We introduce a fast algorithm for anisotropic smoothing and segmentation of hyperspectral imagery. Anisotropic smoothing reduces the spatial and spectral variability within uniform regions in the image, while preserving the sharp discontinuities on the image boundaries, which in turn improves the segmentation by increasing the separability between the different regions in the image. The algorithm solves a discretized Partial Differential Equation (PDE) that generates a discrete scale-space, represented by an irregular pyramid of coarser versions of the image. As the image is coarsened, representative pixels are selected at each scale, enabling a multi-scale segmentation of the image. The segmentation is performed in a top-down process that uses the representative pixels identified on each scale, as seeds. The PDE is solved using Algebraic Multigrid (AMG), a numerical analysis technique useful for boundary value problems on highly unstructured grids, with greater accuracy and speed than traditional relaxation techniques. The coarsening step in AMG is based on a modified version of the Iterated Weighted Aggregation method (IWA), tailored to exploit the discrimination power of high dimensional spaces such as those represented by hyperspectral data.


Poster presented at the 2007 Thrust R2C Multi Spectral Discrimination Methods Conference


Hyperspectral Imagery, AMG, IWA

Subject Categories

Imaging systems




Bernard M. Gordon Center for Subsurface Sensing and Imaging Systems (Gordon-CenSSIS)

Publication Date


Rights Holder

Bernard M. Gordon Center for Subsurface Sensing and Imaging Systems (Gordon-CenSSIS)

Click button above to open, or right-click to save.

Included in

Engineering Commons