Analysis of Neural Networks Illogical Behavior in Classification Problems at the Environmental Monitoring of Water Bodies

Abstract

The problem of illogical behavior of classifying neural networks, arising in the tasks of environmental monitoring of reservoirs, is considered. The paper presents the original method of analyzing neural networks, which allows to identify illogical behavior, i.e., false-positive results of classification and generalization. The problem of generalization of observations, belonging to classes not involved in training process, the so-called problem of external generalization problem, is analyzed. In the framework of the proposed technique methodology, a method for verification of training quality for a classifying neural network is considered, based on a comparison of the results of classification with the results of linear discriminant analysis. Implementation of the developed technique was performed for current ecological problem - classification of blue-green algae according to the self-fluorescence spectra during toxical “bloom” of water bodies.