To accomplish that, RCMAP, based at the USGS Earth Resources Observation and Science (EROS) Center in Sioux Falls, South Dakota, is continuously hunting for new ways to reduce error and make the products as usable as possible. Better modeling, faster computers, better data, more data, expanded products—all of these contribute toward the quest of offering the best possible accuracy.

And their efforts are paying off. Compared to the previous release, the RCMAP dataset published in January 2024 reduced error by 5 to 8 percent. Error reduction is the key measure the project uses to describe how it improves year over year.

How RCMAP Works

Sources/Usage: Public Domain. View Media Details

What RCMAP sets out to do is no easy task. It evaluates the reflectance of each Landsat pixel and then subdivides each one into 10 components. They include shrub, sagebrush, bare ground, herbaceous, annual herbaceous, perennial herbaceous, non-sagebrush shrub, litter, tree canopy and, in the most recent release, shrub height. Most importantly, RCMAP tracks how much each fraction changed on a yearly basis since 1985. 

But Landsat doesn’t “see” those fractional percentages of land cover. So to train models to recognize which pixels are made up of, say, 20% sagebrush, 70% bare ground and 10% perennial herbaceous, Rigge and other researchers spent five seasons (2013-2018) gathering training data. For every pixel painstakingly “ground-truthed” as having a particular mix of those 10 components, the RCMAP model learns to identify similar pixels elsewhere as having a similar makeup. 

Every year since those initial field expeditions, RCMAP has been producing datasets with increasingly more information to help rangeland managers make informed decisions. 

And when it comes to rangeland management, the biggest customer by far is the Bureau of Land Management (BLM), which manages one-tenth of all U.S. land, including 175 million acres in the western states RCMAP covers. The BLM is tasked with balancing key land uses, from preserving habitat for species with reduced populations such as the sage-grouse to leasing land to ranchers for cattle foraging. To do so, it needs accurate information about land health.

Sources/Usage: Public Domain. View Media Details

BLM Assessments

Of course, BLM collects its own detailed land health data, called AIM (Assessment, Inventory, and Monitoring), with boots on the ground visiting and assessing the same components as RCMAP. But RCMAP increases the power of those observations in both space and time by using Landsat data.

“With RCMAP, you can look at changes within a watershed over time that shows you how much bare ground was there in 1985, how much bare ground there was in 2023, the dips and valleys,” pointed out Nathan Kleist, a USGS ecologist at the Fort Collins Science Center (FORT) in Colorado who works closely with BLM managers in analyzing remote sensing data. “Do those fluctuations over time correspond to observed climate events like decreases or increases in precipitation, particularly in hot years?”

AIM data provide an important ground-truthing measurement for RCMAP. The closer the RCMAP model is to what AIM’s eyes in the field say is there, the more confident Rigge and his coworkers can be that they’re on track. When measured against AIM data, the most recent RCMAP release is 15% more accurate than the previous RCMAP release, Rigge said.

Better Modeling, Faster Computing

The overall decrease in error of 5-8% compared to all independent data in the latest RCMAP dataset can be attributed to improvements in processing and methods, Rigge said. In technical terms, they used a neural net rather than a regression tree to improve both speed and accuracy.

“One thing we learned right away is that we can actually have multiple predictions coming out of a model,” explained Kory Postma, a data scientist who helps crunch data for multiple projects at EROS. Previously, using regression tree analysis, RCMAP could only calculate the percentage of a single fractional component at a time—for example, sagebrush. “But now with neural nets, we could actually get a probability for every single classification category at the same time. And the first thing we noticed was the data were less noisy and more confident,” he said.

A main upgrade in methods was to more accurately reflect the green and brown periods each year—or in simpler terms, the change between seasons. Before, RCMAP calculated seasonal composites based on defined date ranges per region. Now they use a statistical-based method for selecting Landsat images for compositing. “Our composites have three to four times more difference between the green and brown composites than they did last time around, and that’s huge for detecting components like annual herbaceous that have a short window of active growth each year,” Rigge said.

Sweetening that success was the speed with which it was accomplished—partly due to the ability to process the 100 terabytes of data needed on Hovenweep, the newest USGS high performance computing (HPC) system, which is housed at EROS. “We did the final model run in about a day and a half,” Rigge said, compared to two weeks for the previous data release.

Sources/Usage: Public Domain. View Media Details

Better Data … and More Data

Another strength of the recent release is the increase in both the quality and quantity of the underlying data, especially the use of Landsat Collection 2 data. “Just right out of the box, those data are better than Collection 1,” Rigge said. “They’ve got improved geolocation accuracy and a higher dynamic range of reflectance values.” In other words, RCMAP needs to be sure that each Landsat pixel is in the exact same spot every year, and Collection 2 makes sure that’s true. 

Additional datasets also were added, including the 2023 AIM data from BLM and a set similar to AIM called Landscape Data Commons, which is compatible with AIM data.

And for the first time, RCMAP has gone international, including a sliver of Canada where the sagebrush biome crosses the border north of Montana into Alberta and Saskatchewan. This is part of a collaboration with Colorado State University researcher Julie Heinrichs, who provided the training dataset needed to accurately assess the fractional components for RCMAP.

Sources/Usage: Public Domain. View Media Details

Expanding Products to Protect Habitat

Concern about sage-grouse habitat spurred the move into Canada, where the species is endangered, with only 250 birds remaining. 

“Sage-grouse is not doing well in the U.S., but it’s doing even worse in Canada,” Rigge said. “The Canadian portion of the sagebrush biome was always more marginal habitat to begin with than in the U.S. portion.” RCMAP’s data can help land managers identify the causes for the decline in sagebrush, a necessary habitat for the sage-grouse.

Concern about the species also was behind the addition of shrub height in the most recent RCMAP release. Sage-grouse prefer certain sagebrush heights that provide more cover to hide from predators and for nesting. Adding shrub height is another way that RCMAP seeks to provide rangeland managers key data, similar to the effort last year to add tree canopy to the list of components. Trees give predatory birds a better perch to prey on sage-grouse.

What’s Next for RCMAP

While future growth of RCMAP’s coverage area might include stretching the boundaries eastward into the Great Plains, Rigge is not interested in expansion just to get bigger. Any moves would need sound scientific reasoning behind them.

For now, in addition to the constant preparation for the next data release, the RCMAP team is anticipating publishing some of its results. One planned paper will deal with the impacts of spatial scaling on error in the RCMAP products. Another will focus on how error varies across regions.

“The papers will be through more of a management lens,” Rigge said. “How do users apply these data to their day-to-day activities? What kind of caveats do they need to think about while using these data? How can they best and most accurately apply these data to the decisions that they need to make?”