Abstract
Neural networks are growing in popularity today as a tool for classification of remotely sensed images. One of the major stumbling blocks for neural networks to be accepted for operational use in remote sensing has been the long computational times involved. This is particularly relevant in the case of the well-known back-propagation (BP) algorithm trained multilayer perceptron (MLP) classifier. The long time involved in training the network to realize the relationship between the training data and their assigned classes is due to lack of a sound theory to set the initial network parameters. Errors in classification are partly due to mixed pixels and absence of contextual information, which is almost always used by human interpreters while analyzing images. An approach is suggested here to take advantage of global optimization techniques to assist in proper training of the MLP classifier, and use of relaxation labelling algorithms to refine the MLP output based on the contextual information present within small local neighbourhoods of pixels. The results are illustrated using Indian Remote Sensing (IRS) images.
Additional information
Notes on contributors
B Krishna Mohan
B Krishna Mohan received his MTech and PhD from Indian Institute of Technology, Bombay in 1983 and 1991 respectively. He is currently working as Senior Research Engineer at Centre of Studies in Resources Engineering, Indian Institute of Technology, Bombay. His research interests include remote sensing, image segmentation, neural image classification, image compression, and geographic information systems. Dr Mohan has several publications in these areas to his credit, and supervised a number of postgraduate level dissertations at IIT Bombay on these topics. He also examined PhD theses from India and abroad. He is life member of Indian Society of Remote Sensing.