The influence of the discrepancies in the observers decisions on the process of identification of maceral groups using artificial neural networks

When analyzing the sorption properties of coal in the context of gas hazards in underground mining, focus should be placed on a number of aspects. These include the internal structure of coal, its structural properties, and its maceral composition. The share of particular macerals is most often determined manually, which e due to the huge diversity of petrographic features e can pose some difficulties even for experienced petrographers. Over the past few years, studies have been carried out into the use of artificial neural networks as a means to develop a methodology that would enable the identification of maceral groups based solely on the knowledge which the computer gains from sample macroscopic images and the information about their content provided by the observer. This paper investigates the effect that the selection of the training set, defined by particular teachers e petrographers, has on the effectiveness of selected neural classifiers. The research was carried out with the participation of three expert petrographers, who classified maceral groups in the macroscopic images of a lump sample of coal extracted from the Upper Silesian Coal Basin. Next, the feature space describing particular classes was defined using image analysis methods. The parameters defining that space were determined each time within a certain neighborhood of the studied points. Thus, the obtained sets were used to train neural networks (MLP) and to indicate the optimal network architecture for each expert. The research also included the identification of the influence that the change of the “teacher” e observer has on the identification process of the analyzed objects, as well as the automatic analysis of those measuring points which were classified differently by the observers e petrographers. The results presented concerning the objectivization of the quantitative analyses of coal indicate that modern methods of image analysis and artificial neural networks can contribute to the improvement of these measurements. However, it requires close cooperation between the designer of the neural network and the expert e petrographer.