Field-scale soil organic matter prediction: evaluating machine learning and deep learning models

Abstract

Within the framework of agriculture, soil features maps availability, as for example those of soil organic matter (SOM) facilitates decision-making in agricultural management. Many techniques have been developed to generate these field scale maps. The objective of this study was to compare the prediction capability of machine learning models for field scale SOM mapping. Quantile Regression Forests (QRFI) and Artificial Neural Networks (ANN) were evaluated using soil features and crop yield as covariates. Regression Kriging was used as reference method. Spatial interpolation was performed in seven fields located in Córdoba and Santiago del Estero provinces, Argentina. Models were fitted using information from all fields (global models) or only from the target field (local models). For all fields, ANN models presented better fit, showing lower normalized root mean squared errors. Sample size per field, as well as the joint variability of SOM and covariates affected the performance of global versus local models. ANN models stand as a promissory option for field scale soil variability mapping.