Hyperspectral Image Classification Based On Deep Learning With Small Samples

Hyperspectral image classification is one of the most popular research directions in the field of remote sensing.It has great applications in the fields of mineralogy,agronomy,urban planning,environmental science and military.Hyperspectral image consists hundreds of bands with rich spatial and spectral information.In the case of limited training samples,the high-dimensional spectral data may cause "dimension disaster".How to make suitable and effective measures to use the rich spectral-spatial information of hyperspectral becomes the main content of hyperspectral classification.Because of the high collection cost of hyperspectral data set,spectral-spatial feature extraction in condition of small training samples is a difficult problem in hyperspectral image classification.This thesis starts off the hyperspectral image characteristics of small training samples and abundant information,and makes full use of advantages of deep learning.And this thesis design effective hyperspectral data preprocessing method and spectral-spatial feature learning model to improve hyperspectral image classification accuracy.The main research contents of this thesis are as follows:(1)In case of small training samples,it is difficulty for traditional deep learning model to extract saliency feature from hyperspectral image,this thesis proposes a preprocessing method for hyperspectral image classification base on metric learning.Firstly,the similarity between hyperspectral pixels is learned by using hyperspectral metric learning network,then the target pixel patch is replaced according to the similarity of pixels and the specified replacement strategy,the preprocessing method make hyperspectral more smoother.This method can not only ensure the spatial continuity and diversity,but also suppress the interference of noise to samples.The experimental results show that the method can achieve better region consistency and boundary discrimination ability in different spectral data sets,which verifies the effectiveness of the method.(2)In case of small training samples,randomly dividing training set and test set may lead the training samples are not representative and the traditional hyperspectral classification model can not encode the spatial relationship of hyperspectral objects.This thesis proposes a hyperspectral capsule network based on unsupervised sample selection.In this method,unsupervised features are extracted by convolution autocoder,then representative training samples with large amount of information are selected by super-pixel and clustering methods.The relationship between spatial features and spatial level of hyperspectral is learned by using spectral-spatial capsule.Experimental analysis shows that the method is robust on different hyperspectral datasets and advanced in comparison with other algorithms.(3)In order to solve the problem that the traditional hyperspectral classification network is difficult to extract strong discrimination and has the same characteristic in the case of small training samples,this thesis proposes a hyperspectral classification method based on the deep capsule network.In this method,the deep capsule network is used to extract the abstract features with high discrimination.Firstly,the deep hyperspectral features are extracted by the spectral-spatial residual block.Then,the hierarchical relationship between spectralspatial features is modeled by the deep capsule layer.Finally,the similarity between same class features and the difference between different class features are increased by adding the regular term of loss function,so as to obtain better classification effect.Experimental results show that this method has a great advantage in classification accuracy compared with other advanced hyperspectral classification algorithms.