🚀 go-pugleaf

                             Final Public Oral Examination
                          For the Degree of Doctor of Philosophy

                                     SUSAN M.SCHWEIZER
                         DEPARTMENT OF ELECTRICAL & COMPUTER ENGINEERING
                  


                    Date:  Thursday, September 9, 1999
                    Time:  11:30a to 2:30p
                    Place: Hamerschlag Hall First Floor Conference Room,
Room 1112
    
                                         PIZZA LUNCH 
                      
***************************************************                     
  
     
                                  
                                           TITLE
             ``The GMRF Detector for Hyperspectral Imagery: An Efficient
                       Fully-Adaptive Maximum Likelihood Detector''

                                         ABSTRACT
Hyperspectral sensors collect hundreds of narrow and contiguously spaced 
spectral bands of data organized in the so called hyperspectral cube.  The 
hyperspectral imagery provides fully registered high resolution spatial
and spectral information that is invaluable in discriminating between man-made 
objects and natural clutter backgrounds.  This comes at a cost.  The high 
volume of data in the hyperspectral cube has precluded the development of 
computationally practical Maximum-Likelihood~(ML) detectors of man-made 
anomalies in clutter.

This thesis solves this problem.  We derive the Gauss-Markov random 
field~(GMRF) detector, a computationally efficient ML anomaly detector
that fully adapts to the unknown statistics of the clutter, and fully exploits 
the spatial and spectral correlation of the hyperspectral imagery.  We 
test extensively our clutter adaptive GMRF detector with real imagery from 
several hyperspectral sensors.  Our results show that the GMRF detector is 
significantly simpler computationally and noticeably improves the
detection performance over the benchmark anomaly detection algorithm. 
Our approach 
avoids the costly step of inverting the large sample covariance matrix of 
the clutter.  We parameterize directly the inverse of the clutter 
covariance and develop several alternative methods to match this inverse 
to the actual clutter statistics.