Journal of Spectral Imaging,   Volume 5   Article ID a5   (2016)

Peer reviewed Paper

Rapid hyperspectral image classification to enable autonomous search systems

  • Raj Bridgelall  
  • J. Bruce Rafert
  • D. Denver
  • B. Tolliver
  • EunSu Lee
Upper Great Plains Transportation Institute, North Dakota State University, Fargo, USA

 Search for papers by this author
Upper Great Plains Transportation Institute, North Dakota State University, Fargo, USA

 Search for papers by this author
Upper Great Plains Transportation Institute, North Dakota State University, Fargo, USA

 Search for papers by this author
Management Department, School of Business, New Jersey City University, NJ 07311, USA

 https://orcid.org/0000-0003-1840-9368
 Search for papers by this author
 Corresponding Author
Upper Great Plains Transportation Institute, North Dakota State University, Fargo, USA
[email protected]
 https://orcid.org/0000-0003-3743-6652
 Search for papers by this author

The emergence of lightweight full-frame hyperspectral cameras is destined to enable autonomous search vehicles in the air, on the ground and in water. Self-contained and long-endurance systems will yield important new applications, for example, in emergency response and the timely identification of environmental hazards. One missing capability is rapid classification of hyperspectral scenes so that search vehicles can immediately take actions to verify potential targets. Onsite verifications minimise false positives and preclude the expense of repeat missions. Verifications will require enhanced image quality, which is achievable by either moving closer to the potential target or by adjusting the optical system. Such a solution, however, is currently impractical for small mobile platforms with finite energy sources. Rapid classifications with current methods demand large computing capacity that will quickly deplete the on-board battery or fuel. To develop the missing capability, the authors propose a low-complexity hyperspectral image classifier that approaches the performance of prevalent classifiers. This research determines that the new method will require at least 19-fold less computing capacity than the prevalent classifier. To assess relative performances, the authors developed a benchmark that compares a statistic of library endmember separability in their respective feature spaces.

Keywords: acquisition systems, autonomous vehicles, classification accuracy, endmember separability, hazardous material detection, real-time spectrometer, resolution agility, simple spectral classifier, unmanned aircraft systems, video imaging

Metrics

Downloads:

1,176

Abstract Views:

4,397