The use of artificial neural networks in the determination of soil grain composition

Abstract

AbstractThe paper presents the possibility of using data mining tools — artificial neural networks — in prediction of hydrometer reading after 24 h in order to limit the duration of the test to 4 h. The authors analysed a database of 693 granulometric composition analysis results of genetically different soils with the use of radial basis function network (RBF) and multilayer perceptron (MLP). The calculations performed showed that it is possible to use MLP to shorten the test time without affecting the quality of the results. The presented accuracy of the model, in the range of 0.55–0.72, allows one to determine the content of silt and clay fractions with an accuracy of 0.49% for equivalent diameter (dT) and 1.50% for percentage of all particles with a diameter smaller than dT (ZT). The results were better than that achieved using linear re-gression models with all predictors (REG), stepwise regression models (SREG), and classification and regression trees (CRT). Taking into account the uncertainty of hydrometric determinations, the obtained forecast values is lower than this uncertainty, therefore neural networks can be used to predict the results of this type of research.