The extraction of the training subset for the spatial distribution modelling of the heavy metals in topsoil

Abstract

The choice of the method of raw data dividing into train and test subsets in the models based on the artificial neural networks (ANN) is one of the underexplored problems of continuous Spatio-temporal field interpolation. Selecting the best training subset for modelling the spatial distribution of elements in the topsoil is not a trivial task since the sampling points are not equivalent. Errors and outliers, which are present in the distribution, can be misleading. Most of the points, which contain utility information, should be involved in modelling. The raw data were Chromium (Cr) and Manganese (Mn) contents in the topsoil on the residential areas in Noyabrsk and Novy Urengoy (South Part) cities (Russian subarctic zone). A three-step algorithm for extraction of the raw data dividing into training and test subsets for modelling the spatial distribution of the feature presented. The spatial distributions of the element contents in the topsoil, which constructed by the multilayer perceptron (MLP), consider spatial heterogeneity and training rules. The MLP structure was chosen by the minimization of the root mean squared error (RMSE). For each territory, according to the number of hits in the training subset, the points divided into three classes: the «elite», «middle», and «useless». Considering this information at the stage of dividing the raw sample makes it possible to increase the accuracy of the predictive model.