Classification Of Remote Sensing Imagery With Texture Information

With more and more launch and application of High Resolution Satellite, the texture information in Remote Sensing Imagery becomes much more abundant. High Resolution Remote Sensing Imagery is playing an important role in numerous areas such as surveying and mapping, urban planning, national defense military, land use and investigation, etc. At the same time, it provides additional opportunities for us to solve the problem of forest type identification which is teasing our mind.However, improvement of spatial resolution causes descent of classification accuracy instead of improvement. Depending only on the spatial information leads to much declassification. Traditional pixel-based classification method can not give a satisfied performance for High Resolution Imagery. Its accuracy of classification is far from productive demand.Plenty of experiments indicate that the problem of forest type identification can not be solved depending only on the spatial spectrum information. High Resolution Imagery contains lots of detail information of texture and structure. Almost each crown of the trees can be separated from the others. Different types of forest display different textures. We can make a way through texture information with spatial spectrum and structure information.Texture and spatial structure information accounts for a considerable proportion of High Resolution Imagery. Object-oriented Classification Method can make good use of them during classification process.Two methods are discussed in this paper, including Grey Level Co-occurrence Matrix and Wavelet Transform. They are used to calculate texture features of remote sensing imagery firstly. Then Object-oriented Classification Method was applied to classify the remote sensing imagery of the study area. The classification result shows that key information combined with proper methods leads to a satisfied result. Each forest type in the study area can be discriminated. The classification accuracy is 89%, which is much better than methods based on pixel and spectrum.