Object-Oriented Land Cover Image Classification By Using Support Vector Machines

Object-oriented remote sensing image analysis is one of the technology development of remote sensing images. Object-oriented remote sensing image analysis method, also known as cell-based segmentation method of image analysis, while the object is picked up through a certain image segmentation algorithm for remote sensing image, to get relatively homogeneity within the object, or even higher uniformity degree of image fragments or regional units.In the application of land cover classification domain, this unit is the land use patch.Object- oriented analysis method of main characteristic is its recognition and classification target is object fragments or regions, rather than a single pixel. We make this identification and classification of process which based on segmentation method of unit level is called the object - oriented remote sensing image classification techniques.The superiority of object - oriented classification technology is not only to give full con-sideration to the spectral characteristics of the surface features, and take advantage of the spatial information between surface features. Therefore, this technology has been e?ective promotion in many application fields. However, in object-oriented land cover remote sensing image classi-fication techniques, this methodology based on kernel function method of learning theory need to be in-depth study. In particular, the introduction of support vector machine technology, the application of the technique is lack of comprehensive and in-depth basic research support.The research is focused on as well-known urgently need to be solved for the theoretical problems and key technologies in the new land cover object-oriented technology, which based on kernel learning theory, support vector machine applied research in remote sensing image, according to the related principles of digital image processing, statistics, computing image en-gineering, remote sensing, image interpretation, researching on object - oriented classification based on support vector machine theory, exploring using support vector machine theory on the issue of object - oriented application potential, ability and prospects. Specific research content and its related innovation is carried out as the following three aspects(research contents and some innovations ):1. Starting from the generalized minimum unit which is as pixels of remote sensing image classification perspective, summed up kernel density function in image classification research results, analyzed some notable feature of kernel density function classification, put forward a rapid"The kernel density estimation of Bayes classification"of land cover classification model and established a classification system. As take the preliminary work of object-oriented analysis for the further research.First of all, a reasonable choice for the kernel function is an important part to support vector machine approach, and this step is also an important for scale parameter selection in the nuclear density gradient image segmentation method. Kernel function can easily make support vector machine to achieve non-linear algorithms. By classical Gaussian kernel function improvement, an initial understanding is acquired on the bandwidth parameters of the nuclear density function, as well as its characteristics and physical meaning. The results show that: The rapid transformation of nuclear function is applied to remote sensing images of land cover classification to achieve a quick calculation purposes, Classification accuracy under the same conditions with the standard support vector machines are compared, The experimental results also verified the improved kernel function to track the e?ectiveness of the sample.This section studies as basic research pave the way as next step by using Support Vec-tor Machine technology and the kernel density gradient image segmentation theory of object-oriented classification2. Based on the theory of kernel density gradient of remote sensing image segmentation,on object level, we step by step carried out classification theory and technology research work and system establishment work based on Support Vector Machines. We carried out a series of object-level basic research. This work is classified in a single-scale object-oriented classifica- tion (SOBIA-SVM).(1) We provided an object-oriented remote sensing image classification method based on the standard support vector machine theory, and the development of the corresponding sub-program (S-SVM). Proposed methods of cell characteristics of S-SVM which is based on seg-mentation, its full use of excellent identification feature of support vector machine in small sample feature space. Experimental data analysis results show that the method can be applied to the object-oriented high-resolution remote sensing image recognition category. And with the traditional the nearest neighbor KNN object-oriented technology compared, S-SVM classifica- tion results is superior to the traditional object-oriented methods.(2) Proposed object-oriented classification methods based on fuzzy support vector ma-chine theory, and the development of the corresponding sub-program (F-SVM). Proposed by using fuzzy support vector machine pattern classification method of object-oriented (F-SVM), firstly the object is given a weight using evaluation of fuzzy mathematics, and singular vector of the feature points in the weakening of the weight re-evaluate. The impact is under mini-mum level to the separating hyperplane. Experimental results show that the accuracy of the proposed method compared to standard support vector machine method improves the 5.1% or so to illustrate the proposed method improved the classification accuracy.(3) New object-oriented classification methods is proposed based on least squares support vector theory, and the development of the two sub-programs S-LSSVM and FG-LSSVM.First, direct selection feature objects, using a standard least squares support vector ma-chine model of the work of experimental data analysis results show that the use of S-LSSVM method, image recognition rate of 92.8%, compared to the same conditions of nearest neighbor KNN method is 95.2 %, SVM is 95.2 % increase by about 2.4 % or so. The results show that: A S-LSSVM object-oriented approach is not only a relatively faster computing speed, the detec-tion accuracy is also achieved under the same conditions in other object-oriented classification model.Then, after this S-LSSVM subsystem, put forward a combinational approach, namely FG-LSSVM, hybridizing least squares support vector machines (LSSVM) with fuzzy and grey degree of correlation (FG), which presents a feasible high-precision image classification algo-rithm for land cover. To compare the performance with other object-oriented methods, with original samples, three models were successively verified, which were standard support vector machines (SVM) and the fuzzy nearness improved support vector machines (FSVM), and the traditional K nearest neighbor (KNN) object-oriented methods. A high precision land cover image classification system was established with the proposed approach. The results show the total precision of FG-LSSVM is about 2.4% higher than that of SVM and FSVM, and KNN object-oriented methods in the study area. Compared with S-LSSVM the FG-LSSVM increased by about 4.8 % or so. The proposed method also meets the requirements of land cover image classification in respect of e?ciency and e?ects.3. Through introduction of computer vision recognition technology, level recognition idea is used in object-oriented system. By changing the scale, the large object layer is divided into separated the lowest level, and at this level of recognition model is to be built by using support vector machines to extract the large-scale land cover types. The rest of the small-scale features is established in a single-level, and in this layer another new support vector machine model is created to extract the type of small-scale surface features. Precise mask by computer visual technology, the extracted features in superposition of all levels get the final classification results.Based on the above visual theory, this paper focuses on and proposes a new multi-level object-oriented method (Multi layer Object-oriented SVM, MOBIA-SVM). The classification work proposed theoretical framework, ideas, focusing on practical value and academic sig-nificance, for the algorithm application design and experimental. Classification results show that the object-oriented multi-level support vector classification method is not only suitable for large-scale remote sensing image, and classification accuracy improve significantly for small-scale object features.The research was further deepening and improvement of support vector machine theory in the application of object - oriented remote sensing image classification and conducted useful explorations and providing technical support for land cover classification. At the same time, the promotion of computer vision technology, especially remote sensing image segmentation in the study of the classification has important practical significance.