Super-Resolution And Clustering For Hyperspectral Image

Hyperspectral image is captured by airborne sensors or satellites,which covers a wide range of continuous narrow frequency bands.The rich spectral information makes hyperspectral image quite popular in many areas such as mineralogy,surveillance and astronomy.However,the spatial resolution of hyperspectral image is relatively low due to the limitation of imagery hardware,degrading the performance in practical applications.Inspired by the recent achievements of convolutional neural networks,researchers focus on deep learning-based super-resolution reconstruction for hyperspectral image to obtain underlying pattern.On the other hand,the high dimensionality,heterogeneity,spectral variability of hyperspectral image pose multiple challenges to the application of clustering.In the theory of subspace clustering-based methods,the high-dimensional dataset can be represented as a combination of low-dimensional subspaces.Based on the above investigations,the main work of this thesis is as follows:Firstly,a novel method called Spectral-Spatial Residual Network（SSRNet）is proposed for hyperspectral image super-resolution.SSRNet captures spectral-spatial features by simulating 3?3?3 convolutions with 3?3?1 convolutional filters on spatial domain plus 1?1?3 convolutional filters on spectral domain,which is more economic by decreasing the number of parameters.Residual learning framework is utilized to ease the training of networks as well as boosting the reconstruction performance from considerably increased depth.Since the l₂ loss fails to capture the underlying multi-modal distributions of hyperspectral image,SSRNet uses a robust Charbonnier penalty function to handle outliers.In addition,this thesis copes with multi-scale super-resolution problem in a single network,which can help reduce the computational cost and memory demand.The extensive experiments demonstrate that SSRNet performs better than existing methods in accuracy and visual improvements on CAVE and Harvard datasets.Secondly,this thesis proposes a novel algorithm called Spatial Distribution Preserving-Based Sparse Subspace Clustering（SSC-SDP）.Subspace clustering-based methods like sparse subspace clustering is designed to represent data as an union of affine subspaces,while it cannot capture the nonlinear structure of the given data.According to the corollary of Moser theorem,the local volume preserving diffeomorphism is equivalent to the distribution preserving map.By minimizing the inconsistency between the density estimated in the original data and its corresponding sparse representation coefficient matrix,the latent structure of complex hyperspectral image can be preserved in the sparse representation coefficient matrix as much as possible.Besides,SSC-SDP incorporates the spatial information to the density estimation of hyperspectral image to obtain the local structure,and devises the optimization solution based on alternating direction method of multipliers（ADMM）for SSC-SDP.Three hyperspectral datasets are conducted to evaluate the performance of SSC-SDP compared with other state-of-art algorithms.