Testing Multiple PCA Techniques with Median Normalization to Analyze NIR Spectral Images of the Martian Surface

Dominic Payne


Our ultimate goal in this research is to better understand the ice clouds of the Martian atmosphere using near-infrared spectral images.  To do this, we need a good, a priori spectral model of the surface.  To get that, we use a process known as principal component analysis (PCA); previous work has shown that the PCA eigenvectors can separate the surface and ice components of the spectra.  Some of our data, however, could not be calibrated and thus it was not clear if PCA would have the same effect.  To see if we can still get some use out of this data, we analyzed it across three different types of normalization: data mean, data median, and spot-spectra.  Collaborator results appear to indicate that the most useful method of normalization is the spot-spectra method, while data mean normalization was the least useful.  Even so, we were reluctant to completely disregard the median normalized data as the spot-spectra method requires a significantly bright region with no cloud coverage, and that is not always present in the images.  Using only median normalized data, we performed four different PCA tests: using the full 1.5–4.1 μm wavelength spectrum and the full image; using the full spectrum and disregarding the Martian poles; omitting the 1.9–2.2 µm wavelengths from the spectrum and using the full image; omitting the 1.9–2.2 µm wavelengths form the spectrum and disregarding the Martian poles. The reason for omitting the 1.9–2.2 µm spectral region is that there is a major spectral absorption due to gaseous CO2 abundant in the Martian atmosphere—a non-linear process that could perhaps affect the linear PCA technique. The reason for omitting the poles is that they have a permanent cap of ground ice, which can confuse the modeling technique that assumes all ice spectral features are due to ice clouds.  We are optimistic that these results will be complimentary to those from spot-spectra normalization, but without the limitations that spot normalizing presents.  With this secondary method of normalization, we can potentially analyze trends from the data that were either not present or not obvious through PCA on spot normalization.  This will allow us to get meaningful data on not only days with less than ideal observing conditions, but also those when Mars is especially cloud covered.


PCA, Mars, Normalization,

Full Text: PDF


  • There are currently no refbacks.