A SEQUENTIAL APPROACH FOR CLASSIFICATION OF PAST AND NEW MULTISPECTRAL MEASUREMENTS BY THE CURIOSITY ROVER’S MASTCAM

The multispectral data acquired by the Mars Science Laboratory Curiosity rover’s Mastcam represents an impressive record of the visible to near infrared reflectance (440 to 1015 nm) of materials on the floor of Gale crater and on the slopes of its central mound, Mt. Sharp. Categorizing these multispectral measurements, and placing new ones into the context of materials already encountered, is a challenge. This problem has been addressed by several researchers (Kerner et al., 2019, 2020; Rice et al., 2022) and a new approach is described here. Given a collection of 12 spectral band Mastcam spectra, three endmember detection approaches are applied: n-Findr, vertex component analysis (VCA), and automatic target generation process (ATGP). These three methods provide sets of endmember spectra that overlap between methods, but one method often detects an endmember missed by the others. These methods are run on the 12 band spectra, to provide a candidate set of endmembers. These endmember spectra are compared against the results from a classification carried out with agglomerative hierarchical cluster analysis (AHCA) run against a smaller number of spectral parameters. The result of the AHCA is that an agglomerated set of “branches” from the hierarchical tree represents collections of spectra, and representatives from these branches are examined to find overlap with the endmembers from the endmember detection step. Representative spectra, or in practice, their associated spectral parameters, are used as training sets for the final step of applying an ensemble of machine learning approaches contained in the MATLAB classification learner app. Different machine learning methodologies can be tried and results examined to see which provides the highest classification accuracy. In trials to date a cubic support vector machine (SVM) classifier provides the best results with 90% classification accuracy. Once trained, this classifier is applied against new measurements. Results will be shown from the clay-sulfate transition zone on the slopes of Mt. Sharp that Curiosity is currently traversing over. Relevance to past measurements and to Curiosity’s characterization of the history of aqueous activity within Gale crater will be addressed.