Individual and pair representativeness of sampling points selection in interpolation tasks of the heavy metals distribution in the topsoil

Abstract

The article is devoted to the problem of choosing a representative selection of a subset for an artificial neural network in the tasks of interpolation of the distribution of metals in the topsoil. Environmental data, often used to build artificial neural network models, are datasets at irregular points. The traditional division of the input data into training and test subsets occurs randomly, which transfers to a number of problems. For selection in the training subset, the question of individual and collective representativeness of points is asked, sending them a request for data on the content of the element in the soil in a given area. The most representative in terms of individual representativeness arise with the maximum reference points, their presence in the training subset of the ANN measurement of error and an increase in the correlation between the results of model calculations and natural measurements on the test subset. When assessing the pairwise representativeness of the identified synergy effects, which, when included, achieve high model reliability) and anti-synergy (the parameters of using less information to describe the content of the elements than separately the points of view included in the pair). The various sampling locations have different information and unequal meaning for feature interpolation.