Abstract: Use the QuickBird remotely sensed imagery in 2005 of Futian District, Shenzhen City as the main data source, the spectrum, shape and texture features of different land-use were extracted by the object-oriented multiresolution segmentation. Then we used stepwise discriminant analysis(SDA)with classification and regression trees(CART)to build “multi-segmentation, multi-features” classification model. The results showed that:(1)With SDA, it could choose an advanced feature set. In this study, among the 32 features we totally chose 27 important features and sorted them by the importance in tables. Among them, the spectrum and texture features were mostly sorted forward, while the shape features were sorted later except the LengthWidth was ranked fifth.(2)The CART model could recognize more important features, and calculated separation thresholds for object-oriented classification automatically. In this study, Mean_b3, LengthWidth, Ratio_b3, GLDVE and NDVI features were the important classified nodes.(3)Under the condition of improving or no significantly reducing classification accuracy, SDA combined with CART model achieved the feature dimension reduction. When we chose R²>0.2, the CART₁₆ model obtained an optimal result of training precision 94.44% and test precision 83.37%, while the feature set was the smallest, compared with primary one, the number of features decreased to half.