Eyal Madar (EE, Technion)
Tuesday, 7.12.2010, 11:30
In this research, we address the problem of anomaly detection using remotely sensed
spectral information collected by hyperspectral sensors. Anomaly detection algorithms first model the
abundant material spectra (background). Then, every pixel spectrally different in a meaningful
way from the background is declared to be an anomaly pixel. Two major approaches to statistical
background modeling can be distinguished: “the local approach” and “the global approach”. Local algorithms can tightly fit the background process but are subject to an over-fitting problem, which may produce an excessive number of false-alarms. Global methods are more resistant to over-fitting, however, they have a limited ability to adapt to all nuances of the background process (an under-fitting problem), which may result in high false alarm rates, as well as low probability of detection.
In our work, we propose a combination of the local and global background modeling approaches by introducing the BEVA (Background Extreme Value Analysis) algorithm. In its local part,
the background process is estimated using a greedy sequential method. It is composed of
robust estimation of the Gaussian statistics and a background cluster hypothesis discriminator,
based on Extreme Value Theory results. In its global part, the obtained local background models
are inter-related to reduce the number of false alarms. In addition, we improve BEVA's local part
via a preprocessing segmentation that is based on Spectral Clustering. We also introduce the
NG-BEVA algorithm; a non-Gaussian version of BEVA that combines Extreme Value Theory results with
Gamma distribution fitting. NG-BEVA is found to improve further the performance.
* M.Sc. Research under the supervision of Prof. David Malah and Dr. Meir Barzohar