Research On Hyperspectral Image Classification Methods Based On Feedback Networks

With the rapid development of remote sensing technology,the methods of human observation on the earth are gradually becoming more refined,diversified and intelligent.The development of hyperspectral remote sensing technology,namely,imaging spectroscopy technology,is considered to be one of the two most significant breakthrough technologies since the advent of remote sensing technology in parallel with imaging radar technology.Hyperspectral remote sensing technology combines imaging technology and spectral technology,obtaining the integration of spatial information and spectral information.Hyperspectral images have the advantages of large number of bands,narrow width of bands,and high spectral resolution,reflecting fine features of land cover classes.Hyperspectral images have been widely used in geological exploration,precision agriculture and forest ry,environmental monitoring,disaster prevention and mitigation,urban and rural planning and other civil and military realms.Classification is one of the most important research directions of hyperspectral images.It aims to classify the pixels in the image into several thematic elements according to the diagnostic spectral absorption differences.The existing hyperspectral classification methods are generally feed forward open network systems,in which only information in labeled samples are used,and the usage of unlabeled samples is not sufficient.In real applications,acquiring labels in hyperspectral images is expensive,which makes it necessary to dive deeply the a posteriori information contained in the unlabeled samples.The feedback networks make use of a posteriori information and optimize training process using testing process.This thesis breaks the limitation of conventional deep learning network structure,proposing intelligent feedback networks,fusion feedback networks and a domain adaptation feedback network.The main research contents of this thesis are summarized as follows:To study into the existing feed forward open classification methods,in chapter 2,the basic technologies and evaluation methods of hyperspectral images are introduced.First,three groups of state-of-the-art classification methods are introduced,including spectral-spatial classification methods,convolutional neural networks and detection operators.A novel training sample number allocation method based on class features is proposed,providing a theoretical basis for training sample selection.Three public data sets are classified by the above three methods,and classification accuracies and precisions are used to evaluate and analysis the classification results.Using the above spectral-spatial classification methods as a basic concept,to address the problem of existing methods dependently deal with training and testing processes,in chapter 3,an intelligent feedback network for hyperspectral image classification is proposed.The proposed intelligent close network system uses a feedback structure to feed back the a posteriori information after the spatial filtering to the original data,and classifiers are trained and tested iteratively.Furthermore,a novel intelligent feedback neural network is extended from the feedback close system by adding random spatial filtering layers and a fully connected layer.Experimental results show that the feedback structure can gr eatly improve the classification performance.Based on the intelligent feedback networks adopting single classifier,to address the problem that insufficient a priori information affects classification performance in the condition of small sample sizes,in chapter 4,a novel fusion feedback network is proposed.Fusion feedback networks fuse a posteriori features obtained by different spectral-spatial classifiers in a nonlinear way on the feature level,and the fused features are fed back to the input of the network,then the method above is extended to multiple classifiers,as “one vs one”,“one vs the rest” and “progressive” fusion feedback networks.Experimental results show that compared to intelligent feedback networks,fusion feedback networks achieve effective integration of a posteriori information through feature fusion,and classification results are further improved.Based on the fact that a posteriori information has the ability to improve indomain classification,to address the problem that there exist large number of hyperspectral images without any training samples to be classified,in chapter 5,a domain-adaptive feedback network is proposed.Domain adaptive learning solves the problem of transfer learning where training samples from a source domain and test samples from a target domain with different probability distributions.The proposed domain adaptive feedback network extracts a posteriori spatial information in a changing feature space,and a new sample similarity evaluation criterion is proposed to accurately update the training sample set and the classifier.Experiments prove that the a posteriori spatial information can be used as an invariant feature and transferred between domains.With the help of the new domain adaptive feedback network,samples in the target domain can be accurately chosen as training samples,and the classification results are significantly improved compared to existing domain adaptation methods.In summary,this thesis focuses on the concept of “feedback networks”,and proposes a novel framework to link training process and testing process.The concept of intelligent feedback network is proposed as a generalization of common spectral-spatial methods without feedbacks.Moreover,a multi-classifier fusion feedback network and a domain-adaptive feedback network are also proposed.The accuracies are largely improved both in in-domain and cross-domain classification.