Prediction of Cadmium Content Using Machine Learning Methods

Abstract Heavy metals are the most environmentally hazardous pollution type in agricultural soils, threatening human and ecological health. Cadmium (Cd) is a highly toxic element but distinctively different with its high mobility in soil environments. The study aimed to evaluate the Cd concentration of Konya plain soils with a specific attribute to soil fertilization practices, mainly phosphorous fertilizers. A total of 538 surface (0-20 cm) soil samples were analysed for the routine soil properties and total phosphorus (P) and Cd. Descriptive statistics, machine learning and regression models considered the accumulation of Cd in soils. Among the MARS, Decision Trees, Linear Regression, Random Forest, and XGBoost machine learning methods used in Cd prediction, the XGBoost model proved to be the best prediction model with a coefficient of determination of 98.1%. EC, pH, CaCO3, silt, and P2O5, which are the soil components used in Cd estimation of XGBoost model, explained 56.51% of the total variance in relation to measured soil properties. Therefore machine learning processes could be a useful tool to estimate the nature of an element in the soils of a specific region by using routine soil properties.