Hyperspectral image analysis via subspace clustering
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
This thesis studies Hyperspectral Image Analysis using various unsupervised learning techniques with a special focus on Subspace Clustering methods. Hyperspectral imaging captures images across numerous (up to hundreds) narrow, overlapping wavelength bands, enabling each pixel in the spatial scene to have a high spectral resolution signature, producing a data-rich, high-volume 3D image cube. However, the high-dimensional nature of the data poses challenges in terms of data processing and analysis. This can be addressed through subspace clustering methods which provide a robust approach for uncovering low-dimensional structures lying in a high-dimensional feature space. This thesis first evaluates different classification methods such as Mean-based classification, Subspace-based classification, and Affine space-based classification. Next, a detailed investigation is performed into several clustering techniques such as K-Means, K-Subspaces, K-Affine spaces, and Threshold Subspace Clustering. Subspace clustering techniques are particularly highlighted in this thesis as they are known for capturing the underlying geometric structure of data points, enabling meaningful clustering even in cases where data points are far apart in high-dimensional spaces but lie on the same subspace. This comprehensive study provides insights into how data from different fields utilizing hyperspectral imaging, including benchmark datasets like Pavia Centre Scene, Pavia University Scene, Salinas Scene, and the Egyptian Blue Test Panel, respond to these methods, paving the way for further advancements in the field of hyperspectral image analysis.
Description
Keywords
Hyperspectral image analysis, Subspace clustering methods, K-affine spaces, Egyptian blue test panel, Salinas Scene