Probabilistic Super Resolution for Mineral Spectroscopy

Alberto Candela, David R. Thompson, David Wettergreen, Kerry Cawse-Nicholson, Sven Geier, Michael L. Eastwood, Robert O. GreenAAAI 2020 · DOI

Introduction

In planetary science and geology, identifying surface materials relies heavily on spectroscopy—the study of how materials reflect or absorb light across different wavelengths. Every mineral possesses a unique "spectral signature" or fingerprint. However, researchers face a persistent trade-off: high-resolution instruments like AVIRIS-NG provide detailed signatures but cover limited areas, while low-resolution instruments like ASTER offer global coverage but produce "blurry" spectra that obscure critical absorption features.

Comparison of high-resolution spectral signatures (blue) with low-resolution ASTER bands (red) for kaolinite and muscovite.

To bridge this gap, researchers have developed the Deep Conditional Gaussian Model (DCGM), a neural network designed to reconstruct high-resolution mineral fingerprints from low-resolution satellite data using probabilistic super-resolution. Unlike typical "black box" AI, this model provides a full probabilistic map of its predictions, allowing scientists to see exactly how different chemical features correlate and where the model is uncertain.

Methodology: The Deep Conditional Gaussian Model

Unlike traditional deep learning models that provide a single deterministic output, the DCGM is built for scientific transparency. It treats super-resolution as a probabilistic inference problem, providing a mathematically tractable way to evaluate how likely a specific mineral identification is.

Probabilistic Architecture

Instead of predicting a single high-resolution spectrum, the network predicts a multivariate Gaussian distribution. For every low-resolution input, the model outputs both a mean ($\mu_y$)—representing the best guess—and a covariance matrix ($\Sigma_y$)—representing uncertainty and the relationships between different wavelengths.

The Deep Conditional Gaussian Model (DCGM) architecture, highlighting the prediction of mean and precision matrix components.

The Cholesky Trick

Modeling a full covariance matrix for over 85 spectral channels is computationally intensive and prone to numerical instability. To address this, the authors utilized a Cholesky decomposition ($ \Sigma_y^{-1} = CC^T $). This approach ensures the model remains stable and efficient while capturing the complex correlations between different parts of the light spectrum. The model is trained using Maximum Likelihood Estimation (MLE), optimizing it to maximize the log-likelihood of the ground-truth high-resolution data rather than just minimizing visual error.

Key Insights and Results

The study utilized data from Cuprite, Nevada, a standard test site for remote sensing. By using ASTER data (5 channels) as input and AVIRIS-NG data (85 channels) as ground truth, the researchers demonstrated several key advantages of the DCGM:

  • Superior Accuracy: DCGM consistently achieved lower Root Mean Squared Error (RMSE) than baseline models like Gaussian Mixture Models (GMM) and standard Variational Encoders.
  • Feature Reconstruction: The model successfully reconstructed the distinct double-peak features of minerals like kaolinite and the sharp dips of alunite from ASTER data that originally showed almost no detail.
  • Geological Interpretability: A major breakthrough was the realization that learned covariance matrices are geologically meaningful. High covariance values often appear near known absorption features (such as the 2.2 μm dip for mica), indicating which spectral regions are most informative or ambiguous for specific mineral mixtures.
  • Uncertainty Quantification: The model correctly showed higher uncertainty in regions where mineral mixtures (like mica mixed with calcite) created ambiguous spectral signatures, providing a reliable metric for geological interpretation.

Conclusion and Future Directions

The DCGM represents a significant step toward making machine learning a reliable tool for Earth and planetary science. By prioritizing interpretability and statistical rigor over mere generative power, it provides geologists with a powerful method for mapping mineralogy over vast areas. Future work aims to expand the spectral range to include visible and thermal infrared data and to conduct formal user studies with expert geologists to evaluate how these super-resolved spectra affect real-world discovery and mapping.

Summary updated

BibTeX
@inproceedings{candela2020probabilistic,
  author = {Alberto Candela and David R. Thompson and David Wettergreen and Kerry Cawse-Nicholson and Sven Geier and Michael L. Eastwood and Robert O. Green},
  title = {{Probabilistic Super Resolution for Mineral Spectroscopy}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i08.7030},
}
Figure