The race to protect our cultural heritage from looting is on, and a groundbreaking development in AI technology is leading the charge. A collaboration between Microsoft's AI for Good Research Lab, Iconem, and Planet Labs has resulted in an AI system capable of detecting signs of looting at archaeological sites using satellite imagery. This innovative approach not only highlights the potential of AI in archaeology but also underscores the importance of technology in safeguarding our historical treasures.
The Challenge of Looting Detection
Looting poses a significant threat to archaeological sites, often leaving barely detectable signs. Disturbed soil, new pits, changes in soil patterns, and subtle shifts in ground shape can be easily missed, especially when they resemble natural erosion or agricultural activity. Traditional methods of monitoring these sites, such as manual image analysis, are time-consuming and impractical, especially in remote or conflict-affected areas.
Building a Comprehensive Dataset
To train the AI system, the team created an extensive dataset covering 1,943 archaeological sites in Afghanistan. This dataset was meticulously verified and expanded by expert archaeologists using various sources, including Google Earth, ESRI, and Bing imagery. The sites were classified based on evidence of deliberate damage recorded before the end of 2023, ensuring a comprehensive and accurate representation of the looting problem.
Testing Different Approaches
The researchers tested two distinct approaches to detect looted sites. The first used deep learning models like ResNet and EfficientNet, trained directly on raw satellite image patches. The second approach employed traditional machine-learning methods, including Random Forest and XGBoost, with various feature engineering techniques. Interestingly, the simpler models outperformed the more advanced foundation models, suggesting that looting leaves highly local signs that general-purpose models might overlook.
The Power of Focused Analysis
One of the most significant findings was the impact of focusing the AI's attention on the archaeological site itself. By using manually drawn masks to outline the site's boundaries, the models achieved a substantial improvement in performance, with an F1 score increase of over 40 percentage points. This simple yet effective technique reduced noise and allowed the models to detect subtle signs of looting more accurately.
Timing is Crucial
The timing of the satellite images played a critical role in the system's performance. Models trained using imagery from individual years performed best for sites photographed around 2020. This pattern suggests that the looting occurred primarily before 2021, and the signs of excavation were most distinct during that period. Over time, natural changes can obscure these signs, making detection more challenging.
Looking Ahead
The AI system is envisioned as a monitoring and triage tool, aiding experts in covering more ground by identifying sites requiring closer attention. The next steps include testing the approach in other regions, such as Syria, Sudan, and Egypt, to assess its effectiveness in different contexts. Additionally, the researchers aim to explore semi-supervised and active-learning methods to reduce the need for manual annotation, making the technology more scalable and adaptable.
This groundbreaking development in AI-assisted archaeology not only showcases the potential of technology in preserving our cultural heritage but also highlights the importance of international collaboration in tackling global challenges.
