An evaluation of consensus neural networks and evidential reasoning algorithms for image classification

Empirical investigations of classification algorithms for lithological discrimination were performed in this research using two study areas in northern Canada: Melville Island, Nunavut, and the Cape Smith Belt of northern Quebec. These investigations suggest that a consensus neural network classifier (a majority-vote algorithm that combines the classification results of ten feedforward backpropagation neural networks) is capable of consistently producing results that approximate those produced by the best individual neural network execution, and that are superior to those generated by the maximum likelihood and evidential reasoning classifiers. The majority-vote consensus routine serves to eliminate the effect of the natural variability among individual neural network classification results, by producing strong results without necessitating the manual evaluation and ranking of all individual neural network classifications. Image classifications produced using the evidential reasoning classification algorithm were generally very poor when measures of initial evidence were derived from the proportions of training data that contained the image values being classified. Evidential reasoning classifications that derived measures of initial evidence from output activations generated by multiple neural network classifications were typically comparable to those of the neural network consensus algorithm.