Research On Hyperspectral Remote Sensing Image Classification Based On Deep Belief Networks

With the rapid development of hyperspectral data spectroradiometer of high spatial and spectral resolution.How to extract feature information and classify data has become an important problem in hyperspectral data classification in the face of the feature such as multi-dimensional,nonlinear and large data volume.As the most effective algorithm of the image recognition,Deep Belief Network(DBN)has the small sample study,highdimensional space,nonlinear,etc,and becomes the hot issue because of its superiority of the hyperspectral data classification.In this paper,DBN is used to the classification of hyperspectral data,aimed to solve the problems of classification.First,this paper analyzes the theoretical framework of DBN in deep learning,And the DBN model is built on the basis of Matlab platform,the feature extraction and data classification of hyperspectral data are completed by this platform.Finally,it classify the data fused hyperspectral date with airborne radar data by using DBN,then evaluate the results.Classification method for hyperspectral data have been conducted in this pater based on the DBN analysis.It turns out that accuracy is highest when using information entropy to determine the best number of hidden layers for different data.Aiming at the airborne hyperspectral data with more band numbers,a kind of band selection method based on Markov Distance has been proposed,which can get rid of redundancy bands and diminish data dimension.The hyperspectral data involve spectral and spacial information simultaneously.The data have been analyzed from spectral,spacial and spectral-spacial feature spaces respectively.Compared to the SVM classification results,it turns out that both results of two method are better when using spectral-spacial feature for classification than single feature.The classification accuracy of DBN is 3.78% higher than if of SVM.Besides,DBN has been used in classification of the airborne hyperspectral images and airborne radar images fusion data in order to testify its availability for feature mining.The result shows that the classification effect of DBN is better than it of SVM and the accuracy of fusion data is higher than it of hyperspectral data.