In mountainous area, many factors such as the high spatial heterogeneity in land cover conditions and the topographic effect on remote sensing data acquirement significantly constraint the change detection by using remote sensing method. It has become a great challenge to develop an effective method to accurately map the land cover changes for complex terrain areas. In the traditional change detection methods, the spectral features are the major information. Due to the lack of the consideration of other information from structure, topography and other parameters, many false changes will be introduced in the final detection results with a big uncertainty. To avoid the false change and overcome the influence from terrain shadow and phenology differences, an object-oriented change detection method by combining multi-scale segmentation and decision tree algorithm for complex terrain condition is proposed in this study. In the new method, firstly, a series of spectrum information (i.e. reflectance, change vector intensity, NDVI), shape information (i.e. area, length-width ratio, shape index) and terrain parameters (i.e. elevation, slope, aspect) are calculated for change detection rules construction. Secondly, the unchanged surfaces which are usually detected as the false changes due to the influence of terrain shadow and phenology difference are extracted as training sample. Finally, the change detection rules are built automatically by C5.0 decision tree algorithm. the proposed method is applied to Landsat-TM images acquired in 1989 and 2009 at Panxi region, the detection results indicate: (1) Over the mountainous area, the method can effectively reduce the influence from topographic effect, and has a big improvement on detection accuracy with the overall accuracy of 93.57% and Kappa coefcient of 0.8706 by validating ground reference samples and stratified random samples. (2) The C5.0 decision tree algorithm performs well in terms of robustness and adaptability, which serves well in land cover classification with types only labeled as "changed" and "unchanged" and highly heterogeneous training samples. Considering the false changes result from topographic and phenological as "unchanged" training sample, the accuracy of change detection is improved effectively. (3) Spectral features are the predominant information in analyzing TM images, and the NDVI differences between the reference image and detected image is an effective indicator for change detection of land cover pattern in vegetation coverage region. (4) The changes of land cover pattern in Panxi region from 1989 to 2009 were closely associated with human activities. The typical driving forces of pattern change include returning farmland to forest policy, water conservancy project, mining, urban sprawl and so on. The changed area amounted to 740.2 km², accounting for 2.49% of the total detected area.