Abstract: Based on Chinese vegetation classification criteria described in the book “Chinese Vegetation”, the vegetation classification system suitable for the farming-pastoral ectones in Northwestern China was established with the field investigation. Using the Landsat images, the terrain data and field data of the study area, the vegetation information in the study area was refined according to the implementation strategy as "hierarchical classification and layer by layer verification". During the extraction process, the mixed pixels decomposition of preprocessed remotely sensed images was performed by using the fully constrained least-squares model, and the vegetation fraction of the study area was obtained. We classified the study area into vegetation area where the vegetation fraction is larger than 5% and non-vegetation area where the vegetation fraction is less than 5%. In the vegetation area, it was further classified into 7 main vegetation type groups which include the tree group, shrub group and grassland group using the CART decision tree based on the spectral characteristic, texture characteristic (Mean) and terrain characteristic (Digital Elevation Model, DEM). Each vegetation type group was again further classified into different sub types based on the refined indexes. The tree group and shrub group were categorized into evergreen vegetation type and deciduous vegetation type based on the NDVI difference ratio index which was established using the seasonal variations of their NDVIs of different plants. The grassland type group was categorized into desert grassland, the typical grassland and meadow grassland using the temperature vegetation dryness index (TVDI). After this step, the spatial distribution of each vegetation type was obtained. It was proved that the overall accuracy of the final classification can reach 79.51% and the kappa coefficient is 0.773. The classification method used in this study makes full use of the spectral information and texture information of the remotely sensed images, and cooperates with terrain information. Experimental result shows that the method of using hierarchical classification and multiple indexes could extract the vegetation information efficiently from the images of the farming-pastoral ectones with high accuracy. The classification result in this area provides the basic data for the further research on the relationship between surface hydrothermal process and land cover change, especially vegetation cover change. Meanwhile, it provides reference for the conservation of vegetation area and ecological environment construction in this area.

Key words: the farming-pastoral ectone, vegetation information extraction, LSMA, CART decision tree, NDVI difference ratio (NDVI_DR), temperature vegetation dryness index (TVDI）