Analysis of methods for representing 3D structures in hyperspectral images

摘要

We develop new models for the spectral/spatial representation of regions with three-dimensional structure in hyperspectral images. We show that traditional spectral/spatial models lead to ambiguities when classifying these regions due, in part, to changes that occur as the environmental conditions change. The new models characterize the variation of vectors that are derived using spectral/spatial filters as the scene conditions change. These models are compared with multiband generalizations of feature vectors derived from co-occurrence matrices. A feature-selection technique is used to reduce the dimensionality of the model for detection and classification tasks. The utility of several subsets of combined spectral/spatial features is compared for the classification of thousands of forest regions that are generated using DIRSIG over a broad range of conditions.