A 3d Point Cloud Deep Learning Approach Using Lidar to Identify Ancient Maya Archaeological Sites

Abstract

Abstract. Airborne light detection and ranging (LIDAR) systems allow archaeologists to capture 3D data of anthropogenic landscapes with a level of precision that permits the identification of archaeological sites in difficult to reach and inaccessible regions. These benefits have come with a deluge of LIDAR data that requires significant and costly manual labor to interpret and analyze. In order to address this challenge, researchers have explored the use of state-of-the-art automated object recognition algorithms from the field of deep learning with success. This previous research, however, has been limited to the exploration of deep learning processes that work with only 2D data, which excludes the use of available 3D data. Our research addresses this gap and contributes knowledge on the use of deep learning-based processes that can classify archaeological sites from LIDAR generated 3D point cloud datasets. LIDAR data from the UNESCO World Heritage Site of Copan, Honduras is used as the primary dataset to compare the classification accuracy of deep learning models using 2D and 3D data. The results demonstrate that models using 3D point cloud datasets provide the greatest classification accuracy in identifying Maya archaeological sites while requiring less data preparation. Further, the research contributes knowledge on the efficacy of data augmentation strategies when working with small 3D datasets.

Assessing the lidar revolution in the Maya lowlands: A geographic approach to understanding feature classification accuracy

“…Only verified training data should be used in machine and deep learning models to maximize their effectiveness, especially given that archaeological datasets fall well short of the millions of examples that would ideally be provided for training (Somrak et al, 2020: 8). New DL-models are being published rapidly, including point cloud-based classifications (e.g., Richards-Rissetto et al, 2021), and hold the greatest promise for interpreting large datasets and making broader comparisons, by eliminating inter-analyst bias while simultaneously being able to rapidly classifying thousands of square kilometers of data.…”

Advances in remote sensing of the early Anthropocene in tropical wetlands: From biplanes to lidar and machine learning

“…It is essential to understand from the ground what the features in the remotely sensed data are to avoid misinterpretation that can lead to incorrect conclusions. After enough features of interest have been correctly identified in the data, however, machine learning algorithms promise to facilitate and accelerate our mapping efforts across the broader region (Bundzel et al, 2020; Character et al, 2021, Character et al, in review; Richards et al, 2021; Talukdar et al, 2020). Here, after we have determined the remotely sensed data most useful for identifying the ancient canals, we test a machine learning algorithm for efficiently mapping the networks.…”

“…Other examples of extensive canal features include Stoner et al (2021) who compared large expanses of canal systems with settlement patterns in the tropical lowlands of Veracruz; Inomata et al (2021) in Tabasco, Mexico, Dunning et al in Campeche (2020), and Šprajc et al (2021) in Campeche, Mexico. Several studies have developed and compared various ways to process and visualize LiDAR DEMs for identifying ancient structures and mounds (Bundzel et al, 2020; Inomata et al, 2017; Kokalj and Somrak, 2019; Richards et al, 2021; Thompson, 2020), but no other peer-reviewed sources have applied these techniques for identifying the ancient canals. Here we compare the same processing and visualization techniques commonly used in urban archaeology for canal and raised field identification.…”