Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Unsupervised Super-Resolution of Hyperspectral Remote Sensing Images Using Fully Synthetic Training

Xinxin Xu, Yann Gousseau, Christophe Kervazo, Saïd Ladjal

TL;DR

This work addresses hyperspectral image super-resolution without ground-truth HR data by training entirely on synthetic abundance maps generated with the dead leaves model. It relies on a linear mixing framework with $HSI_{LR}(l,i,j)=\sum_{n=1}^{N} S(l,n) \cdot A_{LR}(n,i,j)$ and reconstructs $HSI_{HR}$ via $HSI_{HR}(l,i,j)=\sum_{n=1}^{N} S(l,n) \cdot A_{HR}(n,i,j)$. A Residual Dense Network variant, called RDN-DL, processes $K=N$ abundance channels and includes a softmax ASC layer to enforce the Abundance Sum-to-one Constraint. On the Urban dataset, the method achieves competitive SR performance (e.g., PSNR around 27.8 dB) and outperforms several supervised baselines, illustrating the practicality of synthetic-only training for HSIs SR and reducing dependence on labeled HR data.

Abstract

Considerable work has been dedicated to hyperspectral single image super-resolution to improve the spatial resolution of hyperspectral images and fully exploit their potential. However, most of these methods are supervised and require some data with ground truth for training, which is often non-available. To overcome this problem, we propose a new unsupervised training strategy for the super-resolution of hyperspectral remote sensing images, based on the use of synthetic abundance data. Its first step decomposes the hyperspectral image into abundances and endmembers by unmixing. Then, an abundance super-resolution neural network is trained using synthetic abundances, which are generated using the dead leaves model in such a way as to faithfully mimic real abundance statistics. Next, the spatial resolution of the considered hyperspectral image abundances is increased using this trained network, and the high resolution hyperspectral image is finally obtained by recombination with the endmembers. Experimental results show the training potential of the synthetic images, and demonstrate the method effectiveness.

Unsupervised Super-Resolution of Hyperspectral Remote Sensing Images Using Fully Synthetic Training

TL;DR

This work addresses hyperspectral image super-resolution without ground-truth HR data by training entirely on synthetic abundance maps generated with the dead leaves model. It relies on a linear mixing framework with and reconstructs via . A Residual Dense Network variant, called RDN-DL, processes abundance channels and includes a softmax ASC layer to enforce the Abundance Sum-to-one Constraint. On the Urban dataset, the method achieves competitive SR performance (e.g., PSNR around 27.8 dB) and outperforms several supervised baselines, illustrating the practicality of synthetic-only training for HSIs SR and reducing dependence on labeled HR data.

Abstract

Considerable work has been dedicated to hyperspectral single image super-resolution to improve the spatial resolution of hyperspectral images and fully exploit their potential. However, most of these methods are supervised and require some data with ground truth for training, which is often non-available. To overcome this problem, we propose a new unsupervised training strategy for the super-resolution of hyperspectral remote sensing images, based on the use of synthetic abundance data. Its first step decomposes the hyperspectral image into abundances and endmembers by unmixing. Then, an abundance super-resolution neural network is trained using synthetic abundances, which are generated using the dead leaves model in such a way as to faithfully mimic real abundance statistics. Next, the spatial resolution of the considered hyperspectral image abundances is increased using this trained network, and the high resolution hyperspectral image is finally obtained by recombination with the endmembers. Experimental results show the training potential of the synthetic images, and demonstrate the method effectiveness.
Paper Structure (9 sections, 1 equation, 4 figures, 1 table)

This paper contains 9 sections, 1 equation, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Proposed method
  3. Overview
  4. Synthetic Abundance Generation using the Dead Leaves model
  5. RDN-DL network
  6. Experimental Results
  7. Dataset & Setup
  8. Comparison with the State-of-the-Art Methods
  9. Conclusion

Figures (4)

  • Figure 1: Structure of RDN-DL
  • Figure 2: Example of Synthetic abundances generation (a) at the initialization (b) after the deposit of leaves to simulate local variations (c) after 10 leaves (d) after 100 leaves (e) at the end of the deposit process (f) after the $ASC$ constraint
  • Figure 3: Comparison between a real abundance map of the Urban dataset (Left); a synthetic Dead Leaves abundance map generated using the local variation layer as initialization (Middle); and a synthetic abundance map generated without the local variation layer (Right).
  • Figure 4: Visual comparison between the Ground Truth, LR, Bicubic, MCnet, SSPSR, HSISR and RDN-DL on one Urban's patch at the bande n°1, 50, 100 and 150