Since forest is an important indicator of global climate change, the way to extract forest changing should be top priority in forest management and utilization. Especially, the extraction of sub-categories of forest vegetation has always been a difficult point in remote sensing image classification. Therefore, it is important to find a suitable method for forest type mapping, especially in regions with diverse climatic conditions and complex terrain. The present study discussed various methods that could be used to improve the accuracy of forest type classification using Landsat Thematic Mapper (TM) imagery data, taking a semiarid mountainous area in Beijing, China as an example. All classification results were compared with confusion matrices and Kappa statistics. The results showed that: 1) The combination of a digital elevation model (DEM), aspect data, TM4 and TM5, and a synthetic band (TM4-TM2) comprised an optimal dataset when using pixel-based classification. 2) Elevation alone could increase the accuracy by 23% in broad-leaved forest, whereas by 4%-5% in coniferous and mixed forest. Meanwhile, aspect alone could increase the accuracy by 21% in broad-leaved forest, whereas by 13% in coniferous forest and 18% in mixed forest, respectively. Aspect can provide more valuable information for forest mapping than elevation. 3) According to the confusion matrices, the accuracy of pixel-based classifications was slightly higher than that of object-based classification. 4) However, the latter seemed to consist with field investigations better. Our findings implied that integrating distributional characteristics of forests in semiarid regions with Landsat TM imagery could improve the accuracy of forest stand mapping at a regional scale.