Can ai map undiscovered ancient landscapes in Scotland?
Historic Environment Scotland believes that tens of thousands of ancient sites lie undiscovered across the country, in a fast changing landscape. But how to find and record them while they are still there? An exciting experiment with artificial intelligence has had some surprisingly good results. Dave Cowley enters the world of convolutional neural networks
Lidar – vast point clouds, processed, filtered and visualised to produce digital landscapes that can be explored with the roll of a mouse wheel, zooming effortlessly from the general to the detail – used to be a wonder. Airborne laser scanning ( als), as lidar is also known, is now familiar and routine. Used for archaeological survey it is revealing digital treasure chests of ancient sites and landscapes around the world.
So it is on Arran, popularly known as “Scotland in miniature” because of its varied landscape types, where Historic Environment Scotland ( hes) has just completed a rapid survey of all 432 km² of the island. The survey drew heavily on als data from the Scottish Remote Sensing Portal, a partnership between the Scottish Government and the Joint Nature Conservation Committee. Arran has attracted archaeologists for decades – it is an easy day trip from Glasgow – yet the hes survey has more than doubled the number of known archaeological sites. The quantities of known prehistoric roundhouses and medieval and later farmsteads and shieling huts, for example, have all significantly risen.
This is exciting stuff, with implications for how we understand the pattern of prehistoric settlement on the island, or the use of upland grazing in medieval and more recent times. Moreover, the survey was completed during 2018 with just six weeks in the field and a little more than that in the office, a dramatically quicker rate of coverage than would have been possible had we attempted to walk the whole island.
Nonetheless, this is a story that has been told many times – we know there are many previously unknown sites and landscapes awaiting discovery, and that als is a hugely effective way of mapping them. But here this is just the beginning. The Arran survey is part of a research and development project run by hes, which is exploring the effects of digital datasets and technology on our work, and how we might recast and greatly speed up our workflows for the future.
Key to this is recognising that, wherever we survey, we can expect to increase the number of known monuments very significantly. How do we address the fact that so much of our archaeological knowledge is inadequate, compromising management and research? Even the rapid rate of coverage on Arran is not scalable to the rest of the country if we wish to create systematic and reliable information for large areas. That is the real challenge for us: to date only some 10% of Scotland’s 78,000 km² has benefited from intensive survey. What can we do?
Automatic archaeology?
Archaeological survey is largely physical, relying heavily on boots on the ground and desk-based human observation – it’s not up to dealing with large complex datasets, or covering extensive areas without massive resources. Perhaps an answer to this problem lies in machine learning and artificial intelligence ( ai). ai is a growing part of the modern world. It’s in our homes, our cars, and increasingly underpins processes such as surveillance, medical diagnostics and
getting to grips with big data. Can it help us analyse 3d landscape data, finding patterns in Scotland’s historic environment that human interpreters would not have time to see, or would miss entirely?
We started exploring this in 2017 with the Norwegian Computing Center in Oslo, a research foundation that has led the use of “computer vision” – trying to automate what our eyes and brains can do – with remotesensed archaeological datasets. Convolutional neural networks
( cnns) in particular have proved very effective for picking out things of interest from images, for example in facial recognition. Such “image classification tasks” are a key part of what archaeologists do when we look at lidar images, making cnns a promising way to speed up survey.
The Norwegian Center applied their cnn, developed for use on their landscapes, to our digital topographic data of Arran. The latter included a version of the digital surface model for the island, and some “image patches”, or cropped images of selected monument types such as roundhouses, that “showed” the network what to look for.
This produced some good results, where the cnn accurately identified roundhouses, for example on Machrie Moor, but also utter chaos in areas of landscape where lumpy, broken terrain created confusion leading to an overwhelming number of “false positives” – natural lumps and bumps incorrectly identified as archaeological remains. This was a promising start. But it also made clear that there was lots of work still to do, and highlighted some general issues.
First, there is a general concern about ai: how do we know if it is working or not? We need to understand the performance of neural networks, including factors such as how they are trained, how they learn, and how and why outputs from one network may differ from another. Just as experience, intuition and skills vary among archaeologists, so these are factors that need to be considered in developing our use of ais – they are not all the same.
The pre-training of neural networks is a good example of why this matters. Many approaches, including the Norwegian study of Arran, rely on ImageNet, a database of over a million tagged generic images of, for example, animals, fungi and plants. But do pictures of kittens and puppies hinder pre-training of an ai designed to look for archaeology?
This is exactly the question that Jane Gallwey, a phd student at the University of Exeter, has sought to address by using lunar lidar for pretraining – neatly sidestepping the issue of how well ImageNet performs in training an archaeological neural network, and instead using explicitly topographic imagery (even if it is off the planet!). Training, whether for a human or an ai, matters: a pressing concern is establishing large archaeological training sets. This is why Iris Kramer, a phd student at the University of Southampton, will shortly be publishing a “benchmark” of the Arran dataset for use by other computer scientists to test their approaches in a common framework, helping us to better understand variability across different networks. This also highlights the range of issues that need to be tackled. Our work at hes contributes to an informal “automation network” of researchers, including colleagues in Edinburgh, Glasgow, Leiden, Southampton and Venice.
We also need to consider the datasets that are input to cnns. As humans, in surveys we work with visualisations – images derived from
the raw data that are designed (hopefully) to show us the landscape and the archaeological sites it contains. So far, inputs to cnns have been variants on such images. When we process a lidar dataset we often downgrade its detail to make it manageable – we reduce “data dimensionality”, and we have no means of comprehending a point cloud without derived visualisations. However, with increasing computational power it becomes realistic to ask if such a process is best for ai. It works for humans, but it may be fundamentally limiting the potential of a cnn.
Then there is the well-known issue of the black box: ai gives us a result, but we have no idea how it got there. This is a legitimate concern, but it is not unique to neural networks – it applies just as much to you and me.
We all see things subtly differently, and while we can hope to explain to each other what we see and how we interpret it, very often much of the process of archaeological interpretation happens “because I say so”, with little explicit explanation. Black boxes are everywhere. However, tools are being designed that look at the operation of a neural network, identifying which part is being activated by which bit of data. This is fascinating – it has the potential for us to define explicitly how a neural network is working, something that we will probably not be able to achieve for human archaeologists for a little while yet! It should also oblige us to examine and explain our own interpretative processes.
Global significance
How do we integrate ais into our work patterns, and what should we expect? An “archaeological ai”, or perhaps a range of them working iteratively together on problems of detection, will never be magic. They will never replace the role of the archaeologist, but careful thought is needed about how these things fit together. For hes, a national body seeking more extensive, reliable and systematic archaeological data to inform management and research, a key issue is bringing ai to bear in exploring the mass of remote-sensing data that has already outstripped our human abilities to investigate. If we are to do more than scratch the surface of such big data, we need ways of detecting objects of interest that can work tirelessly, 24/7, and systematically. We need to understand the character of the outputs – how reliable are they, how “competent” was the system that produced them, and so on? And we have to reflect on our established knowledge-creation processes, and subject these to the same critical review as ai. This is a fascinating prospect for anyone interested in how archaeological sites and features are identified and classified – whether by humans or machines.
Developing the use of ai is vital to our plans to create reliable and systematic archaeological data for Scotland, the need for which is pressing. During the course of the 18th century Scotland’s landscapes were dramatically “improved”, heralding two centuries of increasing intensification of agriculture, industrialisation, afforestation and urbanisation. Moreover, we may be moving into a time of accelerating landscape change – with drivers such as changing climate and agriculture practices. All these factors threaten the survival of archaeological remains, the more
serious if we have no record of them. So, in the expectation that there are tens of thousands of presently unknown ancient sites scattered across Scotland’s landscapes, if we want to bring them on record we need to work at a pace – and that means working with ais.
In the uk we benefit from several centuries of archaeological recording, providing a wealth of information on which we can build. This is not true in many parts of the world. The collective work of human and artificial archaeologists in undertaking survey that contributes to documenting our global heritage is especially important, the more so in the face of ever more rapid and unprecedented landscape changes.
Dave Cowley manages the Rapid Archaeological Mapping Programme at Historic Environment Scotland