基于决策树方法的蒙古高原土地覆盖遥感分类——以蒙古国中央省为例

Land Cover Classification in Mongolian Plateau Based on Decision Tree Method:A Case Study in Tov Province,Mongolia

蒙古高原包括蒙古全部、俄罗斯南部和中国北部部分地区。蒙古高原的土地利用/覆盖格局与变化,对揭示该区域乃至整个东北亚地区的资源、环境和生态特征,促进该区域可持续发展具有重要的现实和科学意义。本文以在蒙古国中央省及其所含首都乌兰巴托市为研究区,利用空间分辨率为30 m的TM影像,采取QUEST(Quick Unbiased and Efficient Statistical Tree)决策树方法,通过图像目视解译,获取了研究区2010年土地覆盖分类数据。结果显示,草地占据研究区总面积的70.88%,其次是森林占14.83%、裸地占10.73%、农田占2.98%、水体占0.31%、建筑用地占0.27%、湿地占0.02%。通过野外实地采集的139个GPS验证点进行精度评价发现,一级土地覆盖类型的总体精度可达72.66%。针对草地的二级分类的总体精度有较明显下降,其主要是由于中蒙科学家对于草地类型分类体系的差异所造成的典型草地和荒漠草地的混分。

Global and regional land use/cover data processed by remote sensing images is a supportive part for the researches related to environment and so on. Mongolia Plateau includes all of Mongolia, parts of southern Russia and northern China. Study on land use/cover changes in Mongolia Plateau has important practical and sci- entific significance to discover the resources, environmental and ecological characteristics of this area and even in northeast Asia area, and also it is beneficial to enhance international cooperation in regional resources utiliza- tion, ecological and environmental protection and sustainable development. According to the land use/cover map- ping requirement in Mongolia Plateau, a land cover classification approach was designed based on QUEST （Quick Unbiased and Efficient Statistical Tree） decision tree method in the representative area of Mongolia Pla- teau, where includes Toy Province and Ulaanbaatar City in Mongolia. Land cover classification dataset in this ar- ea was captured using Landsat TM images, through QUEST automatic classification and visual interpretation. Results show that the total area of steppe occupies 70.88%, followed by forest accounting for 14.83%, barren 10.73%, cropland 2.98%, water 0.31%, built area 0.27%, and wetland 0.02%. 139 GPS verification points for ac- curacy assessment were collected by fieldwork, which was held in August 2013. Accuracy assessment found that land cover overall accuracy is up to 72.66% in class I, while the overall accuracy in class II is obviously descent, mainly due to the confusion of real steppe and desert steppe. This was caused by the different understanding of these two steppe types between scientists from China and Mongolia. In general, the QUEST decision tree meth- od used in this paper is proved to be feasible for higher accuracy land cover classification mapping in this area.