Denoising Of Hyperspectral Remote Sensing Image Based On Multiple Linear Regression And Sparse Representation

Denoising of hyperspectral remote sensing image is different from two-dimensional natural image, and the three-dimensional data mode determines that the denoising method usually considers both one-dimensional spectral information and two-dimensional spatial information. Hyperspectral remote sensing image denoising belongs to the pre-processing stage of the image analysis, which plays an important role for subsequent application. Of course its existence has a very important significance.Due to the strong spectral correlation of hyperspectral remote sensing image, multiple linear regression(MLR) has been widely used in research and practice. It is used according to the fact that the image signal and noise signal present different spectral correlation. Through MLR transformation, image signal and noise signal will be separated in order to preliminarily estimate pure image and noise signal.As it has compact support and its decomposition has multi-scale, sparse representation has been widely used to denoise two-dimensional image. It is the reason that image signal and noise signal through the sparse representation transformation present different laws in different scales. The threshold method is used to shrink sparse representation coefficients in order to remove noise signal from two-dimensional spatial information.Dimensionality reduction of hyperspectral remote sensing image data is essentially the fact that several bands is used to hold on the most information of dozens or even hundreds bands. In this paper, the noise adjusted principal component analysis(NAPCA) is proposed, its main components are arranged in accordance with the signal to noise ratio(SNR) from high to low. So comparing to principal component analysis(PCA) that main components are arranged with the variance, NAPCA retains better image information in denoising process.Multiple linear regression, sparse representation, data dimensionality reduction are as a key tool, the paper proposes one kind of noise estimation method and three kinds of denoising methods. Simulated experiment and real data experiment are performed to show that the proposed methods are Main work is as follows:1. Noise estimation by combining multiple linear regression and median estimation technology of wavelet coefficientsFirstly, MLR is used to remove spectral correlation of hyperspectral remote sensing image, in order to obtain the initial estimate of noise signal. And then the estimated noise signal is transformed into wavelet domain, the median value of the high-frequency wavelet coefficients is as the estimation of noise variance. Simulated results show that the proposed method obtained better results in estimation error. Real data experimental results show that the proposed method is feasibility and applicability.2. Denoising of hyperspectral remote sensing image based on MLRHyperspectral remote sensing image is performed by MLR to obtain the preliminary estimation of image signal and noise signal. Preliminary estimation of image signal still contains little noise, and noise signal contains image information. So the initial estimation of image signal is further denoised, and little image signal should be extracted from noise signal. Denoising the preliminary estimate of image signal, because of high SNR and lower noise level, noise needs to be highlighted in order to help to be removed. Preliminary estimation of image signal needs to be transformed into differential domain, and then BivaShrink function is used to shrink wavelet coefficients of differential domain image. And the processed image is obtained by the integral computation and integral correction. The noise signal obtained by MLR is directly performed BivaShrink shrinkage function to extract image information. The extracted information from noise signal are used to correct the processed image to obtain the denoising result. Simulated results show that the proposed denoising method makes full use of all the information on hyperspectral remote sensing image and obtains a better SNR and RMSE of each band.3. Analysis and denoising of hyperspectral remote sensing image in the Curvelet domainIf hyperspectral remote sensing image is performed directly to remove noise, it is possible that while denoising, some image details are seen as noise to be removed. So it is considered that hyperspectral remote sensing image is transformed into a new representation domain for denoising. The paper proposed that hyperspectral remote sensing image is transformed into Curvelet domain band by band. And then the same direction in the same scale from different bands in Curvelet domain representation are stacked, data blocks of different direction and different scale are obtained, which represents hyperspectral remote sensing image in the Curvelet domain. By analyzing these data blocks, it is known that they maintained the strong spectral correlation of hyperspectral remote sensing image. So MLR can be applied to predict representation of pure data in the Curvelet domain. Simulated results show that it is feasible to remove noise in the Curvelet domain and to obtain a higher SNR. In contrast to the RMSE and MSSIM of each band, it can be seen that the proposed method be superior to literature methods.4. NAPCA for denoising of hyperspectral remote sensing imageThe idea of hyperspectral remote sensing image dimensionality reduction is the fact that tens or even hundreds bands are compressed into a few bands. The different dimensionality reduction methods have their own advantages. As denoising is the purpose of the paper, NAPCA is used whose main components arranged according to SNR. Firstly NAPCA is performed for dimensionality reduction of hyperspectral remote sensing image. And then some bands are chosen to be kept, BivaShrink function is used shrink amplitude of complex wavelet coefficients for denoising remaining bands. The remaining bands data is also performed to remove noise in spectral dimension. SNR is only used for comparison in simulated experiment. The best results are obtained by using dimensionality reduction. Real experimental results show that three kinds of denoising methods proposed in the paper are superior to the reference methods in terms of removing noise and simultaneously maintaining fine features during the denoising process.