Use of artificial neural network to evaluate cadmium contamination in farmland soils in a karst area with naturally high background values

Abstract

In recent years, the naturally high background value region of Cd derived from the weathering of carbonate has received wide attention. Due to the significant difference in soil Cd content and bioavailability among different parent materials, the previous land classification scheme based on total soil Cd content as the classification standard, has certain shortcomings. This study aims to explore the factors influencing soil Cd bioavailability in typical karst areas of Guilin and to suggest a scientific and effective farmland use management plan based on the prediction model. A total of 9393 and 8883 topsoil samples were collected from karst and non-karst areas, respectively. Meanwhile, 149 and 145 rice samples were collected together with rhizosphere soil in karst and non-karst areas, respectively. The results showed that the higher CaO level in the karst area was a key factor leading to elevated soil pH value. Although Cd was highly enriched in karst soils, the higher pH value and adsorption of Mn oxidation inhibited Cd mobility in soils. Conversely, the Cd content in non-karst soils was lower, whereas the Cd level in rice grains was higher. To select the optimal prediction model based on the correlation between Cd bioaccumulation factors and geochemical parameters of soil, artificial neural network (ANN) and linear regression prediction models were established in this study. The ANN prediction model was more accurate than the traditional linear regression model according to the evaluation parameters of the test set. Furthermore, a new land classification scheme based on an ANN prediction model and soil Cd concentration is proposed in this study, making full use of the spatial resources of farmland to ensure safe rice consumption.