Adversarial feature learning for improved mineral mapping of CRISM data

The Compact Reconnaissance Imaging Spectrometer for Mars (CRISM) has proven instrumental in the mineralogical analysis of the Martian surface. An essential tool for mineral identification for this dataset has been the CRISM summary parameters–which use simple mathematical functions to measure the presence/absence of specific spectral features. While the CRISM summary parameters and browse products (combinations of specific summary parameters) have proven valuable in guiding manual analysis, these hand-crafted representations are not well suited for automated analysis, as CRISM spectral artifacts and noise negatively affect their performance, making these parameters prone to false alarms. We propose an unsupervised technique based on Generative Adversarial Networks (GANs) to learn a more discriminative representation from CRISM data, such that simple metrics in the representation space are sufficient to discriminate between the various mineral spectra present in the data. We describe a simple pipeline using GAN based representations to map mineral signatures of interest across the CRISM image database. We show that the features learned by the GAN are better suited to discriminate mineral signatures in the CRISM database compared to the summary parameters and classical similarity metrics. Finally, we validate the technique over a subset of CRISM images over the Jezero crater and NE Syrtis regions of Mars. •We present a novel feature extraction technique for improved mineral discrimination in the CRISM image data.•The Generative Adeversarial Network based feature extraction technique, leverages a large amount of unlabeled data to learn a discriminative feature set capable of differentiating various mineral spectra.•We illustrate the improved discriminatory power of the feature space over the original space and classical metrics like summary parameters.•A quantitative baseline for improved discriminatory power of GAN based representation is shown on a CRISM-like simulated data•We will also describe an automated mineral pipeline for mineral mapping in CRISM image datasets.•The results of such mapping over CRISM images of the candidate landing sites for the 2020 rover are shown.