DEVELOPMENT OF A SOFTWARE FOR LAND COVER CLASSIFICATION FROM SATELLITE DATA USING ANN ALGORITHM

Abstract

Usually the classification of land cover classes is done through conventional surveying method based on field survey observations. But, due to advancement of space technology, now, Satellites data are available for this purpose. Earlier, classification of these data was done by using traditional statistical classifier. This dissertation suggests the use of Artificial Neural Network (ANN) for classification of remotely sensed data. For this purpose a software named Neuro-Net is developed using ANN algorithm in "C" language. It is based on back-propagation learning network (BPLN). Neuro-Net is developed for supervised classification technique. If data is very complex then a large number of connections are required to discriminate the different classes easily. Neuro-Net provides facility to do so by increasing the number of hidden layers upto three. Neuro-Net is tested for its performance using real remotely sensed data. Results obtained suggests that training time is dependent upon some network's parameters like learning rate, momentum factor, number of hidden layers, number of nodes in each layer etc. Accuracy of classification is also dependent upon above mentioned parameters. Neuro-Net has some limitations like, it does not include adaptive learning of the network, it is designed for a maximum number of input nodes equal to six (six bands for remotely sensed data)