A Comparison between a Modified Counter Propagation Network and an Extended Self-Organizing Map in Remotely Sensed Data

  • PDF / 744,751 Bytes
  • 16 Pages / 439.37 x 666.142 pts Page_size
  • 46 Downloads / 149 Views

DOWNLOAD

REPORT


A Comparison between a Modified Counter Propagation Network and an Extended Self-Organizing Map in Remotely Sensed Data Classification Yongliang Chen · Micha I. Pazner · Wei Wu

Received: 3 February 2006 / Accepted: 12 February 2007 / Published online: 6 September 2007 © International Association for Mathematical Geology 2007

Abstract A modified counter propagation network model and an extended selforganizing map model have the same three-layer network architecture while employing slightly different learning rules. Their network architecture comprises an input layer, a Kohonen layer and an output layer. The neurons between two neighboring layers are fully connected and the neighboring neurons within the Kohonen layer also have neighborhood connections. The modified counter propagation network model employs the Kohonen algorithm to train the Kohonen layer while using the Widrow– Hoff rule to train the output layer. However, the extended self-organizing map model applies a modified Kohonen’s learning rule to train both the Kohonen layer and the output layer. This paper compares the performances of these two models in supervised classification of remotely sensed data. The training results show that compared to the extended self-organizing map model, the modified counter propagation model has faster learning speed but larger output errors. The classification results indicate that the extended self-organizing map model has a faster classification speed and a much higher classification precision than the modified counter propagation model. Y. Chen () Comprehensive Information Institute of Mineral Resources Prediction, Jilin University, 6 Ximinzhu Street, Changchun, Jilin Province 130026, China e-mail: [email protected] Y. Chen Northeast Institute of Geography and Agricultural Ecology, Chinese Academy of Sciences, Changchun, Jilin Province 130012, China M.I. Pazner Geography Department, Social Science Center, The University of Western Ontario, London, Ontario N6A 5C2, Canada W. Wu Research and Design Institute of Urban and Rural Planning of Changchun, 1893 Tongzhi Street, Changchun, Jilin Province 130021, China

560

Math Geol (2007) 39: 559–574

Keywords Artificial neural network · Remotely sensed data · Supervised classification · Counter propagation · Kohonen layer

Introduction Counter propagation network (CPN) was initially introduced by Nielsen (1987). The CPN consists of three layers—an input layer, a competitive layer and an output layer. The input layer is responsible for receiving, preprocessing, and propagating input signals. The competitive layer, fully connected with the input layer, is also called a Kohonen layer. Training of this layer is based on Kohonen’s (1981) self-organization algorithm. The output layer, called a Grossberg layer, is fully connected with the Kohonen layer and trained on Grossberg’s (1980) outstar algorithm. This neural network model provides a practical approach for the pattern mapping tasks because learning is fast in this network. Several articles documented various applica