A Free and Open-Source Tool for Spectral Unit Mapping Through Clustering of NIR Hyperspectral Images on Mars

Abstract

Spectral and hyperspectral data provide essential information on the surface composition of rocky planetary bodies, through which we can reconstruct and understand their origin and geologic evolution. In this field, applications of advanced tools, such as machine learning-based algorithms, are steadily increasing. This study presents a newly developed, open-source, Python-based tool and workflow for the identification and mapping of spectral units with CRISM near-infrared (NIR) data using unsupervised machine-learning (clustering) techniques. Data pre-processing, dimensionality reduction, clustering, and visualization are integrated into an interactive user-friendly environment. The tool employs two different established clustering algorithms: K-means and Gaussian Mixture Models. Dimensionality reduction methods include the use of projection-based and manifold-based dimensionality reduction techniques, respectively Principal Component Analysis (PCA) and Uniform Manifold Approximation and Projection (UMAP). We showcase the abilities of the tool on a set of selected CRISM images. We identify multiple spectral units, from which we retrieve spectral signatures ranging from mafic minerals, with broad absorption bands, to Fe/Mg phyllosilicates and sulfates, which exhibit multiple narrow absorption features. Results highlight how clustering methods can support and enhance the findings derived from classical analysis of spectral parameters, for instance, by highlighting subtle spectral variations. Moreover, with minor tuning and modifications, the tool can be applied to different spectral ranges as well as to spectral or hyperspectral data from other planetary surfaces.