On the use of consensus algorithms to address variability in the results of neural network classifications: Preliminary tests involving two northern study areas

In past work it has been recognized that variations in parameters such as learning rate, momentum, and network architecture can influence the results in neural network classifications of satellite images. New tests suggest that variation in the results of neural network classifications, caused solely by differences in weight initializations, can also be substantial. This issue has the potential to limit the applicability of neural networks in remote sensing classifications. The negative effects of variation in neural network results can potentially be reduced or eliminated through application of consensus algorithms in which the outputs of multiple neural network classifications are combined. Research results presented here were based on training and test data with low sample sizes for many classes and, accordingly, the results must be interpreted with caution. Early results using majority-vote and evidential-reasoning consensus algorithms, however, suggest that near-optimum neural network classification accuracies can be achieved through application of these algorithms.