A group of international researchers has harnessed the strength of artificial intelligence to reveal previously unknown reservoirs from the Angkorian period in Cambodia. Using a deep learning model called DeepLab V3+ for semantic segmentation, archaeologists have made tremendous strides in mapping one of the world’s most significant ancient civilizations with greater efficiency and precision.

Many of Angkor’s temples, including Angkor Wat, remain today as iconic symbols of Cambodia. Credit: Daniel Lienert

The Angkorian Empire, which lasted from the 9th to 15th centuries CE, had its capital in what is today Siem Reap province. Although many of Angkor’s temples, including Angkor Wat, remain today as iconic symbols of Cambodia, much of the surrounding urban infrastructure was, for a long period, buried beneath dense forest and was difficult to fully explore. Whereas satellite imagery and LiDAR technology had already helped reveal over 25,000 features within the Greater Angkor Region, mapping beyond this core area has been extremely time-consuming when done manually.

To accomplish this, researchers trained an AI model to recognize one key feature of Angkorian landscape engineering: reservoirs. These vast, rectangular, embanked water structures—part of the civilization’s water management system—stand out in satellite images and were therefore ideal targets for automation.

Training the AI tool involved collecting over 11,000 examples of reservoir shapes, of which around 3,600 high-quality samples were utilized for training. Every sample was a 512×512 pixel patch of satellite imagery. The model exhibited an F1-score of up to 45% in good areas, the same level of performance as that of an archaeology student who had been trained.

Three reservoirs whose mapped outline (in yellow) does not correspond to the imagery due to changes in climate and land use conditions. Credit: J. Landauer et al., PLoS ONE (2025)

As the authors of the research, published in PLOS ONE, said, “Integrating AI as a tool for reservoir identification into large-scale mapping projects offers promising opportunities for expediting research. However, it remains crucial to underscore the importance of human validation in verifying results.”

The model not only saves time, but it can reduce the workload of mapping tasks by as much as 90%, according to the researchers. While preliminary results are promising, there are limitations that the team acknowledges. For instance, the algorithm struggled to detect temples since they vary in shape and size, so the researchers only concentrated on reservoirs for this study.

The researchers said that the ultimate objective is to further improve prediction quality and automate the mapping of more remote and less-studied regions of Cambodia. They also see broader applications for the model: “This AI model… can be trained on other types of archaeological and non-archaeological features worldwide.”

Future research aims to incorporate domain adaptation techniques, enabling the AI to adapt to new, unseen landscapes it has not been trained on. Such improvements could enable the model to contribute to broader archaeological discoveries, including reservoir boundaries in detail and even temple detection.

More information: Landauer J, Klᴀssen S, Wijker AP, van der Kroon J, Jaszkowski A, Verschoof-van der Vaart WB (2025). Beyond the Greater Angkor Region: Automatic large-scale mapping of Angkorian-period reservoirs in satellite imagery using deep learning. PLoS ONE 20(3): e0320452. doi:10.1371/journal.pone.0320452