A Machine-Learning Framework for the Spatial Distribution Modeling of Potentially Toxic Elements in Urban and Peri-Urban Soils

Abstract

Despite significant advances in knowledge about the spatial distribution of potentially toxic elements (PTEs), the accurate determination them in urban areas remains a challenge due to multiple sources. We investigated concentrations and spatial patterns for six PTEs, namely Pb, Cu, Ni, Zn, Cd and As in the top layer of urban and peri-urban soils. The study was conducted on the territory of the Ufa city (Russia), with a population of more than 1 million people and an area of more than 700 km². For these purposes, a total of 250 soil samples were collected at 0–20 cm depth. Random Forest algorithm, in combination with environmental and anthropogenic variables, was applied for the spatial prediction of PTEs. The covariates were represented by distance from river, topographic attributes, remote sensing data, geology and soil properties, distance from highways and railroads, their density, distance from combined heat and power, and refineries. Results showed that Pb, Cu, Ni, Zn, Cd and As contents ranged from 1 to 98.1 mg/kg, 1.5 to 360 mg/kg, 1.7 to 110 mg/kg, 1 to 336 mg/kg, 0.2 to 1 mg/kg, and 0.1 to 7.4 mg/kg, respectively. The average values of elements did not exceed the maximum and approximate permissible concentrations (MPC, APC). Terrain and anthropogenic-related covariates were estimated as the most important predictors. The generated maps revealed geographic trends and hotspot concentrations of Pb, Cu, Ni and Zn. Our findings and generated maps can provide useful information for future digital soil mapping of PTEs in urban areas.