OS029-0004
A new estimate of carbon sequestered within the global methane hydrate stability zone using machine-learning predicted inputs
Friday, 11 December 2020
Poster
Taylor Runyan Lee1, Benjamin J Phrampus2, Warren T Wood1,3 and Adam D Skarke4, (1)US Naval Research Laboratory, Geology and Geophysics, Washington, DC, United States, (2)US Naval Research Laboratory, Washington, DC, United States, (3)Naval Research Laboratory, Stennis Space Ctr, MS, United States, (4)Mississippi State University, Department of Geosciences, Mississippi State, MS, United States
Abstract:
Recent estimates indicate that methane hydrate in seafloor sediment is the largest reservoir of free carbon on earth and sequesters approximately 1/5 of total global carbon. These estimates are calculated using various empirical and deterministic models, which require several global scale inputs to calculate the thickness of the hydrate stability zone (HSZ) and the transformation of organic carbon to methane in marine sediments. Parameterizing these global scale input variables (e.g. total organic carbon, heat flux, sedimentation rates) has traditionally been based upon spatial extrapolation and/or interpolation from sparse datasets to globally complete datasets. Here, we use machine learning algorithms (MLAs) to produce data-driven, globally-complete input datasets, that are more statistically optimal that those resulting from previous approaches. The final predicted values are then used as input into empirical and deterministic models to approximate a lower limit on the quantity of methane hydrate contained within the global HSZ.
For each of the seafloor properties aforementioned, we use a k-nearest neighbor (KNN) MLA to estimate each respective seafloor property at a 5 x 5 arc-minute where it has not been directly observed. KNN uses global geologic grids which are spatially continuous (i.e. predictors) such as distance to coast to determine correlations between observed data and locations where we want to predict. For each observed location and location that we are trying to predict, there exists discrete values from a predictor grid set. Predictors that are correlated with observed data are used to determine a distance in parameter space. The smallest distance in parameter space between observed values and locations where the prediction is being made indicates the observed values to average together to determine a final predicted value.
Our 5 x 5-arc minute global estimate of the HSZ and the quantity of organic carbon sequestered within is consistent with and/or an improvement over previous estimates highlighting known regions of hydrate accumulation (e.g. continental margins, Gulf of Mexico) Final results provide an updated spatial distribution and global methane hydrate estimate, both of which are relevant to the broader global carbon community.
