Machine Learning Approaches to Archaeological Predictive Modeling in the Age of Wildfire, Lake Tahoe Basin Management Unit, California and Nevada

Abstract

Machine learning is a powerful tool for archaeological prediction mapping. This thesis compares machine learning approaches to Middle and Late Archaic archaeological prediction in the Lake Tahoe Basin Management Unit, California and Nevada. Specifically, the analysis seeks to answer whether logistic regression, Random Forest, or Maximum Entropy models perform better at archaeological prediction. The explanatory variables used to predict site presence include elevation, slope, aspect, distance to streams, land cover, soil, and geology. Of all three models, Maximum Entropy produced the most accurate predictive models based on combined diagnostic metrics. Predictive modeling is a valuable tool in preventative archaeology, where identifying and mitigating adverse effects to archaeological sites in a time-efficient manner is critical. Environmental challenges such as uncontrolled wildfires provide an impetus for indigenous communities, management agencies, and researchers to employ predictive modeling approaches in preventative cultural and heritage resource management applications.