Single image super resolution for spatial enhancement of hyperspectral remote sensing imagery

Student thesis: Doctoral Thesis

Abstract

Hyperspectral Imaging (HSI) has emerged as a powerful tool for capturing detailed spectral information across various applications, such as remote sensing, medical imaging, and material identification. However, the limited spatial resolution of acquired HSI data poses a challenge due to hardware and acquisition constraints. Enhancing the spatial resolution of HSI is crucial for improving image processing tasks, such as object detection and classification. This research focuses on utilizing Single Image Super Resolution (SISR) techniques to enhance HSI, addressing four key challenges: the efficiency of 3D Deep Convolutional Neural Networks (3D-DCNNs) in HSI enhancement, minimizing spectral distortions, tackling data scarcity, and improving state-of-the-art performance.The thesis establishes a solid theoretical foundation and conducts an in-depth literature review to identify trends, gaps, and future directions in the field of HSI enhancement. Four chapters present novel research targeting each of the aforementioned challenges. All experiments are performed using publicly available datasets, and the results are evaluated both qualitatively and quantitatively using various commonly used metrics.The findings of this research contribute to the development of a novel 3D-CNN architecture known as 3D Super Resolution CNN 333 (3D-SRCNN333). This architecture demonstrates the capability to enhance HSI with minimal spectral distortions while maintaining acceptable computational cost and training time. Furthermore, a Bayesian-optimized hybrid spectral spatial loss function is devised to improve the spatial quality and minimize spectral distortions, combining the best characteristics of both domains.Addressing the challenge of data scarcity, this thesis conducts a thorough study on Data Augmentation techniques and their impact on the spectral signature of HSI. A new Data Augmentation technique called CutMixBlur is proposed, and various combinations of Data Augmentation techniques are evaluated to address the data scarcity challenge, leading to notable enhancements in performance.Lastly, the 3D-SRCNN333 architecture is extended to the frequency domain and wavelet domain to explore their advantages over the spatial domain. The experiments reveal promising results with the 3D Complex Residual SRCNN (3D-CRSRCNN), surpassing the performance of 3D-SRCNN333.The findings presented in this thesis have been published in reputable conferences and journals, indicating their contribution to the field of HSI enhancement. Overall, this thesis provides valuable insights into the field of HSI-SISR, offering a thorough understanding of the advancements, challenges, and potential applications. The developed algorithms and methodologies contribute to the broader goal of improving the spatial resolution and spectral fidelity of HSI, paving the way for further advancements in scientific research and practical implementations.
Date of Award4 Oct 2023
Original languageEnglish
Awarding Institution
  • University Of Strathclyde
SupervisorStephen Marshall (Supervisor) & Jaime Zabalza (Supervisor)

Cite this

'