|Contrast based band selection for optimized weathered oil detection in hyperspectral images|Levaux, F.; Bostater, C.; Neyt, X. (2012). Contrast based band selection for optimized weathered oil detection in hyperspectral images. Proc. SPIE Int. Soc. Opt. Eng. 8532: 16. dx.doi.org/10.1117/12.2007322
In: Proceedings of SPIE, the International Society for Optical Engineering. SPIE: Bellingham, WA. ISSN 0277-786X, more
Feature Detection and Discrimination; Contrast Algorithms; ShorelineMonitoring and Surveillance; Hyperspectral Imaging Systems;Hyperspectral Image Acquisition; Platform Movement Correction; LittoralZone; Environmental Surveillance; Optical Stage; Mobile PlatformSensing; Optimal Band Selection; Matched Filter
|Authors|| || Top |
- Levaux, F., more
- Bostater, C.
- Neyt, X., more
Hyperspectral imagery offers unique benefits for detection of land and water features due to the information contained in reflectance signatures such as the bi-directional reflectance distribution function or BRDF. The reflectance signature directly shows the relative absorption and backscattering features of targets. These features can be very useful in shoreline monitoring or surveillance applications, for example to detect weathered oil. In real-time detection applications, processing of hyperspectral data can be an important tool and Optimal band selection is thus important in real time applications in order to select the essential bands using the absorption and backscatter information. In the present paper, band selection is based upon the optimization of target detection using contrast algorithms. The common definition of the contrast (using only one band out of all possible combinations available within a hyperspectral image) is generalized in order to consider all the possible combinations of wavelength dependent contrasts using hyperspectral images. The inflection (defined here as an approximation of the second derivative) is also used in order to enhance the variations in the reflectance spectra as well as in the contrast spectrua in order to assist in optimal band selection. The results of the selection in term of target detection (false alarms and missed detection) are also compared with a previous method to perform feature detection, namely the matched filter. In this paper, imagery is acquired using a pushbroom hyperspectral sensor mounted at the bow of a small vessel. The sensor is mechanically rotated using an optical rotation stage. This opto-mechanical scanning system produces hyperspectral images with pixel sizes on the order of mm to cm scales, depending upon the distance between the sensor and the shoreline being monitored. The motion of the platform during the acquisition induces distortions in the collected HSI imagery. It is therefore necessary to apply a motion correction to the imagery. In this paper, imagery is corrected for the pitching motion of a vessel, which causes most of the deformation when the vessel is anchored at 2 points (bow and stern) during the acquisition of the hyperspectral imagry.