Drilling Into Earth's Past

Experiment With Technologies to Identify the Mineral Composition of Earth's Crust and Mantle Pixel by Pixel.

Hackathon Findings

• tk

Hackathon Projects

• Drilling Earth's Past and its presentation

Background

Please note: Aspects of this challenge overlap with the Meteorite Mineral Mapping challenge. Teams working on either challenge may wish to collaborate or share ideas with one another as they pursue their solutions.

The study of the terrestrial geology of the Earth (our home in the solar system) is critical for improving our understanding of how geological processes on other planets might work. In particular, Mars has massive volcanoes and large regions covered by lava flows, and also has ice caps at both of its poles, and Jupiter's moon Europa may be experiencing subduction, where one tectonic plate slides beneath another.

The crust and upper mantle of a planet is referred to as its lithosphere, and the Earth's lithosphere has seven or eight major plates, and many minor plates. Interactions between the plates can cause earthquakes, volcanism, mountain formation, or oceanic trench formation along their faults (the boundaries between plates). At the points on Earth where subduction occurs, the portion of the crust moving beneath the other plate is always oceanic while the other may or may not be oceanic, and these areas experience a high rate of earthquakes and volcanism. Of course it would be difficult to study active subduction zones in the depths of an ocean or below actively volcanic areas - but this is where ophiolites come in! An ophiolite is a section of lithosphere which has been uplifted above sea level and exposed on top of other crustal rocks. Essentially, an ancient ocean floor is raised to the surface, and effectively gives us a window that views directly into Earth's past with geology that we can analyze.

Image of Semail Ophiolite taken from "Thermal and Transport Properties of Mafic and Ultramafic Rocks of Oman Ophiolite", Nov 2016, S. Arafin, R. N. Singh - https://www.researchgate.net/publication/316707699/

Drilling the Semail Ophiolite

Located in Oman, the Semail Ophiolite covers an area of about 100,000 km², and began to form somewhere around 96 million years ago during the Late Cretaceous period. This was ancient oceanic floor, in a time when plesiosaurs were still roaming the oceans! This geological wonder is one of the best places on Earth to get a better handle on what happens when portions of the crust interact with ocean, as it allows us to access complete cross-sections of sea floor.

The Oman Drilling Project is an enormous international scientific research project comprised of 166 scientists representing 30 countries from 5 continents. Among the goals of the Oman DP are resolving long-standing questions regarding formation of oceanic lithosphere at mid-ocean ridges, hydrothermal alteration of the sea floor and subsequent mass transfer between the crust and the oceans and recycling of volatile elements in subduction zones. To answer these questions, the project has undertaken a drilling program that has drilled over 5.4 km worth of core samples, with 3.2 km of cores recovered completely intact. These intact cores were split down their centers (longitudinally), and every bit of it was surface-scanned with infrared spectroscopy at a resolution of 250 microns (micrometers).

Animated image showing ophiolite formation above adapted from a graphic in [Searle, M & Waters, David & Garber, J & Rioux, M & Cherry, A.G. & Ambrose, Tyler. (2015). Structure and metamorphism beneath the obducting Oman ophiolite: Evidence from the Bani Hamid granulites, northern Oman mountains. Geosphere. 11. 10.1130/GES01199.1.]

---

Solutions

We have standard spectra of known minerals, and we want associate each pixel in a scan with a mineral identity. We know that some minerals cannot occur with other minerals - groups of minerals will likely appear in the same small region of core, to the exclusion of other minerals. We also have manual analysis of small regions of IR maps. We are wondering if it is possible to use computer vision and machine learning driven image analysis to detect interesting features from the billions of scans of IR spectra. There are 31 terabytes of raw data! Can these technologies be applied to all 4000 sections of the intact 3.2 km core?

These are some of the types of things we hope can be achieved:

1. Automatic detection of veins. Veins are quasi-linear features where new minerals fill fractures within rocks. By analyzing how veins cut through each other, we can see how temperature, pressure, fluid volume, and fluid composition change through time.

Example of a portion of a core with many veins outlined. The background image is a false color infrared image.

2. Identifying pure minerals and rare phases. Dozens of minerals are present in these cores. We have spectra from libraries of many of the pure mineral components, but rocks generally have multiple minerals present at a microscopic scale. We are seeking better solutions for identifying the minerals present in each pixel, especially when features in the reflectance spectra due to absorption of light at particular wavelengths overlap with spectral features of other minerals that may also be present. In addition, there are minerals within the cores that are not in our libraries, but these are important to find and keep track of. The results will help us understand how the rocks have changed through time as they have reacted with fluids of different chemical compositions in different environments.

Reflectance spectrum of the mineral calcite from the USGS spectral library and an unknown reflectance spectrum. Calcite is a good match for this unknown, based on the similarity in the shapes and positions of dips in the spectra at 2000+ nm. (Minor differences in the strength and shape of features near 1400 and 1900 nm is due to differences in the amount of water trapped within the mineral.)

Reflectance spectrum of the mineral prehnite from the CRISM spectral library and an unknown reflectance spectrum. Some of the features and dips within the spectra are at similar wavelength positions to prehnite, but the overall slope is different and there are some differences between the unknown and prehnite spectra. The unknown spectrum likely includes some prehnite as well as one or more other minerals, possibly chlorite.

Some possible solutions might be:

• A trainable computer vision system. A system that uses spectral libraries with pure minerals and results of other analyses (for example, x-ray diffraction) to map minerals

• Machine learning to identify and outline veins.

• Advanced feature fitting algorithms. Match absorption features (dips) in reflectance spectra with libraries in a more robust way than existing algorithms.

• Clustering algorithms. Can we cluster similar pixels to reduce processing time, target other laboratory analyses, and limit the number of pixels that we need to manually analyze?

• A crowdsourcing system. If machine learning isn't an option, can we use crowdsourcing to help detect the features we are looking for?

• Anything else you can think of!

---

Resources

Be sure to check the Online Resources and Data Sets page to see if there might be any general purpose code or utilities you might use. What other computer vision resources are you familiar with that you could utilize? We'd love to learn about them!

• Oman Drilling Project Website

• QGIS: for opening IMG files.

• Spectroscopy background - USGS chapter

Datasets

SPECIAL NOTE ABOUT DATA: Most data for this challenge is located on a hard drive on site at the hackathon. This data is considered unpublished and must not be taken offsite or remain on outside systems after the completion of the hackathon, per the agreement all participants sign and in line with the Code of Conduct for this event.

The data types are 3D image cubes, .img files with associated headers (.hdr). The library spectra will be either .sli files (with headers) or ASCII. The headers are created by ENVI/IDL with whatever information we put in. Data are stored in BIL format (Band Interleaved by Line). To date IDL has been used to process the images, but a switch to Python is in process; experiments with the Spectral Python (SPy) tools have been shown to be useful, but we are open to any solution.

• Google Drive link for some files: This link contains files related to this challenge that are not posted on GitHub, including documentation for all data files.

---

Challenge owner: Rebecca Greenberger