As widely used today, high resolution image becomes a useful data source for forest inventory because it can show detailed information of land-cover types which is so helpful in interpreting process. And in the application of high-resolution images, the toughest problem is to find the effective characteristics group to separate each class accurately. In this paper, we tried an object-based method to get the whole forest distribution of the study area. Combining segmentation and decision tree feature selection tool, we tried to find a convenient and effective way to select useful information from such many characteristics brought by ―super-pixels‖ after segmentation. Compared with the traditional pixel-based classification method, we found that object-based method was more appropriate not only for its nearly 10% higher classification accuracy but also providing with more detailed information lying in the image data that help.