The Research Of Dimensionality Reduction For Hyperspectral Image Based On LPP And TWSVM-RFE

Hyperspectral images provide enormous amount of information,resulting in great difficulties in data transmission,storage and processing.Thus,it becomes an advanced research hotspot in the field of hyperspectral data processing technology to find the way to compress the size of data and reduce the dimensionality of images.In this study,band selection method is used to reduce the dimesionality of hyperspectral images.In view of the nonlinear structure of hyperspectral data,the manifold learning method is applied to dimensionality reduction in this study,and two improved algorithms based on locality preserving projection(LPP)have been proposed.Firstly,in order to improve the robustness of the original algorithm,an improved method called ILPP has been proposed.It incorporates the approach of fusing spectral and spatial information to construct weighted matrix,which has decreased the effect of spectra on the size of neighborhood and the sensitivity of the neighbors to the algorihtm.Experiments on two datasets have proved that ILPP is less sensitive to the size of neighbor than LPP.Secondly,to enhance the classification performance,an improved method based on ILPP called TWILPP has been proposed.It combines the idea of the Schr?dinger eigenmaps of making discrete analogue of the Laplacian matrix and introduces similarity weighted matrix and difference weighted matrix to design a new feature mapping model,further strengthening the separability of the algorithm.The experimental results show that TWILPP outperforms other dimensionalty reduction methods in classification performance.Thus it is confirmed that the improvements on LPP are active and practical in dimensionalty reduction.Considering the contribution of different band to the classification of hyperspectral images,ranking-based band selection method is applied to dimensionality reduction in this study,based on which the research of improving the algorithm called twin support vector machine recursive feature elimination(TWSVM-RFE)is carried out and the method TWSVM-RFE-MRMR has been proposed.Incorporating the advantage of the algorithm called max-relevance and min-redundancy(MRMR)that can effectively eliminate the redundant features,to tackle the drawback that the original algorithm is unable to reject the redundant bands.Therefore,more representative bands can be selected and more information can be preserved after dimensionality reduction.The efficacy of embedding of MRMR in TWSVM-RFE has evidenced by improved classification and feature selection performance on two datasets.It is proved that the proposed algorithm is effective and feasible on band selection.