1 (YROXWLRQDU\$SSURDFKWR'LVFRYHU\ RI&ODVVLILFDWLRQ5XOHVIURP5HPRWH6HQVLQJ,PDJHV   J. KORCZAK, A. QUIRIN 8QLYHUVLWp/RXLV3DVWHXU/6,,7&1566WUDVERXUJ)UDQFH {korczak,quirin}@lsiit.u-strasbg.fr $EVWUDFW In this article a new method for classification of remote sensing images is described. For most applications, these images contain voluminous, complex, and sometimes noisy data. For the approach presented herein, image classification rules are discovered by an evolution-based process, rather than by applying an a priori chosen classification algorithm. During the evolution process, classification rules are created using raw remote sensing images, the expertise encoded in classified zones of images, and statistics about related thematic objects. The resultant set of evolved classification rules are simple to interpret, efficient, robust and noise resistant. This evolution-based approach is detailed and validated based on remote sensing images covering not only urban zones of Strasbourg, France, but also vegetation zones of the lagoon of Venice. ,QWURGXFWLRQ The design of robust and efficient image classification algorithms is one of the most important issues addressed by remote sensing image users. For many years a great deal of effort has been devoted to generating new classification algorithms and to refine methods used to classify statistical data sets [Bock, 1999]. At the time of this writing, relatively few workers in the machine learning community have considered how classification rules might be discovered from raw and expertly classified images. In this paper, a new data-driven approach is proposed to discover classification rules using the idea of evolutionary classifier systems. The unique source of information is a remote sensing image and its corresponding classification furnished by an expert. The images have been registered by various satellites (e.g. SPOT, LANDSAT, DIAS) that use different cameras having various spectral and spatial resolutions [Weber, 1995]. These types of remote sensing images generally contain huge volumes of data. And, sometimes they are very noisy due to coarse spatial resolution or unfavourable atmospheric conditions at the time the images are acquired. In addition, data may be erroneous due to inexperienced operators of the measurement devices. The aim of this research is to elaborate an evolutionary classification method that, in contrast to classical algorithms, will allow for supervised creation of autonomous classification. In general, classification rules are discovered from the established classifier system ([DeJong, 1988], [Wilson, 1999]). As we said, the system is data-driven because the formulated classification rules are able to adapt themselves according to the available data, environment, and the evolution of classes. In remote sensing, the initial population of classifiers is randomly created from images and given classes, and then evolved by a genetic algorithm until the acceptable solution is found. In the remote sensing literature, several classification approaches are presented, namely: - pixel-by-pixel: each image pixel is analysed independently of the others according to its spectral characteristic [Fjørtoft, 1996],