Hyperspectral Imagery Classification Based On Deep Learning

Hyperspectral imagery,which is an important data source in remote sensing area,records the spatial information of the scene,and collects the spectral information of the target at the same time.Therefore,it is widely used in areas such as precision agriculture,environment monitoring,modem military and so on.Classification is the most fundamental problem of hyperspectral image analysis,its accuracy directly affects the application and development of remote sensing technology.Solving hyperspectral image classification problem by machine learning methods has become a mainstream.However,due to the issues of large data redundancy,low spatial resolution,and small training set,accurate and stable classification is still an urgent problem.To deal with these issues,based on deep learning,which has been successfully used in computer vision,this dissertation carried out researches from two aspects,information extraction and information utilization of hyperspectral image.The contributions of this dissertation are summarized as follows:Firstly,based on deep learning,spectral-spatial feature extraction method specialized for hyperspectral imagery is developed and analyzed.A deep network is designed to solve the prob-lem of information redundancy,and low spatial resolution,which uses single convolutional k-ernel to construct a spatial contextual information layer,utilizes auto-encoder with sparse prior to build spectral extract layers hierarchically,then all parameters are fine-tuned by minimizing an improved energy function,therefore,realizes spatial and spectral feature extraction.A deep feature extraction method based on relative distance prior is proposed to solve the problem of fine-tuning under small training set,which use intra-class compactness to aggregate the features of the same class,takes inter-class sparseness to disperse the features of different classes.With these priors,the network is able to achieve accurate feature extraction with small training set.Secondly,in order to solve the problem of using the information of labeled samples in supervised classification,representation based classifications are studied,then a spectral-spatial classification method based on spatial regularized collaborative representation is proposed.All deep features of training samples are used to construct a representation dictionary to represent each test sample,and representation residuals of each classes are utilized to built a probability-like vector,then the vector is dealt with spatial regularization to complete the segmentation of a probabilistic graph.This method can achieve higher accuracy of classification under small training set.Finally,in order to solve the problem of using the information of unlabeled sample infor-mation in semisupervised classification,spatial neighbor information based labeling methods are studied.A semi-supervised classification method based on multi-decision labeling is proposed,self-decision is achieved by extracting the attribute information using deep network,local deci-sion is made by using local information from weighted spatial neighborhood scoring strategy,and global information is taken by global similarity learned from a deep network.This method selects the samples with consistent decisions as the training candidate,optimizes the candidate training set via active learning,and extends the training set to improve the classification accuracy.To sum up,by analyzing the information of hyperspectral imagery,this dissertation studies the information mining and utilizing strategy based on the machine learning methods,which is an enrichment of data analysis methods of hyperspectral imagery,and valuable exploration of applications in related fields,has certain theoretical significance and practical value.