Close Window

Segmentation of Hyperspectral Imagery: Level Set Methods that Exploit Spectral Information

John Ball, Mississippi State University, jball@ece.msstate.edu
Lori Mann Bruce, Mississippi State University, bruce@ece.msstate.edu (Presenting)

When conducting automated segmentation of hyperspectral imagery, e.g. delineation of ground cover classes, it is clear that one should exploit the substantial spectral information that is available. Typical methods that are often used include parallelepiped or maximum-likelihood classification, both of which have their limitations. Parallelepiped methods often produce an overabundance of commission errors due to their relative sensitivity to within-class variances. Maximum-likelihood methods often encounter problems with hyperspectral imagery due to their requirement of large amounts of labeled training pixels. In this poster, the authors present level set segmentation as an alternative approach for ground cover mapping when using hyperspectral imagery. Advantages and disadvantages of level set segmentation are presented. A new method of supervised level set hyperspectral image segmentation is presented, where the spectral information is utilized to optimize the level set speed functions, i.e. the functions that control front propagation of the individual ground cover classes during the segmentation process. The new speed functions are composed of a spectral similarity term and a stopping term. The spectral similarity term is used to compare pixels to class training signatures and is based on best spectral band selection procedure developed previously by the authors. The stopping term is created from a new best spectral band selection algorithm, which uses a scaled spectral angle mapper. A case study is presented, and level set segmentation results are compared to those from maximum-likelihood classification.

Presentation Type:  Poster

Abstract ID: 191

Close Window