Spectral mixture process conditioned by spatially-smooth partitioning
First Claim
1. A method for enhancing an image containing endmembers, comprising:
- selecting at least one measure of disparity;
partitioning at least one first set of said endmembers in said image into at least one second set of said endmembers, each said second set incorporating at least one site having some spectral content, wherein said at least one site is a generic element of a lattice, and wherein spatial consistency is imposed on said spectral content of each said site so that each said at least one second set is associated with a spatially smooth region in said image; and
using said at least one measure of disparity, applying a linear mixing model to said at least one second set to globally label said at least one second set, wherein said method yields improvements in;
assessment of the type and amount of materials in a scene, unsupervised clustering of a scene, and post-processing smoothing operations suitable for application to diverse techniques for terrain categorization and classification.
2 Assignments
0 Petitions
Accused Products
Abstract
A method that facilitates identification of features in a scene enables enhanced detail to be displayed. One embodiment incorporates a multi-grid Gibbs-based algorithm to partition sets of endmembers of an image into smaller sets upon which spatial consistency is imposed. At each site within an imaged scene, not necessarily a site entirely within one of the small sets, the parameters of a linear mixture model are estimated based on the small set of endmembers in the partition associated with that site. An, enhanced spectral mixing process (SMP) is then computed. One embodiment employs a simulated annealing method of partitioning hyperspectral imagery, initialized by a supervised classification method to provide spatially smooth class labeling for terrain mapping applications. One estimate of the model is a Gibbs distribution defined over a symmetric spatial neighborhood system that is based on an energy function characterizing spectral disparities in both Euclidean distance and spectral angle.
36 Citations
20 Claims
-
1. A method for enhancing an image containing endmembers, comprising:
-
selecting at least one measure of disparity;
partitioning at least one first set of said endmembers in said image into at least one second set of said endmembers, each said second set incorporating at least one site having some spectral content, wherein said at least one site is a generic element of a lattice, and wherein spatial consistency is imposed on said spectral content of each said site so that each said at least one second set is associated with a spatially smooth region in said image; and
using said at least one measure of disparity, applying a linear mixing model to said at least one second set to globally label said at least one second set, wherein said method yields improvements in;
assessment of the type and amount of materials in a scene, unsupervised clustering of a scene, and post-processing smoothing operations suitable for application to diverse techniques for terrain categorization and classification. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17)
-
-
18. An efficient and accurate method for extracting features from hyperspectral data representing a scene by implementing a spectrally-optimal supervised classification algorithm that has been smoothed by a post-processing routine that imposes spectral/spatial constraints defined by the Gibbs prior probability distribution, Pr(XP).
-
19. A method for extracting features from hyperspectral data representing a scene by initializing a partitioning algorithm with the results of a classifier to improve classification by providing initial estimates based on labels that, as applied under the proper conditions, are spectrally optimal, comprising:
-
a) performing a classification of the scene using a supervised classifier that selects the labeling {circumflex over (x)}sP at each said site s that maximizes the probability Pr(Xsλ
|XsP) of a process that is approximately of the form ofwhere;
b) initializing by the results in Step a), performing a spectral/spatial partitioning of said scene using the relationship;
where;
Us(xsP,g)=energy interaction of site s∈
SP(σ
) with the neighborhood
-
-
20. A method employing a simulated annealing method of partitioning hyperspectral imagery for extracting features from hyperspectral data representing a scene, comprising:
-
using a Bayesian framework to develop a 2-step supervised Gibbs-based classification algorithm, wherein said algorithm is capable of performing high quality spatially smooth labeling of said hyperspectral data; and
using a linear classifier to initialize said Gibbs-based partitioning algorithm, wherein said initializing results in improved label accuracy and smoothness as compared to using only said linear classifier without said Gibbs-based partitioning algorithm, and wherein global labeling accuracy is increased as compared to a stand-alone randomly initialized said Gibbs-based partitioning algorithm, and wherein said initializing said Gibbs-based partitioning algorithm with said linear classifier also reduces the computation by eliminating the need for a multi-grid process and by allowing said simulated annealing to start at a cooler temperature.
-
Specification