Research On The Technology Of Unsupervised Band Selection Of Hyperspectral Image Cube

The unsupervised band selection algorithms of hyperspectral image cube do not need to know some relevant priori information about the image itself and they can be performed in an unsupervised manner. In the real applications of hyperspectral image processing, the priori information of the hyperspectral image is usually unknown or very hard to obtain, but the unsupervised band selection methods can still perform well in these situations and pick out the desired bands, thus leading to a reduced hyperspectral image cube. In this paper, we mainly conduct the research on the technology of unsupervised band selection methods of hyperspectral image and the main work can be summarized as follows:1. This paper studies the common unsupervised band selection algorithms and groups them into three main categories as sequential forward selection methods, band prioritization and decorrelation methods and band clustering methods.2. This paper studies various ways to evaluate the performance of band selection methods, such as target detection or target classification. The target detection methods include orthogonal subspace projection, constrained linear discriminant analysis and fully constrained least squares while the target classification method uses support vector machines with samples for training and testing.3. After comparing the orthogonal subspace projection band selection method with the multiple linear regression band selection method, both of which belong to the sequential forward band selection method, we propose a method called maximum of minimum spectral information divergence band selection which evaluates how much a new band resembles a selected band subset and decides whether it is absorbed or not. Besides, we also propose anew way to decorrelate the prioritized subset without relying on the predefined threshold to generate the desired number of selected bands.4. We perform all these band selection algorithms and performance analysis methods in MATLAB and both on the computer simulated hyperspectral image and the real hyperspectral images.At last, the summary of this paper and the trend of unsupervised band selection methods of the hyperspectral image would be given at the last chapter.