Ground-penetrating radar, electromagnetic induction, terrain, and vegetation observations coupled with machine learning to map permafrost distribution at Twelvemile Lake, Alaska – Wiley – S. Campbell et al.

Seth William Campbell1,2,3| Martin Briggs4| Samuel G. Roy1,5|Thomas A. Douglas6| Stephanie Saari6

1University of Maine, School of Earth and Climate Sciences, Orono, Maine, USA  2 University of Maine, Climate Change Institute, Orono, Maine, USA  3U.S. Army Cold Regions Research and Engineering Laboratory, Hanover, New Hampshire, USA  4 U.S. Geological Survey, Earth System Processes Division, Hydrogeophysics Branch, Storrs, Conneticut, USA  5 University of Maine, Senator George J. Mitchell Center for Sustainability Solutions, Orono, Maine, USA  6 U.S. Army Cold Regions Research and Engineering Laboratory, Fort Wainwright, Alaska, USA   Correspondence Seth Campbell, University of Maine, School of Earth and Climate Sciences, Orono, ME, USA.Email: Funding information National Science Foundation; Strategic Environmental Research and Development


Program Abstract: We collected ground-penetrating radar (GPR) and frequency-domain electromagneticinduction (FDEM) profiles in 2011 and 2012 to identify the extent of permafrost relative to surface biomass and solar insolation around Twelvemile Lake near Fort Yukon, Alaska. We compared a Landsat-derived biomass estimate and modeled solar insolation from a digital elevation model to the geophysical measurements. We show correspondence between vegetation type and biomass relative to permafrost extent and seasonal freeze–thaw. Thicker permafrost (≥25 m) was covered by greater biomass,and seasonal thaw depths in these regions were minimal (1 m). Shallow (1–3 m depth)and thin (20–50 cm) newly forming permafrost or frozen layers from the previous winter occurred below northward oriented slopes with thin biomass cover. South-facing slopes exhibited permafrost when there was enough biomass to shield incoming solar energy. We developed an artificial neural network to predict permafrost extent across the broader region by mapping GPR-observed instances of permafrost to FDEM, biomass, and terrain observations with 90.2% accuracy. We identified a strong linear correlation (r=−0.77) between permafrost probability and seasonal thaw depth, indicating that our models may also be used to explore thaw patterns and variability in active layer thickness. This study highlights the combined influence of biomass and terrain on the presence of permafrost and the value of evaluating such parameters via remote sensing to predict permafrost spatial or temporal vari-ability. Incorporating diverse geophysical datasets with in-situ validation into machine learning models demonstrates a useful approach to upscale estimated permafrost extent across large Arctic expanses.