Learning Based Super Resolution Application for Hyperspectral Images

Authors : Hüseyin Aydilek, Nihat Inanç

Pages : 210-217

Doi:10.47897/bilmes.1049338

View : 22 | Download : 7

Publication Date : 2021-12-31

Article Type : Research

Abstract :Due to its spectral properties, hyperspectral imaging is superior to other types of imaging tools in identifying, distinguishing and classifying objects. Hyperspectral imaging instruments can detect light reflected from certain wavelengths between infrared and ultraviolet, apart from the wavelength that the human eye can distinguish on the electromagnetic spectrum. While this feature provides detailed information about the spectral feature of the object under investigation, it causes its spatial resolution to be low due to the technical overlap between spatial resolution and spectral resolution. Today, applications of hyperspectral images are increasing in important fields such as agriculture, mining, medicine and pharmacy, especially for military purposes. In order for applications to produce more precise results, high spatial resolution is required, as well as high spectral information. Hardware solving of low spatial resolution problem is a difficult and costly method. Therefore, software solution is an interesting area in the field of image processing. In this thesis, a hybrid solution method based on deep learning and sparse representation is proposed to increase the low spatial resolution of hyperspectral images. The method obtains a super-resolution image from a single hyperspectral image with a low spatial image with a deep convolutional mesh. Later, the super-resolution image obtained and the original low-spatial-resolution hyperspectral image are fused with the dictionary learning method, resulting in a new super-resolution image with high spectral and spatial resolutions. The application results show that our method achieves successful results compared to many super resolution applications in the literature.
Keywords : Remote Sensing, Hyperspectral Imaging, Super-Resolution, Deep Learning, Convolutional Neural Networks, Sparse Representation, Dictionary Learning