• 责任者：Pei Huan (College of Information Sciences and Engineering ; Yanshan University ; Hebei 066004 ; China) ; Fang Shifeng
• 刊名：Scientia Geographica Sinica
• 出版年：2013
• 卷期：33(11)
• 关键词：高光谱遥感 ; hyperspectral remote sensing ; 绿洲 ; oasis ; 土地分类 ; land classification ; 新疆 ; Xinjiang
• 页码：p.1395-1399
• 图表：1 illus., 2 tables, 16 refs.
• 分类号：1600
• issnNo：ISSN1000-0690
• isbnNo：CN22-1124/P

     Due to the presence of same object with different spectra and different objects with same spectrum, the accuracy of remote sensing classification is limited, which is especially obvious in arid region. Turpan Ba- sin, which is the typical oasis-desert interlaced area in eastern Xinjiang, is selected to be the typical research ar- ea in this article. The data set in this study includes SPOT images with a resolution of 10 m, DEM and field da- ta. At first, the classification system was built according to land use/cover characteristics of research area. Then common methods such as principal component transformation, tasseled cap transformation and minimum noise fraction transformation were used to extract spectral feature. Vegetation index and wetness index were al- so calculated based on SPOT image, and 12 spectral features were built using these methods. Finally a classifi- cation framework which contains 14 features was implemented by combining DEM and slope. Inter-class sepa- rability method was applied to choose the optimum feature combination. Based on spectral and spatial charac- teristics of different mixed ground objects, and combined with supervised classification results, a decision tree model was built to abstract the land use/cover information. The results showed that decision tree classification based on multi-parameters of land surface could make full use of terrain spectral information and spatial infor- mation, and distinguished confusion features effectively, such as different types of desertification land and ur- ban construction land. Accuracy of the method is 88% and the kappa coefficient is 0.76. Its accuracy has been improved significantly compared with traditional classification methods, which have increased by 7%, 10% and 11% than maximum likelihood method, markov distance method and the minimum distance method. The whole process may provide reference for land use/cover real-time monitoring in oasis of arid areas, and it also has certain significance to desertification study. The novel of this article is that the classification system was built according to land use/cover characteristics of oasis. That is different type and different level desertifica- tion land was contained in classification system. So wind erosion desertification and salinization information were classified at the same time with other land types. It has certain significance for desertification prevention and control. Due to time restrictions, we only used traditional classification methods to classify spot raw data, and compared its accuracy with the results of multi-parameters decision classification. In the future studies, tex- ture features and other features may be added, and different feature selection approaches can be adopted to compare the influence of different features on classification accuracy.