Utilizing machine learning algorithm for finely three-dimensional delineation of soil-groundwater contamination in a...

Intensive industrial activities cause soil contamination with wide variations and even perturb groundwater safety. Precision delineation of soil contamination is the foundation and precondition for soil quality assurance in the practical environmental management process. However, spatial non-stationarity phenomenon of soil contamination and heterogeneous sampling are two key issues that affect the accuracy of contamination delineation model. Taking a typical industrial park in North China as the research object, we constructed a random forest (RF) model for finely characterizing the distribution of soil contaminants using sparse-biased drilling data. Results showed that the R² values of arsenic and 1,2-dichloroethane predicted by RF (0.8896 and 0.8973) were greatly higher than those of inverse distance weighted model (0.2848 and 0.2908), indicating that RF was more adaptable to actual non-stationarity sites. The back propagation neural network algorithm was utilized to establish a three-dimensional visualization of the contamination parcel of subsoil-groundwater system. Multiple sources of environmental data, including hydrogeological conditions, geochemical characteristics and anthropogenic industrial activities were integrated into the model to optimize the prediction accuracy. The feature importance analysis revealed that soil particle size was dominant for the migration of arsenic, while the migration of 1,2-dichloroethane highly depended on vertical permeability coefficients of the soil. Contaminants migrated downwards with soil water under gravity-driven conditions and penetrated through the subsoil to reach the saturated aquifer, forming a contamination plume with groundwater flow. Our findings afford a new idea for spatial analysis of soil-groundwater contamination at industrial sites, which will provide valuable technical support for maintaining sustainable industry.