Artificial Intelligence to Classify Minerals and Rocks with Raman Spectra and Image Analysis
Abstract
Introduction: We are developing an Intelligent Data Understanding (IDU) system that uses neural networks to classify minerals and rocks. The IDU will impact geological surveys carried out by rovers or suited astronauts; which seek to detect biosignatures, mine resources, and analyze Lunar and planetary surfaces or remote and harsh environments on Earth.
Raman Spectra and IDU: We used a Multi-Layer Perceptron (MLP) and a novel type of convolutional neural network, called a "Module Neural Network" (MNN). Locally normalizing each spectrum, by setting the highest peak to 1, accounts for Raman scattering intensity variations. Taking the first derivative of intensity with respect to wavenumber accounts for base-line offset variations. Compacting the spectra with Principal Component Analysis (PCA) mitigates high-dimensionality problems. Results: The IDU discriminated nine major rock-forming minerals: quartz, potassium feldspar, plagioclase, mica, olivine, pyroxene, gypsum, carbonate, and halite. Previous methods classified the first six of those minerals using 852 nm Raman spectra with 83% accuracy and Infrared spectra with 80% accuracy [1]. Our newest methods used an MNN to combine information from two Raman excitation sources: 532 nm and 785 nm. With these spectra, the AI discriminated the nine minerals with 95% accuracy, outperforming MLP's and other tested networks which used a single excitation source. Image Analysis: Additionally, we applied Gabor filters to capture sinusoidal patterns seen in the texture and color of rocks. An MLP was able to use this information to classify igneous rock samples as Felsic, Mafic, or Intermediate with 91% accuracy. Acknowledgements: Funding was provided by the NASA Ames Internship program and the NASA Astrobiology Institute Program Grant # NNX15BB01. References: [1] Ishikawa S.T. and Gulick V.C. (2013) Computer & Geosciences 54: 259-268.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.P43E3507J
- Keywords:
-
- 1942 Machine learning;
- INFORMATICS;
- 6094 Instruments and techniques;
- PLANETARY SCIENCES: COMETS AND SMALL BODIES;
- 5794 Instruments and techniques;
- PLANETARY SCIENCES: FLUID PLANETS;
- 5494 Instruments and techniques;
- PLANETARY SCIENCES: SOLID SURFACE PLANETS