Eyes on Earth Episode 107 – EROS 50th: Land Cover, Part 2

Jane Lawson:

Hello, everyone, and welcome to another episode of Eyes on Earth, a podcast produced at the USGS EROS Center, which celebrates its 50th anniversary this year. Our podcast focuses on our ever-changing planet and on the people at EROS and around the globe who use remote sensing to monitor the health of Earth. My name is Jane Lawson, and I'll be hosting today's episode, where we're talking about the future of land cover mapping in the second of two episodes focused on land cover. EROS has a long history and strong reputation for land cover mapping, including the National Land Cover Database, or NLCD, which is the definitive U.S. land cover product. In addition, the Land Change Monitoring, Assessment and Projection, or LCMAP, effort at EROS has more recently provided a line of products as well. The two projects, both based on 30 meter resolution Landsat satellite data, are currently being combined to bring the advantages of both to future data releases under the NLCD name. Our guests today are here to talk about how the integration of NLCD and LCMAP will benefit the vast variety of people who use the land cover and change data. USGS research physical scientist Terry Sohl has 30 years of land cover experience at EROS, including development of the first version of NLCD. And USGS physical scientist Jon Dewitz has worked with NLCD at EROS for more than 20 years. Welcome, Terry and Jon, to Eyes on Earth. 

Jon Dewitz:

Thank you. 

Terry Sohl:

Thanks, Jane. 

Lawson:

First, let's talk about NLCD. Terry, do you want to explain to us why NLCD was so important to develop and what benefits have been seen? 

Sohl:

Back in the early 1990s, there was a consortium of federal agencies, the Multi-Resolution Landscape Characteristics Consortium, and at the time it was really a groundbreaking effort to even think about trying to map land cover at Landsat scale across the entire U.S. But you know, from an agency perspective, you can definitely see the need for it in that land cover and land use, what's happening on the surface of the Earth, has such a huge impact on what's happening with ecosystem processes and societal processes. And so if you think about things like biodiversity and habitat, carbon greenhouse gases, regional weather and climate, water quality, human health, each one of those has some impact that land use and land cover can help to address. So NLCD was really the first effort to do that in a concerted way across the whole U.S.

Lawson:

And how does EROS fit into that bigger picture MRLC multi-agency collaboration that you talked about? 

Sohl:

Well, at the time, you know, we've always been the home of Landsat. Back in the 1990s when this started, part of it was cost. You know, Landsat scenes themselves were expensive, and federal agencies really were the only ones that had the financial backing as a consortium to be able to purchase the raw data that went into something like a national scale land cover mapping. But EROS itself played the key role in developing NLCD as a project in that it started as a prototype in the mid-Atlantic area. And so there were three of us that helped to develop that first prototype using mosaics of Landsat imagery, not really knowing how that might work at a regional scale, much less a national scale. So EROS kind of led the way, not only from the perspective of, you know, serving as the source of that Landsat data, but also serving as the center for scientific expertise for helping develop the first NLCD. 

Lawson:

Terry, why do you think NLCD succeeded so well as a source of land cover details? 

Sohl:

To be honest, part of it was opportunity. I mentioned the cost. When Landsat scenes could cost anywhere up to a couple of thousand dollars each, and you have over 400 for the United States, now even trying to purchase one Landsat scene and map land cover for the U.S. was a pretty daunting task, just financially. Computationally, too, it was such a challenge in that, you know, trying to take one Landsat scene and goeregister it, trying to place it on a map grid that could actually be used for mapping. That was a very laborious process back in the early 1990s that would involve manually picking ground control points, running a georegistration that usually would run on the machine anywhere between 2 and 4 hours for each scene. So part of it was resources. We had the resources, we had the capability to do it when many others didn't. But the reason it succeeded is it was a good product. And, you know, that's what has sustained NLCD to the current day, that it's maintained its status as the gold standard of land use and land cover for the U.S.

Lawson:

And Jon, you were part of the quality when you started at EROS and have continued to be. Do you want to talk anything about the standards and quality that you maintain? 

Dewitz:

I think one of the the biggest pieces of NLCD that's persisted through time is just trying to do the best land cover that we can. When NLCD 2001 started, we had kind of an evolution as we're having right now with LCMAP and NLCD. You know, that revolution was moving towards the first start of machine learning algorithms from cluster-based processing that NLCD 92 had, and with more of an automated methodology in 2001, that allowed us to really bring our partners in and have a partner-based mapping strategy. For 2001, there were seven active partners mapping along with the USGS. Some of those were states. Some of those were other federal agencies, and some of those were other USGS projects. Now having a somewhat standardized methodology allowed us to do training with those partners and allowed consistency that didn't have to have an eyeball on everything like we had in the past. That really allowed the methodology to be adopted across a larger scale and really some cost savings.

Lawson:

Let's talk a bit about LCMAP and why that was important to develop. And a brief rundown of how it's different from NLCD. 

Sohl:

LCMAP has its genesis going back about eight or nine years, and part of the rationale for taking on LCMAP was some innovation from the methodological side. When we started the first NLCD, it was two Landsat scenes for every location, one with leaves on, one with leaves off. And we reached a stage computationally, data accessibility and affordability, and from a practical standpoint now where we can look at every single pixel in the Landsat archive and use that to our advantage to look at how vegetation greens up and how it senesces in the fall, use that information to look where things are anomalous, where they look abnormal compared to normal. And that's something that, from an LCMAP perspective, was really innovative, something that allowed us to greatly expand the utility of the Landsat archive and also go back further in time than we've traditionally done with NLCD. So, you know, both from a methodological perspective and from the perspective of providing an expanded database of products, LCMAP has been revolutionary.

Dewitz:

Moving from single date change comparisons to the full spectrum of the Landsat record with LCMAP, I think is the next evolution for NLCD. Just like we changed our methodology in 2001 to make it more easily adoptable and transferable, the next evolution of NLCD combining with LCMAP will give us other benefits. Because we were doing that single date temporal change comparison, we did miss some change. And with this full analysis of the Landsat record and the harmonic regressions that LCMAP uses, we really hope to find more of that change and make sure that change is present on the landscape in every instance that it appears. I am really very hopeful that this will provide greater change mapping and more consistency throughout the mapping record than NLCD could have done before this merger. 

Lawson:

Do you want to give us some examples of the types of change that you might be referring to? 

Dewitz:

So some change is only present for a very short time in the landscape. One instance is shrub change in the West. One of the things that the West is dealing with now is invasive species, and cheatgrass is one of them, for an example. That type of grass promotes a lot of fire events, and those fire events can change the shrub structure of the Western U.S. Now, those fire events are only present and really visible on the Landsat imagery for a period of months. And with the previous NLCD methodology, we sometimes have 2 to 3 years between change detection. So most of the time we missed that change and had to rely on other maps to accurately map that change. We're very hopeful that the LCMAP methodology combined with NLCD methodology going forward will allow us to capture more of that change, and hopefully all of that change, and do a better job representing that on the landscape.

Lawson:

So in addition to this specific fire change that you talked about, what other types of research or uses are you hoping that this merger will benefit? 

Sohl:

One of the things that I think I'm most excited about with the LCMAP and NLCD is bringing together the best of both worlds, both the increased thematic resolution, the increased number of classes that NLCD has over LCMAP, but having that long temporal record that LCMAP has going back to the '80s. We're in a time now where Landsat's not the only game in town, and we have, you know, Sentinel and other satellite systems that are up there now that are increasingly being used. But the one thing that none of them are ever going to be able to touch is that long temporal record. And so combining LCMAP and NLCD, that long temporal record with the higher thematic resolution of NLCD, really opens up a lot of research opportunities, particularly on climate change, and trying to look at long-term changes on the landscape, some of those subtle shifts. So what might be happening on the landscape that Jon mentioned and how those might be impacted on climate. I think that's what, to me, is probably one of the most exciting areas that the new products will be very useful for. 

Lawson:

Do you want to just explain the thematic classes that you were referring to a little bit. 

Sohl:

So from a thematic class perspective, LCMAP has something that we call Andersen Level 1, or something close to that. That's a classification scheme that goes back to something that was developed by USGS decades ago, but it's a fairly simple classification scheme with about eight or nine land cover classes. So for example, there's one forest class, there's one urban class. NLCD, on the other hand, has more detail. And so from an urban perspective, there are four classes that are differentiated by the amount of imperviousness that is present on the landscape. For forests, there are three classes. And so it's just a greater definition of actually what's on the landscape. And for certain applications, that really has an impact on the applicability of those source data for addressing something like surface water runoff and what kind of impervious surface amount is present. Or from a biodiversity perspective, what forest type is present.

Dewitz

One of the other evolutions starting in 2001 was continuous field mapping. The present land cover structure is a thematic class, which means that it's simply a type of forest, or a type of developed. With continuous field mapping, we have a range from 0 to 100%, and that was used to map the density of impervious surfaces and the density of forest canopy. Those continuous field maps, they really allowed a lot more information to be known about what's happening on the landscape. For instance, if you think of a pixel inside your city, it's not just all developed. You look down at a house, and you can see most of the time over half of that area is probably lawn, things that are not developed surfaces. And knowing how much of that is there is very important for things like stream runoff, understanding urban forests, and that evolution has really allowed more people to use it and more people to use our classes accurately. As we've moved through time, we've had other partners take over some of those things. The Forest Service took over canopy mapping because, well, they're the stewards of the forest. They're working closely with us as a partner to provide that forest canopy. And NLCD will continue mapping percent impervious surface. As we look through time, I think the consistency of mapping is going to be improved through this merger. Having every pixel of Landsat be analyzed should provide more consistent mapping and provide that mapping across a much broader time frame. That time frame is something that's becoming more and more important to researchers so that they can look back through time and really understand what the ecosystem looked like in the past. And with NLCD currently only mapping back to 2001, there is a big chunk of the Landsat record that's missed. That consistency, moving back through time, will allow researchers to understand the ecosystems and landscape more than was possible previously.

Lawson:

This sounds like a challenging effort to integrate these two big projects together. How is that going? Do you have a timeframe or anything for when the next release will be, or the first in this integration? 

Sohl:

It's a challenge. There's no doubt. And you know, both from a methodological perspective, we're trying something new, I mean, we're using new deep learning approaches that are new to both NLCD and LCMAP. We're trying to take the best of both worlds of what we've already done in the past with NLCD and LCMAP. And those two mapping paradigms have been a little different over the past few years. And so it has been a challenge to try to put those together. I think it's actually going well. We do have a wonderful research team, both on the USGS side and the contract side. We have some really great integration going on in the personnel that were previously on those two projects. And then one of the biggest challenges that doesn't get talked about enough is that it's not just from a methodological perspective, it's from a practical perspective. Again, we're touching every pixel in the Landsat archive. We have Landsat Collection 2 that's now based in the cloud. So to be able to efficiently process all those data, we need to bring the algorithm to the data in the cloud. And so setting up the infrastructure for the new effort, to be able to take those algorithms into the cloud has also been a little bit of a challenge. But we're going to get there, and we're going to have a product by the end of 2024.

Lawson:

Now, taking a big picture view as the world continues to change, and you've hinted at that as well, how crucial will land cover and change data like this and the data in the future, be? 

Sohl:

Yeah, it's going to be critical, Jane, and not just from a cover perspective. You know, one of the visions that we have in the science branch at EROS is trying to move beyond being viewed as a source of land cover. We have other projects that are going on that are looking at changes in landscape condition. So whether that's, you know, the potential impacts of drought on vegetation condition, or how drought can impact agricultural land, or looking at something a little more subtle in terms of how vegetation communities are changing over time, we're trying to expand the type of information we're providing. And so I view the science branch at EROS and I view our role as land cover is expanding in the future to provide even more information on land cover, land use, land condition and land management. And I think it's really that more comprehensive view of what's happening on the landscape that's really going to be critical and set us apart and maintain our role as the gold standard of land change monitoring. 

Lawson:

Jon, do you have any closing thoughts about EROS or land cover going forward? 

Dewitz:

I think probably the biggest strength that NLCD has brought with it is our land cover. That land cover hasn't been tied to a single methodology. We've evolved over time to incorporate the latest, and what we hope are the best, methodologies. But the core thing that NLCD has always brought is that quality in the land cover. Previous NLCD work did rely still on a lot of hand edits for some things, and those hand edits really benefited the accuracy and consistency of the map. Bringing that forward and combining that with LCMAP's ability to see the entire record, bringing that as training data into this new structure using machine learning algorithms, we're really very hopeful that this will provide that same level of consistency, quality and accuracy in a methodology that allows us to do yearly land covers through the entire Landsat record. So that's what I'm really hopeful NLCD will bring to this new evolution.

Lawson:

Thank you, Terry and Jon, for joining us for this episode of Eyes on Earth, where we looked at future land cover and land change work at EROS. And thank you to the listeners. Check out our Facebook and Twitter pages to watch for our newest episodes. And you can also subscribe to us on Apple and Google 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.