Research On Classification Recognition And Extraction Of Buildings By RS Image Assisted By LiDAR Data

In recent years, with the rapid expansion of cities, the number of buildings increases dramatically. How to automatically and rapidly extract buildings from high resolution RS images has significant academic and practical significance in city planning, the updating of GIS databases and digital cites and so on. The development of LiDAR provides a new solution to this problem. In this paper, we explored the classification recognition and extraction of buildings by using Support Vector Machine (hereafter SVM) classification method, along with LiDAR data and high-resolution RS image. We use LiDAR data because of its altitude feature and high-resolution RS image for its spectral and texture.SVM is a new machine learning approach based on Statistical Learning Theory, when it is applied in remote sensing image classification, it has significant advantages in small sample learning,learning efficiency and getting global optimum. Nowadays, SVM is being extensively researched into high resolution RS image classification. On the basis of the previous research, unlike other building extraction methods, this paper adopted a new method, that is, SVM classification technology which uses the fusion of LiDAR altitude data and high-resolution RS image as its experimental date, to distinguish buildings from other objects. Then by using digital graphic processing technique, we processed the results of classification, finally, we extracted buildings. The layout of the thesis is as follows:1. We offered a primary study about the pretreatment of original LiDAR data, and explored on the generation of DSM and DEM based on TerraScan classification, furthermore, we got the nDSM which is out of terrain elevation information, meanwhile, the nDSM is used as aided data in building extraction.2. To begin with, We elaborated four typical Supervised Classification Theory, especially SVM classification theory, and then made experiments on high resolution image, we compared the classification accuracy about this four supervised classification methods through experiments,at the same time, a comparison of building classification extraction accuracy is made. At last, we made SVM high resolution image classification experiment based on different kernel functions, and then we compared the classification accuracy of these four different kernel functions and the building extraction accuracy.3. We presented classification flow chart about the building classification and extraction by SVM and focused on how to ascertain kernel function and kernel parameters, and made building classification recognition experiments by using obtained kernel function and kernel parameters, at last, we made some analyses on the result of building classification.4.We discussed the final processing technique about post classification of image. The main methods are small size removal method to wipe off error categories and mathematical morphology opening-and-closing operation to process building edge, and edge detection operator to extract building edge, finally, we achieve automatic buildings extraction from high resolution image.