Yield and nutrient composition of fertigated staked tomato in southern Brazil

Abstract Fertigated staked tomato (Solanum lycopersicon) is a highly productive crop grown in Santa Catarina State (SCS), Brazil. The timing and dosage of different inputs have been tested but their integration into a crop model is still pending. Our objective was to decrypt the tomato database using machine learning (ML) methods and to generate nutrient standards at high yield level. Managerial, edaphic, physiological and climatic features were documented at experimental sites from 2006 to 2020 in Caçador, Southern Brazil. Features were related to yield using the Random Forest and Gradient Boosting ML models. The models were accurate (R2 = 0.852-0,855). Tissue nutrients and fertilization were the most important features, followed by climate and soil features. Soil management and previous crops showed little importance. The tissue concentration values were centered-log-ratio (clr) transformed to compute nutrient standards at high yield level. While N is known to impact the incidence of pests, tissue N was loosely related to its clr value where tissue N was adjusted to the geometric mean of all nutrients including those that impact pest incidence. To assess the capacity of ML models and nutrient standards to generalize to unseen cases, universality tests should be conducted in farmers’ fields before general use.