Smart Optimization Methods For Hyperspectral Remote Sensing Classification

With the development of remote sensing technology, hyperspectral remote sensing has been widely used in many fields as a new remote sensing technique. Now, hyperspectral remote sensing has been a hot spot of research in application area of remote sensing. Hyperspectral remote sensing image has many bands, enormous data, and the acquisition of the samples is difficult, this brings us new challenges and opportunities. For hyperspectral remote sensing images, the research about classification techniques speedly and efficiently is always one of the hot spots of research fields.Bionics is founded in the mid 1950’s, people get enlightenment from the principle of biological evolution and have proposed many smart optimization algorithms which are used to resolve complicated problems. In recent years, the smart optimization algorithms are used in remote sensing image processing and analysis, and they can get good effect. The thesis studies the classification of hyperspectral remote sensing, the main research works and contributes of this thesis are as follows:1) Traditional classification techniques of remote sensing are systemically summarized, the principle, advantage and process of Support Vector Machines (SVM) are elaborated, and the methods of selecting kernel parameters of SVM are discussed. Methods of dimension reduction of hyperspectral remote sensing are concluded from feature selection and feature extraction. In addition, the measures of classes separated are discussed also.2) Aim at the prominent problem of reparability between band subset selection and SVM parameters selection which ignores the interrelationship, therefore this results in low classification accuracy. The thesis proposes some smart optimization algorithms (Genetic Algorithms,Particle Swarm Optimization, Artificial Immune Algorithm, Simulated Annealing) to select band subset of hyperspectral remote sensing images automatically and SVM parameters simultaneously. Accordingly, more effective SVM classification model for hyperspectral remote sensing is built up. Experiment shows that the proposed classification model can produce higher classification accuracy than traditional SVM classification method. Furthermore, the proposed method has more automatization and intelligence.3) Two improved PSO algorithms are used in band selection and SVM kernel parameters optimization. First, Simulated Annealing-Particle Swarm Optimization is proposed to achieve the complementation between global and local optimization. Accordingly, the optimum solution is achieved in larger areas. Second, an improved PSO algorithm is proposed to achieve positive band number selection, in this way, the selection of PSO has more extensiveness and flexibility.4) Traditional classifications of hyperspectral remote sensing have made use of spectral information mainly. Texture features based on grey level co-occurrence matrix (GLCM) and wavelet feature based on discrete wavelet transform are proposed to classify hyperspectral remote sensing.Aim at the condition, the thesis proposes smart optimization algorithms to optimize feature selection in texture feature, wavelet features and spectral features. Compared with traditional spectral feature space classification, the new feature space has larger features information. Experiment indicates the classification method which combines texture feature, wavelet feature and spectral feature can get better results than only using spectral feature classification.