Research On Dimensional Reduction And Classification Of Hyperspectral Remote Sensing Image

Hyperspectral remote sensing was founded on electrical-magnetic spectral theory, ground rule, electrical technology, computer science and spatial technology, it is developed rapidly as a new independent integrated technology. Because of the special high resolution of hyperspectral remote sensing, its latent usability has been paid attention on. Many methods have been researched aiming at multispectral image, which tend to perfect. But the large data and high dimensions of hyperspectral image disabled implementing methods of multispectral into hyperspectral directly, so it is important to explore methods of hyperspectral image. This paper started with existing algorithms and theories of related principles, then put its emphasis on dimension reduction and classification methods, the main research and innovation contents are as following:1. Dimensional reduction methods had deeply researched in this paper and they are categorized into 4 types: band selection, data partition, feature extraction and fusion, etc. On the basis of band selection methods, a new band reductionmethod was proposed------Adaptive band selection. High relation and highredundancy exist in hyperspectral bands, it is can not only reduce data dimension but also decrease computation greatly by selecting informative bands. Band reduction is necessary from the viewpoint of real-time information processing. ABS method fully considered spatial relation and spectral relation among bands, after implementing ABS method on hyperspectral data set, sorting the index from big to small, bands are selected automatically according to threshold set by system. ABS method selects bands after examining total characteristic of image, so it overcomes shortcomings of transformation method, original feature can be maintained perfectly.2. Factors effecting classification accuracy of hyperspectral remote sensing image was researched, which can be categorized into 5 types: number of training samples, dimension of data, discrimination function, supposed probability modeland class separability. Among these factors, class separability represents the nature of data set and decides the optimal performance of classifier. The better the separability of data set, the higher the classification accuracy. In common, class separability are considered internal and pre-determinated, on the condition of other 4 factors are set, research of class separability are especially important. In this paper, gaussian lowpass filter are used to improve class separability, it can smooth image. For hyperspectral remote sensing image containing multipixel homogenous objects, gaussian lowpass filter can reduce distance within classes and increase distance between classes, so it is more suitable to classification. Bhattacharyya distance under multi-dimension normal distribution are used to scale class separability before and after filtering. Experiments proved that gaussian low pass filter can increase class separability thereby increase classification accuracy.3. On the basis of adaptive band selection, a new structure second generation wavelet are introduced into fusion of hyperspectral remote sensing image. The main idea of the algorithm is decomposing the original image simply into two subsets, which can be expressed as red points and black points in a chessboard visually, it is simpler and highly close to split step in lifting scheme. Then predicting and updating the subsets in rectangular and quincunx grids in turn, approaching to a kind of characteristic. Fusion results of the new structure second generation wavelet is good and it provides informative data set for following classification.4. Hyperspectral remote sensing image has low spatial resolution, probability of mixture pixels is big, so it is not reasonable assigning a pixel to a certain class. In this paper, cluster analysis and supervise classification are combined in classification of hyperspectral image, support vector machine based on fuzzy clustering is proposed. This combined method selects training samples on the basis of fuzzy clustering, it overcomes the blindness of selecting samples. Support vector machine was implemented on hyperspectral image, which resolves the problem of high dimension and little samples in hyperspectral image.Inclusion, deep research of dimensional reduction and classification ofhyperspectral remote sensing had been done in this paper, and new algorithms had been proposed, experiments had tested that the algorithms can get good results.