Improved fast mean shift algorithm for remote sensing image segmentation

Image segmentation plays a crucial role in object-based remote sensing information extraction. This study improves the existing mean shift (MS) algorithm for segmenting high resolution remote sensing imagery by adopting two strategies. First, a pixel-based, fixed bandwidth and weighted MS algorithm is applied to cluster the image. In this process, the space bandwidth is selected according to the resolution of remote sensing images, and the range bandwidths of each band are calculated based on grey feature and the plug-in rule. Gaussian kernels are used for clustering. Second, a region-based MS algorithm is applied to globally merge modes which are obtained in the first step. The spatial and range bandwidths are adaptively adjusted based on the clustering result of the first step. Experimental results with two Quickbird images show that the improved algorithm is superior to the typical MS algorithm, producing high precision and requiring less operation time.