A knowledge-based image analysis system for Landsat TM data, called ELIA, is presented. Topics considered include: how to extract water systems, such as coastlines and river networks, from Landsat TM data and how to represent the extracted information in a form that is convenient for further processing. The extraction process consists of two subprocesses: primary and secondary extraction. During the first phase the system detects relatively large-scale water regions. Later, narrow rivers can be detected by further analysis of the results. The extracted information is described in a graph. Experimental results of the river-finding process are shown.