Sparse Tensor Feature Extraction Based Hyperspectral Image Classification Methods

Hyperspectral remote sensing image analysis has become an important means of earth monitoring and space exploration.Hyperspectral images are three-dimensional(3-D)data,as a complete entity of image and spectrum,including two spatial dimensions and one spectral dimension.Each pixel in the space corresponds to a spectral band,and each pixel corresponds to a ground object.Different kinds of ground objects have different corresponding spectral reflection curves,thus,hyperspectral images can achieve pixel-wise classification.With the continuous development of hyperspectral imaging technology,the spatial resolution and spectral resolution of hyperspectral images are continuously improved,and the information contained in hyperspectral images is also increasingly rich,which promotes the wide application of hyperspectral images,including agricultural detection,military reconnaissance,marine detection,land-use analysis and environmental monitoring,etc.However,at the same time,the hyperspectral image also has the characteristics of large amount of data,strong band correlation and large information redundancy,which brings many challenges to the classification technology.Because hyperspectral images are collected on airborne or spaceborne platforms,and the cost of manual pixellevel labeling is huge,the number of labeled hyperspectral data is limited.Therefore,this paper focuses on tensor-based sparse representation to extract 3-D tensor features from3-D hyperspectral data,in order to minimize information redundancy while preserving the joint spatial and spectral features,and improve classification accuracy.At the same time,in view of the characteristics of the artificial feature extractor and the intelligent learning feature extractor,the algorithm combining the sparse tensor representation with the neural network is studied,which can not only extract the spatial-spectral joint sparse tensor features,but also improve the classification performance by using the intelligent algorithm.Aiming at the problem of small sample data,this paper studies the self-taught learning classification method,to ensure the classification accuracy and reduce the use of hyperspectral training data.The main research contents are as follows:Aiming at the problem of how to fully mine joint spatial-spectral information from 3-D hyperspectral images and improve the classification accuracy,a support vector machine classification method based on 3-D joint spatial-spectral sparse tensor feature extraction is proposed in consideration of the joint spatial and spectral structural features.Firstly,the3-D hyperspectral data is modeled as the sum of the intrinsic spatial-spectral tensor and the corresponding variation tensor.The intrinsic spatial-spectral tensor is decomposed into a core tensor and three factor matrices based on the Tucker decomposition model.The sparse tensor dictionary learning algorithm is used to train the three factor matrices,which are used to extract the joint sparse spatial-spectral tensor features.Using tensor techniques,the 3-D joint spatial-spectral tensor features are extracted directly from 3-D intrinsic data,which includes spatial features,spectral features and joint spatial-spectral features.Support vector machine(SVM)is used as the classifier to classify the extracted joint sparse spatial-spectral tensor features.The enforced sparse constraint can solve the problem of information redundancy in hyperspectral image.Within-class variation can be alleviated by extracting features from the intrinsic spatial-spectral tensor.The extracted joint sparse spatial-spectral tensor features can improve the classification performance.Aiming at the problem of how to mine the joint spatial-spectral information,simplify the design of intelligent classifier,and improve the classification accuracy,an atomsubstituted tensor dictionary learning enhanced convolutional neural network classification method is proposed.Firstly,an atom-substituted tensor dictionary learning algorithm is proposed.When the dictionaries obtained by this algorithm are used for sparse representation,the sparse representation is more accurate and the size of the sparse coefficient tensor is fixed,which is convenient for the design of subsequent classifiers.A simple two-dimensional(2-D)convolutional neural network is used to further extract the features from the joint sparse spatial-spectral tensor extracted by the dictionaries,and then the pixel-wise classification is carried out.Combining sparse tensor representation with convolutional neural network,the feature extractor can be designed by using the prior knowledge of human beings,and the simple neural network can be used to further extract the neglected potential features to improve the classification accuracy.At the same time,the design of the neural network can be simplified to reduce the requirement of the neural network for labeled data.Aiming at the problem of how to use a small amount of hyperspectral labeled data to achieve effective classification,a vector-based online dictionary self-taught learning classification method for hyperspectral image is proposed.A small amount of unlabeled data that differs from the target task is used to improve the classification results on a small sample data set.Firstly,an online sparse vector dictionary learning algorithm is used to train the feature extractor on a small amount of unlabeled data unrelated to the classification task.Secondly,the feature extractor is used to extract features from the target data set,and the SVM classifier is trained with the features of a small amount of data.This method is a vector-based method using the spectral information of every pixel,which has low computational complexity,small computation amount and fast running speed.It is verified that the self-taught learning method based on sparse representation is effective in hyperspectral image classification,and it is proved that the accuracy of supervised classification with a small amount of samples can be improved by using a small amount of unlabeled data.Aiming at the problem of how to fully mine the joint spatial-spectral information from a small amount of hyperspectral labeled data to improve the classification accuracy,a tensor-based dictionary self-taught learning classification method is proposed.Firstly,a sparse tensor dictionary learning algorithm for 3-D samples is proposed.Secondly,the dictionary learning algorithm is used to train the tensorial feature extractor on a small amount of unlabeled data which are unrelated to the classification task.The feature extractor is used to extract the joint sparse spatial-spectral tensor features from the target training set,and the tensor features of a small amount of labeled data is used to train the SVM classifier.The target training set and the unlabeled training set can be collected by different sensors over different scenes,thus the method can carry out the cross-scene and cross-sensor classification task,and greatly improve the classification accuracy.