Figure 14. Mean spectral
SSA plots from FRT00021C92 are shown for the regions of interest for the
Mount Desert Island and Sands of Forvie. ROI coverages are shown
overlain onto a portion of the HiRISE color mosaic. The Mount Desert
Island ROI covers 3709 pixels whereas the Sands of Forvie ROI covers
1709 pixels at 12 m/pixel. Thick lines for the spectral plots were
derived from the full log maximum likelihood processing of the data,
whereas thin lines represent the means for the data before the
processing. The wavelength regions between 0.62 and 0.72 µm are not
included in analyses because they correspond to the instrument’s
blocking filter region. RGB labels correspond to the bands used to make
the CRISM false color images displayed in Figures 1-3. Wavelength
coverages for HiRISE and Mastcam color mosaics are also shown.
Both the MDI and SoF mean spectra converge for wavelengths shorter than
~0.75 µm result from the spectral dominance of ferric
iron oxides and oxyhydroxides at shorter wavelengths (e.g., He et al.,
2022). This is consistent with the observed reddish-brown colors in
Mastcam images from both sites, given that both Mastcam and HiRISE color
images are associated with these wavelengths, as shown on Figure 14.
Johnson et al. (2004) show that thin coatings of iron oxides become less
opaque at longer wavelengths over the spectral range shown in Figure 14.
This phenomenon explains why the longer wavelength segments of the
spectra are dominated by broad absorptions associated with ferrous
silicate minerals. Thus, for detailed analyses of the mineralogy of MDI
and SoF ROIs, we restrict the wavelength range to 0.9 to 2.5 µm where
the data are controlled by the spectral properties of ferrous silicates
within the deposits.
6 CRISM Spectral Endmembers Using Checkerboard Unmixing
The next step in analysis of MDI dune and SoF sand sheet spectra was to
consider the spectral heterogeneity. There were two reasons for pursuing
this topic. First was to understand the areal heterogeneity of the sand
deposits for purposes of understanding the dynamics of deposition and
transport. The second was to evaluate further contributions of local
bedrock grains on the margins of the sand deposits, providing a much
larger areal extent than obtained using APXS measurements. To pursue
these objectives, we used a linear checkerboard unmixing approach with
only L data (1.03 to 2.5 µm, 219 bands). This is because calculations
with the combined S and L data showed that subtle pixel-to-pixel
misregistration offsets at the joins of the two spectra strongly
affected endmember selections. As explained in Section 2, the SMACC
method takes the spectral data set and searches for endmembers that
bound the hyperdimensional swarm. The program then treats each pixel and
its respective spectrum as a linear combination of the endmember
spectra.
The ROIs for the two sand deposits were run through the SMACC procedure
as one data set under the simplifying assumption that the two deposits
have the same mineral phases in different proportions to explain the
differences for the long wavelength portions of the spectra. In fact,
three endmembers explain 90% of the variance in the combined data set.
The remaining 10% was found to be associated with aberrant spectra with
unrealistic spectral shapes, presumably associated with ill-behaved
portions of the 2D CRISM detector array (Murchie et al., 2007). Figure
15 shows the three endmember spectra, along with a color-coded relative
abundance map. The endmember spectra include one (red trace) that is
unlike the mean spectra shown in Figure 14, with a rise from 1.03 to 1.5
µm that curves to flat values at longer wavelengths. The abundance map
indicates high proportions of this endmember at the margins of the two
deposits and is interpreted to be a mixture of local bedrock and
regolith incorporated into the basaltic sands, as shown in Figure 15.
Examples of this local contamination can be seen as partially buried
bedrock in Mastcam color images from MDI and SoF sand deposit locations
(Figures 5 and 8).