Application of a set of heterogeneous neural networks to modelling soil classification in mining regions

摘要

The paper presents using the set of heterogeneous neural networks to model the soil classification. It is base for assessment of the effects of changes in soil transformation conditions in mining regions. Digitised cartographical-soil and topographic documentation was the data source. In case of modelling morphological-classification interrelations for relatively small areas, which are generally located within a single basin, certain homogeneity (both lithological and morphological) allows to build a classification model giving correct indications for more than 95% of validation data. Large areas of strong lithological and morphological diversification are characterised by higher complexity of classification principles. In this situation we may consider using a set containing greater number of classifiers as a tool for modelling and predicting. The paper presents modelling results for data from the Upper Silesian Industrial Basin region (Poland), where a heterogeneous set of classifiers has been employed (2 MLP type classifiers, 2 classifiers representing an example of probabilistic network, linear classifier and classifier with the RBF units). The MLP and FSM (Feature Space Mapping) networks have been used as the first level classifiers (processing the results of indications by zero level classifiers). Employing a set of classifiers and FSM networks gave better results and considerably improved model quality compared to single classifiers.