TLDR: A deep learning AI model was re-trained using declassified CORONA satellite imagery from the 1960s to detect archaeological sites, specifically “tells,” in the rapidly changing Mesopotamian floodplain. The model achieved 90% accuracy and successfully identified four previously unknown archaeological sites in the Abu Ghraib district, confirmed through field verification, demonstrating the critical role of historical imagery and AI in preserving vanishing cultural heritage.
Archaeological research in regions like Mesopotamia faces a significant challenge: the rapid transformation of landscapes over the past five decades has led to the destruction or concealment of countless ancient sites. Traditional methods of identifying these sites, often relying on remote sensing, are labor-intensive and can miss evidence that has vanished from modern view.
A recent research paper, titled “AI-ming backwards: Vanishing archaeological landscapes in Mesopotamia and automatic detection of sites on CORONA imagery”, presents a groundbreaking approach to overcome this hurdle. Authored by Alessandro Pistola, Valentina Orrù, Nicolò Marchetti, and Marco Roccetti, the study introduces an upgraded artificial intelligence (AI) model that leverages historical satellite imagery to uncover lost archaeological treasures.
The Power of Historical Imagery
The core innovation of this research lies in its use of CORONA satellite imagery. These grayscale images, captured between 1960 and 1972 as part of a U.S. reconnaissance program, offer a unique snapshot of the Mesopotamian landscape before extensive modern development and agricultural expansion. By training a deep learning model with this historical data, the researchers aimed to identify “tells” – artificial mounds formed by centuries of human habitation – that are no longer visible on contemporary maps.
How the AI Works
The team upgraded an existing deep learning model, a convolutional neural network designed for “semantic segmentation,” which essentially teaches the AI to classify every pixel in an image. The original model, trained on modern Bing imagery, was re-trained using CORONA images for the Abu Ghraib district, west of Baghdad. This process involved “transfer learning” and a “two-stage fine-tuning” procedure, which are advanced techniques for adapting an existing AI model to new data and improving its performance.
The researchers prepared a dataset of 208 images, carefully balancing those containing known tells with those without. They also employed extensive “data augmentation” – manipulating images (e.g., rotating, mirroring, adjusting brightness) to create a larger and more diverse training set, preventing the AI from simply memorizing the examples.
Remarkable Results
The re-trained AI model demonstrated significant improvements in detecting archaeological sites. Specifically, the model that combined Bing and CORONA imagery (BingCORONA_BingCORONA) achieved a tell detection accuracy of 90%, a notable increase from previous results that barely surpassed 80%. This improvement was particularly evident in models that incorporated CORONA imagery, underscoring the value of historical data in recognizing vanished features.
Unearthing New Discoveries
Beyond improved accuracy on known sites, the AI model proved its worth by identifying entirely new archaeological locations. The AI generated “heatmaps” – visual representations indicating the probability of archaeological remains – which were then reviewed by archaeologists. Out of eight highly probable sites suggested by the AI, four were confirmed through systematic field surveys conducted in January 2023 and January 2024. These sites had never been identified using traditional methods, and despite being completely destroyed on the surface, the presence of ancient ceramic sherds confirmed their archaeological significance and allowed for dating.
Also Read:
- AI Models Offer New Insights into Bridge Health Through Non-Destructive Evaluation
- Unlocking AI’s Memory: How Synthetic Images Combat Forgetting in Learning Systems
A Collaborative Future for Archaeology
This research highlights the profound impact of combining AI with historical satellite imagery. It demonstrates that AI can serve as a powerful support tool for archaeologists, helping them pinpoint areas of interest that might otherwise be overlooked. The study also starkly illustrates the rapid loss of archaeological heritage in Mesopotamia, with 38% of surveyed sites completely destroyed and 23% largely destroyed over recent decades. By enabling the detection of these vanishing sites, this AI-driven approach offers a critical new avenue for cultural heritage protection and research, paving the way for future advancements that could integrate technologies like LiDAR and super-resolution imaging.
