The Effect of Splitting of Raw Data into Training and Test Subsets on the Accuracy of Predicting Spatial Distribution by a Multilayer Perceptron

Abstract

The paper discusses the influence of various methods for splitting raw data into test and training subsets on the accuracy of the prediction of the spatial distribution of the variable for the model based on a multilayer perceptron. A comparison of models was performed taking into account both spatial heterogeneity and the spread of values of the modeled variable with a completely random splitting option. The study was based on the results obtained during soil screening of urbanized territories in the Russian subarctic zone, Novy Urengoy city and Noyabrsk city. The spatial distribution of the chemical element chromium (Cr) in the surface layer of the soil was modeled. For both surveyed areas, the models in which a controlled splitting method was used for training demonstrated more accurate results than models trained using completely random splitting.