Segmentation of remote-sensing images by incremental neural network Mehmet Nadir Kurnaz, Zu ¨ mray Dokur * , Tamer O ¨ lmez Department of Electronics and Communication Engineering, Istanbul Technical University, 80626 Maslak, Istanbul, Turkey Received 7 January 2003; received in revised form 3 June 2004 Available online 23 November 2004 Abstract In this study, a novel incremental neural network (INeN) is proposed for the segmentation of remote-sensing images. The data set consists of seven images acquired by the Landsat-5 TM sensor. Two feature extraction methods are comparatively examined for the segmentation of the remote-sensing images. In the first method, features are formed by the intensity of one pixel of each channel. In the second method, intensities at one neighborhood of the pixel are used to form the feature vectors. In this study, the INeN and the Kohonen network are employed for the segmentation of the remote-sensing images. The INeN is proposed to determine the number of classes automatically. Ó 2004 Elsevier B.V. All rights reserved. Keywords: Artificial neural network; Image segmentation; Remote-sensing; Self-organised map 1. Introduction The constitution of the right data space is a common problem in connection with segmenta- tion/classification. In order to construct realistic classifiers, the features that are sufficiently repre- sentative of the physical process must be searched. In the literature, it is observed that different trans- forms are used to extract desired information from remote-sensing images or biomedical images. In the literature, it is observed that artificial neural networks are widely used for the segmenta- tion of remote-sensing images (Berberoglu et al., 2000; Jozwik et al., 1998; Bruzzone and Fernan- dez, 1999; Giacinto et al., 2000; Serpico et al., 1996; Chen et al., 1999; Villmann et al., 2003). Also, the combination of the neural networks and statistical algorithms is used for the segmenta- tion of the remote-sensing images (Giacinto et al., 2000; Serpico et al., 1996). In most studies, single pixel intensity of each channel is used to determine 0167-8655/$ - see front matter Ó 2004 Elsevier B.V. All rights reserved. doi:10.1016/j.patrec.2004.10.004 * Corresponding author. Fax: +90 2122853565. E-mail address: zumray@ehb.itu.edu.tr (Z. Dokur). Pattern Recognition Letters 26 (2005) 1096–1104 www.elsevier.com/locate/patrec