Training algorithms for artificial neural network in predicting of the content of chemical elements in the upper soil layer

Models based on Artificial Neural Networks (ANN) in recent years are increasingly being used in environmental studies. Among the many types of ANN, the network type Multilayer Perceptron (MLP) has become most widespread. Such networks are universal, simple, and suitable for most tasks. The main problem when modelling using MLP is the choice of the learning algorithm. In this paper, we compared several learning algorithms: Levenberg-Marquart (LM), LM with Bayes regularization (BR), gradient descent (GD), and GD with the speed parameter setting (GDA). The data for modelling were taken from the results of the soil screening of an urbanized area. The spatial distribution of the chemical element Chromium (Cr) in the surface layer of the soil was simulated. The structure of the MLP network was chosen using computer simulations based on minimization of the root mean squared error (RMSE). The model using the LM training algorithm showed the best accuracy.