines of a different variety which were monitored with sap flow sensors and daily measurements of gs were performed. The data-driven models were obtained through artificial neural networks (ANN), which were trained considering different variables as input. The input variables were: vapor pressure deficit (D), photosynthetically active radiation (PAR), leaf and air temperature and sap flux density (Js) and soil water content (SWC). The data-driven model with less input variables (D, PAR and Js) presented low coefficient of determination (R2≤0.26). The highest correlation coefficient and R2 were obtained by the ANN trained with D, PAR and leaf and air temperature (r≥0.901, R2≥0.810). When tested against the data obtained in 2024, the performance of these models was much lower, although obtaining correlation coefficients greater or equal to 0.693. These results demonstrate the potential use of data-driven models to estimate gs.