Urban land cover information extraction combining spectral feature and oriented-object analysis

It is well known that land cover change has important implication on urban planning, urban environment evaluation and so on. However, urban land cover information is not easy to be extracted because of its inner complexity resulting from human impact. Under this circumstance, the accurate recognition and classification of different land use types on urban area from remote sensing data has become an important task. This paper introduces an effective method, which relies on the typical ground objects´ spectral responding features and interacting mechanisms, to undertake classification hierarchically. It can, effectively, reduce spectral variation in the scene origin from complex pattern of urban fields. During the process, the shadow areas caused by buildings in the research area were extracted, and a 60-m buffer zone can be served to distinguish low density urban from high density urban. It is indicated that the density of high-building in the northern part of Beijing City is higher than the counterpart in the southern part. Band math was also used to enhance the reflectance difference between water bodies and other objects, and (Band1+Band 2) / (Band 3+Band4) has been tested as an effectively band calculation in the result (ratio value) of which is above 1 only in the areas of water. As a consequence, the water bodies have been extracted successfully using an appropriate threshold value (1.4-2.3) and mask technique. Further more, SAVI based on the spatial analysis can basically be beneficial to discriminate farmland from woodland. In order to improve the classification result and extract the rest features on urban fields including roads, object-oriented method has been applied integrating spectral, texture and structure information (mean, variance, homogeneity, contrast, entropy and dissimilarity). An evaluation of the method has been performed, and the experimental results, compared with those given by conversational classification methods including purely supervi- sed classification, show the effectiveness and great potential advantages of this approach.