Research On Classification Of Hyperspectral Remote Sensing Imagery Based On Data Fusion

The rapid development of the hyperspectral remote sensing technology, which made it possible to acquire object information on the surface of the Earth, is very helpful for achievement of the quantitative analysis on remote sensing. Hyperspectral remote sensing imagery generally consists of dozens or hundreds of narrow, also contiguous, spectral bands, which accounts for the computational burden and the phenomenon where the response of bands tends to be highly correlated. Consequently, advanced techniques are needed to exploit the extensive information contained in hyperspectral data. Classification of hyperspectral remote sensing imagery plays an important role in its application. These years, a large amount of algorithm on classifying multispectral data is accomplished by researchers, but the characteristics that hyperspectral data possessed restricts its application on hyperspetral imagery owing to a huge computational burden. To solve this problem, this dissertation focuses on hyperspectral remote sensing imagery classification methodology based on information fusion algorithm after a thorough study on the characteristics of hyperspectral data. The major research are as follows:Firstly, dyadic ridgelet transform is brought in to achieve information fusion on hyperspectral remote sensing imagery, and a brand new fusion algorithm is put forward which based on the feature of the data gained after ridgelet transform. This method applies finite Randon transform to all the sub images which were classified into the same band set at first, which can change line singularity in the image into point singularity, then dyadic wavelet transform is implemented to deal with the point singular data. While choosing fusion algorithm, the data character gained by wavelet transform is taken into consideration: normal variance is weighed on those data which represents the outline information of the image; as to those data that contains details and texture information, the pixels with biggest absolute value is chosen to represent the pixel value of fused image. After pixel level fusion to AVIRIS image was accomplished, texture classification is done based on the fused image. Experimental results show that this method can effectively improve fusion result, and achieve high overall accuracy of classification.Secondly, further study is processed to allay the influence of "wraparound effect" brought in by finite ridgelet transform. The study shows that while the larger size of the sub image gained by segmentation of image, the "wraparound effect" takes more important role; while the smaller, the more advantage that ridgelet can embodiment. Nevertheless, the smaller size of sub image also results to clearer block effect, and the fusion results achieved by ridgelet will also approximate to wavelet transform. Thus, compromise should be made due to different consideration while choosing the size of segmentation block.Thirdly, a kind of digital ridgelet transform based on true ridge functions and fast slant stack algorithm is researched on its application in hyperspectral data fusion in order to eliminate "wraparound effect". Since this algorithm avoid using finite randon transform to discretize ridgelet, this digital ridgelet transform can thoroughly eliminate "wraparound effect", thus achieving better fusion results, but it also brings in data redundancy. In order to improve classification accuracy and computational speed, a new neural network called sample weight neural network (SWNN) was combined with this digital ridgelet. Because of the advantages which SWNN possessed (such as, not so sensitive to initial value, no partial minimum value, with a rapid convergence speed, et al), the combination strategy can achieve higher accuracy and less computational burden than other neural network such as BP neural network, RBF neural network.Finally, fusion of hyperspectral data on decision level is studied based on digital ridgelet, sample weight neural network and majority voting rule algorithm. In this algorithm, pixel level fusion is achieved by using digital ridegelet transform, then local classification is processed by using SWNN; then decision fusion is attained by synthesize the results of all the local classifiers with majority voting rules. Experimental results show that this algorithm can achieve high classification accuracy even with very limited training samples, and the decision level fusion algorithm based on SWNN network is better than traditional neural networks such as BP, RPF.