UNVEILING LINKS BETWEEN GRAIN-SIZE PARAMETERS AND DEPOSITIONAL ENVIRONMENTS AND PROCESSES: INITIAL RESULTS EXPERIMENTING WITH MULTIVARIATE STATISTICAL AND ARTIFICIAL NEURAL NETWORK APPROACHES

Developing methods to reliably recognize depositional processes and environments in unconsolidated materials is of broad interest to various Earth science fields, archaeology, engineering, and life-sciences. Despite early efforts to link grain-size characteristics to distinct environmental settings, interpreting grain-size data has proven inherently complex and often contradictory. Early work employed elementary statistical approaches, such as the first and second statistical moments (mean, mode, skewness, and kurtosis) or percentiles (e.g., D50) and often utilized bivariate scatterplots to differentiate past depositional environments, mostly with limited success. Recent researchers have experimented with additional approaches such as multivariate statistics, end-member modelling, decision trees, etc., with some promise. Here I report on initial attempts to utilize multivariate statistics (e.g., Principal Components Analysis) and machine-learning Artificial Neural Network (ANN) analysis to distinguish broad environmental settings (e.g., fluvial, glacial. eolian, coastal) from <2-mm grain-size data in western New York. Preliminary results are promising, especially with regards to ANN analysis which routinely returns a success rate of >80% for individual model runs. The ANN model is then applied to the McKendry precontact archaeological site to better evaluate if artifact-bearing sediments reflect eolian or fluvial processes. Paleoenvironmental implications of these results are considered.