TY - GEN
T1 - A comparative study of loss functions for hyperspectral SISR
AU - Aburaed, Nour
AU - Alkhatib, Mohammed Q.
AU - Marshall, Stephen
AU - Zabalza, Jaime
AU - Al Ahmad, Hussain
N1 - © 2022 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting /republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
PY - 2022/10/18
Y1 - 2022/10/18
N2 - The spatial enhancement of Hyperspectral Imagery (HSI) is a popular research area among the community of image processing in general and remote sensing in particular. HSI contribute to a wide variety of industrial applications, such as Land Cover Land Use. The characterstic that distinguishes HSI from other type of images is the ability to uniquely describe objects with spectral signatures. This can be achieved due to the sensor's ability to capture reflectance in narrowly spaced wavelength bands, which yields an HSI cube with hundreds of bands. However, this ability compromises the spatial resolution of HSI, which must be improved for practicality and usability. There are several studies in the literature related to HSI Super Resolution (HSI-SR), especially using Convolutional Neural Networks (CNNs). Nonetheless, the investigation of the most suitable loss functions to train these networks is necessary and remains as an area to investigate. This paper conducts a comparative study of the most widely used loss functions and their effect on one of the state-of-the-art HSI-SR CNNs, mainly 3D-SRCNN. The paper also proposes a hybrid loss function based on the comparative results, and proves its superiority against other loss functions in terms of Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measurement (SSIM), and Spectral Angle Mapper (SAM).
AB - The spatial enhancement of Hyperspectral Imagery (HSI) is a popular research area among the community of image processing in general and remote sensing in particular. HSI contribute to a wide variety of industrial applications, such as Land Cover Land Use. The characterstic that distinguishes HSI from other type of images is the ability to uniquely describe objects with spectral signatures. This can be achieved due to the sensor's ability to capture reflectance in narrowly spaced wavelength bands, which yields an HSI cube with hundreds of bands. However, this ability compromises the spatial resolution of HSI, which must be improved for practicality and usability. There are several studies in the literature related to HSI Super Resolution (HSI-SR), especially using Convolutional Neural Networks (CNNs). Nonetheless, the investigation of the most suitable loss functions to train these networks is necessary and remains as an area to investigate. This paper conducts a comparative study of the most widely used loss functions and their effect on one of the state-of-the-art HSI-SR CNNs, mainly 3D-SRCNN. The paper also proposes a hybrid loss function based on the comparative results, and proves its superiority against other loss functions in terms of Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measurement (SSIM), and Spectral Angle Mapper (SAM).
KW - 3D SR-CNN
KW - CNN
KW - hyperspectral
KW - loss function
KW - super resolution
UR - https://ieeexplore.ieee.org/document/9909827
M3 - Conference contribution book
AN - SCOPUS:85141010183
SN - 9781665467995
T3 - European Signal Processing Conference
SP - 484
EP - 487
BT - 30th European Signal Processing Conference, EUSIPCO 2022 - Proceedings
CY - Piscataway, NJ
T2 - 30th European Signal Processing Conference, EUSIPCO 2022
Y2 - 29 August 2022 through 2 September 2022
ER -
