Research Of High Resolution Remote Sensing Image Segmentation Based On Modified Watershed Transformation

Nowadays, more information and more complex target structures are provided by remote sensing images. Especially with the emergence of high resolution remote sensing images, such as SPOT5, QuickBird and IKONOS, the growing amount of observational data broadens our horizon for cognizing nature and brings great challenges on how to effectively deal with the data in the actual work. How to extract required information quickly and accurately from high resolution remote sensing image is an urgent problem to be solved. Remote sensing image segmentation is regarded as a solution, which can divides the remote sensing image with high resolution, large amount of information, rich details and complex noises into several sub regions, and then extracts the interested target areas for further analysis and processing, so as to realize the intelligent processing of remote sensing information. A case study of high resolution remote sensing image segmentation based on modified watershed transformation is conducted.The study focuses on the approaches of remote sensing image segment-ation based on watershed transformation. According to the characteristics of high resolution remote sensing image features, we develop a morphological nonlinear filter which can better consider the geometric characteristics of features and is suitable for high resolution remote sensing image filtering process; Based on h-minima transformation technology, an image segment-ation algorithm combined with morphological filtering and marked watershed transformation is proposed to improve the segmentation precision and restrain the over-segmentation problem; Based on the multispectral characteristic, we develop a watershed transformation image segmentation algorithm combined with color gradient characteristics which has application values for later classification and feature extraction. The main research work includes:Firstly, we analyze the image segmentation difficulties on high resolution remote sensing images according to its characteristics. High resolution remote sensing image owns advantages of high spatial resolution, clear geometric texture, wide applications and abundant three-dimensional information, which has been successfully applied in many fields. The emergence of high resolution remote sensing data makes it possible to extract more detail information and it also brings new challenges for automatic extraction and identification information work. The difficulties of high resolution remote sensing image segmentation includes:significantly increased data, high details of information and scale dependence during the geographical spatial process.Secondly, a remote sensing image segmentation algorithm combined with morphological filtering and marked watershed transformation is proposed. Due to the image noises which seriously affect the gradient image quality, a mixed morphological opening&closing reconstruction filter is designed instead of Gaussian filter to smooth the remote sensing image according to image and noise characteristics. Mark extraction is the key to success while using the marked watershed transformation. So the h-minima transformation technology is used to limit the number of minimal areas. Experimental results show that the proposed algorithm efficiently overcomes the over-segmentation problem.Finally, a watershed transformation image segmentation algorithm comb-ined with color gradient characteristics is proposed. Considering the lumi-nance component of the gradient is incomplete, we calculate the color vector gradient combined with color information of the multi-spectral image. Extending gradient concept into vector space and implementing the gradient operator in RGB vector space directly on the multi-spectral remote sensing image. Using closing reconstruction to do the gradient correction and obtaining the marked gradient image with adaptive h-minima transformation, then implementing the watershed transformation. Experimental results show that the proposed algorithm can obtain edges which can meet the requirements of later classification and feature extraction.