Guarino, Giuseppe (2025) Deep Learning for Hyperspectral Pansharpening and Beyond. [Tesi di dottorato]

Documento PDF ITEE38_PhD_Thesis_GiuseppeGuarino.pdf Visibile a [TBR] Amministratori dell'archivio Download (40MB)

Available Versions of this Item

• Parchi Agrari Periurbani: una nuova proposta di paesaggio. (deposited 29 Apr 2010 13:10)
  • Deep Learning for Hyperspectral Pansharpening and Beyond. (deposited 10 Dic 2025 22:30) [Attualmente visualizzato]

Abstract

High spatial and spectral resolution are among the most desired properties in remote sensing imagery, as they enable precise object detection and accurate material discrimination. Hyperspectral satellite imagery provides rich spectral information; however, hardware constraints limit its spatial resolution, reducing its effectiveness in fine-scale analysis. To overcome this limitation, data fusion techniques such as pansharpening fuse low-resolution hyperspectral data with high-resolution panchromatic images, generating products that combine both spatial and spectral detail. Despite recent advances in deep learning–based pansharpening, most existing solutions overlook the unique challenges of hyperspectral imagery, such as massive spectral dimensionality, scarcity of real hyperspectral–panchromatic pairs, and limited spectral overlap with the panchromatic band, hindering their scalability and practical applicability on large-scale datasets. This thesis addresses these gaps through three main contributions. First, it provides a critical review and unified implementation of state-of-the-art hyperspectral pansharpening methods, both model-based and deep learning, released as an open-source toolbox for fair and reproducible benchmarking. Second, it introduces a large-scale dataset of real hyperspectral–panchromatic pairs, enabling a realistic evaluation of this fusion task. Third, it presents three deep learning methods designed to efficiently address the intrinsic challenges of hyperspectral data fusion via innovative band-handling schemes and tailored loss functions. Finally, the thesis investigates the practical potential of hyperspectral pansharpening in the context of cross-sensor scene classification, experimentally demonstrating that fused hyperspectral data can enhance domain adaptation and classification performance across heterogeneous sensors.

Downloads

Actions (login required)

Modifica documento