Remotely sensed data based land use/cover classification, especially in mountainous areas, is a difficult problem that has long drawn attentions among researchers. This paper presents a synthetic approach using C4.5 algorithm to automatically derive classification knowledge with the purpose of constructing a model of decision tree for the final classification of the image. Statistical relationships of the land-use pattern with DEM were analyzed through spatial analysis function of GIS to provide extra knowledge for the post classification processes, which improves the precision of final classification by enhancing the characteristics of the trial zones in the image. According to a classification experiment on the rugged terrain over the upstream of Hanjiang River Basin where the land use/cover ground survey data are available, the proposed approach is far superior to the traditional maximum likelihood classification method.