Analysis of Imaging Spectrometer Data Using N-Dimensional Geometry and a Mixture-Tuned Matching Filtering Approach

摘要

Imaging spectrometers collect unique data sets thatare simultaneously a stack of spectral images and a spectrumfor each image pixel. While these data can be analyzed usingapproaches designed for multispectral images, or alternatively bylooking at individual spectra, neither of these takes full advantageof the dimensionality of the data. Imaging spectrometer spectralradiance data or derived apparent surface reflectance data canbe cast as a scattering of points in an n-dimensional Euclideanspace, where n is the number of spectral channels and all axesof the n-space are mutually orthogonal. Every pixel in the dataset then has a point associated with it in the n-d space, withits Cartesian coordinates defined by the values in each spectralchannel. Given n-dimensional data, convex and affine geometryconcepts can be used to identify the purest pixels in a given scene(the "endmembers").N-dimensional visualization techniques permithuman interpretation of all spectral information of all imagepixels simultaneously and projection of the endmembers back totheir locations in the imagery and to their spectral signatures.Once specific spectral endmembers are defined, partial linearunmixing (mixture-tuned matched filtering or "MTMF" ) can beused to spectrally unmix the data and to accurately map theapparent abundance of a known target material in the presence ofa composite background. MTMF incorporates the best attributesof matched filtering but extends that technique using the linearmixed-pixel model, thus leading to high selectivity between similarmaterials and minimizing classification and mapping errors foranalysis of imaging spectrometer data.