Mapping plant functional type distributions in Arctic ecosystems using WorldView-2 satellite imagery and unsupervised clustering

Thursday, 18 December 2014
Zachary Langford1, Jitendra Kumar1, Forrest M Hoffman2, Victoria L Sloan3, Richard J Norby4 and Stan D Wullschleger1, (1)Oak Ridge National Laboratory, Oak Ridge, TN, United States, (2)University of California Irvine, Department of Earth System Science, Irvine, CA, United States, (3)ORNL, Bristol, United Kingdom, (4)Oak Ridge National Lab, Oak Ridge, TN, United States
The Arctic has emerged as an important focal point for the study of climate change. Arctic vegetation is particularly sensitive to warming conditions and likely to exhibit shifts in species composition, phenology and productivity under changing climate. Modeling of Arctic tundra vegetation requires representation of the heterogeneous tundra landscape, which includes representation of individual plant functional types (PFT). Vegetation exhibits unique spectral characteristics that can be harnessed to discriminate plant types and develop quantitative vegetation indices, such as the Normalized Difference Vegetation Index. We have combined high resolution multi-spectral remote sensing from the WorldView-2 satellite with LiDAR-derived digital elevation models to characterize the tundra landscape in four 100m X 100m sites within the Barrow Environmental Observatory, a 3021 hectare research reserve located at the northern most location on the Alaskan Arctic Coastal Plain. Classification of landscape PFT's using spectral and topographic characteristics yields spatial regions with expectedly similar vegetation characteristics. A field campaign was conducted during peak growing season (June - August) to collect vegetation surveys from a number of 1m x 1m plots in the study region, which were then analyzed for distribution of vegetation types in the plots. Statistical relationships were developed between spectral and topographic characteristics and vegetation type distributions at the vegetation plots. These derived relationships were employed to statistically upscale the vegetation distributions for the landscape based on spectral characteristics. We will describe two versions of PFT upscaling from WorldView-2 imagery: 1) a version computed from multiple imagery through the growing season and 2) a version computed from a single image in the middle of the growing season. This approach allowed us to test the degree to which including phenology helps predict PFT distribution. Ground-truthing was performed using both sets of PFT estimates to characterize uncertainty. Early results show uncertainty exists in wet and inundated areas where bryophyte moss are overestimated. Further investigation will be done for areas of uncertainty and improving our upscaling algorithms.