This　paper　presents　a　method　to　select　optimal　feature　subset　from　object-orientated　image　segmentation　according　to　the　maximal　mutual　information　to　improve　classification　accuracy　of　high　spatial　resolution　imagery　over　urban　area.　The　proposed　method　is　a　three-step　classification　routine　that　involves　the　integration　of　1)　image　segmentation　with　eCoginition　software,　2)　feature　selection　by　maximal　mutual　information　criterion,　and　3)　support　vector　machine　for　classification.　Experiment　is　conducted　on　Quick-Bird　image　in　Fuzhou　city.　Furthermore,　the　proposed　method　with　the　well　known　feature　selection　methods,　namely　Tabu　greedy　search　algorithm　and　fisher　discriminate　analysis,　are　evaluated　and　compared.　The　experiment　shows　that　the　mean　error　ratio　significantly　decreases　with　feature　selection.　It　also　demonstrates　that　the　proposed　maximal　mutual　information　feature　selection　with　support　vector　machine　classifier　significantly　outperforms　the　classification　method　accompanied　with　eCoginition　platform　in　terms　of　Z　test.