Research On The High Spatial Resolution Remote Sensing Image Segmentation Based On The Nonblurring Mean Shift

A large number of remote sensing data has been accumulated with the rapid development of several technologies, such as the aerospace technology, the automatic data acquisition and transmission technology, the data compression technology and the database technology, et al. However, the processing of these data is far from satisfying the needs of practical applications. The application of remote sensing data lags far behind the development of space technology. The appearance of high spatial resolution remote sensing images has brought new opportunities for the development of remote sensing technology and increased the depth of traditional remote sensing applications. The development of intelligent remote sensing image data analysis and understanding techniques, the simulation of remote sensing image interpretation of the human brain's cognitive processes by computer, getting the semantic information from high-resolution images based on the purpose of practical applications are important tasks of image interpretation. And all of these techniques are built on region-based image segmentation algorithms.During image interpretation process from the pixel space through the feature space to semantic space, the region segmentation plays an important part; there are many issues worth studying. As a non-parametric probability density estimation algorithm, the nonblurring mean shift algorithm which can be used for feature space analysis of a variety of occasions has a good theoretical basis and is suitable for parallel computing. In my work, the nonBlurring mean shift algorithm is used as a mathematic tool for the research of high spatial resolution image segmentation. The work can be summarized as the following:1. The nonBlurring mean shift algorithm is a statistical iterative algorithm; the convergence of the algorithm is the theoretical basis for its application. However, there are some mistakes and imperfections in the proof procedure of its convergence in some classical literatures. The research firstly emphasized the relation between kernel function and profile function, kernel function and the shadow of kernel function. Secondly, the convergence of the probability density estimation sequence is proved strictly according to properties of the profile function and the shadow of the kernel function. Finally, the convergence of nonBlurring mean shift procedures with different forms, such as the bandwidth matrix, the variable bandwidth, the variable weight, is discussed.2. A pixel-level and a region-level segmentation algorithm based on spectral and texture information fusion are presented for high spatial resolution remote sensing images. Gabor filter banks are used to extract texture features and a weighted minimum distance classifier is designed on feature variances in both methods. In order to obtain pixel-level spectral features, the paper presents an adaptive bandwidth mean-shift filtering algorithm using pixel numbers as input and using output as selected spectral features, which determines bandwidth based on Gauss assumption. The spectral part of the region mode is used as the region-level spectral feature. The experiments show that:the proposed adaptive mean-shift filtering algorithm is effective; the weighted fusion algorithm has higher segmentation accuracy than methods only using spectral features; the region-level method is superior to the pixel level method in both the objective and subjective evaluation3. Considering the bandwidths of the traditional mean-shift algorithm is not easy to control and the lack of stability of the segmentation results, a wavelet domain multiscale mean-shift segmentation algorithm is presented. The algorithm integrates the segmentation result in different wavelet resolutions by wavelet transform and the region boundary in the finest scale is optimized by morphological operation. Airborne image and synthetic image are used for validating the algorithm. Experiments show that:the segmentation algorithm is superior to the compared four algorithms, has a simple parameter setting and low time complexity.4.The landscape has a multi-level hierarchical structure, land objects only can be observed completely in a appropriate scale range. The methodology that constructing a multilevel hierarchical region structure, which is associated the multi-level hierarchical structure of the real word is studied, and a two-step framework for multi-level segmentation is built. Under this framework, an oil tank extraction and oil depots location algorithm is presented firstly and validated by experiments. Tank extraction in high resolution images can be treated as an example of the segmentation problems from the same semantic level to a different image spatial scale. The large-scale region partition of images will benefit further analysis. Since the large-scale high resolution remote sensing images have demonstrated a strong texture property, a multi-level segmentation algorithm based on region texture modeling and stepwise merging is presented. The method models the region with the joint histogram of local binary pattern and the labels of K means algorithm, and G-statistic is used to measure the difference of two histograms. Compared with MRF based segmentation methods using synthetic images, the method's accuracy has been validated. Finally, a multi-level segmentation procedure for adaptive feature extraction algorithm is presented and used for urban land cover classification. The method associates meaningful segments with land objects and features are extracted from the different size scale of the same land object. Two hyperspectral data sets are used for validating the method. Compared with other methods using the same training set, testing set and classifier, our results have show superior results in both visual evaluation and classification accuracy.