Share Email Print

Proceedings Paper

Hyperspectral data preprocessing improves performance of classification algorithms
Author(s): Suresh Subramanian; Nahum Gat; Jacob Barhen
Format Member Price Non-Member Price
PDF $17.00 $21.00

Paper Abstract

Neural networks (NN) have been applied to hyperspectral image classification when traditional linear statistical classifiers have proven inadequate. The nonlinear and non- parametric properties of NN have often been cited for their apparent success. It has also been known that data preprocessing techniques such as principal component analysis (PCA) greatly improves classification accuracy. While PCA finds the axes of maximum variance in the data it does not guarantee increased separation between an arbitrary pair of classes. A transformation that is sensitive to class structure is obtained by solving the generalized eigenvalue problem of the amongst and within class covariance matrices of the data. Using this transformation, we demonstrate a case where the performance of linear statistical classifiers is comparable to that of NN classifiers for hyperspectral image classification.

Paper Details

Date Published: 31 October 1997
PDF: 9 pages
Proc. SPIE 3118, Imaging Spectrometry III, (31 October 1997); doi: 10.1117/12.283829
Show Author Affiliations
Suresh Subramanian, Opto-Knowledge Systems Inc. (United States)
Nahum Gat, Opto-Knowledge Systems Inc. (United States)
Jacob Barhen, Oak Ridge National Lab. (United States)

Published in SPIE Proceedings Vol. 3118:
Imaging Spectrometry III
Michael R. Descour; Sylvia S. Shen, Editor(s)

© SPIE. Terms of Use
Back to Top