|M.Sc Student||Eyal Madar|
|Subject||Combined Local-Global Background Modeling for Anomaly|
Detection in Hyperspectral Images
|Department||Department of Electrical Engineering||Supervisors||Professor Emeritus Malah David|
|Dr. Bar-Zohar Meir|
|Full Thesis text|
The detection of materials and objects using remotely sensed spectral information collected by hyperspectral sensors has many military and civilian applications. Detection algorithms exploit the spectral information present in hyperspectral data to detect and discriminate localized man-made targets. In anomaly detection, no prior knowledge on the target spectral signature is assumed. Therefore, anomaly detection algorithms first model the abundant material spectra (background process). Then, every pixel spectrally different in a meaningful way from the background process is declared to be an anomaly.
According to the hyperspectral literature, two major approaches to statistical background modeling can be distinguished. In the first approach, named "local", the background is modeled by a large number of local independent distributions, each of which is responsible to represent a different local region in the image. Local algorithms can tightly fit the background data; however they are subject to an overfitting problem. The second background modeling approach, denoted "global", is based on a global representation of the background process in the whole image. By design, this approach is more resistant to the overfitting problem. However, it has a limited ability to adapt to all nuances of the background process (an underfitting problem).
In this research, we propose a combination of the local and global background modeling approaches by introducing the BEVA (Background Extreme Value Analysis) algorithm. In the local part of BEVA, the local background is approximated using a greedy sequential estimation process. It is composed of a robust estimation of the Gaussian statistics and a background cluster hypothesis discriminator, which is based on Extreme Value Theory results. Then, in its global part, the obtained local background models are inter-related to reduce the number of false alarms. BEVA has the ability to adapt to all nuances of the background process like the local approach but avoids overfitting.
In the next part of the research, we propose several improvements to the BEVA algorithm. We improve BEVA's local part via a preprocessing segmentation that is based on Spectral Clustering.
The Gaussian model used in BEVA, although efficient and mathematically tractable, is only partially adequate to represent real hyperspectral data. In order to overcome this drawback, we introduce the NG-BEVA (Non-Gaussian BEVA) algorithm, which replaces the Gaussian assumption with Gamma distribution fitting.
The results strongly prove the effectiveness of the proposed "local-global" approach. On real hyperspectral data, our local-global algorithms perform better than other examined global or local anomaly detection techniques.