Rewriting history: New AI can spot overlooked archaeological sites

The model was implemented using pre-trained models for semantic segmentation, fine-tuned on satellite images and masks of the site shapes coming from a dataset containing almost 5,000 examples. But they encountered several problems.

“The dataset, while may be considered a very large one for near eastern archaeology with its almost 5,000 sites, is hardly sufficient for training a model as large as the state-of-the-art ones we see in use today and, perhaps more significantly, contains many cases that are visible only on certain old imagery,” explained the researchers.

Using a human-in-the-loop approach

To solve these issues, the team used a human-in-the-loop approach to integrate domain expertise during their experiments' training and evaluation phase. That was crucial in improving the dataset used and, in turn, the model.

The outcome of this iterative process was a model capable of obtaining a detection accuracy of around 80 percent.

“Based on these egregious results, we envision a tool for human-AI collaboration to support the archaeologists in the remote sensing operations (rather than replace them) and propose a new kind of workflow, enhancing both their task and the model by providing improved data after every use,” announced the researchers.

The scientists now claim the outputs of their model can be used both for high-speed automatic detection while being aware of the mistakes this could introduce or combined to generate a graphical overlay to direct the user’s attention towards specific areas.

This will hopefully make the search and discovery of archaeological sites much faster and more efficient, leading to new finds in the sector that will forever change our understanding and knowledge of our planet’s rich history.