Accession Number:
AD1000735
Title:
Algorithms on Flag Manifolds for Knowledge Discovery in N-way Arrays
Descriptive Note:
Technical Report
Corporate Author:
Colorado State University Fort Collins
Personal Author(s):
Report Date:
2015-11-20
Pagination or Media Count:
24.0
Abstract:
We proposed an approach for hyperspectral imagery classification that exploits the geometric framework of the Grassmannmanifold i.e., a parameterization of k dimensional subspaces of n-dimnsional space. The algorithm is particularly well suited to applications where sets of pixels are to be classified. Multiple pixels from a data class characterize the variability of the class information using a subspace representation. We use two metrics defined on the Grassmannian, chordal and geodesic, and one pseudometric, to compute pairwise distances between the points--subspaces. Once a distance matrix is generated, we use the classical multidimensional scaling to find a configuration of points with preserved or approximated original distances, thus realizing an embedding of the Grassmannian into Euclidean space. A sparse support vector machine SSVM trained in the embedding space simultaneously classifies embedded subspaces and selects a subset of optimal dimensions of the embedding for subsequentmodel reduction and data visualization.
Subject Categories:
- Numerical Mathematics