A　method　for　urban　major　road　extraction　from　IKONOS　imagery　was　proposed.　The　texture　features　of　the　image　were　first　analyzed　in　three　different　levels.　The　first　level　calculated　the　Mahalanobis　distance　between　test　pixels　and　training　pixels.　The　second　level　was　the　calculation　results　of　Bhattacharyya　distance　between　the　distributions　of　the　pixels　in　the　training　area　and　the　pixels　within　a　3x3　window　in　the　test　area.　The　third　level　employed　cooccurrence　matrices　over　the　texture　cube　built　around　one　pixel,　and　then　Bhattacharyya　distance　was　used　again.　The　processed　results　were　thresholded　and　thinned,　respectively.　With　the　assistance　of　the　geometrical　characteristic　of　roads,　the　three　resultant　images　corresponding　to　three　levels　were　computed　using　fuzzy　mathematics　for　their　likelihood　belonging　to　road　and　then　merged　together.　A　knowledge-based　algorithm　was　used　to　link　the　segmented　roads.　The　result　was　finally　optimized　by　polynomial　fitting.　The　experiment　shows　that　the　proposed　method　can　effectively　extract　the　urban　major　roads　from　the　high-resolution　imagery　such　as　IKONOS.