Combining remote sensing and machine learning to predict current and future vegetation community distributions on the Seward Peninsula, Alaska

Venkata Shashank Konduri, Northeastern University, Boston, MA, Venkata Shashank Konduri and Forrest Hoffman, Computational Earth Sciences Group, Oak Ridge National Laboratory, Oak Ridge, TN, Jitendra Kumar, Environmental Sciences Division, Oak Ridge National Laboratory, Oak Ridge, TN, Verity G. Salmon, Environmental Sciences Division and Climate Change Science Institute, Oak Ridge National Laboratory, Oak Ridge, TN, Colleen Iversen, Climate Change Science Institute and Environmental Sciences Division, Oak Ridge National Laboratory, Oak Ridge, TN, Amy L. Breen, International Arctic Research Center, University of Alaska Fairbanks, William W. Hargrove, Eastern Forest Environmental Threat Assessment Center, US Forest Service, Asheville, NC, Auroop R. Ganguly, Civil and Environmental Engineering, Northeastern University, Boston, MA

Background/Question/Methods

A number of remote sensing and field-based studies point towards a greening trend across much of the circumpolar Arctic region. The shrubification of the Arctic could have a huge impact on the biogeochemistry of tundra ecosystems through changes in surface albedo, soil nutrient availability, atmospheric carbon and energy fluxes and permafrost thaw depth. The existing vegetation maps often lack spatial resolution to capture the diversity and heterogeneity of the plant functional types found in this region. Knowledge of the environmental drivers of vegetation distribution could improve our understanding of various plant physiological processes and also help estimate potential changes in distribution under future climate scenarios.
This study aims at understanding the community composition, landscape-scale configuration and environmental drivers of plant community distribution across different watersheds in the Seward Peninsula region of Alaska, USA. Using data collected as part of field vegetation surveys and airborne hyperspectral imagery from NASA AVIRIS-NG, we created high-resolution (5m) maps of plant communities across different watersheds. We also developed an environmental niche model to understand the various topographic and climate drivers of distribution which was then used to estimate potential changes in spatial distribution for each community through the 21st century.
Results/Conclusions
We developed a deep neural network-based approach that achieved a plant community classification accuracy exceeding 80%. Analysis of various landscape patterns shows that plant communities like the Alder-Willows and Tussock-lichen tundra are more aggregated and occupy a greater proportion of the landscape compared to others such as Mesic Graminoid Herb Meadow and Sedge-Willow Dryas Tundra communities. Using a Random Forest-based environmental niche model, we found that microtopographic features (such as elevation) and soil moisture were stronger drivers compared to climate variables in determining plant community distribution. High resolution maps of vegetation types would improve our knowledge of above-ground trait variability in tundra ecosystems and could serve as datasets for Earth system model parameterization, benchmarking and validation. Insights from niche modeling could help improve our knowledge of mechanisms and environmental drivers of vegetation distribution.