It is a key problem to select optimal features from the total set where spectral, geometric, shape, texture features and some other features are extracted by the process of image segmentation in object-oriented classification. In this paper, the authors present a method for selecting good features from object-oriented image segmentation according to the maximal statistical mutual information dependency criterion so as to improve the classification accuracy of high spatial resolution image. The proposed method is a three-step classification routine that involves the integration of (1) image segmentation with eCoginition software, (2) feature selection by mutual information minimum redundancy and  maximum relevance criterion, and (3) support vector machine for classification. The experiment was conducted on QucikBird image in Zhangzhou city, Fujian province. Furthermore, the proposed method and the well-known feature selection methods such as Tabu search algorithm and fisher discriminate analysis are evaluated and compared with each other. The result shows that the mean error ratio decreases by 4% with the proposed method and that the proposed method for feature selection outperforms the other methods in terms of McNamara’s test.