Spectral-spatial Kernel Learning For Hyperspectral Image Denoising

Noise contamination is a ubiquitous problem in hyperspectral images (HSIs),which is a challenging and promising theme in many remote sensing applications. Alarge number of methods have been proposed to remove noise. Unfortunately, mostdenoising methods fail to take full advantages of the high spectral correlation andsimultaneously consider the specific noise distributions in HSIs. To overcome thesedrawbacks, a spectral-spatial kernel method for HSIs denoising is proposed in thispaper. The proposed method is inspired by the observation that the spectral-spatialinformation is high redundant in HSIs, which is sufficient to estimate the clear images.In this paper, a spectral-spatial kernel regularization is proposed to maintain thespectral correlations in spectral dimension and to match the original structure betweentwo spatial dimensions. Besides, an adaptive mechanism is developed to balance thefidelity term according to different noise distributions in each band. Therefore, it can’tonly suppress noise in the high noise band but also preserve information in the lownoise band. The reliability of the proposed method in removing noise isexperimentally proved on both simulated data and real data.The main contributions of this thesis are three-fold:1) A spectral-spatial kernelregularization is proposed to maintain the spectral correlations in spectral space and tomatch the original structure in spatial spaces. To the best of our knowledge, theproposed method is the first to propose the spatial-spectral kernel regularizationtechnique into denoising for hyperspectral images.2) To suppress different noise levelin different bands，an adaptive strength is considered to adjust the estimating error ineach band. The strength is designed on the bias between the band and the meanestimation. By using the adaptive strength, the noise level in high noise intensitybands can be decreased and the details of images in low noise intensity can bepreserved.3) The proposed method is tested on two kinds of noise, including randomnoise and striping noise. Compared with existing methods, the proposed method canachieve robust and accurate results.