High variation topsoil pollution forecasting in the Russian Subarctic: Using artificial neural networks combined with residual kriging

Abstract

The work deals with the application of artificial neural networks combined with residual kriging (ANNRK) to the spatial prediction of the anomaly distributed chemical element Chromium (Cr). In the work, we examined and compared two neural networks: generalized regression neural network (GRNN) and multi-layer perceptron (MLP) as well as two combined techniques: generalized regression neural network residual kriging (GRNNRK) and multi-layer perceptron residual kriging (MLPRK). The case study is based on the real measurements of surface contamination by Cr in subarctic city Novy Urengoy, Russia. The networks structures have been chosen during a computer simulation based on a minimization of the root mean square error (RMSE). Different prediction approaches are compared by a Spearman's rank correlation coefficient, the mean absolute error (MAE), and RMSE. MLPRK and GRNNRK show the best predictive accuracy comparing to kriging and even to MLP and GRNN, that is hybrid models are more accurate than solo models. The most significant improvement in RMSE (15.5% compared to kriging) is observed in the MLPRK model. The proposed hybrid approach improves the high variation topsoil spatial pollution forecasting, which might be utilized in the environmental modeling.