Estimating Fractional Cover of Arctic Tundra Plant Functional Types on the North Slope of Alaska Using Sentinel and Harmonized Plot Observations
Authors
Jitendra Kumar1* (kumarj@ornl.gov), Tianqi Zhang1, Morgan Steckler1, Amy Breen2, Forrest M. Hoffman1, William W. Hargrove3, Colleen Iversen1
Institutions
1Oak Ridge National Laboratory, Oak Ridge, TN; 2University of Alaska–Fairbanks, AK; 3USDA Forest Service, Asheville, NC
URLs
Abstract
Plant functional type (PFT) fractional cover (fCover) information is used to assess vegetation composition and diversity, which is crucial for understanding the health and functioning of Arctic ecosystems. Satellite data provide a means for scaling local, plot-measured fCover to regional scales. The team harmonized open-access vegetation plot data repositories that provide fCover observations in Alaska: the Alaska Vegetation Plots Database, ABR Inc.—Environmental Research and Services, the Arctic Vegetation Archive, and the National Ecological Observatory Network. Leveraging these plot datasets, researchers generated a wall-to-wall map of fCover for tundra PFTs on the north slope of Alaska. While the plot fCover provide invaluable information about the tundra vegetation, they contain large inconsistencies in terms of sampling scale (species- or PFT-level), plot size (0.5 to 55 m in radius), collection year (1950 to 2021), field sampling approaches (point- or quadrat-intercept), and fCover measurement (percent or Braun-Blanquet codes). The team harmonized the plot fCover using a consistent PFT schema: nonvascular, forb, graminoid, deciduous shrub, evergreen shrub, and two non-PFT classes (bare ground and water, suitable for parameterizing terrestrial land surface models). Researchers then linked the plot fCover with satellite-derived explanatory variables (Sentinel-2 spectra, phenology, and Sentinel-1 polarization bands) and trained random forest regression models to estimate fCover. To reduce the spatiotemporal inconsistencies between plot and satellite observations, the team adopted a multivariate outlier detection approach—Cook’s distance—to identify high-quality plots for model training and validation. The results suggest that the random forest regression models can estimate fCover with good accuracy. The correlations between plot-observed and satellite-derived fCover are substantially improved when using high-quality plot samples. The mapped fCover characterizes the spatial patterns of different PFTs across the tundra biome at a high resolution, providing information needed for improved representation of Arctic tundra vegetation in Earth system models to better quantify climate-vegetation feedbacks in the Arctic tundra ecosystem.