Machine learning-based identification of geomorphological units in Quintero Bay (32°S) and its implications for the search for early drowned archaeological sites on the western coast of South America

High-resolution predictive modeling of submerged landscapes has successfully allowed the detection of early archaeological sites that are presently underwater. These models have traditionally relied on geophysical techniques, which can be both time-consuming and expensive, especially for extensive survey areas. In contrast, geomorphological mapping using Machine Learning (ML) techniques has emerged as a rapid and accessible alternative with numerous advantages over conventional methods. In this work, we employ ML algorithms (Random Forest, Support Vector Machine, Partial Least Squares, and Principal Component Analysis) trained on land to analyze the seabed of Quintero Bay to identify relic landforms that characterize the paleolandscape within which the submerged early site GNLQ1 formed. The methodology also included a multicriteria analysis that integrated geological (geomorphological, tectonic, eustatic) and archaeological (attributes of non-submerged records in the region) approaches to delineate potential areas of archaeological interest. The findings of this work can guide and enhance future archaeological research. The results underscore the importance of possessing a comprehensive understanding of the study area and its associated variables to the successful application of ML techniques. This also applies to modeling drowned paleolandscapes. Nevertheless, despite these challenges, ML-based modeling of drowned paleolandscapes can provide an overview of the distribution of geoforms comprising the paleolandscape, which in turn can help identify future geophysical survey areas to focus on in the search for archaeological evidence, thereby improving our understanding of the relationship between early human groups and these landscapes.