Comparing computer algorithms for identifying archaeological mounds in South Carolina

What is it about?

The saying goes, "there's more than one way to skin a cat." This applies to science as well: there are often many methods that can be used to address a research problem. In the field of remote sensing (studying the earth through non-invasive means, such as images), processing gigantic datasets is a growing problem. While LiDAR, a technology that allows researchers to map ground surfaces, has proven to provide detailed topographic information in heavily forested environments, tools for efficiently processing these data are varied.  Machine learning algorithms offer ways of efficiently and accurately identify archaeological deposits in these massive datasets. However, which algorithms work the best for archaeologists? In this article, we compare four different methods for automatically identifying archaeological mound and ring deposits from LiDAR datasets. Specifically we test segmentation, inverse depression analysis, template matching, and a new method. Segmentation refers to a computer algorithm that divides an image based on color, texture, and other properties. In so doing, the segmentation procedure creates "image objects", which often represent real features on the ground. To understand inverse depression analysis, we must first explain depression analysis. Scientists studying geography and Earth Science often need to locate sinks in the ground surface (or depressions). To do this, researchers have developed ways of automatically scanning LiDAR datasets for areas with sinks in the ground surface. For detecting archaeological mounds, we created an inversed version of these algorithms, so that instead of identifying sinks, we identify rises in the ground surface, which can signify mounded features. Template matching is also known as pattern recognition. Simply, the method works using statistical assessments of similarity; in other words, it tries to match a known sample of information with a new dataset. In this paper, we created a training dataset of known archaeological mounds in the study area. These images became used to statistically evaluate the entire LiDAR dataset covering over 800 square miles or land. Lastly, we created a new method to identify mounds in LiDAR data. This method uses statistical assessment from template matching and segmentation. By combining both of these methods together, mounds could be detected based on both statistical similarity, as well as characteristics such as elevation, size, and shape. Our findings show that this algorithm was the most successful in identifying archaeological deposits, followed by the inverse depression analysis.

Featured Image

Photo by Greg Nunes on Unsplash

Photo by Greg Nunes on Unsplash

Why is it important?

The area studied (Beaufort County, SC), much like the majority of coastal regions, is increasingly under threat of urban development as well as flooding and erosion caused by climate change. The techniques applied here can be implemented in other parts of the world, and this paper provides some insights into each respective method's strengths and weaknesses for identifying mounded archaeological features. By comparing different ways of accomplishing the same task, future work can avoid weaker methods and focus on using and developing stronger methods. Additionally, by demonstrating the strengths and weaknesses of these different approaches, the archaeological record can be better protected using these technologies in the future. Detecting cultural materials can be beneficial for monitoring sites that are at risk of damage from climate change and urban development. Additionally, new sites that we didn't have records of can be discovered.