Mineral Identification from spectral data is performed based on the presence/strength of specific absorption features at designated spectral windows. Since a large portion of the spectral data only contains background information which is not used for diagnostic analysis, standard metrics are unable to accurately measure similarity. We are currently developing an unsupervised feature learning technique based on Generative Adversarial Networks (GANs) that can "measure" spectral similarity. Even simple similarity metrics such as SAD do a far better job of assessing spectral similarity in this feature space. We will leverage this feature space representation for mineral (spectrum) mapping in CRISM images.

We are developing a new method to enhance the quality of the reflectance signals acquired by Compact Reconnaissance Imaging Spectrometer for Mars (CRISM) to help scientists identify and map minerals on Mars. Our focus covers not only denoising, but also improving atmospheric compensation and calibration.