A novel methodological framework for predicting and mapping agriculture-related soil attributes using Euclidean...

Recent advances in statistical and machine learning (ML) methods have improved the prediction of soil attributes at fine spatial scales, yet the comparative performance and reliability of these techniques remain unclear. This study compared Ordinary Kriging (OK), Inverse Distance Weighting (IDW), and ML algorithms in predicting and spatializing soil attributes, while also evaluating prediction uncertainty and computational processing time. Conducted in Minas Gerais State (Brazil), the analysis used Euclidean distance based predictors derived from X-Y coordinates and regular grids with 5, 7, and 10 divisions. Soil attribute maps (CEC, phosphorus, sand, and clay) were generated using OK, IDW, Random Forest (RF), Cubist, Support Vector Machine (SVM), and Earth. Model performance was assessed using R2, RMSE, MAE, and the coefficient of variation. IDW and OK showed the lowest predictive accuracy (R2 = 0.52-0.58), whereas ML methods, especially RF and SVM achieved superior performance (R2 = 0.62-0.70). Among ML algorithms, Earth performed worst, while RF produced the highest accuracy for all attributes except sand, for which SVM performed best. Processing time was shortest for IDW, followed by OK; among ML models, Earth was fastest, followed by RF, SVM, and Cubist. Larger regular grids improved ML prediction and spatialization but increased computational cost. ML methods thus outperform traditional geostatistical interpolators, benefiting from the use of numerous covariates and flexible algorithmic structures, although requiring greater computational time. These findings demonstrate the robustness and practical potential of ML approaches for soil attribute mapping.