The advent of remote sensing was a boon to the ancient science of phenology.
Satellite data offered a global view of nature’s seasonal life cycles that historical tabulators of budding trees and buzzing bees could scarcely imagine.
From the start, however, remotely sensed phenology has come with caveats.
Where on-the-ground observations have granular accuracy without broad context, remote sensing via satellite offers broad insight but misses subtle details.
Detail is defined by the satellite sensor, and there are tradeoffs from one sensor to the next.
Opt for finer spatial readings—say, the 30-meter pixels of Landsat’s Operational Land Imager (OLI)—and scientists lose the near-daily coverage offered by a sensor like NASA’s Moderate Resolution Imaging Spectroradiometer (MODIS).
Choose MODIS—avoiding the cloud cover that contaminates a Landsat scene and leaves users waiting 8 to 16 days for the next one —and the resolution widens to a grittier 250 meters per pixel.
Some users need a closer look to glean usable information, and researchers at the U.S. Geological Survey’s (USGS) Earth Resources Observation and Science (EROS) Center have spent years working toward that goal.
The need for both detail and speed is most pressing in dryland ecosystems, where a flush of precipitation can feed rapid growth in species like the invasive, fire-feeding cheatgrass that has bedeviled land managers in the western U.S. for decades.
What if there were a way to offer reliable 30-meter readings to land managers with MODIS-level frequency? Moreover, what options will remain for the multitude of MODIS-reliant phenology projects across the globe when the instrument is decommissioned a few years from now?
New research from EROS on dryland phenology ponders those questions and offers promising results.
“When MODIS gets decommissioned, we’ll have to figure out a way to map these phenologies across large, heterogeneous landscapes,” said Neal Pastick of SGT, Inc., a contractor to the USGS and the study’s lead author. “Anything that has to do with vegetation dynamics, drought—anything related to those subjects—we’ve got to figure out a way to monitor these systems. Our approach was kind of the first step in that direction.”
Most in-season phenological modeling has come to rely on some combination of data from MODIS and two other sensors with lower spatial resolution: The Advanced Avery High Resolution Radiometer (AVHRR) and Visible Infrared Imaging Radiometer Suite (VIIRS) from the National Oceanic and Atmospheric Administration (NOAA).
With more testing, the new approach could offer accurate modeling with finer detail, fully independent of MODIS, AVHRR or VIIRS.
The paper, titled “Spatiotemporal Analysis of Landsat-8 and Sentinel-2 Data to Support Monitoring of Dryland Ecosystems,” was published in May by the journal “Remote Sensing.”
The paper was co-authored by Bruce Wylie of USGS EROS, and Zhuoting Wu of USGS’ National Land Imaging Program in Flagstaff.
Landsat and Sentinel data, harmonized in a single season
The research hinges on a novel approach to the harmonization of Landsat and Sentinel satellite data.
The team used a regression tree model that combined single-season readings from Landsat 8’s Operational Land Imager (OLI) and Sentinel-2’s Multispectral Instrument (MSI) to accurately predict seasonal variation in vegetation at 30-meter resolution across a 12,000 square kilometer swath of the Great Basin between Utah and Nevada.
Most studies combining data from the U.S. and European Space Agency satellites have looked to calibrate readings from Landsat 8’s OLI and Sentinel’s MSI by combing scenes from previous years.
Those methods rely on the time of year—the length of day and the sun’s position in the sky— to fill in the gaps created by cloud cover over individual scenes.
That’s useful in areas with established forests and heartier vegetation, said co-author Bruce Wylie of the USGS, but misses change in areas than green up rapidly in response to in-season rainfall.
“Here’s a moisture limited system where days can be as long as they want, but if you don’t have water, it doesn’t matter much,” Wylie said. “We have a system that’s driven less so by day of year and more so by water availability.”
Combining data from both sensors’ infrared and near-infrared bands produced Normalized Difference Vegetation Index (NDVI) predictions that closely mirrored measured signals.
The regression tree model used in the paper made it possible to create 30-meter synthetic images for any day of the growing season across the study area.
If popcorn-shaped clouds taint scenes collected on a day when observers were in the field, the model can use readings from the days before or after it to fill in the gaps and correlate the observations.
The results suggest that it might even be possible to offer 30-meter projections in a host of situations, including for the days immediately before or after a fire.
“We have yet to see how well our synthetic images capture near instantaneous disturbances like fire, but we nailed it on this small study area,” Wylie said.
New approach could improve land cover products
The results hold promise for a range of projects that currently rely on courser sensors to produce NDVIs, such as fire danger and biomass mapping or leaf-area indexes.
The model also could become a useful tool for the National Land Cover Database (NLCD). That project’s shrub and grass vegetation classifications generally rely on three readings across a season.
If further testing proves out the harmonization approach across multiple geographic areas, the near-daily synthetic readings could be used to fine-tune those classifications, said Collin Homer, the USGS lead for the NLCD.
Green cheatgrass doesn’t look much different from more desirable native vegetation in a single scene, but its life cycle is shorter. A model showing in-season change could help Homer’s team sort out those differences.
“Based on that performance, you know that it’s cheatgrass, as opposed to something that doesn’t green up and green down as quickly,” Homer said. “This adds a temporal dimension to the data.”
With a loss of MODIS data on the horizon, Homer is pleased to see new approaches emerge.
“MODIS is 250 meters, and it’s kind of on the way out, so it’s pretty cool that we’ll have new tools to distill the data,” he said.
The next steps for Pastick and Wylie are to expand the time series into other years, incorporate more bands of data, and add data from earlier Landsat satellites.
“We did it for NDVI from two spectral bands, but the potential is there for doing it with all the bands,” Wylie said.
Land managers see value in finer resolution
A more immediate goal is to test the near-real time model across more of the American West. That’s something land managers have wanted for years, and badly enough to pay for it.
Lindy Garner of the U.S. Fish and Wildlife Service (FWS) signed a Statement of Work last year with EROS to downscale its current 250-meter maps to 30 meters for the Eastern portion of the sagebrush ecosystem, where cheatgrass has yet to gain a permanent foothold.
“There’s a lot of effort being put at a 250-meter scale, which is useful when you’re looking at the big picture,” Garner said. “But from the land management side, we need it at a finer resolution to make those decisions.”
It might not make sense to pour resources into eradication if an invasion is widespread, Garner said.
But if remotely-sensed vegetation maps have enough regional detail to highlight invasion-free areas for prevention and low invasion areas for early detection and rapid response, they can help guide allocation of resources and seasonal management efforts.
Land managers look at resilience and resistance to determine the response to invasive grasses, Garner said, with 2016’s Integrated Rangeland Fire Management Strategy and DOI Science Framework for Conservation and Restoration of the Sagebrush Biome as guides.
Accurate data is critical to making cost-effective choices to protect communities, wildlife habitat and hunting opportunities.
If an area has low resistance to invasion and high levels of cheatgrass, the fire risk jumps significantly, and land managers must act accordingly.
“If we could deal and treat these invasive grasses and get in front of the game, we could possibly reduce fires in local areas,” Garner said.
The initial plan was to downscale EROS’ existing cheatgrass maps to 30 meters using multi-year NDVIs, with a focus on the Betty Buttes area in Oregon and blocks of public land in Colorado and Wyoming.
Pastick and Wylie intend to hone the harmonization technique for dryland mapping and use it to produce near-real time maps for Garner and other land managers by 2020.