Research On Hyperspectral Reconstruction Based On Interferometric Imaging Mechanism And Error Characteristics

Interferometric hyperspectral imaging technology offers significant advantages in obtaining accurate spectral information for earth remote sensing and deep space exploration due to its advantage of multiple channels and high throughput.The conversion of interference images to spectral images through spectrum reconstruction is a key research element in interferometric spectral imaging to obtain precise spectral information of the target.However,several mechanical factors in the imaging process,including ghost,mixed noise,and maximum optical path difference,inevitably degrade the quality of the original interferometric data,thus limiting the quantitative applications of hyperspectral images.Therefore,investigating theoretical methods of spectrum reconstruction based on imaging mechanism and error characteristics and developing techniques that can effectively reconstruct clear and spectrally accurate hyperspectral data from degraded interferometric data are essential.This dissertation aims to address the challenges encountered in spectrum reconstruction for interferometric hyperspectral imaging and proposes solutions through theoretical simulation and error analysis,with a primary focus on issues related to ghost correction,hybrid noise removal,and high-precision spectral reconstruction.The first aspect of this study focused on the simulation and error analysis of the interferometric imaging system.The study developed forward "spectral->interference" and inverse "interference->spectral" models based on the imaging mechanism of HJ-2A/B and conducted a comprehensive analysis of the noise,quantum efficiency,smear,apodization,and ghost in the imaging system.These analyses provided a fundamental reference for spectrum reconstruction and interferometric imaging spectrometer design and optimization.Secondly,the dissertation proposed a novel algorithm for ghost correction in temporally-spatially modulated Fourier transform interferometric hyperspectral imaging.The method utilized a mirror symmetry model and a SURF(Speeded Up Robust Features)matching algorithm to accurately align ghost images,and obtained the correction coefficient of the ghost image based on region consistency.The proposed method effectively addressed the problem of inherent ghost in the imaging system.Experimental results showed that after ghost correction,the signal-to-noise ratio substantial increased in clearly reflected image areas,and visual quality of hyperspectral images significant improved.The proposed method has been applied to several systems for interferometric hyperspectral data processing,demonstrating its effectiveness in correcting ghost artifacts and improving the quality of hyperspectral images.Thirdly,the dissertation proposed an interferometric hyperspectral denoising algorithm called LR-Pn P(Interferometric Hyperspectral Image Cascaded Denoising via Low Rank and Plug-and-Play Regularization).The method constructed a cascade denoising model by considering the global low-rank and non-local similarity features of spectral data,as well as the noise features of interferometric data.The method also used discrete cosine transform(DCT)to fuse these features into a unified framework,and further improved the denoising effectiveness by directly inserting existing nonlocal features into the denoising model using the plug-and-play(Pn P)approach.In push-scan imaging data experiments,the LR-Pn P method achieved the best performance,at the same time,it outperformed the comparison method on real data of HJ-2A/B satellite.The experimental results demonstrated that the proposed method effectively removes noise in interferometric hyperspectral data.Finally,this dissertation proposed an algorithm for interferometric hyperspectral recovery using a fully connected residual U-Net(FCUN).The study demonstrated the potential of deep learning methods for spectral recovery by analyzing the similarity between fast Fourier transform(FFT)and multilayer neural networks,then a fully connected residual U-Net(FCUN)based on residual connectivity and U-Net architecture was proposed.Compared to FFT based restoration methods,the resolution of single pulse spectral restoration has been improved;At the same time,the FCUN also achieved good results in simulated data of pushbroom imaging and real data of HJ-2A/B satellite.These findings suggest the potential of deep learning methods for improving the performance of interferometric hyperspectral imaging.Overall,the three core algorithms established in this dissertation have solved the typical problems of interferometric hyperspectral imaging,including inherent optical ghost correction,random noise removal,and high-precision spectrum reconstruction.The research has enriched the methods and theories of interferometric hyperspectral reconstruction and improved the efficiency of quantitative remote sensing for interferometric hyperspectral imaging.These findings provide valuable contributions to the field of hyperspectral imaging research.