AI allies archaeologists: find new sites in satellite photos

Seven out of ten. Tells Nicholas Marchettiarchaeologist at the University of Bologna, who when artificial intelligence began to “guess” where the traces of ancient settlements are hidden or buildings by analyzing satellite images, with a 70 percent success ratethe engineer who programmed the machine, Marco Roccetti, he shook his head. Instead he was stunned: “She’s already better than us”. If you add to these performances thehelp of man, we arrive at 80. For Marchetti it is a turning point, which from finding traces of civilization sites, Halaf, Abbasid, Ottoman, in the plains of the Mesopotamia south, could be replicated in many other regions of the world, and make life much easier not only for archaeologists, but also for administrators. All using freely available images, such as Bing maps, and open source software. Very important thing: the tool has already been made public on GitHub and presented, in open access, in the magazine Scientific reports of the group natures.

The archaeologist’s first dilemma has always been know where to dig. Sinking the bucket, the pick and then the trowel costs time, effort and resources, it is not done by chance. Instead, that of the authority who wants to build an infrastructure is know where not to go, in order not to intercept finds from the past, perhaps interesting, but which would slow down the work and add costs for the survey, removal or, in exceptional cases, revising the layout to preserve them. Thus the automatism of a machine that has learned to make predictions on the presence of potential sites of archaeological interest, combined with human experience, can be a game changer.

Especially in the plains such as the Tigris and the Euphrates, the first cradle of human civilization: “The alluvial plains are the densest and most threatened archaeosystems on the globe – argues Marchetti, professor of Archeology of the Ancient Near East at the University of Bologna – they present a maximum density of evidence and at the same time of danger, because the great civilizations of the past, but also modern ones, arose around the great rivers, such as the Yellow River, the Nile and the Volga”.

Lean diet for AI

To find tracks, ground markers, archaeologists inspect satellite images and areas and then verify in the field (ground truthing) if indeed there is what they expected. Remote sensing is a scan of an area looking for anomalies, or recurring patterns, in a context. It feels like a task made just for a machine. And the Fertile Crescent, inhabited by Sumerians, Babylonians, Assyrians, Akkadians and Persians, is the right terrain from which to start but it was certainly not child’s play.

First of all for the quantity (it would be better to say, la scarcity) of data with which to train the AI: “Six or seven years ago Nicolò presented himself to me, with foresight, realizing that remote sensing was a job that could be done by algorithms – says Marco Roccetti, professor at the Department of Computer Science – Science and Engineering at UniBo – but the corpus of annotated images (those in which the marker of an archaeological site was known and verified ed) were just 5,000, I expected at least 50-100,000″.

In short, the machine had little to learn from. The starting point was a webGIS where – thanks to the Eduu and Kalam research projects of the University of Bologna – the data of 16 previous surface archaeological surveys had been collected and geo-referenced, on documents that go as far back as the 1960s. Although the algorithms have already become very powerful and precise in recognizing the shape of a face, objects or animals, for archaeological sites the variety of geometries and contexts complicates everything. It took years to train the algorithmwith successive steps: “We cooked those 5,000 examples in all possible ways, using the segmentation, therefore dividing the images, reflecting them, so that a single one became ten – continues Roccetti – to have sufficient examples to achieve accuracy. The transfer learning: I only have 5,000 potential sites recognized as such, but I have an algorithm that has acquired expertise by looking at and recognizing many other things and this gives it an advantage, a bit like a human who also has other experiences. The turning point in my opinion came with the introduction of the self attention: substantially, it is like transforming the visible image into another space, in which the machine sees things, recurrences, that the human eye would not see. And the mathematical transformation of a photosomething similar happens with the jpeg image compression algorithm”.

The human factor

But even after all this, the accuracy of the machine did not exceed 70 percent. The next step was then include in the process of analysis and confirmation, the contribution of the archaeologist: “human in the loop“.

“We understand that the machine didn’t have to work by itself – underlines Roccetti – so we used an already known method. The machine proposes solutions, the archaeologist notes and after a few cycles the machine has reached an accuracy of 80 percent”, a progressive learning process supervised by the young researcher Luca Casini. Annotations and corrections by the PhD student Valentina Orrùwere so decisive that even an error led to the recognition of a new site, guiding the archaeologist’s eye to the right area, just a kilometer further on.

Future

Emerging behaviors, when artificial intelligence surprises even its creators

by Andrea Daniele Signorelli

30 March 2023

But what does artificial intelligence “see” and “see”? A shape, a stain? The answer, which would be simple for a human, is not at all and has something to do with the mystery that the functioning of the AI ​​carries with it: “Because in the end it recognizes that place there and not another, we almost don’t know – admits Roccetti – the machine is trained to recognize mathematical occurrences, functions and relationshipswhich occur with a certain probability in certain contexts, e that the human eye cannot see. Like ChatGpt which predicts the next word based on a mathematical model, in latent spaces where it has transformed the set of words”.

But it works, with an accuracy which, Roccetti underlines, would not be sufficient to recognize a tumor in a CAT scan, but for archaeologists it is a formidable aid. And that, from Mesopotamia, could be brought to other parts of the world.

An instrument given to the world

The context, we said. Now that the algorithm has learned how to recognize hidden archaeological sites in Iraq, it could do so in similar environments in other parts of the globe, according to Marchetti: “Floodplains have common characteristics, while hilly and pre-desert areas are much more specific, so from a machine learning investment point of view is more complex. In our opinion, this large investment of time and resources which has been a very long process, will pay off n times. It can be replicated in archaeologically similar areas in the world. We have seen that in the case of Uzbekistan we will have to adapt the tool to the specifics of the case, and with the necessary adjustments it could also work in the Po Valley”.

The plains are fragile environments, because they are attacked by man, agriculture and economic activities. In Iraq, 50 percent of the sites have been lost for this reason, Marchetti explains. The algorithm developed by the University of Bologna will be a tool available to all scholars around the world who want to use it: “We have decided to make it available open source because we are convinced that third world countries also have the right to access science cutting edge, not only to second-rate products – concludes the archaeologist – in order to be able to compete at the same level. But let’s take the example of a state that has to build a road, a railway or a canal: without a map, you draw a line. But if we do a run with our algorithm, the territory is populated by spots. Which need to be confirmed, but you already know that 80 percent are sites, you have easier control on the ground. First of all you pass the infrastructure where there is less damage, and that damage that you cannot avoid you go to check it and modify according to the importance. This advocacy thing is revolutionary.”