Using Mixed Samples And Support Vector Machines To Extract Urban Landcover Information

City is a combination of natural and artificial environment, as well as the center of all humanactivities, thus, it is of great significance to monitor urban environment. Remote sensingtechnology has the advantages of acquiring data of large area at one time, and expressing them inboth image and non-image way. What’s more, remote sensing technology can arrive at placeswhere few people can get and has little influence on the environment. Thus, it is used extensivelyin urban environment research. But, the complexity of urban areas and the difference between thesame class lead to the phenomenon of mixed pixels in mediate and low spatial resolution image,which limits the accuracy of using remote sensing data. Thus, many studies focus on improvingthis accuracy in order to extend the usage of mediate and low spatial resolution image. This studyfocuses on sub-pixel and mixed class in order to improve the accuracy of extracting information ofurban areas.This study selected Beijing City, which is one of the classical cities in China and its data isthe first image of Landsat8data on Beijing area, as study area, extracted Beijing City’s land covercomposition at a sub-pixel level through mixed pixel decomposition based on a new strategy. First,a new classification system is put forward to name those extracted classes. It is named W-V-I-Smodel (W for water body, V for vegetation, I for impervious surface, S for soil). It is expressedpicturesquely by a regular tetrahedron. The vertexes, edges and facets and inside space of theregular tetrahedron have different meanings. The support vector machines is used to extractBeijing land cover information under this model and the result is analyzed by high spatial resolution data, expert experience and field survey data. The main contents and conclusions are asfollows:1. A classification system with mixed classesThe classification system of W-V-I-S model is promoted under the theory of V-I-S, and it isbuilt based on a regular tetrahedron. The four vertexes of regular tetrahedron stand for water body,soil, vegetation and impervious surface, respectively. The edges of the regular tetrahedron standfor mixed classes of two elements of water body, soil, vegetation and impervious surface, thefacets of the regular tetrahedron stand for mixed classes of three elements, and inside space of theregular tetrahedron stands for mixed classes of those four elements. In different locations, theproportions of each element are different. This classification system can depict the class ofmediate and low spatial resolution image more accurately and roundly, and mixed class is firstlyintroduced into the classification system.2. The sample library based on the W-V-I-S modelFirst, the range of the study class should be extracted, which includes the class and itsneighboring region. Second, the mask is clustered by unsupervised method and some samples arerandomly selected from those clusters. Then the true class names of those samples are tested byhigh spatial resolution data, expert experience and field survey data. These class names are addedto the library and the library can be used in the extraction of other classes. However, the librarywill be improved by the way of human-computer interaction. The function of computer is toevaluate the similarity between samples. And human beings should recognize the new class’sname.3. Extraction of sub-pixel information Combining the sample library and SVM based on radial kernel function, the sub-pixelinformation of urban areas is acquired. The result shows that my method has better performancethan SVM and linear spectral unmixing trained by pure samples. That is to say the nonlinearmodel is more suitable to depict the urban areas. What’s more, in this study, impervious surfaceinformation is not extracted and according to the W-V-I-S model, if water body, vegetation andsoil are extracted, then the left class belongs to impervious surface. In addition, the city of Beijinghas both places of interest and modern buildings, which is a very complex area. In order to extractits impervious surface information accurately, more data, such as high spatial and spectral dataand other auxiliary data, may be acquired.