Compositional and Machine Learning Tools to Model Plant Nutrition: Overview and Perspectives

The ceteris paribus assumption that all features are equal except the one(s) being examined limits the reliability of nutrient diagnosis and fertilizer recommendations. The objective of this paper is to review machine learning (ML) and compositional data analysis (CoDa) tools to make nutrient management feature specific. The accuracy of the ML methods averaged 84% across the crops. The additive and orthogonal log ratios of CoDa reduce a D-parts soil composition to D-1 variables, alleviating redundancy in the predictive ML models. Using a Brazilian onion (Allium cepa) database, the combined CoDa and ML methods returned crop response patterns, allowing feature-specific fertilizer recommendations to be made. The centered log ratio (clr) diagnoses plant nutrients as a compositional nutrient diagnosis (CND). Using a Quebec database of vegetable crops, the mean variance of clr variables (VAR¯) allowed comparing total variance among species and growth stages. While clr is the summation of equally weighted dual log ratios, dual nutrient log ratios may show unequal importance regarding crop performance. The RReliefF scores, gain ratios or gini inequality coefficients can provide weighting coefficients for each dual log ratio. The widely contrasting coefficients of weighted log ratios (wlr) improved the accuracy of the ML models for a Quebec muck onion database. The ML models, VAR¯ and wlr, are advanced tools to improve the accuracy of nutrient diagnosis.