Forecasting of Spatial Variable by the Models Based on Artificial Neural Networks on an Example of Heavy Metal Content in Topsoil.

Retrieving additional information about a spatial distribution of chemical elements in topsoil is an important part of environmental protection. Artificial neural networks (ANN) such as generalized regression neural networks (GRNN) and multilayer perceptrons (MLP) are used to predict the spatial distribution of chemical element Copper (Cu) and Nickel (Ni) in topsoil. The initial data are soil survey data sets on surface content of Cu and Ni at a specific location in subarctic Novy Urengoy, Russia. The network structures were chosen during computer modeling, based on the minimization of the root mean squared error (RMSE). Each network has advantages and limitations for the content of the different elements in the various soil types. The hybrid models combining the techniques of ANN and geostatistics consists of several steps: 1) comparing of kriging estimates and network ones; 2) building residues; 3) residual kriging and combination with ANN estimates. The work confirms that trained ANN is suitable for modeling both the normal (Cu) and abnormal (Ni) spatial distribution of pollutants. The best prognostic hybrid model was model of the multilayer perceptron with residual kriging (MLPRK), which allows improving the predictive accuracy for the considered elements. [ABSTRACT FROM AUTHOR]