Abstract

Improving the spatial resolution of hyperpsectral image (HSI) has traditionally been an important topic in the field of remote sensing. Many approaches have been proposed based on various theories including component substitution, multiresolution analysis, spectral unmixing, Bayesian probability, and tensor representation. However, these methods have some common disadvantages, such as that they are not robust to different up-scale ratios and they have little concern for the per-pixel radiometric accuracy of the sharpened image. Moreover, many learning-based methods have been proposed through decades of innovations, but most of them require a large set of training pairs, which is unpractical for many real problems. To solve these problems, we firstly proposed an unsupervised Laplacian Pyramid Fusion Network (LPFNet) to generate a radiometrically-accurate high-resolution HSI. First, with the low-resolution hyperspectral image (LR-HSI) and the high-resolution multispectral image (HR-MSI), the preliminary high-resolution hyperspectral image (HR-HSI) is calculated via linear regression. Next, the high-frequency details of the preliminary HR-HSI are estimated via the subtraction between it and the CNN-generated-blurry version. By injecting the details to the output of the generative CNN with the low-resolution hyperspectral image (LR-HSI) as input, the final HR-HSI is obtained. LPFNet is designed for fusing the LR-HSI and HR-MSI covers the same Visible-Near-Infrared (VNIR) bands, while the short-wave infrared (SWIR) bands of HSI are ignored. SWIR bands are equally important to VNIR bands, but their spatial details are more challenging to be enhanced because the HR-MSI, used to provide the spatial details in the fusion process, usually has no SWIR coverage or lower-spatial-resolution SWIR. To this end, we designed an unsupervised cascade fusion network (UCFNet) to sharpen the Vis-NIR-SWIR LR-HSI. First, the preliminary high-resolution VNIR hyperspectral image (HR-VNIR-HSI) is obtained with a conventional hyperspectral algorithm. Then, the HR-MSI, the preliminary HR-VNIR-HSI, and the LR-SWIR-HSI are passed to the generative convolutional neural network to produce an HR-HSI. In the training process, the cascade sharpening method is employed to improve stability. Furthermore, the self-supervising loss is introduced based on the cascade strategy to further improve the spectral accuracy. Experiments are conducted on both LPFNet and UCFNet with different datasets and up-scale ratios. Also, state-of-the-art baseline methods are implemented and compared with the proposed methods with different quantitative metrics. Results demonstrate that proposed methods outperform the competitors in all cases in terms of spectral and spatial accuracy.

Publication Date

7-7-2022

Document Type

Dissertation

Student Type

Graduate

Degree Name

Imaging Science (Ph.D.)

Department, Program, or Center

Chester F. Carlson Center for Imaging Science (COS)

Advisor

David Messinger

Advisor/Committee Member

Nathaniel Barlow

Advisor/Committee Member

Charles Bachmann

Campus

RIT – Main Campus

Share

COinS