First results of Unsupervised Learning techniques applied to CRISM dataset on Mars

Kurzfassung

Spectral and hyperspectral data from remote sensing instruments provide essential information on the composition of planetary surfaces. On Mars, high resolution hyperspectral data are provided by the CRISM instrument, onboard NASA’s MRO spacecraft. CRISM collects hyperspectral cubes in the 0.4-4 micron range, with a spectral sampling of 6.55 nm/channel and a spatial resolution up to 18.4 meter/pixel. A CRISM scene is traditionally explored through RGB maps of spectral parameters, such as band depth. To guide the user in this work, the CRISM team provided a set of 60 standard spectral parameters, identified based on the known spectral variability of the planet. After a first assessment with this method, extraction of single or mean spectra from selected ROIs (regions of interest) is usually performed. This is a solid approach, however, as it focuses on a few portions of the available spectral range at once, it does not fully exploit the potentials of a hyperspectral dataset. Machine Learning techniques can help us explore CRISM data more efficiently. Here we present the results from the development of a Python framework that allows the application of two different Unsupervised Learning techniques (k-Means and Gaussian Mixture Models, GMMs).