Background Suppression and Feature-Based Spectroscopy Methods for Subpixel Material Identification
August 2012
Robert S. Rand, National Geospatial-Intelligence Agency
John M. Grossmann,The MITRE Corporation
Roger N. Clark, U.S. Geological Survey
Eric Livo, U.S. Geological Survey
Thomas Parr, BBN Technologies
ABSTRACT
Feature-based imaging spectroscopy methods are effective for identifying materials that exhibit specific well-defined
spectral absorption features. As long as a pixel contains a sufficient amount of material so that the absorption retains its
predominant shape, a feature-based method can work well. However, there are occasions when a background material
can mix with a material of interest, and significantly distort and maybe even remove the absorption. In such cases, the
material identification capabilities of these methods are likely to be degraded. This effort proposes an approach to
accommodate these conditions. The parameter values to determine fit of an absorption feature are selected to be more
tolerant of distortions and the signal contributions of any detected sub-pixel backgrounds are removed by making use of
a physically-constrained linear mixing model. This mixing model is used to remove any detected background spectra
from the image spectra within the bounding locations of the spectral features. However, an expected consequence of
loosening the parameter values and performing sub-pixel subtraction is an increase in false alarms. A statistically-based
spectral matched filter is proposed as to reduce these false alarms. We test the individual and combined approaches for
identifying full-pixel and sub-pixel Tyvek panels in an experiment using a HyMAP hyperspectral scene with ground
truth collected over Waimanalo Bay, Oahu, Hawaii.
Additional Search Keywords
Hyperspectral, feature-based methods, imaging spectroscopy, sub-pixel, mixed-pixel analysis, spectral
match filters, ACE, scene segmentation, background suppression
|
