Study Of Hyperspectral Image Fusion And Classification Method Based On Variational Probability Model

Hyperspectral image is a three-dimensional image of map combination,which contains abundant spatial information and spectral information.Therefore,hyperspectral image can provide more valuable information in remote sensing image analysis.In real life,hyperspectral technology has been widely used in feature recognition,environmental monitoring,mineral exploration,military target and so on.However,due to the limitation of the accuracy of hyperspectral imaging spectrometer and the cost of hardware equipment,the spatial resolution of hyperspectral images collected at present is still very low,and images with high spatial resolution and high spectral resolution can not be obtained at the same time,which limits the acquisition of hyperspectral images.Therefore,it is the current research trend to obtain high resolution hyperspectral images(HR-HSI)by fusing low resolution hyperspectral images(LR-HSI)and high resolution multispectral images(HR-MSI).This paper will focus on two aspects of hyperspectral image fusion and classification.By analyzing the characteristics of hyperspectral images,an algorithm framework that can effectively fuse the target images is proposed.Further classification research is carried out on the obtained fusion images,and a model that can improve the classification accuracy is proposed.The research contents of this paper include:(1)Hyperspectral and multispectral image fusion based on the variational probability model.Considering that most of the existing fusion methods are constrained by linear spectral decomposition,a variational probability model(VPM)based on nonlinear decomposition is proposed for unsupervised hyperspectral and multispectral image fusion tasks.The model is composed of two variational autoencoders,which are used for LR-HSI and HR-MSI input images respectively,and learn jointly by sharing parameters θ.Due to the advantages of convolutional neural network(CNN)in image processing,the model uses CNN to extract the deep features of the image.In order to improve scalability and efficiency,two variational inference models parameterized by neural network are used in the VPM network to infer the distribution of hidden variables,so as to realize the nonlinear decomposition process.In addition,the joint likelihood function of LR-HSI and HR-MSI pixels parameterized by neural network is modeled to realize the nonlinear fusion process.Finally,the mini-batch gradient descent method is used to optimize the network parameters.The experimental results on the open data set of hyperspectral images show that the proposed fusion method based on VPM is better than other traditional fusion methods,and the effectiveness and efficiency of the model are proved from the aspects of fusion evaluation parameters,image quality and running time.(2)Joint model of hyperspectral fusion and classification based on variational probability model.On the basis of the fusion experiment of the hyperspectral images,the high resolution hyperspectral images(HR-HSI)obtained from the fusion experiment were further classified and studied.SVM,1D-CNN,2D-CNN and 3D-CNN models were used respectively for classification processing.However,the classification performance of hyperspectral image is highly dependent on spatial and spectral information because of its map combination.Therefore,this paper proposes a combined CNN model,which combines 3D-CNN and 2DCNN to combine spatial and spectral features.In addition,the number of training samples is increased by means of flipping the original data,adding noises and improving of brightness.The results of experiments on two kinds of hyperspectral public data sets show that the combined CNN method not only improves the accuracy of classification,but also shows superiority in the small sample recognition task through data augmentation.In addition,this method is combined with the image fusion method to make the model not only capable of generating hyperspectral images,but also able to identify hyperspectral images,which has more practical application value.