The race to protect our cultural heritage is on, and a groundbreaking development in technology is leading the charge. A team of researchers has harnessed the power of artificial intelligence (AI) to detect looted archaeological sites from space, a feat that could revolutionize how we safeguard our history. This innovative approach, detailed in a recent study, showcases the potential of AI to identify signs of looting, even in the most remote and dangerous areas.
A Looted Legacy
Looting archaeological sites is a grave threat, often leaving little trace of its occurrence. Disturbed soil can easily blend into the landscape, making detection challenging. The team, comprising experts from Microsoft's AI for Good Research Lab, Iconem, and Planet Labs, has developed an AI system that can identify these subtle signs using satellite imagery.
The challenge lies in the fact that many cultural heritage sites are located in remote or conflict-affected regions, making traditional monitoring methods impractical. Satellite images offer a safer alternative, but manual analysis is time-consuming and prone to human error. This is where AI steps in, providing a rapid and efficient solution.
Building the AI's Eye
The researchers created a vast dataset, covering 1,943 archaeological sites in Afghanistan, with 898 confirmed as looted and 1,045 preserved. This dataset was meticulously compiled using the Archaeological Gazetteer of Afghanistan and verified by expert archaeologists. The team then utilized PlanetScope satellite imagery, capturing monthly mosaics with a resolution of 4.7 meters per pixel, covering the period from 2016 to 2023.
A crucial aspect of the study was the manual marking of site boundaries by archaeologists. These spatial masks allowed the researchers to test the AI's performance when focusing solely on the archaeological area, significantly improving its accuracy.
AI vs. Traditional Methods
The team employed two distinct approaches: deep learning with ResNet and EfficientNet models, and traditional machine-learning methods like Random Forest and XGBoost. The results were striking. The deep learning approach, using ResNet-50, achieved an impressive F1 score of 0.926, outperforming traditional methods.
Interestingly, the newer, more advanced foundation models did not consistently outperform simpler models. The researchers attributed this to the local nature of looting signs, primarily small texture changes, which simpler models could detect more effectively.
Uncovering the Subtle Signs
SHAP analysis revealed the model's key predictor: the strength of edges in the near-infrared band. These sharp boundaries, created by excavation, were crucial in the model's detection process. The model's ability to identify subtle texture changes, such as contrast, entropy, and homogeneity, from small patches of the red, green, and near-infrared bands, was remarkable.
Focusing on the Target
The study's most significant finding was the impact of providing the model with clear direction. By using manually drawn masks, the AI's performance improved dramatically, achieving a 40% increase in F1 score. This simple yet effective approach reduced noise and allowed the model to focus on the archaeological site, enhancing its detection capabilities.
Timing is Everything
The timing of satellite images played a pivotal role. Models trained on imagery from 2020 performed best, as most looting occurred before 2021. Over time, natural changes can obscure signs of looting, making detection more challenging.
A Monitoring Tool for the Future
The researchers emphasize that their system is a monitoring and triage tool, aiding experts in their efforts. It can help identify sites requiring immediate attention, covering vast areas that would otherwise be impossible to monitor. The team aims to expand its application beyond Afghanistan, targeting regions like Syria, Sudan, and Egypt, where looting is a significant concern.
Overcoming Challenges
A major hurdle is the reliance on archaeologists for site boundaries and labeled examples. To address this, the researchers are exploring semi-supervised and active-learning methods, reducing the need for manual annotation. The team's open-source code and dataset methodology offer a valuable resource for further research and development.
In conclusion, this AI-driven approach to detecting looted archaeological sites is a significant step forward in heritage preservation. By leveraging technology, we can better protect our global cultural heritage, ensuring that the legacy of our past remains intact for future generations.
