Mapping past forests in the Atacama Desert, Chile: A remote sensing and machine learning approach for archaeology

The last decades have seen a significant increase in the use of spatial and remote sensing analytical techniques in the field of archaeology, leading to new perspectives and research questions. However, the full potential of these methods remains largely unexplored. Albeit the focus on site detection and resource mapping has been gaining traction, we have yet to examine the great potential these techniques could have to support various tasks, such as the design of targeted field campaigns, process optimization, and the identification of new findings, mainly in large and difficult-to-access areas. This paper presents a case study focused on mapping sub-fossil botanical remains (wood, leaf litter, charcoal and tree imprints) from past forests in the Pampa del Tamarugal (PdT), Atacama Desert, Chile (21°07′36″ S, 69°26′49″ O). These areas have proved key for the lives of pampean populations since the region's first settlements around 12.000 calibrated years before present (cal yr BP), being incorporated not only into subsistence strategies but also into their cosmologies. In this study, we present an experimental remote sensing model using Sentinel-2 free multispectral satellite imagery, which, when combined with ground truth data, allows for the detection of spectral proxies associated with past forests' spatial distribution. The results validate the use of remote sensing methods for the documentation of these past natural or silvicultural forest patches. Through the discussion of this case study, this paper highlights the remarkable potential of combining machine learning algorithms with traditional archaeological and paleoecological research, emphasizing their implications for field campaign planning and development, as well as resource optimization.