Automated Analysis of Mars Multispectral Observations

Multispectral and hyperspectral imagers are now commonly used to obtain remote sensing measurements for the study of Mars, and many more such measurements are planned for the future. These techniques present a number of data collection, processing, and analysis challenges for planetary scientists. For example, CRISM, the spectrometer that will fly on Mars Reconnaissance Orbiter in 2005, is expected to return about 9 terabytes of data over the mission duration. Each mul-tispectral map will be 5120x5120 pixels in size (25 MB). It is not obvious how to easily browse this data, much less perform detailed analyses of it. More data means more information and the opportunity for new insights about Mars, but it carries with it a heavier and heavier burden for the analysis process. Conventional methods for analyzing multispectral data include techniques such as examining absorption band depths at specific wavelengths or plotting two-dimensional histograms of radiance at different wavelengths. Selecting wavelengths and band depths that will yield the most compositional or min-eralogical insight requires a significant amount of expertise about the object being observed. More critically, the process can be very time-consuming, with each histogram providing a single two-dimensional slice of the data for interpretation. For spectral data with only 10 wavelengths, there are 900 such histograms. Instruments such as the HST Space Telescope Imaging Spectrograph (STIS) have observed Mars at 1024 wavelengths; there are 1,047,552 corresponding possible histograms. This is not an upper limit, however; there are several other histograms that can provide insights, such as plotting radiance at one wavelength against the radiance ratio at two other wavelengths, or plotting radiance against a band depth or slope feature. Of course, not every pair of features will yield interesting results when plotted against each other. Often areas of interest are already known, such as the 900 nm band depth or the radiance at 440 or 750 nm for Mars observations. However, our knowledge of Mars is certainly not comprehensive. The ability to discover additional informative relationships is critical. In this abstract, we describe the result of applying an automated clustering algorithm to two data sets composed of Mars observations. One data set was collected by STIS on the Hubble Space Telescope; the other was obtained by the Mars Pathfinder Lander. We find that the results are comprehensible and, when a manual analysis is available for comparison, there is a good amount of agreement between the two sets …