 Hey, folks, my name is Michael Chang. I am on Facebook's open source team. I'm going to kick this off by doing a quick introduction to our open source program. You may be familiar with some of our more well-known projects, including React and React Native, which is the front-end framework for web and mobile, PyTorch for ML, GraphQL for APIs. In terms of numbers, across the company, our open source program supports more than 500 projects, which have received a combined total of 300,000 contributions from 16,000 contributors. Our open source program also sits within a broader ecosystem of initiatives that seek to address and focus on some of the more difficult problems that we face, both as a company and as a society. And that includes Facebook connectivity, which aims to make the internet more accessible to everybody on the planet, Facebook AI research, which engages in AI and ML research across a broad range of disciplines, both of which regularly make most of the tools, data, and information that we use and generate available for public consumption. And that is why, earlier this year, together with Google and Uber, Facebook announced the Urban Computing Foundation, which we see as a compelling platform for organizations to get together to address the impact of technology on urban landscapes. Even though the UCF is brand new, Facebook has been active in the geodata space for many years. And that is for both humanitarian reasons and for our data creation purposes. Here to talk about two examples of that is my colleague, Mike Mogersky, who will talk about disaster maps, and also Dristee Patel, who will talk about open street maps. Mike. Thanks, Michael. Hi, everybody. My name is Michael Mogersky. I'm an engineering manager at Facebook in the Spatial Computing Group. We're a group that largely deals with questions that start with the word where. And we perform a lot of mapping and geospatial data services within Facebook. Myself and my colleague Dristee are going to talk about two of those efforts today. And I'm going to focus specifically on a project called Disaster Maps, which we've been running for approximately the past year or so. And I'll talk about what disasters are in the context of a project like this, who's interested in those, and who around the world cares about mitigating disasters. And then finally, what we at Facebook are doing to play a role in that ecosystem. This is an aerial image of the campfire from November of last year up in Northern California. It happened just a few hundred miles northeast of the Facebook campus, and actually blanketed my first week at the company in thick, dangerous smoke. But it was much, much worse for the people on the ground in Paradise, California, the town most affected by this. 85 people lost their lives, 19,000 buildings were destroyed, many of them single-family homes. It was the single largest wildfire in California history, most destructive wildfire. And we think that this is indicative of a very large number of similar crises that are increasing around the world at this time. Cyclones, hurricanes, earthquakes, all these kinds of events we think are being exacerbated by the oncoming climate crisis. And we at Facebook are interested in participating in some sort of response to this. But there's a lot of folks that are already out there responding to crises. From first responders, such as fire departments, all the way out to a large cluster of organizations that perform relief activities during and then long after a crisis hits a community. And what are some of the concerns that these folks have? Well, they're really trying to answer several critical questions when they're looking at response either immediately during a crisis or a while after. First of all, they just want to know what areas are affected. This is a core geospatial question. Where is the actual problem? Second, they're interested in the human impact. They want to know if people are evacuating, where they're going, where they're coming from, has power or connectivity been disrupted? These are important concerns for people who use mobile phones to communicate, which is basically all of us. So you have to know things about power and connectivity in order to understand whether folks are able to talk to their friends, families, loved ones, that kind of thing. And then finally, if people leave and don't come back, do they settle someplace else? Where might that be? We know from the hurricane Katrina disaster in 2005 that many people just left New Orleans for good at that point. And so part of the response to that disaster involves understanding what that kind of human geography and human landscape looks like over the long term. Some of the organizations that we partner with are these. There's about 30 of them. And they have a lot of different concerns. Direct Relief is a large organization that performs sort of direct delivery of relief supplies into hard hit zones of crisis. IDMC focuses on internal displacement monitoring, essentially internal refugees within countries. Net Hope is one that focuses on network connectivity. So they bring network and cellular hardware to areas where network has been disrupted, and use that in order to reestablish cell phone communication so that people impacted by a disaster can continue to contact and communicate with people outside of the area. All of these are organizations plus dozens of more that partner with Facebook on the disaster maps efforts. And so what can we do about this? Well, Facebook happens to have access to a lot of information about how people move around the world, especially from folks who share their location with us via the Facebook mobile application and others. And so if you can have access to this information and run it through important privacy protections, such as anonymization to ensure that nobody's individual information is accidentally released, the addition of random noise to help mask further the individual identities or movements of specific people, spatial smoothing to blur out information so that you're looking at kind of neighborhood and town scale data rather than individual house scale data. And finally, the removal of small counts in order to kind of attenuate the amount of information you're giving out about very underpopulated rural areas. All these things together add up to a very valuable source of information that many of those partners on the previous slide use in order to respond to disasters more effectively. So what form does our help take? Well, there's kind of five types of mapping that we provide to these organizations. Those of you who are familiar with Facebook's safety check know that it's a way for people to essentially let their friends and family know that they're okay in a disaster, but we kind of tail onto that and start providing data in areas impacted by safety check. One type of map that we provide to disaster responders is a population map. Very simple map showing broad change in patterns over space and time of how the Facebook user population has moved. So this map is showing us Paradise California, which was primarily affected by the campfire last year, as well as neighboring towns like Oroville, Chico and Willows. And what you can see from this map is that there's a kind of blue cooling effect people leaving Paradise, California, who used to be there and a red kind of growing effect in places like Chico and Oroville. So an organization like Direct Relief might look at this and say, okay, it's realistic that the folks who need N95 masks or other relief supplies might be found in Chico since they're not gonna be in Paradise during this fire. So this might be an important hint that they use in order to direct relief efforts. We produce a type of map called movement maps. This is something a little bit more precise, whereas the previous one gives you kind of large, absolute numbers of population change, movement maps gives you kind of unusual movements, even if they're very small scale. So if there's 40 people who normally move between point A and point B, and then during this disaster, you're seeing a few hundred people do that move, it'll show up as a thick line on a map like this, which might help a disaster responder understand that that's where they should focus their energy. Zooming out a little bit, we're also thinking about large scale and long-term displacement. So these types of maps will show that for example, people might have left Paradise, California and settled with friends or family in other towns or rented apartments elsewhere or just picked up and gone to a different state. This is important for understanding what the kind of long-term population dynamics in these hard hit areas are. And then the last two maps really focus on the kind of technology dimension of disaster response. So because we know kind of big patterns about how Facebook is used on mobile phones, we can start to infer things like network coverage maps. This is extremely useful for NGOs like NetHope for example, who might wanna patch holes in network coverage. It might also help early responders understand where they're not seeing information. Maybe there's not coverage in a particular area and that might be a hint that that's a place they might go look and see if there's people there who aren't showing up on some of the other maps. And also due to the behavior of kind of charging devices and powering up Android devices, we're able to also infer broad patterns and power availability by making inferences from essentially battery level. So this might help us understand in a remote area whether there's a power outage to an area hit by a hurricane or an earthquake or similar. So these five types of maps are the core of Facebook disaster maps. They're a privacy protected way for us to share broad aggregate user patterns with responding agencies and NGOs that need to be able to get to these populations and understand what they're gonna see when they get there. But there's a sixth type of map which I haven't talked about, which is basically what is the actual existing infrastructure like roads and land use and so forth in those areas. And so I'm gonna hand it over to a colleague of mine, Drishti Patel, also in spatial computing. We'll talk about how we work with the open street map community on that sixth type of map. So I'm Drishti, taking over from Michael and Michael over here. And I'm gonna be talking about maps at Facebook. So a lot of you might be wondering where we even use maps at Facebook and how it kind of all ties in. Turns out high quality maps are really valuable to our users and they're actually used across a variety of products. So you can see here in examples of search, commerce, events, travel, some of the humanitarian causes that Michael mentioned, as well as even things like looking for a job. Our base map at Facebook mostly consists of data from open street map. How many people are familiar with open street map? All right, perfect. This is an open source community, so that's not surprising at all. But basically open street map is a free and editable map of the world that's largely built by volunteers, which is pretty amazing. So think Wikipedia for Geo. It powers thousands of map data, websites, mobile apps and hardware devices. And over time it's gained a lot of usage in various sectors. Traditionally data collection for open street map was pretty manual. And this includes tracing satellite imagery, data collection from mobile devices, as well as GPS devices. A lot of just really hobby mappers putting their GPS device on bikes and actually tracking things over time. And also low tech field maps, which is something that still happens today. So collecting data by hand. As you can see in this map, even though OSM has all of these amazing benefits and has gained adoption in various sectors, there's still a large part of the world that is in mapped. And you can see very clearly in these areas where it's highlighted in pink. These are not underpopulated areas of the world. They are high population, high dense areas, but it's still lacking significant map data. And in some places there's no data at all showing that people exist in those areas. Data collection is complex, it's difficult, especially in some of the more rural parts of the world where there's no power, there's no connectivity. You're ridden with political and environmental constraints to go around and do this work. And so it's very hard to scale up into these different sectors and areas of the world. So this is why Facebook has been doing some research to try and figure out what are the ways that we have that we can speed up this process in improved data coverage and filling these missing gaps. One of the benefits that we have inside of Facebook is all of the power that we have around AI and machine learning. Things that have helped us get there that have changed in the industry as well is access to high resolution satellite imagery is much easier. You can basically bite with a credit card. And then there's been really major advances in deep learning. And you can see in this image right here, these are basically two different data sets that are currently open at Facebook. The yellow that you see is actually our high resolution population layer. And it's available on season and a couple of different open data spaces where you could just download this data and use it for whatever you want really. And then the white roads that you're seeing are the roads and that's what I'm gonna focus on today. So you can see in this image, this is an image of Jakarta, Indonesia, one of pretty much the most highly dense urban areas in the world. And over the past two years, there's a team at Facebook that has been working to improve this model. And so we can extract features from satellite imagery at a global scale. And that's the magenta layer that you're seeing showing up. And because we've made this training data at home, the model actually works across various terrain types. And this is what you can see in the example in Ethiopia where the colors and textures all look different, but the model still works pretty well. This is another example of Nigeria, a more hilly kind of rural area as well. It seems to be doing okay. And can anyone guess where this is? So this is Boston, where our engineering team for this actually sets. And you can see it does pretty well in urban city areas as well. This tends to be easier because, you know, there's a much more defined pattern for these areas. So now that we have these models at global scale and you have roads for a whole country, what does that really mean? How can you really get it into the hands of the community? And so this is what we've been thinking through as well. So there's an editor called the ID Editor that was created by Mapbox, that is open source as well, that we took in-house and kind of expanded on first to test things out of how can we get this vectorized data back into the hands of the community and making sure that we keep the integrity of OSM good as well because we don't want to do these massive changes over time because it is a really community-driven project. And so essentially we used what was already available, expanded on it, and then allowed users to access our AI data through current tools. This tool is now open because we spent a lot of time testing it, working with local communities, making sure it was okay over the past year, and it is now fully open sourced and you can find it on the Facebook incubator under the tool named Rapid, which is what we're calling the Facebook version of ID. So secondly, we also partnered up with Humanity on OpenStreetMap. This is also a huge part of the OpenStreetMap community, a tasking manager that essentially lets NGOs of various types and sizes come and create tasks for humanitarian projects. And this is everything from long-term health programs or development projects to disaster response. And people all over the world, usually in the thousands, usually show up to help map these areas during the times of crisis. And so we've worked to create a tasking manager that is ML enabled. So it allows people to access all of this AI data from various places. It doesn't necessarily have to be Facebook and integrate these tools into the tasking manager. And so in the time of disasters, they can access this data in a more easy and efficient way. Lastly, to make all of this kind of work seamlessly because we're running our models on specific type of imagery, we've partnered up with Maxar. And this is the vivid layer that we work off, which is pretty much the mosaic, the most recent mosaic they have of the world that stitches everything together nicely. And so it's free of cloud cover. This is also available on OpenStreetMap as you can see. And the image that we're showing is actually the place we're at right now. And lastly, this is kind of the walkthrough of what this would look like for a user. So essentially you would select a task on the tasking manager, open it up an ID, which this is a very typical user workflow. We're gonna start off by showing what happens when an editor manually draws. You would pick a point line or area feature, go through and make the edits manually by clicking nodes, and then go ahead and tag the features to whatever appropriate tagging it would be. The other option is using the rapid workflow. So essentially when you open it up, all the roads are already available. And then you would just basically add it to the graph if you wanted to use it. You can add the different feature and then go ahead and add the tags. As you can see, one of the other things we've added into our version of Rapid are lots of validation and data integrity checks. So users get real-time feedback if there's something wrong with the data. If you're drawing a road over a building or the connection doesn't make sense, we have suggested edits and warnings. And so people can get real-time feedback and train because most of the people that are doing this mapping are completely remote and usually may not have someone around for guidance. So this real-time feedback is really helpful. So getting to the heart and soul of OSM is these local communities all around the world that basically self-sustained set up and choose to map up their roads. It's really amazing. You'll see some of the countries and if you look at the user edits, it's pretty much one or two people that have built out entire highway systems for their whole country. And all of this is on a volunteer basis, so in addition to whatever else they're doing in life. So it's absolutely fantastic to see all of this stuff happening. And we've teamed up with these local groups all over in different regions to figure out what's the best way to collaborate, not just throwing tech out, but figuring out the most efficient way to work with them. And this is an example of the work we've been doing in Indonesia. We teamed up with a local mapping team in Indonesia recently. And what we found is that even when there's very strong communities, it's really helpful to have some of the digitizing done already because they get to spend more time on the information that you can only capture on the ground. For example, telling what type of bridge it is, whether it's something I can walk on or drive a car on or may not be good for travel at all, these are the kinds of information that you can only get on the ground. So lastly, I'm going to leave you with a video that basically summarizes this process, which was also an example of Thailand. All of this is now open sourced and available, and we're going to end off with some links later on so you can find it. But yeah, I'll just get that started. Thanks.