Hyperspectral image super-resolution for remote sensing using high-resolution multispectral image

Abstract

Hyperspectral imaging (HSI) has emerged as a pivotal technology in remote sensing, offering unparalleled spectral resolution that enables detailed analysis of the Earth's surface. However, the inherent trade-off between spatial and spectral resolutions in HSIs often limits their practical utility. This thesis addresses the critical challenge of enhancing the spatial resolution of HSIs using state-of-the-art deep learning approaches, with a particular focus on remote sensing applications. The investigation primarily explores the efficacy of Convolutional Neural Networks (CNNs) and Transformers in performing super-resolution on HSIs. Through extensive experiments conducted on a remote sensing dataset, the research assesses these models based on their ability to reconcile the high spectral fidelity of HSIs with enhanced spatial details. Subsequently, the selected methods are compared in terms of six different performance measures, to quantitatively measure the spatial and spectral fidelity of super-resolved images. Additionally, visual analysis techniques such as mean absolute error maps and spectral angle maps are utilized to offer qualitative insights into the models' performance. The findings reveal that Transformer-based models, owing to their proficiency in capturing long-range dependencies, significantly outperform CNNs. This superiority is consistent across different scales of downsampling, underscoring the robustness of Transformer models to resolution degradation. The results of this study provide valuable insights into the current state-of-the-art in hyperspectral image super-resolution, offering guidance for future research in this domain.