Anomalies in the Soil: Enhancing Archaeological Discoveries with Machine Learning using CoeusAI
In 2024 the Netherlands eScience Center started working on the collaborative project DroneML with the 4D Research Lab of the University of Amsterdam, the department of Archaeological Sciences of Leiden University and the Netherlands Forensic Institute (NFI). The project involved processing archaeological sensor data recorded with drones using machine learning. This led to the development of CoeusAI, a plugin for QGIS, a major open-source geographic information system (GIS) to analyze and edit spatial information.
We visited an archaeological site in Weesp and spoke with Dr. Jitte Waagen, archaeologist at the University of Amsterdam, and Dr. Maurice de Kleijn, former Research Software Engineer at the Netherlands eScience Center and now Spatial Information and Digital Infrastructures expert at the Dutch Heritage Agency.
Can you explain what you are doing with the drones?
Jitte: We use drones in a lot of different projects to see if we can detect archaeological traces of human activity in the past, whether visible at the surface or as buried remains, using sensors, such as thermal cameras, multispectral cameras, and so on. Such traces can sometimes still be observable with the naked eye, but often much more information can be retrieved using advanced sensors. The identification of such traces can be described as a process of screening all anomalies in the collected data, which we define as patterning that is unlikely to have natural causes or be explained by modern land use (e.g. lines, squares). In a subsequent step, such anomalies are analyzed from an archaeological interpretative framework. The anomaly identification can be very time-consuming and complicated, especially in large areas and with many different data layers. In a typical drone remote sensing fieldwork project we collect a lot of data from an individual site in different times of the day and year, for instance thermal data and height data, which we provided for the DroneML project at the eScience Center. Research software engineers Maurice, Ou Ku and Christiaan Meijer developed an interactive machine learning tool to assist with this process of anomaly identification. It works in a GIS environment and through machine learning distinguishes between interesting anomalies and irrelevant features.
What are the key archaeological challenges these days?
Jitte: Archaeological sites are under a lot of pressure. This is partly because of climate change: with water levels getting lower wood that has persevered for years can deteriorate quickly in dryer soil. Also, there is a huge need to quickly identify archaeology when it might be under threat from development of infrastructure. And we don’t want absence of evidence to turn into evidence of absence. Working with drones gives us the possibility to quickly gather large amounts of data without excavating. Excavating is very expensive, it takes a lot of time and you only get one chance to do it and do it right.
Can you explain what the tool that you have developed does?
Maurice: We explored multiple approaches. We first met with our project partners to thoroughly dive into their workflow and understand the tools they were using and how we can develop a tool that is actually useful. Part of this process involved joining them in the field to see how the data is collected. Regarding the machine learning component, we researched multiple options. For our initial design, the tool allowed an algorithm to analyze every pixel of the different datasets collected. The algorithm examined various values in relation to the values of neighbouring cells. In this way, we trained a neural network on anomalies that would help us quickly identify features of interest. It was also important that the tool should be fast and provide quick feedback to the user, which made it extra challenging in the development, but we managed.
What has been the most challenging aspect of this project?
Maurice: The most challenging part of this project was fully understanding what our project partners required. We can develop the software in any form they desire, but the key is to make something useful for them, as well as for other archaeologists and even experts outside that discipline. Therefore, we maintained close contact with our project partners and met regularly. Additionally, we shared ou progress with our technology leads at the Netherlands eScience Center and discussed it with colleagues in our Special Interest Groups. This way we got input from multiple perspectives, ensuring that we were moving in the right direction.
What have you enjoyed most about this project?
Maurice: What we liked the most about this project is that what we did is very tangible and has a large potential to have real impact on the archaeological domain. This tooling did not yet exist, and as such can be an asset for fast inspection of large and complex datasets in field archaeology. But not only the archaeological domain can benefit from it, we also see applications in forensics and maybe applications which we did not think of yet.
What has been the most beneficial part of working with Research Software Engineers at the eScience Center?
Jitte: The benefits have been really obvious. We are archaeologists before everything else and in the 4D Research Lab we are technological generalists within the domain of material and immaterial heritage. As the project involved complicated machine learning technology, it was better to work with research software experts.
What impact do you hope this software will have, either in the future of this project or in other projects?
Maurice: We designed a plugin for the free and open-source software package QGIS. This software allows users to explore, analyze, and visualize spatial data and has more than 100,000 users on a weekly basis. The tool we developed is primarily intended for archaeologists but can also be used for other purposes. In fact, Jitte has already been using the tool to analyze an inscription. Essentially, it aids in the process of detecting anomalies in images, so it might also be used for forensic purposes, physical geography, ecology, or land use change research. We therefore hope that our tool will have a big impact and hope that others can use our developments to their advantage.
