TOM ADAMSON: Hello everyone, and welcome to another episode of Eyes on Earth, a podcast produced at the USGS EROS Center. Our podcast focuses on our ever changing planet and on the people here at EROS and across the globe, who use remote sensing to monitor and study the health of Earth. My name is Tom Adamson. One of the things EROS visitors on a tour enjoy seeing is the colorful NLCD map. It's a striking large map of the U.S., and people enjoy pointing out where they are from or where they're heading on their road trip. And we like to brag about how it's an EROS product. The National Land Cover Database has been around for a while, and we have a few Eyes on Earth episodes about it already, but there is now a new release and it's being called the next generation of USGS land cover mapping. It's called Annual NLCD and we're going to talk about what's new and what people will do with it. So today we're talking with Jon Dewitz and Jess Brown, who are leading the effort on Annual NLCD. Jon, let's start with you. Can you start with what is land cover? JON DEWITZ: Land cover is something that describes what's going on on the land surface. Most people's experience with maps, especially us older generation, starts with a map you use to figure out where you're going. And it may seem that a land cover map and a state map are the same thing, but they're very different. Those maps that give you direction and show roads are just a small part of what we do on this land cover product. Land cover maps themselves are really used to categorize and to allow a computer to analyze where things are, how much of them are there, and what's happening on the landscape. That really didn't start happening until 30 years ago. So when people say maps, many people still think of a map as a very basic thing used for navigation. The big difference between an old paper map that some of us have seen and NLCD, is that it's a digital product. Humans are very good at looking at a picture and understanding visually what's going on, but we're very poor at doing large number calculations in our head. So NLCD is digital. It allows us to very quickly use a computer to understand how much forest is in an area, how much developed is in an area. Visually, we can see that, but quantifying it and being able to put that into, you know, actionable things for policy around the nation is really why we make the map. ADAMSON: That brings us to NLCD, the National Land Cover Database. Jess, give us the introduction to NLCD. JESS BROWN: So the National Land Cover Database is a representation of the Earth's surface into something that can be used further in science. We try to understand the changing planet, and we need this information to feed into all these other purposes. But it is much more than just a map. In the National Land Cover Database, we express land cover in 16 different categories. These 16 categories, that have been traditionally used, divide the landscape up, or categorize the landscape, into water, ice/snow, there are four classes characterizing developed -- the developed surface, which is which is really like our urban landscapes, three types of forest, barren land, shrub, herbaceous, two agricultural land cover types, pasture/hay and cultivated crops, and then two wetlands types, woody wetlands and herbaceous wetlands. ADAMSON: Okay. And clearly, the landscape is complicated. The landscape is more complicated than these 16 classes. So this is a simplification. Is that a good way to-- BROWN: Yeah, that's a good way to put it, Tom. It's definitely a way to simplify the landscape. ADAMSON: So let's talk about these fractional products. Even though these land cover classes are something of a simplification of describing land cover, there is this fractional product where you can get a little bit more detail about something like impervious surface. That's one of them right. How does that one work? DEWITZ: When we started NLCD 2001, we saw there was a need for additional discrimination for some of the land cover classes. There are many things in the landscape that may only make up a small fraction of the pixel, but they have an outsized effect on everything surrounding it. Developed is one of those. We started this as a database concept where developed and forest canopy inform the land cover. So when I initially wrote up the land cover definitions, we worked to figure out how we would split those developed classes from this fractional impervious surface. Impervious surface has a very outsized effect on the landscape. It affects drainage. It affects air quality. It really shows you where people are. And as we know, the more people there are in a location, the more effects we have on the local vegetation. So as we went through, we split those classes into four developed categories. But those four classes are derived directly from that percentage product that 0 to 100%. In that pixel, if it's only 15% developed impervious surface, it's great to know that because there can also be 50% canopy, 20% grass. And those are things that we can't understand with the simple-- It's this type of developed in the land cover. ADAMSON: Well, if I look out my front window and I see a street, that's an impervious surface, but there's also trees overhanging that street. There's also a little grass strip next to the street. Is that the kind of thing you're talking about? DEWITZ: That's exactly right. We also have another product, percent forest canopy layer that also works hand in hand with land cover and with impervious surface to help a user figure out those percentages, because developed is the dominant feature on the landscape. Even if a pixel is only 15% developed and 50% tree canopy, developed is what comes in the land cover. But because it's a database, you can use these other pieces to understand more of what's happening in that specific location. There's a lot of things on the landscape that are just this or that. ADAMSON: Yeah. DEWITZ: Agricultural fields, you know, they're not they're not 30% agriculture in a pixel. A whole field is agriculture, but developed and canopy those things, they can be very small pieces of each pixel. And that's what we really wanted our users to be able to tease out and understand. There's also places that analytically don't make sense. Atlanta, for instance, can have, say, 50 to 70% developed surface underneath that forest canopy and also have 70% canopy cover. And we know percentages only go to 100. And understanding that relationship is really important for urban foresters, local land managers, urban development, things like that. ADAMSON: Let's get into the product itself. What does Annual NLCD provide? Like, what should users expect when they open the box? BROWN: The Annual NLCD is a new product suite that we're releasing, covering 39 years of of history for this country. So this database consists of a product suite with six products. And we're in those products we're describing many different characteristics related to the land surface. Land cover, land cover change, a confidence product called land cover confidence, fractional impervious surface, an impervious descriptor, and spectral change day of year. So these six products together comprise the database. So we're pretty aware that the land cover product is the most popular. People are comfortable using that. They're very interested in using that in their modeling or studies. But we've also included land cover change, which is an annual change indicator between two adjacent years of land cover. The 1986 product represents change from 85 to 86. You know, the users could calculate that themselves from the annual land cover, but we did it for them. And basically it provides numerically, the class that was in the prior year and the class that is in the current year. The majority of the landscape is actually very stable. So what we see when we look across the whole record is about 80% of the pixels across the country didn't change, but there's so many changes that that have happened and have occurred during that 39 year period. And this annual land cover change product helps characterize that more specifically through time. Land cover confidence really comes out of our methodology directly, and that's a probability value for the land cover class that was derived in our classification methodology. Gives people an idea of how confident the algorithm was that this land cover call is correct. And some of the things that Jon was talking about earlier about how the pixels aren't pure-- So Landsat pixels do not represent one land cover type often. There's often heterogeneity within that that 30 by 30 meter pixel. And also there are situations where the spectral reflectance that Landsat is measuring is very similar for two different land cover types. For example, shrub and grass might have very similar signals in the spectral reflectance that Landsat collects. So that's the basis for our classification. So land cover confidence might tell the user how well the algorithm was able to identify this particular land cover type. So the final product, the spectral change day of year, comes out of our change detection algorithm, which is the shorthand for that is CCD. That is a change detection algorithm that uses a harmonic modeling approach. And we are really basically running all of this Landsat data through time in order to detect when an abrupt changes happen. Now, not all land cover change is abrupt. Some is more gradual. So this actually helps us pull out the day of year when the Landsat surface reflectance showed a change, you know, basically an anomaly is detected through that harmonic modeling approach. ADAMSON: What makes Landsat uniquely suited to being the main input into NLCD? BROWN: So I always like to say that Landsat's super power is time travel. ADAMSON: Yeah. BROWN: The beauty of the Landsat observing system is that it started in the 1970s. Now we're using data from the 1980s forward. There are some specific reasons based on sensor characteristics why we do that. We start with the Landsat 4 record in 1982, and we start that change detection effort, that change detection algorithm in '82. By the time '85 comes around, which is really the first year of the record of land cover, we have a pretty decent idea. The harmonic model process then can do a better job of determining a change has happened. And we also have fewer observations at the early part of the record. Landsat was young. There were other challenges to collecting and saving that data so that it's part of our archive. But USGS has done a phenomenal job of processing the entire Landsat archive so it can be used for this kind of purpose. Even though we have different instruments throughout that record, the way that we calculate surface reflectance, for example, helps us be pretty confident that when we are seeing a change in time, it's truly a change on the land surface rather than an atmospheric anomaly or something to do with geometric registration of the data. So what Landsat provides us in making the National Land Cover Database is that long record of wall to wall coverage of our country. So we can now utilize that record, and have, to make a time series of land cover. That is one of the unique properties of the new Annual Land Cover Database. This 30-meter record, it's a one of a kind resource. There's nothing else that exists like Landsat for studying our Earth. ADAMSON: What is the value in having annual NLCD versus what it was before, something like every 2 to 3 years? BROWN: So one of the predecessors to this effort, the LCMAP project: Land Change Monitoring, Assessment and Projection, was really the first effort that USGS did to create an annual periodicity of land cover across the U.S. That project ended in order for us to merge that capability with legacy NLCD, which was more on the 2 to 3 year cadence. But LCMAP did not have the detail of land cover classes. So really, the user community was demanding all the detail that they can get and the high periodicity. And a third component was we'd like the longer record. So, you know, we're trying to satisfy a user community that seems to have evolving and ever increasing needs for more complex information. We have a user community that is very interested in performing carbon sequestration modeling, and they want to know where, you know, what is the carbon source and sinks across the country every year. They are thrilled to get annual periodicity. We have users in the water modeling community. And Jon mentioned this to do with impervious surface. So if you're going to utilize land cover to make a model of flooding, for example, you want to know within your watershed what the amount of impervious surface is, what is the amount of developed land cover versus other types of land cover that would absorb the water? And that does change over time. So it just helps some of these models achieve more realistic representation of of the Earth's surface to improve their model accuracy. ADAMSON: How do you know that the data is accurate? BROWN: So at our center, we have a huge long history all the way back to the beginning of NLCD of doing the best job we could at assessing the accuracy of our products. This has evolved over time and it is one of the most difficult things, I think, that we do, although we do a lot of difficult things. The current effort, involves collecting 10,000 reference plots across the country that our team-- members of our team interpret. And these locations, these specific plots, are used to do an independent accuracy assessment of these products. We're in the middle of collecting all of those. It takes a large amount of time and effort to do that. And we almost sequester those folks, like, we can't talk to them because we want those reference data to be truly independent and not to skew our maps. So we have, over the past nine months, that team has collected 5,000 of the ten. And then over the next nine months, we'll collect more to better cover rare classes and land cover change, which is basically rare. And then we'll put together an accuracy assessment and publish it like we have done for many, many years, back to the beginning of NLCD. ADAMSON: It sounds like you already know how to do this. BROWN: We do, but it's still a pretty big job. ADAMSON: Yeah. BROWN: So so we are able to, at this point, take a peek into that with the reference data that we've collected. And, we're not ready to publish numbers, but we're pretty happy with the results. But then I'll say we're never happy with the results because we've always want to do better. And there's always things that we can improve upon. DEWITZ: I don't exactly agree that we know how to do this. This is the first time we're doing land cover accuracy assessment with this complexity. ADAMSON: Okay. DEWITZ: First time it's been done yearly for 40 years with 16 classes of land cover, assessing every single year. So, one point that's been made is even with the 5,000 points, that's multiplied by 40 years, they're assessing every single date. So that's a lot more than 5,000 points. Historically, NLCD has only done that for 2 or 3 dates, which is a lot lower level of effort. So we're transferring the knowledge from doing that 16-class accuracy assessment for just a couple of years across 40 years in the landscape where we don't have high-res imagery for all of the dates like we've had in the past. We're going back into the '80s, where there's only the Landsat record and having to rely on assessors' experience, understanding what Landsat looks like and what the class should be, respectively. ADAMSON: Let me see if I understand what one of these plots is like. What is one of those 10,000 plots? DEWITZ: That's a Landsat pixel. ADAMSON: A plot is a 30-meter pixel. DEWITZ: Yes. ADAMSON: So are you just looking at the imagery? You're kind of doing this remotely? DEWITZ: So you're looking at the imagery. The imagery gives you perspective on the larger landscape. If it's a pixel in the middle of a crop field, that's easier to understand with the context. The same as if it's a pixel in the middle of Lake Erie, the context lets you know you're likely going to get that water pixel right. ADAMSON: Okay. DEWITZ: That water pixel is, say, the little stock pond on the way to work here, it may only be 1 or 2 pixels. That's a lot harder to get the context for what's going on in the landscape, especially if you don't have high-res imagery. ADAMSON: So you're looking at the Landsat image, but you're also looking for some other source imagery to compare it to? DEWITZ: Exactly. ADAMSON: Okay. DEWITZ: If it exists. ADAMSON: If it exists, yeah. DEWITZ: There are sometimes there are mid '90s digital or ortho quarter quads that we can use. Those are black and white and it goes from one in the mid '90s to usually one in the 2000s, one in the mid 2000s and then 2010, you start getting more frequent updates. ADAMSON: NAIP, N-A-I-P, National Agriculture Imagery Program comes to mind. Is that something that's used too? DEWITZ: We use everything that we can get our hands on. ADAMSON: Whatever's out there. Okay. It's easy for me to say, well, why don't you just look at Google Earth and compare it and figure it out? DEWITZ: We use Google Earth. We use Google Earth extensively for accuracy assessment. But again, Google Earth can only provide imagery that was there in history. So you can look in Google Earth to 1985. What you'll find is a Landsat pixel instead of high-res imagery. ADAMSON: Yeah. Sure thing. BROWN: Maybe, maybe one of the important things is how this is all evolution. Like science, our capabilities are evolving, our imagery is evolving. Our tools are evolving. All of this evolves over time. And we're trying to, in our position, stay as close to the cutting edge as we can. But at the same time, Jon has said many times, we can't be on the cutting edge because we need to be repeatable. We need to be transparent. We need to have systems that can be run again and again to produce the data reliably year after year. So we have to strike this balance between trying to to do the best we can and use new methods, updated methods, but yet be consistent. DEWITZ: I would add to that, I would say we're touching the cutting edge, but we can't be bleeding edge. ADAMSON: Yeah, it's very delicate it sounds like. BROWN: We might get cut. ADAMSON: You still might, you still might. Can you name a few examples of benefits of Annual NLCD? What areas is it especially useful for? DEWITZ: For developed areas, for instance, understanding where urban population growth has taken place, especially in large urban centers. The urban sprawl, the spread of infrastructure, all of those things have a profound effect on the local landscape, and that has not been able to be quantified in a way where you could find change over time. This is the first land cover at this scale to go from '85 to current, and have comparable change through that time frame. BROWN: One of the projects, also in this building, has been delving into quantifying and modeling urban heat island effects. And so this data is revolutionary for them. Not only do we have a 39-year record, but we have this detail of land cover and impervious surface, fractional impervious surface over these cities that will help them be able to model and map urban heat island effects through that 39-year record as well. We just had a noon seminar today on projection scenarios of future and past land cover and land use. That's done through a modeling approach called FORE-SCE, and annual land cover feeds into their scenarios so that they can take the land cover record back before Landsat and go back even into the 1600s, and then go forward in time to understand what future land cover might look like. The elegant part of that, too, is bringing in-- they can bring in information related to what we expect out of future climate, and how those land cover types might be affected by climate. Okay, so I just got an email from a person who said he was an environmental economist. I don't even know what an environmental economist really does, but he's interested in using the new products, the Annual NLCD, to help him with modeling flood impacts on underserved communities. DEWITZ: One of the recent questions that came in to the help desk was trying to understand how our data was used to create this map of impaired air quality, and aligning that with the Head Start program to show where children may experience incidences of increased asthma attacks, things like that. ADAMSON: Oh, wow. DEWITZ: So by intersecting NLCD, where developed is and how much developed is in an area, you can get an estimate of potentially poor air quality from cities, intersect that with where Head Start reports locations, and all of a sudden now you have a population estimate for children and how those children may be affected over the long term just by their location on the map. ADAMSON: Wow. DEWITZ: They can mitigate that by putting money into green space in those areas. ADAMSON: Far reaching, far reaching examples and benefits anyway. This product uses a lot of data. How do we deal with so much data? BROWN: To make that 39-year record of land cover, we're actually touching-- Okay, so this is the statistics. Basically 1.4-some million Landsat observations, through history, through time. We are looking at over 3,000 average observations over a spatial area called a tile. We process in something called tiles. ADAMSON: Okay. BROWN: Eight Landsat bands per observation in those tiles. Bands are the different parts of the electromagnetic spectrum that Landsat collects surface reflectance over, collects and processes surface reflectance. And we estimated that we processed or touched over 295 trillion pixels to make this this land cover dataset. ADAMSON: Is that a lot? That seems like a lot. BROWN: We think it's a lot. So we did all of this processing in the commercial cloud. I don't believe this would have been possible, at least not in the time that we processed, if we weren't using cloud compute capabilities. We have a stunningly talented group of scientists and engineers who have set this system up for us. It is designed to be rerun. It's not exactly like, snap your fingers and we're going to do it again, but we are planning on rerunning the system next year, when we can add another year's worth of Landsat data and produce an updated dataset that will cover through time, through 2024. And we want to do that every year into the future. ADAMSON: But it doesn't sound like this is something we could have done a decade ago? DEWITZ: No. BROWN: Absolutely not. ADAMSON: We needed some of these advancements to happen recently, cloud computing, to make this work. BROWN: Cloud computing and artificial intelligence. We're using several deep learning models chained together in our classification architecture. ADAMSON: And what's the output? How many pixels are being released in Annual NLCD? DEWITZ: Over 2 trillion. ADAMSON: That's pretty cool too. Where can people view the map? Is there like an online viewer that people can work with, and how about researchers, can they download the data? DEWITZ: The data will continue to be available as it has in the past on mrlc.gov. We're also adding additional venues. It'll be available for direct processing on AWS. It will also be available on EarthExplorer. And we're currently just finalizing a new website to showcase some of the science and linkages that go into making the National Land Cover Database. It'll function very similar to the previous MRLC viewer, but of course, with 40 years of data, it changes a bit to be more usable. BROWN: Our objective is to offer people as many different ways to get their hands on these data as they want. Some people are not going to want to download 2 trillion pixels. They're just going to want their area of interest for the years they're interested in. The viewer will give them that. If you are a super user, and you're already in the cloud working, you don't have to move the data to your own desktop. You can do processing in the cloud yourself. That's why we ran in the cloud. We did not want to move the Landsat archive to another system to process it, because that's yet more time. ADAMSON: What's next? I assume calling it Annual NLCD means a new release maybe around this time next year? BROWN: Yeah, our hope is actually to do it quicker next year. And our plan is to try and release data in the late spring or early summer. That will be an update for the conterminous U.S., for the lower 48 states, adding the year 2024. So then it truly will be a 40-year record of annual land cover. Our plans are to keep doing that into the future on an annual basis, to try and make improvements to the system. And then in 2026, add Alaska and Hawaii land cover for annual time periods. DEWITZ: The new website is usgs.gov/annualnlcd. ADAMSON: Thank you, Jess Brown and Jon Dewitz for talking with us about Annual NLCD. And thank you, listeners. Check out our social media accounts to watch for our future episodes. You can also subscribe to us on Apple and YouTube podcasts. VARIOUS VOICES: This podcast, this podcast, this podcast, this podcast, this podcast is a product of the U.S. Geological Survey, Department of Interior.