IDENTIFYING GRAIN POPULATION DEPOSITIONAL ENVIRONMENT WITH THE AID OF MACHINE LEARNING (LINEAR DISCRIMINANT ANALYSIS) IN THE ENIGMATIC LOYALHANNA LIMESTONE (MISSISSIPPIAN) OF WESTERN PENNSYLVANIA

The Loyalhanna Limestone (Mississippian) is a lithologically and economically distinct unit in the Central Appalachian Plateau of Pennsylvania. The member, around 20 meters thick, forms the base of the Mauch Chunk Formation. It is widely quarried in the Allegheny Mountain Section as a road aggregate due to its excellent durability and skid resistance stemming from the significant presence of arenitic quartz clasts, which reside in a sparry calcite matrix. The unit is characterized by large, dipping foresets with set heights up to 4 meters. Decades-long contention exists regarding whether the mesoform/dune-scale foresets were deposited as aeolian deposits or in a high energy marine environment (or a combination thereof).

Here, sites were sampled in southern Somerset and Cambria Counties where marine- and aeolian-consistent features have been variously described. Eight samples were thin-sectioned, scaled, and magnified. Randomized areas within each magnified field were randomly chosen from each sample and grain axes measured with calipers. Mean (in mm) and logarithmic-graphical measures (Folk & Ward) of sorting, skewness and kurtosis were calculated from 24 sized regions. Corresponding data from known aeolian and subaqueous dunes were obtained from literature to train the linear discriminant analysis (LDA) model using the “MDA” package in R. LDA is a supervised machine learning method used to delineate groups. Skewness and kurtosis data computed via the same method limited usable subaqeous data (31 samples vs 432 for aeolian dunes).

LDA reported all but one of the 24 Loyalhanna samples as belonging to the subaqueous depositional class (posterior probablity > 99%). Kurtosis had a particularly strong weight on the overall discriminant analysis. The one sampling region classed as aeolian was taken from a larger sample in which two other regions were classed as subaqueous. LDA rarely classifies 100% of groups correctly and the lone subaqeous return exhibited the lowest kurtosis value and second greatest sorting value in the Loyalhanna dataset. LDA results are consistent with Brezinski and Kollar (2021), who report subaqueous features (e.g. algal borings and marine macro- and microinvertebrates), and the authors’ own observations of a likely groove cast and channel-like feature in a quarry highwall.