We specifically used log ratio values of compositions in a correspondence analysis (Aitchison, 1994) to reduce the dimensionality of the compositional data sets to explore the extent of regolith contributions to the sand measurements. The final data set consisted of laboratory-based spectra used in the Hapke (2012) modeling of mineral assemblages to retrieve phase abundances and grain sizes (Tables 1 and 2).

SSA spectra were retrieved from both the S and L CRISM data (Murchie et al., 2007). The processing included explicit modeling of atmospheric gases and aerosols using the DISORT radiative code (Stamnes et al., 1988; Wolff et al., 2009) adapted for Mars conditions and the Hapke photometric function (Hapke, 2012) as a surface boundary condition (Arvidson et al., 2006). SSA spectral cubes were regularized, and the best estimate of surface spectral properties were retrieved in the presence of Poisson noise using a log maximum likelihood approach (MLM) and projected at 12 m/pixel (Kreisch et al., 2017; He et al., 2019). Comparison of the sensor space view shown in Figure 2 with the map projected view shown in Figure 3 demonstrates the effectiveness of this processing pipeline for our two sand deposit study sites (MDI and SoF). Both the S and L 12 m/pixel processed data cubes for FRT00021C92 were registered to the 25 cm/pixel HiRISE mosaic. The separate S and L data cubes were also joined to retrieve full VNIR spectra from 0.50 to 2.60 µm.

CRISM SSA spectra are a representation of a mixture of minerals and amorphous phases at varying abundances and grain sizes. Two techniques were used to infer sand mineralogy, spectrally speaking. First, the Sequential Maximum Angle Convex Cone (SMACC) endmember model (Gruninger et al., 2004) was used to delineate endmember spectra for regions of interest (ROIs) for both sand deposits. The procedure was used to choose endmember spectra from the scene and then model each pixel within the sand deposit as a linear combination of endmembers. This approach helped to evaluate the spectral heterogeneity of the sands. Second, we retrieved mineral phase abundances and grain sizes using the Hapke (2012) model and laboratory spectra (Table 2). We employed a Bayesian implementation of the Hapke model using a Markov chain Monte Carlo (MCMC) algorithm as developed by Lapotre et al. (2017b). The MCMC algorithm produced a variety of best fit spectra over ranges of abundances and grain sizes, and thus we employed violin plots (Hintze & Nelson, 1998) to show the statistical distribution of fits for numerous runs. The runs were conducted on two representative CRISM spectral endmembers identified by the SMACC method, chosen to avoid bedrock regolith contributions to the ferrous silicates within the sand deposits and thereby making model retrievals more robust.

Table 2. Reference information for model endmembers.