Geostatistics or machine learning for mapping soil attributes and agricultural practices

ABSTRACT Applying the upcoming technologies in agriculture has been a major economic, environmental and social challenge for scientists and farmers. In order to overcome such challenge, this study evaluated the advantages and limitations of using geostatistics and machine learning for soil mapping in agricultural practices and soil surveys. The study occurred in Tocantins State, Brazil, and consisted into seven areas with a total extension of 17.24 km2, 222 meters regular gridded resulting in one-point sampling per 0.0493 km2 of five randomly sampled cores within a 1 m circle radius. It was collected 332 georeferenced soil samples at 0-20 cm depth using an auger and then, soil laboratory analyses performed. Afterward, liming rate maps were originated from the predicted soil attributes clay, cation exchange capacity and base saturation comparing four methods: ordinary kriging, random forest, cubist, support vector machine and the best model results of each soil attribute. Evaluating the methods, the Pearson’s index presented strong results for soil attributes predicted by random forest and ordinary kriging. Machine learning methods can be successfully applied for soil mapping in agricultural practices and soil surveys using less soil samples rather than geostatistical framework.