Prediction of arsenic in tropical soils by machine learning algorithms

Physical–chemical soil properties are influenced by soil formation factors. Although these variables are valuable for soil mapping, they are poorly explored when mapping heavy metals in soil. Machine learning has shown satisfactory results in predicting heavy metal distribution in small, homogeneous sites, but modeling in highly diverse regions remains challenging. In this work, we evaluate ML algorithms for predicting arsenic (As) in the soils of the state of Minas Gerais—an area with dense soil data and high geo-pedological diversity. We combine spatial data analysis and predictive modeling to investigate soil–As relationships, characterizing spatial structure via autocorrelation, variograms, and multiscale Moran’s eigenvector maps. We fit both a standard Random Forest (RF) and a spatial RF (RFsp). The RF baseline predicted poorly when considering different spatial folds (oob R² = 0.02), whereas RFsp achieved an out-of-bag R² of 0.70 and a full-dataset R² of 0.98, demonstrating that explicitly modeling spatial autocorrelation vastly improves arsenic content prediction. We furthermore show that silt, clay, and pH KCl are the strongest non-spatial predictors, and we illustrate how spatial patterns and interactions drive model accuracy. RFsp therefore offers a powerful, parsimonious tool for mapping As contamination across heterogeneous landscapes.