Automating Design Workflows Using ArcGIS Machine-Learning Tools and Aerial Location Intelligence

DAVID GARRIGUES: Hi, everyone, and good day. Welcome to our class. And we're going to show you how to use AI to extract features from high-resolution imagery. Let's get started with a few introductions. I'm David Garrigus. I'm the head of engineering applications here at Kimley-Horn. Been here for about 17 years. And a fun fact about me, I'm a big time foodie. My wife and I, we love to cook, and we love to dine out. So if you are one of those, hit me up. I'd love to talk about it. Jeff.

JEFF SAUNDERS: Thanks, David. Hi, my name is Jeff Saunders. I'm a director of product management at Nearmap. I've spent over the last two decades building products and solutions for the built environment. And although I know many of you have beaten this number, this is my 12th Autodesk University. So excited to be here, excited to share our presentation with you. Brett, unto you.

BRETT HEIST: Thanks, Jeff. Yeah, Brett Heist. I'm a solutions engineer here at Esri. And this is actually my first AU. And excited to be here supporting and presenting along with these two wonderful gentlemen here. I've been in these sectors for a little over a decade. And a fun fact about me is I once got the opportunity to serve John C. Reilly ice cream. Back to you, David.

DAVID GARRIGUES: Excellent. Thanks, Brett. And an honorable mention, he couldn't be here today, but his name is Kyle Starkes. What a phenomenal developer. He's a solution engineer over there at Nearmap. And he couldn't be here today, but he's the one who helped build the tool, and design it, and make it all happen. And so he couldn't be here, but I did want to make sure he got an honorable mention.

With that said, let's talk about what the agenda is. So first of all, I'm going to start off with, what was the big idea? How did we get here? What happened behind the scenes? And then what's going to happen is we're going to move on to Jeff. And Jeff's going to talk about laying the groundwork with Nearmap, and high-resolution imagery. And making AI extractable features possible. Lastly, we're going to end up with Brett. Brett's going to expose the opportunities that we have with AI with Esri. And then we'll have some time for Q&A. Or you guys can email us at any time. So we'll show that at the very end.

So the big idea. So yes, I've been with Kimley-Horn for over 17 years. And am responsible for all of our engineering software, regardless if it's quality, deployment, programming, training. But what I'm most proud of is the opportunity to be able to turn a software vendor into a software partner. Partnerships involve work. And they involve the work towards a common goal. So when we talk about these things, we need to understand that we have to keep working together, and you have to dedicate time to each other.

And that's what you're seeing here today. This is the result of time well spent. So when we take a look at Nearmap, we had to go and enter into a new type of partnership, a venture, if you will. And then, later on, do the same thing again with Esri. Now, Nearmap can absolutely deliver high-end imagery that is needed for the work that we're going to need to have start with. To be a great partner, we simply cannot satisfy our own needs. We must truly work to be a spokesperson for the industry at large.

So starting with Nearmap, we needed two things. We needed to be able to extract features from an image, such as pavement, car spaces, landscaping. But then we also needed a faster way to download high-end imagery in larger scales. So let's take a look at that.

So what we're seeing here is-- currently, if you go down-- if you go up to Nearmap's website, you can see that I'm starting off with this 0.075 pixels over here. The problem is that as I increase the size, so does the resolution drop. So you can see now I'm at 0.299. And the bigger I get, now at 0.597. So how can I get the highest resolution possible? Well, currently what we had to do was we had to go and set it, and then download the files.

Then move it over a little bit. And then get that overlap going. And then download the files. And then move it over a little bit. And then get the overlap. Then download-- you guys get the picture. This is over and over again is what we had to go do. Is there a simpler way? And so the answer is, yes. But we have to do some programming. Now. I've done a lot of programming in my time, but this is a little bit above my head. But if you really wanted to know how all this happens and what it does, well then, pause here and talk to Kyle.

But this is how it all works. And on our side, Paige Dyer, she couldn't be here either, but Paige was the one who worked with Kyle and got all this working. So with that said, what did we go and do? Well, we had to go and invent this tool. So this tool is allowing us to go and make a square, just like you showed before. So in order to do that, this right here is going to have to allow us to go out and go grab the lat and long. So to do that, just go out to Google. And you can right click on any spot and it will give you the lat and long.

All you do is pick on it, and it will copy that to the clipboard. And then we go back into our tool, and we can paste it in. By moving the cursor down into the next one, it automatically puts anything after the comment to the next one-- next line. And then we get a radius. So while we're defining a radius, we, Kimley-Horn, put a max of 3,000 on our people. Next thing is the EPSG code. So what's an EPSG code? Well, we have to have a translator from web Mercator back to NAD83. So in my case, I just went to Google, typed in EPSG code for North Carolina, and then I ended up with 2264 for NAD83 foot.

You can do the same thing. There's lots of sites to go and do this at. But I come back over here in my tool and I type in the EPSG code. Now, I've got three formats. JPEG, PNG, and TIFF. I can assure you you're going to want to use JPEG. So that's if we wanted to go just to develop a square. I know we said radius, so the radius fills out a square. But what if we want to do an irregular shape, something like a polygon? Again, I still have to put in my EPSG code.

But this time, what I'm going to do is I'm going to use a tool from geojson.io. And what I can do is I can type in Raleigh, which is where I'm located. I can type in Raleigh, North Carolina. I can even type it in wrong, it'll find it. So fortunately for me. And so then what happens is, as I zoom in, what I can do is I get to determine, how do I want to create this new area? I can do it circular. I can do it with a rectangle. I can do it with just a line. Or in this case, what I'm going to do is draw a polygon.

So now, these polygons are editable. So if you miss your pick or whatever, you can go back and do that. Now, here's the reality. Kimley-Horn we put a stipulation that we were not going to allow our staff to go more than two miles by thousand feet, which is going to be equal to about 10,000,000 square feet. All right? So we put that stipulation on people. So because we don't want everybody to go download the Earth.

So now what's going to happen is you can take a look and see what all the info we have. And you can see that I'm underneath 10 million, so I'm in good shape. And that was something that we did, not Nearmap. But we put that stipulation on our people. So now what's happening is I got the code over here, and I can copy it. So little button over there, when I press that, I've copied it to the clipboard. All you got to do is go back into our tool and paste it in.

When I do, I just get to pick where I want to go. And so I'm going to put it in my AU folder over here. And then I can pick the output. Now, because I've got Movie Magic going on, this normally takes about 15 minutes, so-- but hey, I just shortened up the clip for you. So now download complete. What's the next step? Let's talk about Civil 3D.

So when we go into Civil 3D, one of the first things we're going to want to do is we're going to want to go and change our drawing settings over here. So in my case, I happen to know my code. NC83F is my code, so I go and set that for the entire drawing. Now, what I've also done is I went into map aerial, so my geolocation for that right there-- by the way, that's going to be going away soon. And Esri's is going to pick up the tab for us. So thank you again, Esri. I'm using the I Insert command over here. And as you can see, I've pumped this image out several times here.

It only gave out one image, by the way. But I practiced in things, so we've got several here. So I bring that one in. And inside here, it takes a little while to get it going. This is a pretty big image. It's like 100 meg. Hundreds, maybe 115, 120 meg. So I'm going next, next, next. And then I'm going to zoom extents. And then I'm going to see myself down over here. And I can see I've got some black areas. Which is fine, because I didn't need that much data.

You remember how my shape works? I'm going to go over here and set transparency for the image. And I'm also going to set transparency for the color. So I'm going to match that to the black background. So I'm picking that in the black background. And then it takes a while to chum it up a little bit. But then I get the image that I wanted, and I can see transparently. And I can see the images in the background as well, so making it transparent. But how good is this resolution?

Well, if we zoom in a little bit, we're going to see the difference between what Bing has and what Nearmap has. So here, you can see payment cracking. You can see all the paint stripes just as well. I mean the resolution's incredible. So this is where we want to be. So if we take a look at the overview of what happens in a project, we all know we can sit there and take our data and we can go and get images today, and things like that. But now, I've got something else going on in the existing conditions.

I can now kick this out to ArcGIS Pro. I've got some opportunities inside Nearmap as well for exporting out these features. I can go out and go find a building footprint. Maybe go find out some pavement. Go do those kind of things. And now, I can take those-- once those assets of features have been identified, I can move them straight into my design. With that said, let's hear more from Jeff.

JEFF SAUNDERS: Well, thank you, David. This is a great handoff here to really start talking about how we looked at laying the groundwork, and laying the groundwork in developing projects for the workflow that David just showed. So let's start with some of the key foundational data. I think David said it really well in that having high-resolution aerial imagery at the start of the project, making it the foundational start really helps to define what can be done with that imagery. What information on site conditions can be used at the outset.

So let me start with talking a little bit about some of the remote sensing technology. So there are a number of different remote sensing technologies out there, from satellite to drones to ground surveys. Nearmap's approach is somewhat unique in that we design and build our own camera systems. And we capture imagery from manned aircraft. What that gives us is this sweet spot at 2.5 centimeters, 7.5 centimeters GSD, that allows us to see a lot of detail with amazing clarity of that imagery.

And it's important to note that we fly this all of the urban areas in North America on a proactive basis. So we have a program that runs annually. We capture multiple times a year across all of those areas. And that allows us to do a lot of exciting things to provide on-demand, as-needed data to projects as you're starting them. But what is 2.5 to 7.5 centimeter GSD really look like? Let's take a look at a few examples.

So on your right here, that's one of the captures we've done at 4.5 centimeters. And on the left, similar to David's example that he was showing side by side, there's the satellite imagery at a native 30 centimeters. Now, there's a lot of great work being done in satellite AI enhancement. But again, here's a comparison between an enhanced 15-centimeter image and an aerial image from Nearmap on the right-hand side. So you can see the clarity. You can see the crispness of the imagery, and what you can do with that.

But let's dive in a little bit deeper. So if we look closer at each of these at this location, you can start to see the tiles. You can see a lot of different roof-related objects here. You can see a lot of architectural artifacts. So this really does become a very powerful set of data as you're looking to start your projects out to really understand site conditions, to understand what's out there currently. And to be able to use that from the start of your project. And also be a foundational piece for AI-derived extraction, which we'll talk about a little bit more later.

Let's talk a little bit more-- and this speaks to the accident that happened earlier this year that affected the civil infrastructure in the Baltimore area. We worked with responding agencies and provided our imagery. As it turned out, we were proactively flying this area at the time as well. So it did coincide with that, and enabled us to really be quickly supporting the accident as it unfolded. So we talked earlier about capturing vertical imagery and high-resolution aerial imagery from that perspective.

Nearmap also captures oblique imagery, which helps us do a lot of exciting things, and support a lot of events like this one as they show up. So this allows us to really zoom in, see some more detail about what the impact of this accident looked like. And where different structural impacts may be happening, how to best support responding crews as they're addressing the issue at hand. Let's change gears one more time here. David showed this as well.

But here, we're looking at an airport and looking all the way down. As we zoom in further and further, we can see cracks on the runway and on the taxiway here. And that allows us to do a lot-- make a lot of smart decisions early on in the project. So again, getting crisp information on the site conditions, on what exists out there in the field is really valuable and an important starting point. But as we talked about earlier, as David alluded to as well, there's a lot more that we can do with this data.

Since we're capturing not just vertical imagery, but also oblique imagery, it allows us to build out 3D data as well as AI-derived insights and terrain information that can really be key components to starting your engineering projects. So I mentioned vertical capture imagery. We have that. Oblique imagery. We're capturing that as well, which allows us to create a derived panorama view. As well as 3D data, and AI-derived insights. We also have a post-cat focused product that allows us to capture when certain events post-catastrophe events happen. And we're able to help responding agencies and others respond to those claims.

But in terms of 3D data, this allows us to create, essentially, plug-and-play data for your projects around with a textured mesh with point cloud. And you'll probably see some examples of that here at Autodesk University. Our point cloud within recap and InfraWorks, digital surface models, digital elevation models to start to work with contours and terrain at the beginning of your project. And a true ortho 3D-derived imagery, which allows us to correct for some of the lean and buildings that may come from vertical imagery.

David mentioned this earlier, too. But I mean, a key piece of all of our work and the partnerships that we forged with Esri and Autodesk allow us to create this plug-and-play data sets and insights that can be used across the building and infrastructure project lifecycle. And that allows us to bring data directly into Autodesk, to bring data into Esri and share it with Autodesk. Really helps us to support the full gamut of workflows that would come into play for building and infrastructure projects.

So we're going to look at two Kimley-Horn projects as sort of examples. I'm going to talk about an example here at the Orlando International Airport. And Brett's going to tackle the sphere, because it wouldn't be an AU without at least referencing Las Vegas at some point. So we'll get to that in a minute. But let me start with a video. This speaks to the Orlando International Airport. And so here, we're in Nearmap's Map Browser product to look at where the imagery is.

This is very similar to David's starting video. But there's a few things that I wanted to highlight here. So we're looking at an area in the southern portion of the Orlando airport. And this is a project that is in place today. But if we wanted to use a time machine and pretend that we were back in December of 2017, we'd see that there isn't much here on this property area. But we might need to use that as a starting point for our design project. And from this, we get, obviously, the vertical imagery on the site conditions.

Now we can compare them side by side. But in the simpler form of what David showed, I'm going to just pull a set of images and 3D data from that location for that capture area. And you can use this as a way to look back at historical data, or to look at current conditions as they stand today. And so this really allows us to extract this content. And allows us to choose what types of data we want to get, and at what accuracy.

Where does all this come into play? So as you're looking at automating your design workflows and looking at how to get really good site conditions and understand all the information around a project at the start of a-- at the beginning. There are a lot of long lead time deliverables where this imagery can really come into play. Whether it's site suitability studies, environmental studies, traffic studies, or a broader stakeholder review and commentary. This content, whether it's 2D, vertical imagery, oblique imagery, or 3D-related data and AI can all help you look at recent up to date data.

Start with topographic data for your conceptual designs. Look at 3D and oblique contextual information to more broadly assess the environment that you're going to be working in. And ideally, with AI and other tools to reduce and eliminate some of the manual drafting tasks that may exist. Let's take a look. Obviously, the DSM is a nice data set. And it can be really useful. But what is the common question everybody gets when they start a project? Where do I get my topo?

So here's an example, obviously, bringing this data into Civil 3D. And I'm going to use the ArcGIS Pro AutoCAD toolset just to show this example. Because Brett and I spent some time curating this data and we built a project-- a group within ArcGIS Online that allowed us to curate some of the key data we wanted to look at. So I'm going to bring in the aerial imagery, the vertical imagery here. And then I'm going to bring in the topo as well.

And so I sped this up a little bit here to save time. It only took a couple of minutes. But in the interest of the presentation, I wanted to make sure we got to it. But basically, we're going to have contour information showing up there in that purple-magenta color at the bottom. And that really highlights where this project is starting in this case. So we're going to get topo, we're going to create a surface all from the data captured from the aerial imagery. So we get base map data, and we get terrain data, and contours here.

So I could also bring in the satellite imagery using the FDO tools as a WMS, or WMTS as well. If we want to stream that information, that's another option available. But in this case, I think it was important to highlight the Esri, Autodesk, Kimley-Horn, Nearmap partnership. And how this can produce results for you, and how you can use this information. So again, here we have the contours. I'll turn off the base map for this example. And you can see here we have a set of contours. And we can just verify-- well, actually, the first thing we need to do in this case, since we brought the data in from Esri, is we should look at assigning the elevation attribute field to the contours itself.

So I forgot to turn on the pop-up menus here when I ran this one. But here, I can assign the attribute of elevation in there, and assign a Scale Factor 1. And basically now what I'm doing is assigning that attribute to each of the contours so they'll each have their own elevation. What this allows me to do-- and I'll do a quick list to show you that it actually did work. But I'm going to create a surface first, and then I will prove to you before I create contours that I'm actually pulling data that has elevation associated with it. So we have a z-value associated with that contour line.

Now I'm going to add those contours in. This is Nearmap data from December 2017. So we'll start with that. And then we'll let Civil 3D do its-- Civil 3D do its magic, and just approve that we now have-- once we select all the contours in this case. Once we do that, I'll just prove to you that we actually do have a surface, we have a TIN. And we can use this as a starting point for our project. All in pretty quick succession without even having to leave your desk, in this case.

So I'll turn on the TINs here. Just make sure that the TIN is visible through the display manager. And I'll give you just a peek at that TIN in a 3D view. So here we go. We have that TIN we have a 3D surface. Maybe not super exciting with a lot of rises and falls, but a pretty good data set to get started with. And with this, now we can start our projects and start to do some of our conceptual design as needed.

But wait, there's more. So David mentioned this from the beginning. But some of the other things that the Nearmap provides is mature machine learning models that allow you to derive deeper insights. And ideally, save you time from some of the work of digitizing or some of the work of identifying certain conditions or certain material types. So all of this plays into helping to automate and speed up the design process.

We also are able to provide-- and this showed up as in the imagery, but ability to identify pavement damage, looking at different pavement-- looking at both raster and vector perspective here. This is information you could pull into your design, into your planning model in Esri or Autodesk as key starting information. And then here's another example of using the vector detected pulls, or the vector version of the detected poles. And that can be, obviously, a starting point as you're looking at utilities and other public works related projects.

But there's a lot of other layers that we provide. This is just a short list of some of the main ones. There may be different use cases that you're looking to solve. And these can be these can come in really handy. But we know, at the same time, these don't solve all-- or answer all of the questions you may have. And that's where we want to do the handoff with Brett and have him show you some of the exciting work that the aerial imagery from Nearmap and the content from Nearmap, as well as the AI detections, plus the work that Ezra's been doing, really come into play to automate your design workflows.

Brett, over to you.

BRETT HEIST: Thank you so much, Jeff. I just wanted to start off by just expressing again, how excited I am to be a part of this presentation today. And am truly excited by the work that Nearmap and Kimley-Horn have done. And this tool that has been born of this partnership to allow Kimley-Horn to more easily access and download and import the high-res imagery that Nearmap provides. I think that this is really exciting, again, because it is going to provide a lot of opportunity downstream to do more with this imagery.

And it's becoming not only just a base map and a core element of the project and design lifecycle, but also it's becoming a source of data. I think this technology has finally advanced to the point where it's gone from being interesting to valuable. And that's what I wanted to continue to explore here today and see how we can leverage GeoAI and ArcGIS to again, continue on with these workflows and add value to them.

But before we hop into that, I did just want to take a brief moment just to stop and define GeoAI, and what it means when we say that. So when we talk about GeoAI, we're talking about two different concepts here. We have AI and this ability to have a machine, or teach a machine to learn and do human-like tasks that were traditionally not accessible through a machine to do things read, and see, learn, analyze, and create.

And when we look at this now, there's the subsets where we're able to take that concept and that framework and these capabilities, and do more specific tasks. Like machine learning, where we can feed specific data sets into a model to have it learn specific patterns. And deep learning, which is even a more specific subset of those previous two. Which you can think of it as like functioning like a human brain, where the computer is really learning complex patterns and concepts by piecing together simpler concepts.

And it's really not until we then marry these two together with spatial analysis that we get GeoAI. So it's leveraging these capabilities of artificial intelligence with spatial analysis to not only generate and do things like feature extraction, but also do something with those as far as analysis goes. That can help us with making decisions, and asking questions, and getting answers. This is also where the analytic engines that are available through ArcGIS, again, provide that added value for us to be able to do something with this data that we're extracting.

So whether that's further image and raster analysis, network analysis, connecting to real-time feeds, the platform really lends itself to continuing on with that feature extraction again and providing additional value to what we're actually getting out of that imagery. So if any of David like I do, when I first started talking to him about this, he was like, this is all well and cool, Brett. But what can I do with this? What can I do with this today? That's the real value for me is if I can actually use this and not just talk about it.

And so that's where I wanted to start as we transitioned into the use of GeoAI within ArcGIS. And that's with our pre-trained models. And so as you can see, we have a lot of models. We're continually adding to these. And we support a lot of different sectors and use cases for these from public safety and transportation to utilities. And these are really great because, just like we saw in Nearmap and their mature AI and machine learning models, that a lot of the work has already been done for you to where you're just able to leverage those results. And point these at some imagery and get back some features or some specific classifications, or things that you would like to have.

And to get started with these, it's really easy. We have all these deep learning packages available to you. And you can go here to the Living Atlas right now and type in those magic keywords of DLPK, and you're going to get a return. As you can see, we have 79 current models that are available. This is a great starting point to get a better understanding of not only what's available, what's possible. So from feature extraction to pixel classification, object detection. These models can do a whole lot of different things.

And once you've maybe identified one that you're interested in, this is also where you can find more specific information about that model, as far as the description and what it's really intended to do and its use case. The licensing requirements that are required for you to be able to even run this model within our platform in ArcGIS. We'll have some links to things like step-by-step guide on how to use the model. And even explanations for the parameters and the arguments that it can take. So you can better understand not only how to use this model, but how it's going to work.

And then most importantly, the input here. As you can see, we need that high-resolution imagery. And that's where this partnership with Nearmap is really going to shine and give us access to that. We'll also get some information about the output, the applicable geographies where it's been trained. So you can expect where to be successful with it. And with that, even some accuracy metrics to know how good of a results and return that we're going to get when we're running this model, along with some samples. So we can again, have level expectations going into this to not only know what we're going to get, but how well we're going to get it.

I think right now, the top four models that-- you saw, we had 79 within there. I think currently the top four are these that you see on your screen right now. We have the Segment Anything model, which if you're not familiar with, was integrated and born out of Meta. And we brought that into our platform and made it available as a deep learning package. And this model does exactly what the name says. It segments anything. And is a really powerful tool for being able to, again, get objects out of imagery.

We also have Building Footprint Extraction, Land Cover Classification. And the one that I love and most excited about and we'll talk about here in a little bit is the Text Sam, which is a spin off of the Segment Anything model. But it's been integrated with a large language model-- large learning model, LLM, to provide us with a text prompt. So we can be more specific in what we're segmenting out of the imagery. So Sam is a great model, but because of its ability to segment anything, sometimes that provides a lot of noise. And the Text Sam really provides us with an opportunity to be more specific in what we want to have segmented out of that imagery.

So let's move on and hop into a demo, and see what this actually looks like to use, how it works. Some expectations of, again, what the results are. And some tips and tricks on how we can improve the results when they're not necessarily maybe what we were expecting. I think this is also a good time to let you know that I think a good word to keep in mind here as we move forward with this-- and as we look at the application of deep learning models, and machine learning, GeoAI in these workflows, and I think that word is accelerate.

This isn't going to replace your current workflows, but it is going to really help accelerate them and help you get to 70% or 80% of the way where you need to. These things aren't going to work perfectly every time. And I think it's just good to be honest and transparent about that. So with that, let's hop into a quick demo to see this in action. So I'm starting here in our desktop application, ArcGIS Pro. And as you can see, we're starting with the aerial imagery that was extracted out earlier from Nearmap.

So as I zoom and pan around here, you can see we're still receiving and getting access to that really nice, crisp, high-resolution imagery where you can see fences, and pavement markings, and objects and everything with contained within that imagery. And then from there, it's really easy to get started. So let's hop over to this northwest corner where the parking lot is. And earlier in that video that we were just watching, we saw that I stopped on the car detection model. And so we're going to use that one here and see what it looks like to actually run.

So it's as simple as opening up our detect objects using deep learning tool. We just point that toward our imagery. We get the option to give it a specific name if we want. Or you can just leave it with the default. And then from there, we can pull in our model. And what's really nice about this is can download and use these locally. But we can also just connect directly to the Living Atlas again. So those keywords of DLPK are going to get us the return for all the models that are currently stored there in the Living Atlas. And then from there, we can find the one that we want here, car detection.

And then from there, we can just click OK. Depending on your internet connection, this is going to take maybe just a few seconds, maybe a minute to load. And then we're presented with our arguments. Which in this case, we're just going to run default and come back to these in a second. Just so we can see what these results look like by just, again, running it with the default and kind out of the box. So once I'm ready, I can click Run, which I have gone ahead and done ahead of time. And you can see now, we got our results.

So just like that, we've gotten all of the cars that were in the image of this parking lot back. But you can see we did miss some. So again, just level setting and being transparent about expectations here. It's not always going to work perfectly. But there are some ways that we can easily improve these results before we need to really panic, or go down any other route. And before we actually look at that, it's important to understand how this tool is actually working. So when this tool is running, it is splitting the image into tiles, like we can see on the screen here.

And then it's parsing through these tiles. And based upon the imagery it was trained on and the patterns it was trained to detect, it's going to look for those within the pixels of those tiles. And so that's really important to know, because then we can come in here and change the cell size to help affect the results that we are going to get here. So depending on the imagery we're working with, the size of the feature, if we just go in there and make that simple adjustment to account for that, we can get much better results. So just by changing the cell size to a specific size, I rerun that tool. And then you can see I get a lot, lot better results.

So again, think of this as just a way to really help accelerate your workflows. And you're not necessarily going to be able to completely replace them, but it's going to certainly help you get there a lot faster. And then from there, it's really just a rinse and repeat type cycle. So we're able to continue leveraging this tool that we have over on the right. We can pull in those specific deep learning models and packages from the Living Atlas, whether that's Building Footprints, like we can see here. Or maybe we want to look for other things like parking lots. We are just going to continue to do that workflow, and knowing that we can always change those parameters.

But inevitably, we're going to run into a case where we don't have a deep learning model. And so we can either go back to Nearmap and leverage their deep learning packages and their robust library. Or we can also explore this opportunity that we have with the Text Sam that's available through the Living Atlas as well. And like I said, this just gives us the ability to have a prompt available within the tool to be specific about the feature or the object that we're really wanting to detect.

So just like I did before, I just navigate to the Living Atlas. I load in that deep learning model. Again, it's just going to take a couple of few seconds-- to a few seconds. And then you can see, I get a text prompt. And then from there, I'm able to put in descriptive text to look for certain things. So here's a blown up version of the tool just so we can see. So I have a text prompt. And then you can type in singular or multiple values. And again, you can get really creative here as far as like what you want this tool to segment within that imagery.

And the possibilities here are really endless. So in this case, I might have the need to identify some of the vegetation around this parking lot. So I can just type in tree, or as you can see, like I'm doing here, you can be super OCD and make sure that you try to type in every word that you can think of that's associated with greenery, or shrubbery, or trees. And then we can run that tool and get those results back pretty quickly. And again, there's limitations here. Depending on the imagery that was used to train the model, the imagery that we have, and even shadows, you can see they all have an effect on the results that we're going to get.

But again, we're able to accelerate at least getting to this point, and then can clean it up from there. And again, think about getting really creative with this. You could do things like parking islands. You could do wetlands. We can do utilities, like maybe manholes and catch basins. And even things like light poles. And those can even include the shadows that might be beneficial depending on the angle of the imagery if we're working with nadir or oblique.

So again, this tool can be really powerful. So let's hop over to another site and explore more of that geo side of GeoAI. So again, I'm here in Pro. And I'm starting with that really great high-res imagery from Nearmap that we have, and that Jeff extracted earlier. And I also have the contours that their AI model extracted for us. And so this is an opportunity to turn this into a surface within here, and maybe do some hydraulic modeling. Or we could leverage those analytic engines that we have available, specifically the raster.

So we might want to know-- we have a beginning image and an image of some construction after a certain amount of time. We can point A tool toward both of these images and say, hey, tell me, what are the differences between these images here. And return that in into the map in a visualization like we have here. So it can be really easy to detect change and see where construction is happening, and maybe more importantly, not happening between these two time periods of imagery that we've extracted, again, from Nearmap.

The dark green areas are going to be a significant change. And the deep purple's are going to be of even more significant change, things like buildings popping up. Lastly, none of this is really any good to us as far as these results that we're getting if they just live within this application in the desktop. So a next step would be to take this information and publish it to the web. And in doing so, we're creating a live service that is more easily accessible, and really helps increase sharing and collaboration amongst stakeholders, both internally and externally.

So I can take all of that imagery or all those features that we've extracted from that imagery, including the imagery itself, and share that up into the web. And like you can see here now, this is available through a web browser that I can share out to anybody in the team. And just to show, just like Jeff did, that this is real and live, we can toggle these layers off and on here. And then this is also an important last step because this is going to set us up for what we will end with and how we can start enriching our designs with this information with a pipeline directly into our design drawings.

So before we move on to that, some next steps that you can take with deep learning, you can continue to learn these pre-trained-- or continue to leverage these pre-trained models, both available within ArcGIS and within Nearmap. Or you can start fine tuning an existing model. Within ArcGIS, we have a ecosystem or a framework that can allow you to not only to fine tune or repurpose an existing model, but you can also train a deep learning model from scratch.

And this is, for me, where the real-- another place where of excitement comes from this tool that's been developed through this partnership of Kimley-Horn and Nearmap, is that this whole process is predicated and starts with in imagery. So with this ability now to extract out this high-resolution imagery at will and through different geographies, it's really going to set up Kimley-Horn for success when they want to move into training their own models. To have that rich resource of imagery to start that process and be able to quickly train models in a robust manner is going to be really, really beneficial to their workflows. And provide them with the ability to really be innovative in how they deliver projects in the future.