| The detection of objects in hyperspectral imagery has many military and civilian applications. One approach to object detection is change detection. Change detection is the process of using images acquired at different times for the detection of objects of interest. The use of change detection algorithms provides a reduction in false alarms over standard one pass algorithms. Numerous change detection algorithms have been proposed and this thesis provides a taxonomy of such algorithms, which can be divided into two classes: direct detection and estimation based detection. Based on mathematical tractability and physical phenomenology, the linear prediction algorithms provide the best option for the detection of objects with no information about the target required. This thesis provides a detailed examination and comparison of different linear prediction approaches to change detection, specifically the chronochrome and covariance equalization algorithms. Additionally, a third linear prediction technique in the form of whitening is proposed. Real hypersectral data is used to compare the algorithms in three different scenarios. First, a well controlled scene is used to assess performance with registered images in the presence of illumination changes. Second, an aerial image pair that is not co-registered is used to compare the algorithms performance when the targets are in the overlapping region of two images. Lastly, an aerial image pair of an airport scene is used to test the algorithms in the presence of significant man made clutter. |
