Schemes for integrating multiple neural networks for Multispectral Data Classification

Supervised classification of remotely sensed data using artificial neural networks (ANNs) is largely limited by the number of nodes and architecture of the ANN, given the limited training samples. The best practice is to train a number of networks and choose the one which gives maximum classification accuracy over the data other than the training data set. Recently, consensus schemes using multiple classifiers have been attempted to overcome this problem. In the present study, four recent techniques proposed for combining multineural networks are examined for classifying the satellite imagery. Of these, selector net approach proposed by Partridge and Griffith is found to yield the best classification accuracy.