Research On Feature-level Domain Adaptation For Hyperspectral Image Classification Arithmetic

Hyperspectral sensor images at adjacent spectral wavelengths,which can explore the intrinsic properties of ground objects.The practical application of hyperspectral image depends on the development of classification technology to a great extent.Hyperspectral image contains rich spectral-spatial information,but with it comes the contradiction between big data and less labeled samples.Solving this contradiction effectively is of great significance and research value for the accurate interpretation of subsequent images.Therefore,based on the characteristics of hyperspectral image,this thesis discusses the joint extraction of spatial feature and spectral feature and the reuse of labeled samples.Firstly,in order to solve the problem of within-class reflectance heterogeneity in a single hyperspectral image,the intrinsic image decomposition is introduced to enhance the reflectance homogeneity by removing the shading component.In addition,in order to realize the extraction of spectral-spatial features,the thesis constructs a 3D convolution neural network based on the deep learning,removing the phenomena of within-class "salt and pepper noise" and cross-edge "over-smoothness".The spatial continuity of the result is enhanced,and higher classification accuracy is obtained.Secondly,in view of the spectral shift phenomenon caused by samples with different distributions,the domain adaptation theory in transfer learning is introduced to realize the distribution alignment between source domain(training)and target domain(test).Since the first-order statistics represent the central tendency of the spectral data,and the secondorder statistics characterize the spectral variation of each band and correlations between bands,the first-order and second-order statistical feature alignment method is adopted to reduce the distribution differences and improve the classification performance between different domains.Finally,in order to solve the problem of the cross-scene classification where there are larger distribution differences between source domain and target domain,the thesis proposes a deep domain adaptation network based on statistics characterize.Combining deep learning and domain adaptation,a loss function measuring the distributions difference is added into the network to realize the transfer between different scenes based on the 3D CNN.The thesis studies the deep domain adaptation under small samples and unsupervised condition,which makes the existing samples reused and enhances the adaptability of the samples from source to the target scene.What’s more,the method effectively reduces the burden of manual labeling and obtains significant results.