Machine Learning Prototype for Spectral Classification and Regional Mapping of Thermal Emission Spectrometer (TES) Data

A proof of concept machine learning spectral analysis system has been developed utilizing a Mars Global Surveyor (MGS) Thermal Emission Spectrometer (TES) dataset. Early results are presented. Atmospherically corrected surface emissivity spectra (covering 180° W - 180° E; 70° S - 70° N) were analyzed. All spectra were averaged in 4 pixels / deg. resolution bins to reduce noise [Rogers et al. 2007]. This resulted in > 56,000 spectra from low-dust regions across the Martian surface in 73 surface-sensing spectral bands (~232 - 1300 cm-1). Spectra were imported to the Matlab ® scientific computing environment and augmented with geolocation data. We developed software codes for unsupervised clustering (k-Means) [MacQueen 1967] of the entire global emissivity dataset. Unsupervised clustering methods enable discovery-level science since codes include no a priori knowledge of spectral features. We employed unsupervised clustering in our system to simulate searches for spectral features on bodies that have not been well-explored and/or for instances without obvious spectral features, but where clustering may reveal subtle patterns.

Spectra were treated statistically, as N-dimensional vectors, and clustered based on the usual Euclidean vector norm [Duda et al. 2001; Richards 1993]. Cluster number is user-defined with an optimal number determined, as in previous analyses [Marzo et al. 2006, 2008], by examining the Calinski-Harabasz (C-H) metric [Calinski-Harabasz 1974]. Based on this metric, our data tended to group in either 2 primary or between 5 and 7 clusters, depending on the geographic clustering area. Due to the statistical averaging inherent in the clustering process, global clustering showed ubiquitous Surface Type 1 but did not allow discrimination of spatially limited spectra that are of potentially high scientific interest. Regional clustering revealed the presence of olivine and Surface Type 2 (Fig.). Hematite was also detected in the Sinus Meridiani region, as in Christensen et al. [2000]. We also developed a generalizable (quasi-supervised) method to scan for hematite and other minerals. Our system and methods are extensible to other spectral datasets, covering various spectral bands, allowing machine learning based analysis of data from a variety of current and future planetary missions.