Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Noise2Noise Denoising of CRISM Hyperspectral Data

Robert Platt, Rossella Arcucci, Cédric M. John

TL;DR

The paper tackles denoising of CRISM hyperspectral data in the absence of clean ground-truth spectra. It proposes Noise2Noise4Mars (N2N4M), a self-supervised 1-D U-Net that denoises spectral signals using noisy targets, avoiding the need for noise-free references. Experiments on synthetic and real CRISM data show N2N4M achieves lower reconstruction error and improves downstream mineral classification metrics compared with benchmark methods, while preserving diagnostic absorption features. The findings support more reliable Martian mineral mapping and site analysis, with potential applicability to other domains lacking pristine ground truth.

Abstract

Hyperspectral data acquired by the Compact Reconnaissance Imaging Spectrometer for Mars (CRISM) have allowed for unparalleled mapping of the surface mineralogy of Mars. Due to sensor degradation over time, a significant portion of the recently acquired data is considered unusable. Here a new data-driven model architecture, Noise2Noise4Mars (N2N4M), is introduced to remove noise from CRISM images. Our model is self-supervised and does not require zero-noise target data, making it well suited for use in Planetary Science applications where high quality labelled data is scarce. We demonstrate its strong performance on synthetic-noise data and CRISM images, and its impact on downstream classification performance, outperforming benchmark methods on most metrics. This allows for detailed analysis for critical sites of interest on the Martian surface, including proposed lander sites.

Noise2Noise Denoising of CRISM Hyperspectral Data

TL;DR

The paper tackles denoising of CRISM hyperspectral data in the absence of clean ground-truth spectra. It proposes Noise2Noise4Mars (N2N4M), a self-supervised 1-D U-Net that denoises spectral signals using noisy targets, avoiding the need for noise-free references. Experiments on synthetic and real CRISM data show N2N4M achieves lower reconstruction error and improves downstream mineral classification metrics compared with benchmark methods, while preserving diagnostic absorption features. The findings support more reliable Martian mineral mapping and site analysis, with potential applicability to other domains lacking pristine ground truth.

Abstract

Hyperspectral data acquired by the Compact Reconnaissance Imaging Spectrometer for Mars (CRISM) have allowed for unparalleled mapping of the surface mineralogy of Mars. Due to sensor degradation over time, a significant portion of the recently acquired data is considered unusable. Here a new data-driven model architecture, Noise2Noise4Mars (N2N4M), is introduced to remove noise from CRISM images. Our model is self-supervised and does not require zero-noise target data, making it well suited for use in Planetary Science applications where high quality labelled data is scarce. We demonstrate its strong performance on synthetic-noise data and CRISM images, and its impact on downstream classification performance, outperforming benchmark methods on most metrics. This allows for detailed analysis for critical sites of interest on the Martian surface, including proposed lander sites.
Paper Structure (9 sections, 4 figures, 1 table)

This paper contains 9 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Works
  3. Methods
  4. Preprocessing
  5. Model and Training Scheme
  6. Results
  7. Discussion and Future Works
  8. Conclusion
  9. Appendix

Figures (4)

  • Figure 1: Comparison of a low-noise and high-noise pixel spectra of the same location extracted from two CRISM images (FRT0000A053 and FRS000364CA) acquired 7 years apart. Absorption features to identify the surface material as an Mg/Fe-Smectite are highlighted. Shown in green is the low-noise spectra with synthetic Gaussian noise added.
  • Figure 2: Qualitative comparsion of results on unseen spectra from two different mineral classes. Shown are the high-noise input, low-noise target, and high-noise data denoised by CoTCAT, and N2N4M (ours). Absorption features for mineral identification are highlighted.
  • Figure 3: Two pairs of test images, from top to bottom: low-noise reference image (LN), high-noise image (HN), denoised image (DN). Images shown as the strength of absorption features (listed at base) in each pixel, which highlight outcrops of hydrated minerals identified in the reference image.
  • Figure 4: Noise2Noise 4 Mars (N2N4M) Network Architecture. All convolutional kernels were size 5.