Urban Building Extraction Based On Multiple Features From High Resolution Remote Sensing Images

In recent years, China’s economy develops rapidly. In order to conduct the urbanplanning, we need to fully master the city building information. The remote sensingimagery has developed rapidly in recent years. QuickBird, Worldview and etc RemoteSensing Satellites can provide a large amount of high resolution images, which are moresuitable for building extraction from urban areas. But the high resolution remote sensingimages contain more noises than low resolution images. The noise will produce a badimpact of results. So it is not suitable to use the traditional pixel-based image processingmethods to analysis high resolution images. Meanwhile a single feature is not enough.We should select multiple features to analysis the image. In this paper, buildings areextracted from high resolution remote sensing images using multiple features andobjected-oriented method.The thesis first focuses on the research of image segmentation, including themethods based on Normalized Cut and improved watershed algorithm. Normalized Cutand improved watershed algorithm are combined to develop a novel method for highresolution remote sensing image segmentation. There is a serious over-segment problemin watershed method. The small segmented regions we get through traditional watershedalgorithm have no practical significance. The Normalized Cut algorithm is a NPproblem, and the more pixels in the image, the larger the generated graph of nodes is.So using Normalized Cut algorithm to segment high resolution images directly will bevery complicated and consume plenty of time. In the combined method, we use thesegmented regions, instead of pixels to form the weight matrix W of Normalized cut.This method can exploit the advantages of the Normalized cut and watershed algorithmboth. The cost has greatly reduced and can get a relatively accurate segment result.Then feature extraction and selection methods are studied.36features are extractedincluding shape, texture and spectrum information. ReliefF algorithm is used to choosethe most useful feature set. ReliefF method selects features by computing the weight ofeach feature. We improved it by removing the feature which is highly relevant to otherfeatures and at last we get the optimum feature set.Finally we apply object oriented classification technology and Morphologicalshape optimization method to get the building areas and outline. The object-orientedimage processing technique shows significant advantages than traditional methods based on pixels. It can effectively avoid the paper effect. And after morphologicalfiltering and removing small areas which are small than the minimum building areaallowed. The building areas and outline are extracted completelyand accurately.