Generally, land uses are the ways in which lands are used by human activities. Especially in cities with dense population, land use is the direct results of various living requirements by the people. Therefore, urban land use is a very important indicator that can be used to measure the development of urbanization and their impacts on ecosystems. Hence, accurate extraction of urban land uses is critical for urban planning, land management and environmental protection. It has been proved that remote sensing technique is most efficient to extract land covers with high accuracy since different land covers have the specific physical properties which can be discriminated easily by the spectral bands of remote sensing data. However, considering the difference in land use and land cover, it is still a big challenge to extract urban land uses directly from remote sensing data with fast and satisfied accuracy because every land use is usually composed of different land covers. This study developed a new approach on urban land use classification by introducing landscape ecology theory into the application of traditional remote sensing techniques. According to the landscape ecology concepts, the landscape patterns within the same land use types are similar. In landscape ecology, many landscape metrics have been developed to measure landscape patterns on different spatial levels, such as patch level, class level and landscape level. On each spatial level, the landscape patterns have their own characteristics. Thus, some significant landscape metrics might be found to characterize the specific landscape patterns of different land uses, which can be identified as the important variables to discriminate different land uses. Then, by integrating these variables and traditional remote sensing approach, urban land uses were classified. The study area is the city center of Beijing, where dense population is living and amount of residential land use developed intensively. Considering the characteristics of landscape pattern of the study site, residential and non-residential land uses are proposed to be extracted by using the medium-resolution Landsat data. Firstly, the land covers mainly including built ups and non-built ups (vegetation) were extracted directly from the remote sensed data by the traditional approach based on the spectral response features. Then, the land parcels that are the basic land use units extracted mainly by the road segments as the boundaries. Totally, by the significant test, there are 27 significant landscape metrics identified to discriminate residential and non-residential land uses. On the class level, there are 12 significant landscape metrics measuring the distribution patterns of built ups within different land uses, while on the landscape level there are 15 significant landscape metrics measuring the total landscape patterns. Based on these significant landscape metrics, the Fisher linear discrimination model was developed to discriminate different land uses. By applying this model, the residential land non-residential land uses within the 5^th ring of Beijing were finally classified. The total classification accuracy was 79.7%, and the statistical Kappa coefficient was 59.8%. This study developed a new approach for urban land use classification, which enriches the application techniques of remote sensing. The approach on land use classification could be potentially applied to other urban area.