Abstract: The process of extracting urban area from remote sensing images is an important step for monitoring and measuring urban change.In the existing literatures,many statistical approaches were developed to solve this difficult task,but there are few successful reports because the reflectance of an object usually does not characterize its nature.In this paper,we propose a structural method for automatic detection of urban area from satellite panchromatic remote sensing images.Urban area is a kind of complex combinatorial object and each urban area has its particular combinatorial mode,therefore it is difficult to describe an urban area model in detail.Through analyzing urban area structures and observing many real satellite panchromatic city images,we firstly summarize four common properties for most of an urban areas as follows:(1) an image of an urban area contains variation in pixel intensities;(2) an urban area has a large size;(3) for most of local images of an urban area,there are two main peaks in the orientation histogram of short lines which are extracted from a local image,and the orientation lap between the two peaks is about 90 degrees;(4) there are also two peaks in the orientation histogram of long lines,and the two peaks account for a majority of the histogram.Secondly,we follow these properties to devise a hierarchical method for extracting urban area from satellite panchromatic remote sensing images.In the first level,some candidate regions of urban area are segmented from low-resolution image based on the two former statistical properties.In the second level,the candidate regions are verified by computing the two latter complex geometric features of urban area from high-resolution image.We divide three steps to detect the candidate regions of an urban area.An input image is firstly sampled to create a low-resolution image,and then the sampled image is processed using histogram equalization and quantization and is segmented by computing texture energy features and exploiting split-and-merge algorithm.Finally,some large areas with high texture energy are selected as the candidate regions.After acquired the candidate regions,we use the two latter features of urban area to make further analysis.Because the two properties are based on the local range of an urban area image,therefore the candidate regions must be firstly divided into many sub-images,and then the procedure of urban area confirmation is implemented on each sub-image.The verification procedure is implemented into three steps.All straight lines are firstly extracted and binarized into short and long line sets using a defined threshold,and then two feature parameters are computed:one is the orientation lap between the two highest peaks of the orientation histogram of short line set;the other is equal to the sum of the two highest peaks of the orientation histogram of long line set.Finally,when the first parameters is close to 90 degrees and the second one is larger than 0.5,the sub-image is labeled as CITY,otherwise as NONCITY.In a candidate region,if the ratio of sub-images with CITY label is larger than 0.7,then it is an urban area.In order to smooth the boundary of an urban area,we propose an edge point density(EPD)-based method:first,EPD of all pixels are estimated to get an EPD image;in the second step,a two-threshold scheme is implemented to segment the EPD image;finally,two mathematical morphological operators are used to smooth the boundary of the urban area.We have embedded the proposed algorithm into a system developed by our laboratory for automatic complex object recognition and have tested many real SPOT remote sensing images.Experimental results have shown the feasibility of our algorithms.