Essential Maintenance: All Authorea-powered sites will be offline 9am-10am EDT Tuesday 28 May
and 11pm-1am EDT Tuesday 28-Wednesday 29 May. We apologise for any inconvenience.

loading page

Investigating High-Latitude Permafrost Carbon Dynamics with Artificial Intelligence and Earth System Data Assimilation
  • +7
  • Bradley A Gay,
  • Andreas E Züfle,
  • Amanda H Armstrong,
  • Neal J Pastick,
  • Jennifer D Watts,
  • Paul A Dirmeyer,
  • Kimberley R Miner,
  • Konrad J Wessels,
  • John J Qu,
  • Charles E Miller
Bradley A Gay
Jet Propulsion Laboratory, California Institute of Technology

Corresponding Author:[email protected]

Author Profile
Andreas E Züfle
Emory University, Department of Computer Science
Amanda H Armstrong
University of Maryland, Earth System Science Interdisciplinary Center
Neal J Pastick
United States Geological Survey, Earth Resources Observation Science Center
Jennifer D Watts
Woodwell Climate Research Center
Paul A Dirmeyer
George Mason University, Department of Atmospheric and Oceanic Sciences
Kimberley R Miner
Jet Propulsion Laboratory, California Institute of Technology
Konrad J Wessels
George Mason University, Department of Geography and Geoinformation Science
John J Qu
George Mason University, Department of Geography and Geoinformation Science
Charles E Miller
Jet Propulsion Laboratory, California Institute of Technology

Abstract

It is well-established that positive feedbacks between permafrost degradation and the release of soil carbon into the atmosphere impacts land-atmosphere interactions, disrupts the global carbon cycle, and accelerates climate change. The widespread distribution of thawing permafrost is causing a cascade of geophysical and biochemical disturbances with global impact. Currently, few earth system models account for permafrost carbon feedback mechanisms. This research identifies, interprets, and explains the feedback sensitivities attributed to permafrost degradation and terrestrial carbon cycling imbalance with in situ and flux tower measurements, remote sensing observations, process-based modeling simulations, and deep learning architecture. We defined and formulated high-resolution polymodal datasets with multitemporal extents and hyperspatiospectral fidelity (i.e., 12.4 million parameters with 13.1 million in situ data points, 2.84 billion ground-controlled remotely sensed data points, and 36.58 million model-based simulation outputs to computationally reflect the state space of the earth system), simulated the non-linear feedback mechanisms attributed to permafrost degradation and carbon cycle perturbation across Alaska with a process-constrained deep learning architecture composed of cascading stacks of convolutionally layered memory-encoded recurrent neural networks (i.e., GeoCryoAI), and interpreted historical and future emulations of freeze-thaw dynamics and the permafrost carbon feedback with a suite of evaluation and performance metrics (e.g., cross-entropic loss, root-mean-square deviation, accuracy). This framework introduces ecological memory components and effectively learns subtle spatiotemporal covariate complexities in high-latitude ecosystems by emulating permafrost degradation and carbon flux dynamics across Alaska with high precision and minimal loss (RMSE: 1.007cm, 0.694nmolCH4m-2s-1, 0.213µmolCO2m-2s-1). This methodology and findings offer significant insight about the permafrost carbon feedback by informing scientists and the public on how climate change is accelerating, strategies to ameliorate the impact of permafrost degradation on the global carbon cycle, and to what extent these connections matter in space and time.
23 Dec 2023Submitted to ESS Open Archive
26 Dec 2023Published in ESS Open Archive