GRSG 34th Conference 2023

Title: Denoising Hyperspectral Images of Mars with Machine Learning

Author: Robert Platt

Abstract:

The Compact Reconnaissance Imaging Spectrometer for Mars (CRISM) has produced the most detailed and comprehensive dataset to date regarding the mineralogical composition of Mars’ surface. Throughout its operational lifespan, CRISM has unveiled the presence of numerous minerals, marking their first-time identification on Mars (1). However, as the operational duration of CRISM extended, a noticeable decline in the signal-to-noise ratio (SNR) of the images ensued.

This diminishing SNR is attributed to successive failures of the cryocooling systems onboard the instrument (2). Consequently, discerning minerals with highly similar spectral signatures has become progressively more challenging in recent observations made by CRISM. To address this issue, we propose a novel signal denoising technique inspired by the Noise2Noise image reconstruction methodology rooted in Deep Learning (3). Our approach leveraged a dataset sourced from a series of meticulously characterized CRISM images and pixels possessing high SNR (4). Synthetic noise is then introduced to this dataset, using a Gaussian distribution with an additional randomly scaled uniform factor.

The model’s performance is quantitively assessed using synthetic as well as real data, extending the evaluation to encompass downstream classification tasks. Comparative analysis against alternative CRISM signal reconstruction techniques substantiates that our proposed model excels in the accurate reconstruction of spectra. This is especially notable in the case of narrow absorption features that are prevalent in hydrated mineral groups and phyllosilicates, resulting in a 30% increase in downstream classification performance compared to prior methods. Our model allows for analysis and confident mineral identification in CRISM observations which have previously been unusable.

This opens avenues for characterization of mineralogy in as yet unstudied areas of the Martian surface. 1. Viviano CE, Seelos FP, Murchie SL, Kahn EG, Seelos KD, Taylor HW, et al. Revised CRISM spectral parameters and summary products based on the currently detected mineral diversity on Mars. J Geophys Res Planets. 2014;119(6):1403–31. 2. Bultel B, Quantin C, Lozac’h L. Description of CoTCAT (Complement to CRISM Analysis Toolkit). IEEE J Sel Top Appl Earth Obs Remote Sens. 2015 Jun;8(6):3039–49. 3. Lehtinen J, Munkberg J, Hasselgren J, Laine S, Karras T, Aittala M, et al. Noise2Noise: Learning Image Restoration without Clean Data. arXiv; 2018. Available from: http://arxiv.org/abs/1803.04189

The Compact Reconnaissance Imaging Spectrometer for Mars (CRISM) has produced the most detailed and comprehensive dataset to date regarding the mineralogical composition of Mars’ surface. Throughout its operational lifespan, CRISM has unveiled the presence of numerous minerals, marking their first-time identification on Mars (1). However, as the operational duration of CRISM extended, a noticeable decline in the signal-to-noise ratio (SNR) of the images ensued. This diminishing SNR is attributed to successive failures of the cryocooling systems onboard the instrument (2). Consequently, discerning minerals with highly similar spectral signatures has become progressively more challenging in recent observations made by CRISM. To address this issue, we propose a novel signal denoising technique inspired by the Noise2Noise image reconstruction methodology rooted in Deep Learning (3). Our approach leveraged a dataset sourced from a series of meticulously characterized CRISM images and pixels possessing high SNR (4).

Synthetic noise is then introduced to this dataset, using a Gaussian distribution with an additional randomly scaled uniform factor. The model’s performance is quantitively assessed using synthetic as well as real data, extending the evaluation to encompass downstream classification tasks. Comparative analysis against alternative CRISM signal reconstruction techniques substantiates that our proposed model excels in the accurate reconstruction of spectra. This is especially notable in the case of narrow absorption features that are prevalent in hydrated mineral groups and phyllosilicates, resulting in a 30% increase in downstream classification performance compared to prior methods.

Our model allows for analysis and confident mineral identification in CRISM observations which have previously been unusable. This opens avenues for characterization of mineralogy in as yet unstudied areas of the Martian surface. 1. Viviano CE, Seelos FP, Murchie SL, Kahn EG, Seelos KD, Taylor HW, et al. Revised CRISM spectral parameters and summary products based on the currently detected mineral diversity on Mars. J Geophys Res Planets. 2014;119(6):1403–31. 2. Bultel B, Quantin C, Lozac’h L. Description of CoTCAT (Complement to CRISM Analysis Toolkit). IEEE J Sel Top Appl Earth Obs Remote Sens. 2015 Jun;8(6):3039–49. 3. Lehtinen J, Munkberg J, Hasselgren J, Laine S, Karras T, Aittala M, et al. Noise2Noise: Learning Image Restoration without Clean Data. arXiv; 2018. Available from: http://arxiv.org/abs/1803.04189