Ju Hyun Back, Hong Gi Kim

The KIPS Transactions:PartB , Vol. 8, No. 4, pp. 335-342, Aug. 2001
10.3745/KIPSTB.2001.8.4.335,   PDF Download:

Abstract

In this paper, we propose a multi layer neural network for recognition of remote sensing data by a Voronoi diagram. The proposed multi-layer neural network distinguishes the classes with the Voronoi polygon area and initializes error back propagation''s initial learning of connecting weights, thresholds and the number of hidden layers'' nodes with the coefficient of hyper plane equation. The proposed method systematically decides the initial information of error back propagation, so that it have the advantage of improving the slow convergence and local minima of error back propagation. We calculated the hyper plane equation of boundary edges of the Voronoi diagram by converting the mean of each class of training sets into a Mathematica package. To measure the performance of the recognition power by the image classifier, which uses the multi-layer neural network proposed above, we compared it with minimum distance classification and maximum likelihood classification, which is frequently used in the recognition of remote sensing data. The results of minimum distance classification and maximum likelihood classification were 81.4%, 87.8% respectively for the majority of land cover features, while the an accuracy of the proposed method was 92.2%.

Statistics
Cite this article