From archaeologist to data scientist
Iris Kramer
I'm a phd student in computer science at the University of Southampton. I started after undergraduate and masters study in archaeology, when I became interested in research on automated methods for the detection of archaeological sites. I like a challenge, so I focused my research on the most advanced method out there – deep learning. This is a big step on from telling computers what to do, using systems that can perform specific tasks, such as image classification, without explicit instructions, instead relying on patterns and inference.
If that sounds familiar – it is because these are the kinds of processes that we humans are quite good at. The challenge in computing is to move beyond the usual kinds of classifications of cats vs dogs, for example, to detecting archaeological sites on remote sensing data. This is not straightforward. Archaeological remains are found in different types of land use and in different states of preservation – they can look different in remote-sensed data. So, in my research, which includes working with Historic Environment Scotland on Arran, I have been looking to improve the performance of convolutional neural networks ( cnns) with transfer learning. This is about using knowledge gained in addressing one problem, and applying it to related questions – a little bit like how we draw on our range of experiences when we tackle a new activity. For cnn- based archaeological site detection, the balance between the background landscape and the “site” is a big challenge, mainly because the dispersed nature of most site distributions means that the background can be overwhelming – like a needle in a haystack. I am trying to fuse information drawn from different remote-sensed data and different site manifestations, so that the networks can learn the cumulative character of sites, moving beyond unique expressions.
This is starting to work well, though the network is still generating some false positives that an expert human interpreter would be able to disregard. Here, we are now thinking about how aspects of human interpretative processes can feed into the design of our computer system. For example, a field archaeologist will instinctively draw on their observations of context, like contemporary land use, in directing their observation and interpretation. They will know, for instance, that the earthwork remains of a Bronze Age roundhouse are not going to survive in what is now heavily improved farmland. Thus, when a roughly circular slight earthwork turns up in such a location – we know it is probably very recent, with the characteristic circular site of a cattle feed-bin being a likely explanation. In such cases we, as human beings, bring contextual knowledge to bear, something that we can also feed into our computer systems – improving the “experience” and background knowledge of the network to better distinguish “noise” from archaeological sites. These kinds of advances in deep learning are building blocks for an ai archaeologist – not science fiction, but a tool that will have a dramatic impact on how we develop our capacity to work with big, remote-sensed datasets, and rapidly improve our knowledge base and understanding of the landscape.