Unsupervised Band Selection For Hyperspectral Image Based On Multiobjective Optmization

In recent years, the study of hyperspectral remote sensing data band selection(BS) has been widely concerned by scholars owing to its good application in ground materials classification and target recognition. Hyperspectral image comprises hundreds of narrow and contiguously spaced spectral bands. The abundant spectral information provides accurate object recognition potential. However the adjacent bands generally have high correlation. Hyperspectral image is costly in processing time, storage space,and communication bandwidth. Especially when the dimension is high, there is a strong correlation among adjacent bands. A lot of redundant information among bands makes the classification accuracy decreased with increasing dimension of limited training samples. This phenomenon is knowm as Hughes. A lot of statistics classification methods are invalid. So hyperspectral image dimensionality reduction is a key of hyperspectral image application.Due to unsupervised band selection methods have the following advantages: do not require the priori kowledge, do not destroy the physical characteristica of the original image, better to retain the useful information and remove the redundancy. Unsupervised band selection methods have been widely application in ground materials classification and target recognition. This thesis is mainly concerned with the unsupervised band selection. The author’s major contributions are outlined as follows:Unsupervised band selection for hyperspectral data classification based on multiobjective optmization. The so-called unsupervised refers to does not require any priori knowledge, i.e., does not require labels. In order to solve unsupervised hyperspectral image band selection problems. It is modeled as a multiobjective optimization problem in this thesis. We obtain a series of Pareto optimal solutions with different degrees of the amount of information and the differences between the selected bands, and users can choose an interesting solution as required. The proposed method can overcome some shortcomings of the traditional method. The proposed method can destroy the physical characteristica of the original image and better to retain the useful information and remove the redundancy. After dimension reduction, the proposed method is more suitable for ground materials classification and target recognition. Bandselection is a combinatorial optimization problem. So unsupervised band selection for hypersprctral image based on multiobjective optimization can achieve good results. Two popular indexes, overall accuracy(OA) and Kappa coefficient are adopted to assess the quality of classification. The improved unsupervised band selection method can obtain the higher overall accuracy and Kappa coefficient in a short period of time. Crossover operator and mutation operator as the main evolution operators play important roles in the process of search the optimal solution. This thesis adopts non uniform crossover operator and heterogeneous dynamic mutation operator to improve the local search ability. The experiment proved that this method is feasible and effective. The Indian_pines Cp dataset and Salinas dataset as the experiment datas. When the bands number is from 0 to 30, experimental results show the proposed method not only has higher classification accuracy and kappa coefficient but also has lower sensitivity to noise.