Assessment of the Spatial Variability of Metal Contaminants Using Digital Mapping

Abstract

This study utilized the methodology of digital soil mapping (DSM) to investigate the spatial prediction of toxic metals and their environmental covariates in the Ghorveh Plain, western Iran. The environmental covariates are defined as the factors that control the distribution of toxic metals at the geographical scale under investigation. They could be used for predicting the sources and monitoring of pollution. A total of 150 soil samples (0-30 cm) were analyzed for toxic metal concentrations and some soil properties. A comprehensive set of environmental variables was obtained from remote sensing imagery, DEM, and ancillary data, which were identified as likely to control the spatial distributions of toxic metals. The genetic algorithm was utilized to identify "all-relevant" environmental covariates for each toxic metal. Three machine learning algorithms, namely random forests (RF), cubist, and regression trees (RT), were employed to establish the statistical relationships between toxic metals and the environmental covariates. The RF model exhibited the most optimal prediction performance. All three models, particularly the RF, demonstrated robust performance, exhibiting minimal impact on the model's functionality when confronted with alterations in the training and testing data. Consequently, the optimal model, RF, was integrated with a bootstrapping method to generate prediction and uncertainty maps. The soil properties and hydrologic factors were the primary variables influencing the spatial distribution of each toxic metal. This study indicates that the integration of DSM techniques with machine learning models and supplementary datasets offers a viable approach to the generation of maps for monitoring and prioritizing remediation measures in areas contaminated by toxic metals.