Super-Resolution for Hyperspectral Remote Sensing Images Based on the 3D Attention-SRGAN Network

Abstract

Hyperspectral remote sensing images (HSIs) have a higher spectral resolution compared to multispectral remote sensing images, providing the possibility for more reasonable and effective analysis and processing of spectral data. However, rich spectral information usually comes at the expense of low spatial resolution owing to the physical limitations of sensors, which brings difficulties for identifying and analyzing targets in HSIs. In the super-resolution (SR) field, many methods have been focusing on the restoration of the spatial information while ignoring the spectral aspect. To better restore the spectral information in the HSI SR field, a novel super-resolution (SR) method was proposed in this study. Firstly, we innovatively used three-dimensional (3D) convolution based on SRGAN (Super-Resolution Generative Adversarial Network) structure to not only exploit the spatial features but also preserve spectral properties in the process of SR. Moreover, we used the attention mechanism to deal with the multiply features from the 3D convolution layers, and we enhanced the output of our model by improving the content of the generator’s loss function. The experimental results indicate that the 3DASRGAN (3D Attention-based Super-Resolution Generative Adversarial Network) is both visually quantitatively better than the comparison methods, which proves that the 3DASRGAN model can reconstruct high-resolution HSIs with high efficiency.

Hyperspectral Image Super-Resolution Based on Spatial Correlation-Regularized Unmixing Convolutional Neural Network

“…in which η is the balance parameter that controls the approximation and the sparsity of abundances. In this work, we use || || 2 instead of || || 0 to promote the generalization ability of the network [44,52]. Note that the SSTM defined in (6) S ∈ R p 2 ×p 2 is actually a local transform matrix with p 2 × p 2 elements, which transforms each pixel from the HR image patch to the LR patch, rather than the global transform matrix in (3).…”

“…Based on the recurrent neural network (RNN), Fu et al [42] proposed a bidirectional 3D quasi-RNN that combines both CNN and RNN for single HS SR work, and they achieved improvements in terms of both restoration accuracy and visual quality. Shi et al proposed super-resolution approaches by incorporating attention modules [43] and a generative adversarial network [44,45], which achieves better visual quality over some state-of-the-art approaches. A brief analysis and comparison of recent HS image SR methods can be found in [46].…”

Single Image Internal Distribution Measurement Using Non-Local Variational Autoencoder

“…Several reconstruction-based SISR methods have been introduced to solve the SISR problem, utilizing a shallow feature learning process [39], [40]. KernelGAN [41], consisting of a deep linear generator and a discriminator, supports blind SISR.…”

Recent Advances in Multi- and Hyperspectral Image Analysis

“…Since the spectral profiles are specific for different materials, exploiting such high-dimensional data can help to determine the characteristics of the objects of interest. Therefore, the analysis of hyperspectral images (HSIs) has attracted research interest in various fields of science and industry, including mineralogy [3], precision agriculture [4,5], medicine [6], chemistry [7], forensics [8], and remote sensing [1,9,10]. On the other hand, multispectral images (MSIs) contain much fewer bands with larger bandwidths [11].…”