Predictive Soil Pollution Mapping: A Hybrid Approach for a Dataset With Outliers

Abstract

Spatial regression or interpolation is widely used for predictive soil pollution mapping, which aims to estimate all unobserved soil pollution based on a finite number of sample points. However, it may be unreasonable to use spatial regression or interpolation directly for an environmental soil dataset with outliers, because the mechanism generating outlier datasets is always different from that generating normal datasets, which necessitates handling outliers separately. Therefore, a hybrid approach for estimating unknown soil pollution concentrations is developed in this paper. The hybrid approach comprises three main steps: First, spatial outlier detection is used to uncover abnormal sample points and the study area is then divided into the normal and outlier areas. Second, spatial regression and interpolation are applied to analyze the normal and outlier datasets, respectively. Finally, the results of the predictive soil pollution mapping are derived from the prediction combination of spatial regression and interpolation. An environmental dataset recording heavy metal Cd and As concentrations at Huizhou, China was selected to verify the performance of the proposed approach. The numbers of identified outlier points of heavy metal Cd and As concentrations were 16 and 13. For the prediction result of Cd, the mean square error (MSE) and mean relative error (MRE) of the hybrid approach were about 0.028 and 0.332, respectively. For the prediction result of As, the MSE and MRE of the hybrid approach were about 3.834 and 0.366, respectively. All of these values were smaller than those of models used for comparison. The result of the comparative analysis demonstrates the feasibility and effectiveness of the proposed approach.