About the talk
The automated extraction of roads from aerial or satellite imagery at regional and city scales applies to a multitude of long-term efforts such as: increasing access to health services, urban planning, and improving social and economic welfare. Optimized routing via up-to-date road maps is also crucial for such time sensitive efforts as determining communities in greatest need of aid, effective positioning of logistics hubs, evacuation planning, and rapid response to acute crises.
Satellite imagery may aid greatly in determining efficient routes, particularly in cases involving natural disasters or other dynamic events where the high revisit rate of satellites may be able to provide updates far more quickly than terrestrial methods. Existing data collection methods such as manual road labeling or aggregation of mobile GPS tracks are currently insufficient to properly capture either underserved regions (due to infrequent data collection), or the dynamic changes inherent to road networks in rapidly changing environments. For example, following Hurricane Maria, it took the Humanitarian OpenStreetMap Team (HOT) over two months to produce a fully validated map of Puerto Rico, even with a team of thousands of volunteer mappers.
In this talk we discuss the City-Scale Road Extraction from Satellite Imagery (CRESI) algorithm that rapidly extracts large scale road networks and identifies speed limits and route travel times for each roadway, using only satellite imagery as input. Including estimates for travel time permits true optimal routing (rather than just the shortest geographic distance), which is not possible with existing remote sensing imagery based methods.
Furthermore, we explore some of the interesting lessons learned from the recent SpaceNet 5 competition, for which CRESI served as the algorithmic baseline. For example, we show that neighborhood-level details are more important to road network extraction than broader city-scale specifics like: road widths, background color, lighting conditions, etc. Such lessons inform any number of real-world applications, such as how disaster relief organizations should distribute resources and personnel.
Adam Van Etten is the Chief Data Scientist for In-Q-Tel, where he focuses on applied machine learning topics of interest to the US Government. His most recent research has been in the geospatial analytics realm, where he applies machine learning and computer vision techniques to satellite imaging data. Adam currently focuses on managing and assisting with the myriad In-Q-Tel Labs research projects, as well as helping to guide the strategic vision of the lab. Other recent foci for Adam are helping run the SpaceNet initiative, and exploring the limitations and utility functions of machine learning techniques. Prior to In-Q-Tel, Adam was a Solutions Architect and Data Scientist for Data Tactics working at DARPA headquarters developing tools and scalable algorithms for big data analysis on a variety of projects. Adam received his Ph.D. in physics from Stanford University and bachelors in physics and astronomy from the University of Washington.View the profile
I have no way of knowing if anyone I can do this. So, again, qualities. Without a degree in computer version. I'll go and skip the summary. You can read that on the website. So what's the motivation here? We need to accelerate mapping. And so a good day to point here. Is that after Hurricane Maria? It was like 70-plus days over two months to map Puerto Rico, with just an army of volunteers, and it was something that obviously was impressive effort, but it's of optimal. I take this long. So if you cannot accelerate
this, that would be obvious even beneficial. There's a ton of potential impacts of automated routing. Don't spend much time here, but you know, the price of their father. And if we can do this, an automated fashion, something would benefit all these communities. Pull up more about the motivation of why we're doing this. Since the rapid extraction of Road networks and key routes is critical and dynamic. Scenarios such as being exactly like hurricane Maria and you'll come in. The room is what can I already do routing? You know, what's up in that Google
Maps? App store or being or MapQuest? If you remember those days are great options, right? But those data sets are awesome and intensive. The date of the solutions are kind of proprietary, quite often. There are 15 complete. And if there is a dynamic event, like the like a disaster there that I've dated optimize routing on airplane mode, you know, it's the doesn't really work. Usually there's been a lot of research and write. So It's either just segmentation mask, like this is Middle image on the right, or if you do get geometry, geometry and connections. It's only Road
just to just to Geometry, not the properties, but rub properties are crucial for to optimize routing. And here's an example. Why right here. Is this the sample road Network? And if you just do, you know, distance paste? Rally on the left. That's the route, but using you care about time, right? How do you get from point A to point B in the fastest time? And that's what you see in the right. There is very, very different route. And that's something it is important to know rather than just use this optimized time. Alva Street, map is an amazing resource
something that, but I think a lot of people are very, very excited about a supportive of Hazard. We have anything complete though. So, you know, if Google Maps has an option, but that's some damn straight. So here's it up. The street map example of Shanghai and you'll max speed is a tag that you see, but you know, that 90% of those who have no Max Speed tag. So again, it's, it's hard to do is optimized routing with open source Solutions exist currently as well. So they sent this is a,
a non-profit operation run by Cosmic Works where I, where I work, but in collaboration with a bunch of the partners. Related to go to tell advancing foundational mapping. It's pretty good day to say, hole puncher from cities. And then, the newest addition is bleed all over the globe, but we'll be focusing on the road apiece. So what kind of skip on there's two different types of finish, do mapping with labels. Building Footprints as one and then Road Network labels and features. So about twenty thousand kilometers of roads. Have been
manually labeled with features for nine different, Geographic areas. They're kind of Drive interest in space app and show with, state-of-the-art, can be, have been seven different challenges around the space that data sets. We here will focus on space. Next three and five, which are the road State assets. And so, this is what ultimately, for four different cities, for 3, and then three more than five to eight different cities. Total working at Road extraction and routing for those cities.
At the station at 5 and 5:00. Will. The challenge here was to automatic automatically extract roads just from satellite imagery and then get optimal. Right estimation. We shall attempt rejection. So on the left, you see it and put him in style and then in the middle here, this is a ground Truth for a Snapchat over Moscow. And then I'm going to categorize these roads, right? So, get the network and also some more properties the roads. Train in brief. So, what are the outputs for? We had asbell space netsuite, we get
a bunch of competitors algorithms out there open source. And so that's pretty exciting. In this case, will be focusing on that a baseline algorithmic because it had company performance but much much faster. So, we call this crazy City scale, run extraction from satellite imagery. I was working and also noted that this General approach was what all those winning Our Lives also used for her space. That's why I write with a goal here is to let slip to extract the road network, but then also has good speed estimates. How do we
do it? How we see the info. I'll hear spent into a segment, a smelly room and multi-class than sleeping on the ground where we reach different shaped by a speed. And let me lead that together to get a row graphs here. And then we can use these multiple channels to pull out the actual speed estimate for each rose. Quartz gets the travel time to Lisa, multi-step process, but we end up getting out, you know, both the road Network and then the property Road, this case, the property we care
about is the travel time estimate for twice the speed. And then, of course, if you can't do it, over large areas family, very useful. So you can actually do sick, very simplistic approach to do this scales large areas. So we can do this on arbitrary scale. How quickly obviously a City Skate City scale is the goal here, but you can expect it to as large scales if you're looking tired. So how has to do? So here's an example. This is one of the testes for basement V where there was no training done on the city. What's the weather is just testing. This is xe Dar. Es Salaam. So
here's a fairly large and you run this through the first step. Give you a segmentation mask roads. And this is where Corral takes off and stop right here as you're. All right. We're good on. Let's go one step further, and let's pull out the actual graph. Again. This is pretty useful, in this is where a lot of it. The road Network rap called good. But if you really want something useful right now, let's color by Road speed. So he missed case, right? Yellow as slow and ride is faster. And this is an actual row, graph know, sweetheart or
anything, but the imagery. And then what we get is this Road Grass, where we're projecting roads and their speeds and we can use this, you know, for optimized. And here's another shot was kind of interesting. Perfect, but There are examples where we are able to kind of connect behind buildings, right? Which is obviously tricky but important to do. If you have an alternator shot. Something is not part of the overhead. Insecurities just comparing two kind of other.
Other options out there and literature. So Road, Tracer, is a pretty cool algorithm. That was put out a little bit ago and I'm to pull out roads and if we forget speed whatsoever, just look at the road networks. Nothing. But that the rotational Afton in and then our invitations on the right, we can see a significant improvement over that the rotation fermentation about speed. Just look at the geometry and that's it. Then if you look at the speed as well, so this is a prediction here over Shanghai. We actually see the speed,
is actually very high, fidelity so much detail here, but there's a very small gap between the length and the time, the metrics, which means that again, they are very good at explaining speed for the roadway. So the hardest part is getting the network, but then we can get the speed very, very well predicted. Once we get the network. He's just a gif of kind of building up the road at work in Las Vegas for a test area. And you can see right. There's I've seen a pretty good job of organizing residential areas,
which were yellow, so speed limits, 25 miles an hour and then building up to your red being highways. But again, this is doing a pretty good job of pulling up that road Network starting with nothing but imagery. Find a routing. Race is kind of a premise originally of why we're doing this. And once we have this road Network. We can extract the route. So starting in the the red here going to the green. It was really interesting. Right? As you can show that I'm kind of thinking of random two points on the
left is what you would get. If you just typical routing that you get from out, I'll go in now, which is the shortest length. Then I'll know on the right is acting Corporal name of the travel time information. We pull out from this algorithm and it's very very different route to underscore the importance of having this travel time estimation in your in your algorithm, to get the optimal route. And this is something that, you know, once you have this network, you don't eat, no matter what your starting point is. It's pretty quick to
almost instantaneous. Frankly writes it to calculate these alternate routes. So a few more things that that all going to quickly and we started too late but a few lessons learned from space in 5 Min or even going to, this is the kind of demonstrate how this approach. Actually it will apply to new scenario solo. There are blessings of it was interesting. So we're sitting here are the scores for that. The top five competitors, first base in at 5, as well as the Baseline on the far right for each different city. So there
there's four different cities and then purple is totaled and the takeaway here in a bit. So it's kind of process is That's red, which was the Unseen City. No one had seen that before, is not appreciably lower in most submissions than the other cities. So, what that means is that, and when we trained on basic six, different cities, what we did. Then testing on a toe dancing geography. We actually have a ballets real bus to New areas. That's pretty exciting. That means that these models actually have a chance to deploy, you know, if there's a disaster
you can trip retrain on an area. You might still have a chance of having a model that useful, in a new unseen geography. We have to play it and then hopefully meet results. Is also significant variation in sea level predictions. And so we have a small different chips for scoring weeks. Maybe 60 are obviously for a larger pictures. But even though the competitors had very similar scores. If I go back to hear you on that, the purple line is all very similar as the huge variation in predictions. That means is that neighborhood
level details or are far more important than broader things, like no road with background color, lighting conditions are kind of more City scale. So give me kind of microscopic /, neighborhood level features, Roxy very important for Road networks and surprising to see the huge difference since course, depending on these these factors. You can do a lot of correlation between features of roads and cost for this one little bit. But what's interesting is your weirdest doing a,
a rainforest fit based on a few features for modeling. What we see is that we actually have a very very pretty good model for given a certain set of features of Antonin. Six different features, predicting what the Apple scores so happens that we're measuring Road Networks. The most prettiest teacher is the total length of the road in in that the scene Fall by Max Speed. So, that's interesting in it. Again. If you have an unseen area, you can easily compute these features like total meters Max Speed,
etcetera. You have a sense for how well, you're actually predicting that road. That's good to know when you going to Auntie. So you can closings. So what show does open-source methods to extract Road? Networks directly from overhead imagery, we can talk too much about speed but it's extremely fast. So we're getting almost a kilometer per per minute. You got a CPU, we see that even for new locations and they can actually be brought up already with the travel time. Estimates are actually very accurate.
And this combination Road, networks travel time. You can actually do optimize routing and unseen locations or than an experience. So, I will go in there and hear some information channels. For people to bruise.
Buy this talk
Buy this video
Our other topics
With ConferenceCast.tv, you get access to our library of the world's best conference talks.