Hiperspektral Görüntülerde Lidar Destekli Spektral Bölütleme
Özet
In this thesis, hyperspectral images and LiDAR (Light Detection and Ranging) data has been fused by using spectral clustering methods in order to obtain unsupervised hyperspectral image segmentation. Hyperspectral Images, unlike ordinary RGB images, contain hundreds of spectral bands. Because of these high dimensions, it becomes a huge problem to process hyperspectral data. For this reason, dimension reduction and segmentation in image processing have an important position. In this work, spectral clustering methods that do not require a prior information are used for segmentation of hyperspectral images.
Hyperspectral images have many spectral bands as well as high spatial resolution per pixel. Although spatial and spectral information contribute to segmentation, it is difficult to distinguish objects with similar spectral characteristics in the same scene using only this information. If these objects which comprise similar spectral information have different altitudes, it is possible to distinguish them using elevation information. However, hyperspectral images do not contain any elevation information. Therefore, hyperspectral images and other sources of elevation information can be combined to provide a more detailed interpretation of the objects in a given scene. In this thesis, elevation information obtained from LiDAR data and spatial-spectral information are fused to provide hyperspectral image segmentation.
Spectral segmentation has become a popular method in recent years because it does not need any priori information about the image, it is easily solved by standard linear algebra methods and gives better results than traditional methods. Normalized Cut and Schroedinger Eigenmaps, which are spectral clustering methods, have been used in order to segment hyperspectral images. The affinity matrix used by these methods has been constructed using LiDAR and hyperspectral data. In particular, significant segmentation results have been obtained by using the affinity matrix generated by the Pointwise Mutual Information (PMI) method. In addition, the segmentation results generated by different solutions of these spectral segmentation methods have been examined and compared. Besides the spatial and spectral information, the contribution of the elevation information obtained from LiDAR to the segmentation is examined and the segmentation results of the proposed fused methods and the segmentation results of the available methods are compared.
