Spectral Unmixing Applied to Desert Soils for the - Naval ...

and Kruse, 2011; Green et al., 1988)) portion performs noise suppression within the data
as well as some reduction in both spatial and spectral space. Pixel Purity Index (PPI)
determination is where the purest pixels (endmembers) are identified using convex
geometry. Endmembers are then visualized through utilization of the n-dimensional
visualizer in the ENVI software package. The endmembers were identified through the
use of both the ASD spectral library described previously and libraries built by USGS
(Clark et al., 2007). The endmembers were then run through the MTMF spectral (partial)
unmixing process to determine location and abundance in the HSI data.
b. The MTMF Method
MTMF involves three analysis steps. Step one involves an MNF
transform, step two calculates the matched filter, and step three is where mixture tuning
(MT) occurs. MT utilizes convex geometry to measure pixel mixture probabilities that
are composites of both the target spectrum and background spectra. MTMF is an
automated process that only requires the data and predetermined endmembers (Boardman
and Kruse, 2011). In order for step one to occur, pre-processing of the data must be
completed. Pre-processing has two main objectives performed as part of the execution of
the MNF transform (Boardman and Kruse, 2011). During the pre-processing phase,
noise whitening is done via decorrelation of noise present in the data and unit variance
across all dimensions of the spectra. The noise whitening and data characterization step
is a crucial part of the MTMF process because it is a main element of what allows
estimation and target detection within the imagery (Boardman and Kruse, 2011). Three
possible options exist in estimating an MNF transform, estimation using a shift
difference, using a dark current image, or through use of known noise parameters. All
three methods are statistical approaches that are useful under varying conditions. The
end result is to determine the eigenvectors and project the noise whitened data onto them
in order to decorrelate the data using equation (7) from Boardman and Kruse (2011).
43