IDENTIFICATION OF LOESS FROM FLUVIAL DEPOSITS AT FLAGSTAFF RIM, WYOMING, BY MACHINE LEARNING

Grain-size analysis has played a key role in interpreting depositional environments. The traditional analytical methods of plotting and visual identification of various grain-size curves can be time-intensive calling for new analytical techniques. In the western United States, an abrupt transition from fluvial to eolian loess environments occurred during the Paleogene based on the change from a stratified, poorly sorted lithofacies association to massive, well-sorted siltstone or very fine-grained sandstone lithofacies. However, loess deposition may have initiated during the fluvial deposition, thus earlier than the observed lithofacies changes. This study applies unsupervised machine learning methods including K-Means and Spectral clustering, and dimensionality reduction methods, including principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE), to a high feature grain size dataset collected in the Flagstaff Rim area of Wyoming. Our results show unsupervised learning techniques can effectively classify grain-size data into similar groups. By comparing the groups identified by machine learning with those based on field observation and geological interpretation, we found that machine learning methods can assist the interpretation of depositional environments and identify the temporal trend of transport processes. Our results also show the presence of eolian loess transport prior to lithofacies change, suggesting a gradual transition from fluvial to eolian deposition. Our study highlights the effectiveness of applying machine-learning techniques for large grain-size datasets in the interpretation of sedimentary environments and shows the trend for future sedimentological research.