Research On Structural Features Extraction And Object-Based Classification Of High Resolution Remote Sensing Imagery

Compared with the low or intermediate resolution remote sensing images, high resolution ones could provide a large amount of detailed information about land cover. Thus such images open new avenues for classification and pattern recognition of remote sensing images. However, due to the increasing variance within the same class and the decreasing variance between different classes in spectral domain, the problem of land cover’s recognition and classification is increasingly difficult. In order to solve the problem above, this paper aims to investigate a wide variety of methods to improve the classification accuracy which includes extraction of structural features, multiscale segmentation, multifeature fusion and multiple kernel learning. The major works implemented are as follows:1. Basic knowledge about process of classification, classifiers and accuracy assessment methods are illustrated systematically.2. In order to reduce the effect of salt and pepper which is produced by traditional features extraction methods, two kinds of structual feature extraction methods Size-Compactness-Index (SCI) based on super-pixel and Enhanced Morphological Shadow Index (EMSI) based on morphology are proposed. The first step to get SCI feature is to generate super-pixel areas with threshold control criteria and non-maximum suppression criteria. Then SCI shape feature is constructed by the size and compactness of areas. And EMSI is made up by the Differential Morphological Profiles of CFO’s black hat transform. The contrast experiments are carried out using several kinds of feature extraction methods (GLCM, PSI, SCI, DMPCBR, DMPOBR, EMBI, EMSI). The experimental results show that:SCI get the highest classification accuracy followed by EMSI.3. With respect to multiscale segmentation technology, a Minimum Spanning Tree multiscale segmentation algorithm based on edge detection is proposed. Firstly, Canny is applied to extract edge information in panchromatic imagery, then such edge information is combined with spectral information to get the weight of edges based on graph theory, and minimum spanning tree algorithm (Kruskal) is used to complete the initial segmentation of color images. Finally, the objects’ features made up of spectral, shape and edge information are comprehensively analyzed to merge areas, and the segmentation results are generated. The experimental results show that the proposed algorithm is more effective and more efficient than eCognition 8.0. At the same time, the object-based GLCM and object-based PSI features which derived from the proposed segmentation algorithm achieve higher Overall Accuracy than the traditional GLCM and PSI features do in classification experiments.4. In order to make the spectral features collaborated with high dimensional spatial features, Vector Stacking (VS) is used to implement the fusion of multi scale features. The experimental results show that:These structural features extracted in this paper (Object-Based GLCM, Object-Based PSI, SCI, DMPs, EMBI, EMSI) are complementary to each other. Therefore, combining all of them or some of them would probably achieve better results than any single feature. Furthermore, Vector Stacking is a good method to fuse a wide variety of structural features which are expressed in different ways.5. Some disadvantages of single kernel methods are pointed out in this paper when dealing with all samples in complicated conditions. Therefore, the multiple kernel methods are brought into classification problems of high resolution images, and the experimental results show that the multiple kernel methods achieve much higher Overall Accuracy when compared with traditional classification methods.