Multi-Feature Fusion Strategies Of Representation Framework For Hyperspectral Image Classification

The containment of rich information in huandreds of spectral narrow bands makes hyperspectral images widely use in many areas, such as mineral exploration, environmental monitoring, etc. Recently, with the development of the study in hyperspectral classification field, spatial features fusion-based classification models has become a hot issue, and spatial texture features have been performed well for the representation-based classifiers developed in recent years. The main contents are as followed:Firstly, this article illustrates the importance of hyperspectral image classification study in some aspects of hyperspectral data such as the structure, characteristics, application and research status. Meanwhile, the image fusion strategies have been summarized, especially for those in feature level and decision level. Also the principle of hyperspectral data classification has been expounded. After the comparation of the supervised and unsupervised classification, the Evaluation mechanism of classification results is introduced.Secondly, a weighted residual-fusion-based strategy of multiple features with is proposed. Multiple features include local binary patterns (LBP), Gabor features, and the original spectral signatures. In the proposed classification framework all the residuals of different features are added together in different weights and the label of a test pixel is determined according to the class yielding the minimum residual. Experimental results in two real hyperspectral datasets show the superiority that the proposed residual-based fusion with those using only one feature.Thirdly, a feature level fusion-baesd classification strategy has also been proposed in this article, for the proposed residual-fusion-based strategy is limited to classifiers and over-reliance on the representation based classification. To solve this problem and improve the classifiers'performance of hyperspectral images classifiation, the spatial features have been combined in terms of weighted summation kernel. Then the two types of kernel representations are used for the result of feature fusion process, those are kernel collaborative representation-based classifier (KCRC). Experimental results of several real hyperspectral image datasets demonstrate that the proposed spatial features fusion-based strategy is superior to the original KCRC as well as the fusion feature can make more improvement than the sigle feature.