• AVimmerse

New ways of looking for the past

In the third of the Heritage and Technology series, we’re joined by Dr Dave Cowley, Deputy Head of Survey and Recording (Archaeological Survey) for Historic Environment Scotland. He looks at developing archaeological surveys in the 21st century.

Listen to what Dr Dave Cowley had to say here:

Speaker transcription*:

Thank you very much, and apologies for my technical difficulties that caused all the chaos earlier on. So what I really wanted to try and do is just have a look or think about what digital technologies and data sources are doing for archaeological survey in the 21st century. And I wanted to start with just looking at why Archaeological Survey matters. Essentially, our Historic Environment landscapes a very rich sources for understanding our past I've just put up a couple of images there, of intertidal fish traps up in Apple cross and the the remains buried remains of a hill fought in a Slovenian visible fleetingly in crop proxies. And essentially, to identify the in terms of the wider landscape, what we have on record already, is a small proportion of what's out there. And that knowledge of what and where is absolutely crucial to knowledge creation, our understanding of the past, and also our capacity to manage that those material remains of what people did in the past, for the future. And that's obviously crucial to underpinning the kinds of stories that we can tell about the past. And this is a kind of multi scalar approach, something that you can do a very broad brush, like the landscape characterization, that you see a Veyron at the top left. It's a process that significantly improves our understanding of the past at a Island or regional level. But it's also a jigsaw puzzle, which you see the mapping in blue on the right hand map of the Roman military complex and Newstead in southern Scotland. And this is the product 50 plus years of archaeological investigation and aerial photography that has put that map together over a long period of time. And across the bottom, I've illustrated that kind of stepping up or down scales, essentially working from prehistoric landscape in upland pasture, through a Roman temporary camp, and native settlements near Dumfries down to a very, very detailed treatment of an archaeological monument. And before I get into the meat of the talk, I just wanted to show this illustration, which are used as a kind of analogue for the benefits of remote sensing for archaeology, but particularly that elevated view and this is a view of a street, just around the corner from my office in, in Edinburgh. And it's there to as an analogue for you and me walking around the landscape. And how we might experience it, the the lack of detail and the distance, the closeness of view. And I've then taken this picture, standing up essentially to mimic that kind of bird's eye view. And two points I wanted to make here really is one is that that wider context really shows us the pattern, it resolves features that perhaps we weren't really seeing in the past, or in the in the in the ground view. But one absolutely crucial point is that this is a view that tells us stories. So you can see that there is a dribble of paint, at a very basic level, we can read the direction of the slope, but the thing I really like about it, it as an analogy is the it also tells us about a backstory. There used to be a home base DIY store across the road that provides contextual information. But I also like the backstory of the decorator, your flat or whatever it is. But the key point without wanting to press that analogy too far is that when we look at landscape information, whether we're in person or remotely, we're reading it for those stories about the past and not just images, or pretty pictures. And in here, we're at a cusp our are in a period of change, where traditional survey practice, whether it's dealing with detailed site records or entire landscapes has largely been dominated by inherently manual processes.

And in this context, this is not new news. digital data is really a game changer. And but it does beg a series of questions to visualise As shown on the right hand side took me and a colleague, two days or so of walking backwards and forwards, slowly collecting height data for an area that's about 100 metres by 200 metres 10 years ago, incredibly slow. The map on the other side shows the current availability of airborne laser scanning data in Scotland, of comparable some of it of comparable detail, but available there through the Scottish Government. Also, the scale of of data availability is absolutely fundamentally changed over this last decade. And much of what we've been doing really is is experimenting with how that's going to work and how we think about it. And in here, we're feeling our way towards how we best combine born digital workflows. Well, if you like analogue or manual workflows, I'm not going to talk anything really about LiDAR, or airborne laser scanning data, you can look it up, there's lots of good material there. But essentially, the main point about this is that it provides us with dense, precise and accurate terrain models that can be visualised using using software, like the relief visualisation toolbox from our friends in Slovenia, in a whole multitude of different ways. So it really is absolutely fundamentally about different ways of looking, and ways of looking at the landscape. And that cannot be mimicked by us as individuals in the landscape.

And one illustration, which I'll come back to this is a complex of Roman, an Iron Age remains a burns walk. If you're driving north on the 74, around Lockerbie, you see a flat top tilde on the right hand side, next time you go up there, stop and have a look, it's fantastic. And every archaeologist who's ever visited this site, by and large, just coming through a gate down the bottom of the image and has walked up here and wandered on this temporary camp is obviously known this enclosure up here, this enclosure in here, all known. But what's interesting is it's only when we started playing with the LIDAR data that this square enclosure popped up. And I guarantee you that every archaeologist who's ever been there, myself included on multiple occasions, has walked straight across that and never seen it. So that is just one illustration of a pattern repeats itself, repeatedly of this the power of this different way of looking. I'm not going to talk through this diagram, you'll all be pleased to know. But essentially, it's making one key point which I just want to make and move on from, which is that digital data facilitates supports digital workflows. And within that you can create this is something that we've created in QGIS, essentially which enforces things like if the sore eye terms and quality assurance, and allows us to upload data to the national record of historic environment where then becomes accessible through our online portal Canmore without any rekeying. So this is one of the kind of side benefits of working in this Gordon digital environment. But what I've mainly wanted to talk about is automated object detection, the use of AI and machine learning. And essentially, where we've moved to now in this general field is with deep learning and machine learning. There's a real step change and they didn't tasks like image and object recognition and to some degree classification. computational approaches are outperforming human beings as a matter of routine and you can see a few examples there of where it's being applied. We've done a little bit of work in this in partnership initially with the Norwegian Research Institute. And this produced two patterns really quite interesting. Good results, if you like and reassuringly, I think probably it also provided some absolute chaos in parts of the island of Ireland, where the topography is very lumpy, bumpy. So the six images that you see on the right hand side are the footings of prehistoric roundhouses correctly identified by the machine learning approach, along with along the bottom row. A modern agricultural feature Hey, Each and prehistoric burial camp, that essentially, this was enough to convince us that the approach was worth pursuing, along with many other people. So this is kind of where we are now it's a very fast developing field, there is a proven capability to detect. The question now is how not if we use this as an approach as part of our Archaeological Survey approach, but we're starting to learn about the requirements for careful design and implementation. And I think we've also probably moved beyond the inherent suspicion of this as an approach that we're seeking to replace the expert archaeologist. That's certainly where I started. And that's why ATS is involved with Glasgow University on a Lucy Colossians, PhD, which is essentially about exploring how we are going to implement this new see some results from Aris Kramer of our KPIs work on Aaron on the right hand side.

And beyond the fact that this is beginning to work, and we're seeing the benefits of these developments, it's the approach of bringing in machine learning and AI AI is allowing us to look at data and processes in detail. And I think in here, there is some interesting things that we can do to think about the black boxes that are both represented by certain computational approaches, but that are also represented by you and me, we are some of the most profound black boxes in any interpretive and thinking process. And we're also starting to learn a little bit about some how some neural networks for example, may be better. Archaeologists and others, the the Norman Bates, Norman AI example there is a an AI that was trained on the dark web and one that was trained on pictures of puppies and kittens from the rest of the web. And was then given a series of inkblot images to classify and you see a couple of examples in there essentially, it demonstrating as the same with us as people that how you train something actually really does matter. And just to illustrate, again, that variability in the work in the outputs of different approaches to machine learning the work has on Alma, Rouen, Edinburgh, he basically fed the same data into a whole series of different neural networks and demonstrated how very, very different the inputs or the outputs were. Jane Galway and her colleagues also did pre training on lunar LiDAR, which is at least topographic data set rather than the image based image net and produce significant improvements by doing that way, so we're starting to learn about these kind of approaches. And this is starting to help us define really a candidate slightly more mature approach to AI and machine learning for Archaeological Survey. Beyond the hey, let's try it does this work. And I've already mentioned the importance of training and different network architecture. We're also to a degree training are using images that are optimised for ourselves with our own visual cortex as part of the input in the training, and there's a question there is actually, should we be training the computers to look like us? Or should actually be looking at the raw data? And then this key question, which Lucy Cameron, for example, is investigating is about how do we integrate into workflows and I just wanted to spend a couple of minutes talking about how we create those workflows that are both drawn and challenged traditional practice. So in here in traditional practice, for an archaeological knowledge creation, there's this primacy of gaze, I saw it. But in between those kinds of processes, whether it's engagement in the field or desk based mapping, there's this kind of Gordian knot, this black, this black box, where basically by the time we've created the outputs on the right hand side of your screen, it's because we, because I say so, process, we're aware it's inherently unaccountable. And this is an area which people have recognised in many kinds of ways. We all recognise the contingency of how we see different things we see colour, we hear things differently. So in there, there is that issue about how we see things, and for anybody who has any doubt of the kind of relevance To this to archaeology. At the top right, we have an eye tracking exercise, we would all like to kid ourselves that we can look at a scene systematically, not true. fixation is all devouring. At the top left, we see an archaeological excavation plan, which was given to two different people to reconstruct the pattern they saw. And this identifies one of our great strengths as human beings, we're very good at seeing patterns, we're also really, really good at making them up. And then the bottom two images illustrate aspects of interpersonal variation. Essentially experimental work, we don't on our end, to give multiple interpreters the same datasets, the same specification, to let them get on with it and to compare the outputs. And bottom line is that the outputs between different individuals can be quite different.

So in here, one of the things I'm really interested in is that relationship between how we embed a computational approach to data inspection, looking at LIDAR into our workflows. And this is an example again, from Aris top left you'll see a round feature, it's about a metres in diameter, a desk based interpretation identified as an enclosure with a level of confidence to which men, I'm not sure and I want someone to go and visit it. In the field. When we did go and visit in the field, the red lines at the bottom left, there are GPS recorded tracks of our field workers. And they had stopped around on and dismissed as natural. The eras Kramer's earlier version of her AI defined it as almost certainly around house. And when I sat at my desk with the visualisation at the top left, and just anybody who came in through the door, I said, Come over have a look at this. What do you think it is? Everybody said, Oh, looks like a round house. And we went back in March 22, this year and had a look at our group of four of us. And probably three of us were convinced that it was a slightly rough and ready Round House with one sceptic. And I think in there, there, there's probably no absolute rights and wrongs. But there are balances of probability. And I think in there, recognising that we all see things slightly differently experience and visual acuity, and so on will all play a role. Bringing an AI into that mix to me is absolutely fascinating. And where that takes me is that one of the crucial issues about AI and machine learning is that it does offer us opportunities to explore differently to build on that capacity of LIDAR data, for example, to be visualised in all sorts of different ways, which we cannot match. And that to, I suppose, adopt a bit of humility, about the primacy of our gaze, and ourselves as the ultimate arbiters of truth. So that in itself represents some very, very significant challenges to established practice. And there's a question about how we fit that into our workflows. That is all very much about work in progress. Essentially, this is a challenge for archaeal landscape archaeologists survey as a prep field of practice to think about how it embeds this into its routine work in the way that GIS was through the 1990s. I want to just round up with a couple of images, which illustrate where digital topographic data is providing us things that we would not see in any other way. The parallel lines that you see running roughly up and down the screen, indicated by the red arrow, are the very, very slight banks of a cursus monument not far from tour more on the west coast of Ireland. This is a unique find to the West Coast, unique to Aaron. These are mainly monuments that we see elsewhere, Stonehenge, for example. It's a pretty rough and ready object. But we would not have found this without the LIDAR data. And I'm saying that because I walked backwards and forwards across it any number of times without seeing it. So this is one of the things that the LIDAR data and power to visualise in different ways really does represent a step change.

And then going back to that issue about how we integrate AI machine learning and our field practice, this is a bit of work that we did In March, again, GPS recorded tracks, desk based assessments, the outputs from that the classifications that were produced during field visits 2019, the outputs from IRS Kramers AI. And what I just wanted to draw your attention to is here that we have quite a good candidate for a prehistoric the footings of a prehistoric, Round House, we missed it, the desk based assessment, we missed it in first round field visits. And this GPS track is a footpath that, again, I'm saying this about myself, I'm not being critical of other people, I walked within seven metres of this roundhouse a good number of times. And it's not because I'm an incompetent field where for quite a long time, and take care over it. But and when we went back to it in March, and we're gonna stand on it deliberately, there's no question it's there. So this is, this is this is the main point I wanted to try and make in this in this talk is in this aspect of working with digital datasets, the all of those issues about how you see how I see how an AI sees how we can train that how we iterate that feeding back between those processes, I think is an enormously powerful mechanism to bring Archaeological Survey forward. In concluding this, I just want to take us back to the stories. So you know, I think if we're just finding sites, we're doing something useful. They're on record, they inform other people understanding and are can be managed for the future. But actually, it's it's we want to be able to tell about the past. And this is a little bit of LIDAR down near strand rar, in southwest Scotland, which an intern brought to me say, What do you think this is, and I looked at it for a long time. And I'm ashamed to say that the penny didn't really drop that we needed to rotate the image so that North wasn't to the top to see that this was a big no written out. And this was it was it is within the red circle, essentially on the route that the Island ferry takes for Southwest Scotland. And you look at the date of the LIDAR acquisition, late 2000, sorry, in 2015. This is when Indy ref the independence referendum in Scotland was being fought over. And it's a piece if someone has taken a loan cutter on the back of a tractor and cut their campaign, not vote no. On to the side of the hillside. And the point I'm making in there is that there it's again, it's about in itself, it doesn't mean anything. It needs to be thought about in this case, we needed to unpick the map convention of north to the top we need to look at in a different way. And we needed to contextualise it and then feeds into one aspect of the narrative histories of 20/20 century UK politics, and Scotland's politics. So again, there's the story behind it is actually all important. And we live in a in a very in a changing world. And for me, these two landscapes and opposite ends of Scotland, a view of Glasgow in 1947, no question a heavily modified altered landscape, where the Historic Environment has been going through enormous amount of change. The image on the right is in Orkney, and is a submerged landscape. But this is a landscape which in many ways, has been as utterly transformed as Glasgow. And it's in telling those stories that I feel we're on the cusp of a change, which allows us to combine the muddy booted archaeologist with the high performance computer to great effect. Thank you very much.

Keith Myers

Amazing, that's so interesting. Dave Kelly, they're introducing us to the past using new technology. And these discoveries are obviously you being used to rewrite the record really, we've seen quite quite big headlines about the Amazon recently. These kinds of things, so please post your contact details on the on the chat as well Dave, that's great. Thanks very much