00:00.00 archpodnet Welcome back to episode one seventy four the archeotech podcast and we are wrapping up this discussion that got started a little late but basically automated archaeological feature detection using deep learning and on optical uav imagery preliminary results is the name of the article. Check it out in the show notes. Over at http://acpodnet.comforward/archaeotechforward/ 1 ne seven four if you want to see the article and the article links to all the supplemental material like paul mentioned and just like a lot of good stuff. So go check that out. But yeah, when we were headed to the break. We. You know we were talking about I guess the possibilities around this. 00:38.28 Paul Yeah, so you said use the term hitting a button just before we went to break and that that got me thinking That's actually ah, really apt in terms of this article because the authors are keen on accessibility. Okay, they they want. 00:39.36 archpodnet So yeah. Um, yeah. 00:53.73 archpodnet Um. 00:56.75 Paul The imagery to be accessible to be publicly accessible. They want the code the source code of their um of the project to be publicly accessible. They wrote about it even though it's in Preliminary format. They wrote about it in an open access journal. 01:10.30 archpodnet Over. 01:13.50 Paul Accessibility is a really big thing and 1 part of the accessibility that runs right through it is that most of the time when we've looked at like computer vision kinds of projects. It's very code heavy. Um I played with some of these before and you know you you get in the weeds pretty quickly working with. 01:22.35 archpodnet Yeah, yeah. 01:31.62 Paul Whatever packages they've got oftentimes. It's in it's in Python So I have some access to it because that's kind of the way my brain works but um, but it' still it's It's very code based and very code heavy and and depending on the programmer. You know what your different inputs might be what your variables might be that you have to set. 01:38.30 archpodnet Ah. 01:51.50 Paul What the settings are whatever you want to call them. Um might be very obscurely named um and that's not the route that they went down here. They they actually built a gui now. The gui is exactly what stopped me from working with it on my mac but again I'm not afraid of a piece. 01:56.32 archpodnet Yeah. 02:09.41 Paul With a little bit more time I'll either get it work on my Mac or I'll get it working on. You know one of the zillions of linux boxes I've got around I just don't have in front of me today. Um, and they explain what a few of the settings are and say you know you could do a setting of 100 and water. You could do a setting in 50 depending on. 02:16.16 archpodnet Sure. 02:27.95 archpodnet M. 02:27.97 Paul You know these various parameters and that's in the article that they explain not not in the but basically it's just it's picking a number off of ah ah in a dialogue box on a gui and then the training part. The biggest part of the training that happens is human interaction saying. 02:37.11 archpodnet Yeah. 02:47.11 Paul You know, tracing the outline of kind of feature and saying hey this is that kind of feature and here's another image and here's that same kind of feature and here's a new outline and then letting the computer do the work of trying to interpret. What's the same about these various photos. What's the same about the features that have been highlighted with them and so it. 02:49.53 archpodnet Oh. 02:59.96 archpodnet Yeah. 03:06.17 Paul Again, back to the the notion of accessibility. It's trying to be accessible by not hiding because the source code is there but not forefronting the code making it so that with minimal training you or I could go and. 03:17.20 archpodnet Yeah. 03:25.97 Paul Take this and train a set of images and see what kind of data we get back again. It's still in process. So there's I think a lot of refinement that has to be done on the actual process of the machine learning to identify these properly and which is something else. They talk about you know the hits and the misses. Ah, but. 03:38.65 archpodnet Yeah. 03:44.48 Paul That's that's a positive way to go I mean you and I are both Apple users right? And that's where max first became very popular was as the simple things to do that. You didn't have every setting under the sun which really upsets certain people but most of the time most of the settings are there. 03:46.60 archpodnet You know. 04:01.63 Paul In the way that most of the people are going to use it. 04:01.97 archpodnet Yeah, yeah, which is the whole point right? Yeah so and and that's yeah, that's where I would love to see I Hope they really get this done and they can get enough data in there to actually make this thing happen which was kind of the other huge point of this article. 04:16.51 Paul Me yes. 04:21.41 archpodnet There's not enough imagery to to put into these models for that to learn I mean they need they mentioned even at one point I think hundreds of thousands of images of different things in order to get the the neural network the convolutionary neural neural network to actually learn what these are and and we all know that's true because you could have. A hundred pictures of the same thing and it's all going to look slightly different right? So we need different angles. We need different. Um, you know, different conditions different different all kinds of things for that to work out so that was one of the things I like that they mentioned is that they want researchers to share their data and for academic researchers. 04:43.29 Paul River. 04:59.76 archpodnet Even incentivizing them with ah you know I guess credit academic credit for actually sharing the research not just completing the research but you get a little extra if you share it in open repositories. Yeah. 05:08.44 Paul Um, yeah, and then for um, for reference they they looked at 3 different kinds of sites so they looked at structures. 05:20.30 archpodnet Yeah. 05:23.30 Paul They looked at what they called Mound sites which they didn't describe in great detail but I'm assuming they mean tell sites which is you know, particular site formation process. That's very common in the middle East Oh that's the other reason why this interested in me because their examples are all. 05:27.44 archpodnet Ah, yeah. 05:39.78 Paul In Middle Eastern archaeology arabian archaeology so stuff I'm familiar with from my own training and stuff that interests me um and then the other kind of sites they did structures I've mentioned um Kaats Kaats are a kind of. 05:52.80 archpodnet And. 05:59.21 Paul Aqueduct basically that's common on either side of the persian gulf and then also I've seen them in in Yemen so in Southwestern Arabia as well and so you can kind of imagine a sort of East-west band across the southern tip of Arabia over into Iran. 06:08.21 archpodnet Yeah. 06:18.40 Paul And it's an underground aqueduct that every so often you have an access hole that goes up to the surface right? So it'll be a very gradual slope to the to the channel that's underneath the ground and then periodically every fifty feet one hundred feet whatever it is a whole. 06:24.79 archpodnet Um. 06:37.27 Paul That that goes back up to the surface for digging the canal and for keeping it clean and interestingly enough of the 3 different kinds of sites that they were picking. That's one that I thought was going to be the easiest to identify because it's circles. It's holes. So it's a very dark center. 06:37.61 archpodnet So. 06:56.82 Paul And they're linear. Ah and they're linear than they can stretch for miles know and that's one that they had lots of Miss hits right? things that that they had that were getting misidentified as the wrong site not being identified properly so that does suggest that there's um. 06:57.30 archpodnet Yeah, yeah. 07:15.32 Paul There's a lot of room for improvement there on the the actual algorithms of that that neural network How they're building that out. Um, way above my pay grade to be what's going on. Ah um I know some of the terms and if you ask me to actually describe what they mean. 07:21.40 archpodnet Yeah. 07:27.66 archpodnet Yeah, yeah. 07:34.84 Paul All sound like that you know that kid that didn't read his homework and is trying to write the book report. Um, which is basically what I am right now with this? um ah, but but it was interesting to see these different things and so they had. 07:36.63 archpodnet Um, nice. Yeah. 07:49.42 Paul You know, roughly one hundred to 200 photos that they ended up using of each of these different kinds of structures or sites or other sites and structures. Um, so. 07:53.98 archpodnet Um, yeah, which which that in is in and of itself that they were able to gain some you know a little bit of success with with that limited number of photos is pretty encouraging to be honest, Yeah, so. 08:06.21 Paul Um, yeah, and then that brought up me questions about the the difference if you had more photos would you want more photos the same few things so that you can really dial in what you know this particular kind of structure looks like this. 08:11.63 archpodnet I Want to talk? yeah. 08:24.41 archpodnet On her. 08:26.00 Paul Particular got not and or do you want a whole bunch of different things showing the variability which which is the better strategy I don't know they don't go into that but it is something That's always in the back of my mind. 08:36.74 archpodnet Yeah, well in the last few minutes of this podcast I want to talk about because they they briefly talked about you know where future directions should go here but I want to talk about the far future and and what this could mean um and what I mean by that is. You know in the last week or so I was finishing up a book and I might have to look at my audible account and see what that was I don't remember what the name it was but it was about artificial intelligence and what designing artificial intelligence even means right? like what kind of rules do we apply and what are the if you really think about these experiments like what. And and the rules that you could give an artificial intelligence around service to humans and things like that. Well, where could that go wrong and you know not in a weird sci-fi way but in a realistic way where where can this actually just go wrong and it makes me think not really down that road but this is leading to some sort of an Ai. 09:22.86 Paul Um, in. 09:31.27 archpodnet You know a computer program that says I'm looking at all this data and here's all the things that I think are features right? But once we teach it that and it gets really good at that and people start using it more and we just keep feeding it and feed it and feeding it and feeding it like anything you feed too much. It'll just get way too big and then now. It's got all this information. So what's to stop it from starting to you know, stopping people from starting to input information like actual interpretation. You know say well every time we see this it means this and every time we see this along with this it means this so that's not too far fetchched which means. You know at some point what's the archeologist for you know I mean like we send out the drone data like the the the development company here in the United States says we're putting up a ah super mega walmart um, over the entire state of Idaho so we need to ah you know. Go out there and do some survey because that's still a law luckily in 2300 and it sends out the auto drones or these super high resolution satellites that they can access right away and then brings back all that imagery to the computer system. It says we found this this and this it sends out things to excavate or mitigate and. 10:31.30 Paul 50 10:45.95 archpodnet And and then you're done like like the real jobs that Crm archeologists have is really just to help construction companies know what to avoid and if they can't avoid it then we have to excavate it.. That's what an archaeologist really does when it comes down to interpretation and stuff like that. It's not even really technically part of our jobs I mean it kind of is to an extent but not really. And I'm just wondering if the entire job of an archaeologist could be taken up by this in the Future. You know. 11:13.55 Paul I'm going to bring our question I'm not going to even try to answer that I'm going to bring another question that's ah in the closer future because another thing that struck me reading this It's Uav's and we normally think of drones. We think of uafs and we normally think of quadcropter drones in particular and I was like well most of the work I've seen with Quad Crop Quadcopter drones is site level right? It's not to identify where the sites are. It's to look at features on a site. 11:27.82 archpodnet Yeah, yeah. 11:43.17 archpodnet Um. 11:45.75 Paul Um, but this article is clearly geared at larger swaths of of landscape. Yeah and I was wondering how to how to how this works you know? So do we need just like you know old Spy photos. You know the Coronas um. 11:55.24 archpodnet In here. 12:02.72 Paul Would it work with satellite imagery is the is the resolution. The detail good enough now I don't know but most of the drones aren't going to be flying high enough to get a good landscape view and then when I clicked on Mark's ah 12:06.54 archpodnet Of her. 12:12.96 archpodnet Um, yeah. 12:18.26 Paul Twitter account just to see if I could find that article I didn't find it but I'm sure that's where I got it from I saw a photograph that he'd posted on his account of a fixed wing drone and I thought of course duh. That's how that's what he's meaning by the uabs and this this thing was. 12:28.65 archpodnet Um, yeah, yeah. 12:38.10 Paul Terrifying I mean it was in the office there and it had at least a ten foot wing spec supposedly can be up in the air for 4 hours so so when he writes about ua it's it's on a slightly different scale than what than. 12:40.56 archpodnet No my God Oh yeah. 12:54.21 Paul What I've been using so far. No from. 12:54.67 archpodnet Yeah, well landscape um Landscape Imagery should always be done with fixed wing at this point right? because the quadcopter drones. Yeah, the quadcopter drones are just too. 13:02.70 Paul Um, I would think yeah, um. 13:08.35 archpodnet Heavy really? ah you know because they don't have any real flight characteristics. They're purely being held aloft by the force of the engines or the motors I should say and exactly right. It's just a brick but the the really nice u a in fact, trimble I think makes 1 trimble makes a fixed wing Uab that has rtk. 13:13.94 Paul Yeah, it's a brick. 13:27.95 archpodnet Ah, submeter rtk accuracy within within the device and it's got something like a 6 or eight foot wings span I saw it at a thing I was at a few years ago and um, but fixed wing drones are the way to go now they are subject to a lot of the same limitations of course that physics puts on aerodynamic things like there. They're really light usually and they have electric motors not gas motors which that lightness of them obviously makes them really subject to wind and um, you know with fixed wing. You do tend to go a little higher and you go a little farther away and you go for a lot longer period of time so you really have to be cognizant of what the wind is doing to you. 13:48.50 Paul Susan Isn 14:00.67 Paul Um, yeah, you also go with the higher you go faster so you capture more in the photograph but you have to go higher so that at that speed you don't get motion. Blur. 14:08.58 archpodnet Yeah, yeah. 14:16.30 archpodnet Exactly? Yeah, So but I think I think all those problems are going to be solved just by increasing technology in other fields right? And like we get batteries that are lighter and faster. We get. Ah. You know materials that are better so we can probably make the the thing a little bit heavier because we got better batteries and motors and then we get better cameras so you could fly a little bit lower. But maybe the camera's so good that you can fly higher and it's got ah just ah, an amazing resolution that the computer can see and then you're talking about satellites you know is the resolution Good enough. 14:44.30 Paul E. 14:49.27 archpodnet Yeah, it's probably not what the ones we have access to but we know damn well that there are government satellites up there that can you know count your nose hairs right? So um, as soon as we get access to stuff like that because they've got even better stuff then that's going to that's going to change everything I think actually access to satellite imagery. That's not. 1 of the things I mentioned in the article is 1 of the downsides of satellite imagery is. It's not representative of now it could be representative of a year ago or two years ago yeah right so if we have access to real time satellite imagery or at least near real time like within the last few months and like we can. 15:13.37 Paul Oh that was a good point. Yeah yeah. 15:27.30 archpodnet Book time on it and say hey you know it sends you back. An email says great the satellite passed over and took all your imagery from this date then and and it's within the last like three or four months that would be fantastic and then you know we would be able to ah probably not use drones ever again for landscape photogrammetry landscape imagery like that. So I don't know. 15:45.68 Paul Well I'm going to throw another wrinkle that this wasn't mentioned in the article though I did ah really appreciate that that concurrency of the imagery you know your ua v stuff most archaeologists are using them. It's photographs that were taken today yesterday during the project as opposed to whatever. 15:47.00 archpodnet Exciting time. 15:55.20 archpodnet No. 16:01.99 archpodnet Um, yeah. 16:05.20 Paul Other imager you have that might be you know corona stuff from the sixty s it might be you know lances that stuff from the 70 s or 80 s but is definitely not current. Ah 1 big advantage that was not mentioned in the in the article of the ue vs of the drones is that you're generally under the weather. 16:11.76 archpodnet Yeah, yeah, yeah. 16:22.59 archpodnet Yeah, that's good point. 16:24.96 Paul So satellites you know that kind of real-time satellite stuff and that's used a lot for um, for forestry for agriculture for a lot of that sort of thing um is very subject to Cloud cover and so you might have a stretch of time that you don't get good imagery of the ground because of Cloud cover. 16:36.49 archpodnet So. 16:43.42 Paul You don't have that problem with the with the drones generally speaking. So you there's they're pros and cons to everything. 16:44.34 archpodnet Yeah, well, that's and that's probably why as we've said on virtually every single episode. It's all about a suite of tools at your disposal right? So you know use the one that you know. 16:57.59 Paul Oh yeah. 17:02.95 archpodnet Might be the most reliable use try to use that the most and then you know use other things to fill in where that's less reliable like I would imagine high resolution satellite data would be the most reliable thing if we had access to that kind of satellite data. But then also like you said cloud cover different conditions. Um, you know whatever the case may be. Then you can fill it in with other stuff. So or maybe right now drone is the most reliable thing and we can fill that in with satellite imagery data in certain circumstances where I don't know maybe for 1 reason or another we just couldn't get good drone imagery of it. So um, yeah, like for example, places that we want to study that are having. 17:28.76 Paul E. 17:39.99 archpodnet Hostile action right now or or you just can't get in there for various political reasons. But we have good satellite data and we can do that I mean you get into bunch of ethical concerns there that have to be navigated but you know it's um, it does open up different avenues for research so all right. 17:40.25 Paul Are a. 17:47.70 Paul Music. 17:58.14 archpodnet Well I think that's ah about all we have time for again if you have ever had any comments like our friend James please. Ah, you know, send us an email Chris at http://archeologypodcastnetwork.com um paul@lugall.com both of those are in the show notes for every episode. And then so are other contact information like our Twitter handles and um I mean leave us a comment or something on Itunes like a review but don't try to like ask us a question there because nobody reads those like I don't even I look like once every six months and it's not a good place to ask questions I've seen questions in other podcast reviews they're like hey on this episode this this this and this and like that's. These are for reviews not questions. So um, yeah, otherwise anything more to add Paul. 18:40.68 Paul Um, just the 1 thing is that I'm glad that this exists because I'm going to explore it a little further not on the site identification. But like I said at the start I think this might be a good intro for me on doing intrasight. 18:50.44 archpodnet Um, yeah. 18:58.80 Paul Inspection with my aerial photographs. So yeah, it's ah I'm glad it exists. 18:59.75 archpodnet Oh. 19:05.30 archpodnet Yeah, sorry ah I'm just I'm just vamping at the end here. Did you ask a question at all because you broke up for like a good solid 10 seconds just wrapping it up. Okay, okay catch got you? um. 19:17.77 Paul Yeah, yeah, just wrap it up. 19:24.45 archpodnet Yeah, so with that I think we'll be out and Paul you are headed off to Iraq. So this is the last episode with you for a little while we do have an interview coming up talking about Ai and robotics that's going to be pretty cool that's going to be in the end of March we'll we'll release that but I'm sure we'll have some other episodes. 19:31.86 Paul In. 19:40.50 Paul 2 19:43.25 archpodnet Regardless of Paul's absence and then we'll get to talk about Iraq when he gets back. So with that I think we'll see. 19:46.66 Paul Yeah, hopefully I've got some good ah some good new data to talk about. 19:51.53 archpodnet Ah, yeah, absolutely all right? Well with that. We'll see you guys in a couple weeks. 19:56.14 Paul All right take care.