Hyperspectral Image Classification Based On Deep Domain Adaptation

Hyperspectral remote image analysis is an important part of remote sensing technology,and its emergence and development is of great significance to human understanding of the world.In order to meet the recognition rate of hyperspectral image,improve the classification accuracy and cross domain adaptability,this paper uses deep learning and transfer learning to study the hyperspectral image classification algorithm.The main work includes:(1)Aiming at the high cost of hyperspectral image classification and the large gap between different images,a hyperspectral image classification method based on unsupervised heterogeneous domain adaptation Cycle Gan is proposed.This method combines deep learning,transfer learning and generating confrontation network to capture the transferable features between hyperspectral images in different fields.Firstly,the region nearest neighbor representation is used to construct the joint representation of spatial-spectral features of hyperspectral image;secondly,the twoway adversarial transfer is used to process the source domain and target domain features of hyperspectral image,the two-way mapping is used to find the internal relationship between the source domain and target domain data,and the two-way adversarial is used to constrain the source domain and target domain features to complete the source domain and target domain Secondly,by minimizing the difference of second-order statistics between the source domain and the target domain,the problem of insufficient mapping constraints in data extraction from different domains is solved.Finally,the aligned source domain features are used to train the classifier,and the trained classifier is applied to the classification task of target domain data.(2)Due to the characteristics of hyperspectral image “combining images with spectrums”,and the common phenomena of “the same spectral curve represents different objects” and “different spectral curves represent the same object” in heterogeneous hyperspectral image classification tasks,this paper proposes a hyperspectral image classification method based on heterogeneous spatial-spectral domain adaptation.This method combines the spatial information and spectral information of hyperspectral images,and uses transfer learning to reduce the cross domain distribution difference between source domain and target domain.Firstly,the auto-encoder network is constructed to extract the spectral features of hyperspectral images,and the subspace alignment method is used to transfer the spectral features of hyperspectral images in source domain and target domain.Secondly,the spatial features of hyperspectral images based on region nearest neighbor representation are constructed,and the depth migration network is used to reduce the distribution difference between the data in source domain and target domain.Finally,the extracted spectral features and spatial features are spliced,and the classifier trained by the spectral-spatial features in the source domain is applied to the classification task in the target domain.There are 23 figures,13 tables and 116 references in this paper.