Page 4 - i1052-5173-32-9

A High-Resolution Multispectral Macro-

Imager for Geology and Paleontology

Ryan A. Manzuk, Dept. of Geosciences, Princeton University, Princeton, New Jersey 08544, USA, rmanzuk@princeton.edu; Devdigvijay
Singh, Dept. of Geosciences, Princeton University, Princeton, New Jersey 08544, USA; Akshay Mehra, Dept. of Geosciences, Princeton
University, Princeton, New Jersey 08544, USA, and Dept. of Earth Sciences, Dartmouth College, Hanover, New Hampshire 03755, USA;
Emily C. Geyman, Dept. of Geosciences, Princeton University, Princeton, New Jersey 08544, USA, and Division of Geological and
Planetary Sciences, California Institute of Technology, Pasadena, California 91125, USA; Stacey Edmonsond, Dept. of Geosciences,
Princeton University, Princeton, New Jersey 08544, USA, and School of Earth and Ocean Sciences, University of Victoria, Victoria,
British Columbia V8W 2Y2, Canada; Adam C. Maloof, Dept. of Geosciences, Princeton University, Princeton, New Jersey 08544, USA

ABSTRACT morphic history. At points on a map or beds are significant (Solomon, 1963; Neilson and
Accurately assessing the shape, size, and in a stratigraphic section, lithofacies obser- Brockman, 1977), and the small fields of
modality of features in rock samples is a vations from field campaigns form the view available in most microscopes limit
longstanding problem in geology. Recent backbone of geologic study. Throughout the scale of features studied to those only a
advances in machine learning have intro- recent decades, the rise of geochemical few millimeters in size (Higgins, 2000). To
duced the possibility of performing these techniques has increased the value of sam- build on previous petrographic findings and
tasks through automated image analysis. To ples brought back from the field. For exam- contextualize geochemical data, we can
leverage these methods for geological and ple, many measured sections through car- develop techniques to quantify lithofacies
paleontological applications, we first need a bonate stratigraphies now include bed-by-bed over a broader range of feature sizes and
way to acquire high-resolution images of isotope and trace element measurements with more continuous spatial sampling.
polished slabs and thin sections with a field that give insights into local carbon cycling
of view large enough to fit samples con- (Ahm et al., 2021), global marine redox New Potential for Petrographic Data
taining crystals, fossils, bedforms, etc. We state (Dahl et al., 2019), sediment diagene- through Image Analysis
describe a new multispectral setup that can sis (Ahm et al., 2018), and correlations Geologists could outline, count, and mea-
acquire images at ~3.76 mm per pixel spa- within (Hay et al., 2019) and between basins sure all the fossils, grains, or crystals in sam-
tial resolution over a 21 cm field of view, (Halverson et al., 2005; Maloof et al., 2010). ples to extract these data, but manual petro-
2
equipped with 8-band (470–940 nm) spectral However, reliable interpretations of these graphic study is too time-consuming to
resolution, plus a band for ultraviolet (365 geochemical data benefit from knowledge accompany each of the hundreds or thou-
nm) fluorescence. Additionally, we present a of the physical properties of the rock sam- sands of geochemical measurements and
5-band (470–940 nm) light table with auto- ples, such as grain/crystal sizes and modal- observations made on a single map. We can,
mated rotating polarizers, which allows use ities (Geyman and Maloof, 2021), primary however, turn to recent advances in machine
of the camera as a high-throughput transmit- mineralogy, porosity/permeability, and learning that have introduced the possibility
ted light thin section imager. The use of color cross-cutting relationships between fabrics of training models to recognize rock fea-
bands outside the visible spectrum, as well as (Bergmann et al., 2011; Hood et al., 2016; tures (Yesiloglu-Gultekin et al., 2012;
the registration of multiple cross-polarized Corsetti et al., 2006; Dyer et al., 2017)— Koeshidayatullah et al., 2020). The need for
rotations, encode rock properties that data that also serve to refine analyses of automated feature classification is familiar
heighten image contrast and improve the sedimentary environment (Geyman et al., to many fields, and effective solutions now
accuracy of machine learning models. Our 2021). The above examples come from sedi- are being realized in industries such as
setup and methods provide an efficient way mentary geology, but the need to match autonomous vehicles (Tian et al., 2018) and
to (1) build reproducible image archives of geochemical data to quantitative lithofacies biomedical image processing (Li et al., 2018).
rock specimens to complement field obser- also applies to interpretations of igneous A variety of machine learning models
vations, (2) classify and segment those and metamorphic conditions (Higgins, 2000). can be trained to perform these tasks, but
images, and (3) quantitatively compare litho- Workers have developed methods to they all learn to classify image features
facies and fossil assemblages. approximate rock contents from samples, through repeated practice on example images
often by point counting on the stage of a manually labeled by humans (LeCun et al.,
INTRODUCTION microscope (Shand, 1916). Although this 1989), and some of the most effective mod-
Geologists have developed an eye for the technique has brought about many geologi- els for general applications require more
physical rock characteristics that encode cal insights, the uncertainties that stem than 300,000 traced examples (Lin et al.,
Earth’s sedimentary, igneous, and meta- from incompletely sampling a rock’s surface 2014; He et al., 2017). Prior to training an

GSA Today, v. 32, https://doi.org/10.1130/GSATG533A.1. CC-BY-NC.

4 GSA TODAY | September 2022