A KERNEL BASED CLUSTERING OF HIGH RESOLUTION SATELLITE STEREO IMAGES IN URBAN AREA

In this paper, we use Kernel Graph Cut (KGC) algorithm for multiregional clustering of HRSI in urban area. Use of kernel function in clustering algorithm transfers input feature space to high dimension and better results can be produced. Since the input feature space has significant effect in clustering output, to achieve better results especially in elevated objects such as buildings and trees (that are the most important objects in urban area) we use Normalized Depth Map (NDM). The NDM is generated by applying Semi-Global Matching (SGM) to the stereo image and removing terrain effect from depth map. Therefore, slope effect is removed from clustering results. Also for distinguishing between trees and grass lands we used Visible Vegetation Index (VVI) as another feature layer. The VVI is generated using intensity values of Blue, Green and Red channels of visible image by sampling from vegetation area. To show the ability of kernel graph cut for HRSI clustering we compare it with typical k-mean clustering algorithm. Evaluation results showed that kernel graph cut can produce more accurate results than k-means clustering algorithm. Also Results showed that the accuracy of clustering had significantly changed by using specified feature layers.