Research On Hyperspectral Image Classification Methods Based On Spectral-Spatial Feature Domain Adaptation

Due to the advantages of high spectral resolution and image-spectrum merging characteristic,hyperspectral imagery has been widely concerned by researchers all over the world since the 1980 s.As the most important research direction in hyperspectral image(HSI)interpretation,ground object classification has achieved rich fruits in the past decades and played a significant role in geological exploration,urban management,military reconnaissance and other application fields.However,with the “Big Data”trend and the increase of application requirements,the limitations of HSI data properties have brought some challenges for information extraction and pattern recognition of classification technique in many situations: the distortion of spatial and spectral information caused by noise interference severely reduces the reliability of information extraction;the complex structure of ground objects under high-resolution imaging conditions leads to the difficulty in characterizing discriminative information;the disparities of data distribution and feature dimensionality between different images hinder the cross-scene utilization of supervised information.To address these problems,this dissertation starts with the study of HSI restoration,based on which,the discriminative spectral-spatial feature extraction technique is investigated,and the research on cross-scene domain adaptation classification is conducted from the homogeneous and heterogeneous perspectives according to the properties of sample features in different scenes.The main content of the dissertation includes the following parts.First,a regularization model with joint constraints is proposed to tackle the spectral and spatial information distortion problem that limits the classification accuracy.By combining the low-rank and total variation priors,the model is capable of maintaining the intrinsic properties of HSI.In order to overcome the inaccurate measurement of sparse noise matrix in the existing approaches,an equivalent formulation is introduced in this dissertation to rewrite the L0-norm sparsity constraint to remove the impulse noise,stripes,deadlines and other outliers.Simulated and real data experiments confirm that the proposed model can effectively recover the spatial and spectral information from the observed noisy data,thus improve the image quality and enhance the classification accuracy.In terms of exploiting spectral-spatial features of HSI,a multiscale feature extraction method based on joint convolutional sparse decomposition is generated in this dissertation to overcome the difficulty of characterizing the discriminative features of high-resolution images with complex spatial structures.Based on the fact that the spatial information of an image can be decomposed into a structural component that related to the properties of the ground objects at a certain scale and the irrelevant local texture component,a joint convolutional sparse decomposition algorithm is introduced to separate the spatial structural features at different scales from the hyperspectral data by using an inverse problem solving approach.Through extracting principal components of spectral dimension to further enhance the inter-class differences,high-precision classification is achieved on four single hyperspectral scenes,which demonstrate that the proposed method is capable of adapting high-resolution images with different land-cover types and provide effective spectral-spatial features for classification task.In order to address the problem of feature disparities caused by acquisition-related shifts(illumination conditions,atmospheric conditions and sensor properties)when transferring discriminative knowledge of the same class between different scenes,cross-scene domain adaptation is investigated in this dissertation for HSI classification.For the case that the source and target scenes are acquired by the same sensor with the same feature dimensionality,homogeneous domain adaptation method based on distribution alignment of spectral-spatial features is studied.Specifically,a homogeneous domain adaptation model that adapts the joint probability distributions while preserving the neighborhood relationship of samples is proposed to overcome the distribution shifts between different hyperspectral scenes.Through applying the homogeneous domain adaptation model,a cross-scene spectral-spatial classification framework is developed by performing the joint convolutional sparse decomposition on source and target scenes to extract multiscale features at the first stage,and a deep cross-scene classification method based on homogeneous domain adaptation is generated by embedding the adaptation loss into a spectral-spatial unified network.By performing cross-scene classification experiments on three real data pairs,the proposed methods are verified to be effective in aligning the spectral-spatial features between homogeneous hyperspectral scenes and solving the cross-scene transfer learning problem when the target scene is poor-labeled.As for the situation where the source and target scene data are acquired in different feature spaces with different feature dimensionalities,a heterogeneous domain adaptation method with joint distribution alignment and cross-domain structure preservation is proposed for spectral-spatial classification based on the aforementioned research content.First,the multiscale spectral-spatial features of source and target scenes are extracted by conducting the joint convolutional sparse decomposition approach,so that the spectral information and spatial structure information at different scales can be jointly adapted to learn transferable features for classification.Then,in order to overcome the heterogeneity between the two feature sets,a cross-domain local structure-preserving criterion is introduced to combine with the distribution alignment function to build the heterogeneous domain adaptation model.By learning asymmetric projections for source and target feature sets,the model is capable of mapping the spectral-spatial features into a new common subspace.Experiments conducted on four real data pairs demonstrate that the proposed approach can effectively transfer spectral-spatial knowledge between heterogeneous HSIs and yielded superior results for poor-labeled scene to be classified.