Nonlinear Feature Extraction And Classification Of Hyperspectral Data

Hyperspectral remote sensing technology has been one of hot research topics in remote sensing field and has produced good social and econolic benefits in many practical scenarios.Therefore,it is an urgent need to improve and perfect the processing and analysis method of hyperspectral data.Although the traditional feature extraction method has achieved some initial results in the processing of hyperspectral data,it doesn't solve the problems of data redundancy and the lack of artificial mark samples.Considering the nonlinear characteristics of hyperspectral data,firstly this paper improves and optimizes the existing algorithms based on the manifold learning method.Furthermore,in recent years,because of its strong feature learning ability,deep learning method has become more and more attractive in feature extraction and classification of hyperspectral data.In this paper,the problem of limited artificial mark sample in training dataset is discussed.A convolutional neural network model based on multi-layer feature fusion is designed to extract its nonlinear features,and finally realized the classification.The main contents of the study are as follows:(1)In order to solve the noise problem of original hyperspectral data,we pre-process and de-noise the original hyperspectral data,and then transform the corrected three-dimensional hyperspectral data into two-dimensional data.(2)For the nonlinearity of hyperspectral data and the problem of "small sample",a new method called orthogonal exponent discriminant locally reserved projection method(OEDLPP)is proposed based on manifold learning.Compared with princepal component analysis(PCA),locality preserving projection(LPP),discriminating locality preserving projection(DLPP),exqonent discriminating locality preserving projection(EDLPP)and orthogonal discriminating locality preserving projection(ODLPP),the experimental results show that the proposed algorithm has advantages in obtaining the effective information of samples and can achieve higher classification accuracy.(3)A convolution neural network model based on multi-layer feature fusion is designed by using the deep learning method,combined with the typical convolution neural network framework model and OEDLPP algorithm,which makes it perfonm better in the classification of hyperspectral data.It will play a significant role in solving the problem of limited training samples of hyperspectral data.