Hyperspectral Image Classification Based On Deep Stacking Network

The technology of hyperspectral image classification is the most important research direction of hyperspectral remote sensing and has been widely used in military and civil fields.However,because hyperspectral data has a large number of bands,strong correlation with each band and the large amount of computation,the conventional image classification methods in hyperspectral image processing have greater restrictions.How to quickly and accurately mine needed information from a large number of hyperspectral data,which is still an urgent problem to achieve high accuracy classification.Therefore,more and more researchers pay attention to hyperspectral image classification based on deep network.Currently,deep learning methods has been applied to in the field of hyperspectral remote,which includes deep autoencoder network,deep belief network and deep convolution network.In these deep neural network models,thousands of parameters need to be learned,and it lead to time-consuming in training.It is difficult to parallel training on multi CPU to use stochastic gradient descent algorithm based on minimal batch processing.Deep stacking network overcomes the problem amenable to parallel training which can quicken learning and training speed.Based on the analysis of the existing algorithm,this paper extensively studies the structure and algorithm of deep stacking network,combines logistic regression classifier and proposes an effective method for hyperspectral image classification.The main research works and results are as below.(1)The paper analyzes hyperspectral remote sensing image's data characteristics and introduces the features of hyperspectral remote sensing image and studies several classical feature extraction algorithms which comprise PCA,manifold learning,maximum noise fraction transformation,neural network and deep neural network.Next,the paper introduces classification flowchart of hyperspectral image and existing classification algorithm for hyperspectral pixels which include unsupervised classification and supervised classification.(2)This paper focuses on the analysis of deep stacking network model.The structure,learning algorithm and fine tuning algorithm of deep stacking network model are studied.The influe nce of the regularization technique and the depth of the network model is discussed and the model of logistic regression is studied.These provide a new foundation for hyperspectral image classification.(3)The paper studies a new method to extract depth features of hyperspectral images by using deep stacking network model,and raises a classification method which is deep stacking network and logical regression(DSN-LR).The problem of choosing the regularization Dropout value and the depth of the model is studied,and the optimal parameters are selected for the simula t io n experiments.The simulation results show that compared with the traditional feature extraction algorithm,deep stacking network can extract better depth features and that to ensure the optimal conditions,the classification index of DSN-LR has been better than traditional methods.The classification results of confusion matrix and KSC figure further explain the superiority of DSN-LR algorithm.(4)The paper proposes a new hyperspectral classification method base on DSN-LR and the joint of spectral information and spatial information.Classification accuracy can be further improved by adding spatial information on the basis of spectral information.A spatial feature extraction method based on PCA transform and pixel neighborhood is used,which combines spatial features and spectral features to form spectral space information.Using deep stacking network-logistic regression(DSN-LR)extracts the features of spectral spatial information to fulfill classification.The simulation results show that the classification index of DSN-LR method based on spectral space is higher than traditional support vector machine and the aforementioned spectral information based classification.DSN-LR method based on spectral space effectively improves the classification performance of hyperspectral images.