Spectral-Spatial Hyperspectral Image Classification Based On Deep Neural Network

Image classification is one of the most popular areas in hyperspectral remote sensing.Hyperspectral image classification technology has achieved great progress in civil and military fields,and has been widely used.In recent years,with the continuous development of hyperspectral imaging technology and imaging spectrometer,the spectral resolution and spatial resolution of hyperspectral images are also getting higher and higher.Due to the increase of spectral resolution,we can get the rich information of remote sensing,but also bring new pressure and great challenge to the later image processing.How to effectively mine the required information from hyperspectral data with the rich band number,and achieve accurate classification,is still an urgent research problem.In this paper,the characteristics of hyperspectral image data are considered,and the nonlinear characteristics with high discrimination ability are extracted by using neural network model based on deep learning,so as to realize the accurate classification of hyperspectral image.This paper mainly studies the two kinds of deep leaning model:stacked autoencoder and convolutional neural network.The main work and achievements of this paper are as follows:Firstly,this paper systematically introduces the theory basis of hyperspectral image.This paper describes the characteristics of hyperspectral image and some existing hyperspectral image classification method at home and abroad.Secondly,this paper studies two network models of deep learning.Analyzes the model structures of deep stacked autoencoder and convolutional neural network,and introduce the concepts of convolution and pool operation,the advantages of local receptive fields and parameters sharing.These studies provide a strong foundation for hyperspectral image classification.Finally,this paper proposes several classification methods of hyperspectral image.The first proposed classification method is the stacked autoencoder based on boundary discrimination method.The algorithm can utilize stacked autoencoder model to extract hyperspectral image with high discrimination for classification,by selecting the boundary samples for network fine-tuning,improves the classification performance model.We discuss the over-fitting problem of the stacked aotuencoder under small sample.The second proposed classification method is the stacked autoencoder based on a spatial-adaptively constrained and boundary discrimination method.The algorithm discusses that the stacked autoencoder extracts contextual feature of sample by introducing superpixel segmentation,and makes full use of the unlabeled adjacent samples to overcome the small sample size problem.The third proposed classification method is the convolutional neural network based on jointing spatial-spectral features method.The method joints the spatial features and spectral features of samples,and can obtain the features with high discrimination by jointing the shallow and deep features of convolutional neural network.It can effectively improve the feature extraction ability of convolutional neural network.In addition,we introduce a data expansion strategy.This strategy can select the appropriate samples to label in the unlabeled samples,according to the similarity of the spatial and spectral,and then expand the training set.It can effectively overcome small sample size problem and further improve the performance of classification model.