Decoding the Spatiotemporal Complexities of the Permafrost Carbon Feedback With Multimodal Ensemble Learning

• Published in: Journal of geophysical research. Machine learning and computation Vol. 2; no. 1
• Main Authors: Gay, B. A., Pastick, N. J., Watts, J. D., Armstrong, A. H., Miner, K. R., Miller, C. E.
• Format: Journal Article
• Language: English
• Published: Wiley 01.03.2025
• Subjects: artificial intelligence; carbon cycling, modeling; climate change; cryosphere; fuzzy logic; machine learning; neural networks; permafrost carbon feedback; remote sensing
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000402

Complex nonlinear relationships exist between the permafrost thermal state, active layer thickness, and terrestrial carbon cycle dynamics. In Arctic and boreal Alaska, significant uncertainties characterize the spatiotemporal rate and magnitude of permafrost degradation and the permafrost carbon feedback, with increasing recognition of the importance of thawing mechanisms. The challenges of monitoring sub‐surface phenomena with remote sensing technology further complicate the issue. There is an urgent need to understand how and to what extent thawing permafrost destabilizes the carbon balance in Alaska and to characterize the feedback involved. In this research, we use our artificial intelligence‐driven model GeoCryoAI to quantify permafrost carbon dynamics in Alaska. The GeoCryoAI model uses a hybridized process‐constrained ensemble learning framework to simultaneously ingest, scale, and analyze in situ measurements, remote sensing observations, and process‐based modeling outputs with disparate spatiotemporal sampling and data densities. We evaluated prior naïve (a) persistence and (b) teacher forcing approaches relative to (c) time‐delayed GeoCryoAI simulations, yielding the following error metrics (RMSE) for active layer thickness (ALT), methane (CH4), and carbon dioxide (CO2), respectively: 1.997, 1.327, 1.007 cm [1963–2022]; 0.884, 0.715, 0.694 nmol CH4km−2 month−1 [1994–2022]; 1.906, 0.697, 0.213 µmol CO2km−2 month−1 [1994–2022]. Our approach overcomes traditional model inefficiencies and resolves spatiotemporal disparities. GeoCryoAI captures abrupt and persistent changes while introducing a novel methodology for assimilating contemporaneous information at various scales. We describe GeoCryoAI, the methodology, our results, and plans for future applications. Plain Language Summary Understanding how climate change rapidly alters the Arctic is paramount, given the potential to mobilize more than 1,400 Pg of organic carbon sequestered in its permafrost, peatlands, and frozen soils. Nevertheless, data prior to satellite remote sensing are sparse in space and time, with multi‐decadal records existing at only a few locations across more than 14.5 million km2. Satellite remote sensing addressed many of the sampling density issues. However, remote sensing data are often less accurate than site‐level measurements, require extensive validation, do not directly sample the geophysical variables of interest, and are difficult to merge with site‐level data exhibiting different spatiotemporal characteristics. We address these challenges using GeoCryoAI, an artificial intelligence‐driven model. GeoCryoAI simultaneously ingests and analyzes disparate data types to provide estimates of the permafrost state and the associated cycling of methane (CH4) and carbon dioxide (CO2) from thawing permafrost. In this paper, we describe the utility and results of the GeoCryoAI architecture for various permafrost domains across Alaska. We assess the strength of the model and discuss evolving applications. Key Points We quantify nonlinear dynamics of the permafrost carbon feedback and reconcile the multimodal data dichotomy with artificial intelligence GeoCryoAI is a hybridized ensemble learning architecture with stacked convolutional layers and memory‐encoded recurrent neural networks This optimized framework substantially improves the efficiency, scalability, and precision of simulating the permafrost carbon feedback