GeoCryoAI Permafrost, Thaw Depth and Carbon Flux in Alaska, 1969-2022

This dataset provides model code, input data, sample results, and documentation for an artificial intelligence-driven model, GeoCryoAI. The purpose of GeoCryoAI is to quantify permafrost thaw dynamics and greenhouse gas emissions in Alaska.

This a preprint dataset that has not undergone quality assurance nor editorial review.

This a preprint dataset that has not undergone quality assurance nor editorial review.

Project: Arctic-Boreal Vulnerability Experiment

The Arctic-Boreal Vulnerability Experiment (ABoVE) is a NASA Terrestrial Ecology Program field campaign based in Alaska and western Canada between 2016 and 2021. Research for ABoVE links field-based, process-level studies with geospatial data products derived from airborne and satellite sensors, providing a foundation for improving the analysis and modeling capabilities needed to understand and predict ecosystem responses and societal implications.

Acknowledgements:

This study was funded by NASA's Postdoctoral Program at the Jet Propulsion Laboratory, California Institute of Technology, administered under NASA grant 80NM0018D0004.

Spatial Coverage: Alaska

Spatial Resolution: 1 km

Temporal Coverage: 1969-01-01 to 2022-12-31

Temporal Resolution: Monthly

Study Area (These coordinates are the approximate locations of the study sites and may not match the extent of the data files. All latitude and longitude are given in decimal degrees.)

Site	Northern Extent	Western Extent	Eastern Extent	Southern Extent
Alaska	72.0	-170.0	-139.0	54.0

Data File Information

The text file, geocryoai_documentation.txt, provides a detailed description of files in this dataset.

Most data files are organized into three folders, and the content is provided in zip archives. Larger folders were split into separate zip archives

• documentation - - detailed description of model, including architecture and work flow
  • documentation.zip . 
• modeling - model code, sample input data, development, evaluation, and sample results in respective subfolders. 
  • modeling_code.zip -  holds the code, development, and evaluation subfolders. The code subfolder includes a Jupyter notebook to demonstrate model execution. 
  • modeling_data.zip - holds the data subfolder.
  • modeling_results.zip - holds the results subfolder.
• preprocessing - code and data, in separate zip files, for preparing input data files.
  • preprocessing_data.zip - data files for preparing input into the AI model
  • preprocessing_code.zip - scripts for preparing input data files for the AI model

The overall objective was to quantify how the Arctic is changing in response to climate change and what quantifiable evidence of the permafrost carbon feedback may we measure and simulate to better understand and simulate the risks and uncertainty of this chaotic nonlinear feedback and its impact on the earth system.

The quantification of precision for permafrost degradation and carbon release were captured with each iteration of the pre-processing workflow prior to resampling, which required aggregation or interpolation, thus introducing generalization and further uncertainty. The uncertainty benchmarks were evaluated prior to and concluding resampling methods using the spatial average and standard deviation. All uncertainties were computed with the uncertain panda Python library. Collectively, ALT, CH4 flux, and CO2 flux and their derived monthly uncertainties were computed as -0.4238+/-0.134 m, 0.499+/-3.659 nmol C(CH4) km^-2 month^-1, and -5.272+/-127.717 µmol C(CO2) km^-2 month^-1, respectively. 

GeoCryoAI is a hybridized process-constrained ensemble learning framework consisting of stacked convolutionally layered long short-term memory-encoded recurrent neural networks. This architecture leverages a multimodal composite of in situ measurements (Gay et al., 2023), airborne remote sensing observations, and PBM outputs to ensure biogeophysical processes are emulated by physical laws while improving the performance and predictive accuracy of the model. Moreover, this framework integrates optimization equations, regularization cost functions, and a Bayesian Optimization search algorithm to populate a hyperparameter dictionary, reconcile data mismatches, expedite convergence, and minimize error residuals during teacher forcing.