Uncovering archaeological sites in airborne LiDAR data with data-centric artificial intelligence

Mapping potential archaeological sites using remote sensing and artificial intelligence can be an efficient tool to assist archaeologists during project planning and fieldwork. This paper explores the use of airborne LiDAR data and data-centric artificial intelligence for identifying potential burial mounds. The challenge of exploring the landscape and mapping new archaeological sites, coupled with the difficulty of identifying them through visual analysis of remote sensing data, results in the recurring issue of insufficient annotations. Additionally, the top-down nature of LiDAR data hinders artificial intelligence in its search, as the morphology of archaeological sites blends with the morphology of natural and artificial shapes, leading to a frequent occurrence of false positives. To address this problem, a novel data-centric artificial intelligence approach is proposed, exploring the available data and tools. The LiDAR data is pre-processed into a dataset of 2D digital elevation images, and the known burial mounds are annotated. This dataset is augmented with a copy-paste object embedding based on Location-Based Ranking. This technique uses the Land-Use and Occupation Charter to segment the regions of interest, where burial mounds can be pasted. YOLOv5 is trained on the resulting dataset to propose new burial mounds. These proposals go through a post-processing step, directly using the 3D data acquired by the LiDAR to verify if its 3D shape is similar to the annotated sites. This approach drastically reduced false positives, attaining a 72.53% positive rate, relevant for the ground-truthing phase where archaeologists visit the coordinates of proposed burial mounds to confirm their existence.

Keywords

Archaeology Data-centric artificial intelligence Data augmentation Deep learning LiDAR Location-based ranking Object detection

URI

http://hdl.handle.net/10400.1/19891

Publisher

IEEE

DOI

10.1109/ACCESS.2023.3290305

Collections

ICR2-Artigos (em revistas ou actas indexadas)

CC License

cclicense-by

Altmetrics

Full item page