The soil is an informative indicator of the technogenic pollution and the state of the ecosystem. The reconstruction of the spatial distribution of heavy metals in the topsoil presents by the method of arti cial neural networks (ANN) trained on a representative subset. An important factor in creating an ANN model is the construction of a training sample. The various schemes for constructing an ANN training sample for the different tasks have been proposed, but the choice of the method for splitting the environmental monitoring raw data for ANN training has not been fully determined. This work is presented a four-stage algorithm to construct a representative training subset that allows minimizing the forecast accuracy when modelling the spatial distribution of heavy metals in the topsoil. As raw data are the content of chromium (Cr) and manganese (Mn) in the topsoil in the residential area of the city of Noyabrsk, Yamalo-Nenets Autonomous Okrug (subarctic zone of Russia). Spatial distributions of element contents were modeled by a multilayer perceptron (MLP) with sigmoid and hyperbolic tangential activation functions. To assess the model accuracy, the root mean square error (RMSE) was calculated for each element content. The best accuracy was shown by the MLP model with the activation function of the hyperbolic tangent by about 10 %. The proposed algorithm provides lower average RMSE values for the spatial distribution of the element in the presence of outliers. For unimodal spatial distributions without singularities, the four-stage algorithm does not provide high prediction accuracy. The model was trained on representative points that most often fall into the training subset and provide the smallest model error. According to the number of hits in the training subset, the points were divided into three classes: “elite”, “medium” and “useless”.