Machine learning models for predicting soil particle size fractions from routine soil analyses in Quebec

Soil texture and particle size distribution are important soil properties for understanding and interpreting the multiple processes and interactions in soil agrosystems. However, soil particle size analysis is rarely included in routine soil laboratory analyses due to cost but can be derived from other soil physico-chemical properties and from more powerful regressions using machine learning techniques. The performance of multiple linear regression and four machine learning algorithms (MLAs)-K-nearest neighbor (KNN), Random Forest, extra-gradient boosting (XGBoost), and multilayer neural network (NeuralNetwork)-was compared to predict particle size fractions of clay, sand, and silt using routine analyses of soil pH, cation exchange capacity, Mehlich-3 extracted elements, and density of dried sieved soil from 8,364 soil samples distributed across Quebec. Particle size fractions were predicted as compositional data using isometric log ratio transformation. The XGBoost model performed best, with RMSE values of 7, 10, and 12% and R-2 values of .77, .57, and .73 for the prediction of clay, silt, and sand fractions, respectively. Feature importance classification varied from one model to another, but the best predictors remained the same regardless of the model used. Sieved soil density measured in the laboratory and Mehlich-3 Mg, K, Fe, and Zn ranked as better predictors of particle size fractions. Predicting soil particle sizes using routine soil physico-chemical properties and MLAs appears an option for incomplete legacy datasets and a promising economic alternative to current methods carried out in commercial laboratories.