Interpretation of remote-sensed and prospective in-situ spectroscopic data from surfaces of airless bodies relies on a combination of modeling and empirical fitting routines to produce estimates of mineralogy, rock types, and space weathering. Current methods for these analyses tend to be computationally cumbersome, sensitive to noise and other factors, and poorly constrained with confidence limits. Machine learning methods provide a more accurate alternative to physical mixing models such as the Modified Gaussian Model for problems of unmixing spectral signatures of mixtures of minerals. "Whole-spectrum matching" is a technique that enables large-scale analyses of spectra and avoids costly and error-prone peak-fitting and dimension reduction steps. It exploits all available data to compute similarity scores between spectra of unknowns and those in reference libraries. In this project, we demonstrate an application of whole spectrum matching to infrared and Raman spectroscopy of mineral mixtures. This method has applicability to a wide variety of spectroscopic applications where quantitative estimates are needed.
This method expands on the state of the art performance achieved by whole spectrum matching for the identification of pure mineral phases, extending the technique to account for mixtures of unknown minerals without a priori knowledge of a mixing ratio. As with pure mineral identification, the process begins with simple baseline correction and normalization, both for the query spectrum and a target library containing a superset of the pure minerals of interest. This library need not be collected under similar instrumental conditions to the query spectrum, because the whole spectrum match is insensitive to many of types of spectral variation caused by heterogeneous acquisition parameters, such as varying laser power. After the query and targets are preprocessed, several rounds of whole-spectrum matching are performed, ranking combinations of pure mineral spectra according to their similarity to the query.
Unlike existing methods for mixed mineral identification, this whole-spectrum approach does not explicitly model any peaks. This greatly simplifies the procedure, which requires no human expert intervention. The lack of peak modeling also improves the robustness of the matching performance, especially with respect to noisy spectra. Another benefit of the proposed method is its flexibility of application to spectra sampled over variable energy ranges and resolutions. This advantage is especially necessary when searching for mixture matches with an in-situ spectrum against a database of lab-quality spectra. Whole-spectrum matching also considers the relative intensities of all characteristic peaks, allowing all peaks in a spectrum to be matched rather than using only the most intense. Finally, this method does not make any assumptions about the number of pure minerals that make up the unknown query spectrum, and may thus be used to identify arbitrary mixtures.