Quantitative Interpretation of Arctic Tundra Attributes Using Remote Sensing: Leveraging Field Data, Modern- and Legacy Landsat Data, and Commercial Imagery in Northern Alaska

Thursday, 18 December 2014
Gerald V Frost Jr, Matthew J Macander and Peter R Nelson, Alaska Biological Research, Inc., Fairbanks, AK, United States
Integrated analysis of ground-based vegetation data and remote sensing supports vegetation mapping, landscape-change detection, wildlife habitat assessment, and tracking of phenological events such as green-up and senescence. The life-cycles of tundra plants occur within a highly compressed seasonal window, making the quantitative assessment of vegetation and landscape attributes from ≤30m resolution remotely-sensed imagery, such as above-ground biomass, % shrub cover, and % surface water, a difficult task when applied across large study domains. To support mapping of vegetation and landscape attributes across ~100,000 km2 of Alaska’s North Slope, we obtained ground data for tundra vegetation using a point-intercept sampling approach across a network of 107 field plots spanning gradients of bioclimate, landscape position (upland, lowland, riverine), and geomorphic setting (foothills, coastal plain). At each plot, vegetation data were collected along three 50-m linear transects, compatible with 30-m Landsat imagery. We summarized live vegetation, litter, and non-vegetated surfaces using three terms: top cover (uppermost “hit”), percent cover (total areal cover along transect), and hit density (all “hits” at a point). We then evaluated a suite of data models (e.g., General Additive Models, classification tree, clustering) and data-mining approaches (e.g., neural networks, random forest) using midsummer Landsat TM/ETM+ acquisitions since 1985, and OLI acquisitions for 2013–2014. The large size, frequent cloudiness, and interannual variability of the study area necessitated the compositing of a multitude of Landsat scenes. A median NDVI compositing technique was used to select Landsat observations from cloud- and shadow-free pixels that met day-of-year and year constraints. This technique produced seamless, phenologically consistent composites that are largely free of artifacts and suitable for regional-scale analysis. Ground-based training data and an archive of very-high-resolution (< 2-m pixel) commercial imagery provide means to evaluate statistical and data-mining approaches.