A machine learning approach to predicting seafloor properties and their application in estimating a global methane hydrate inventory

• Main Author: Lee, Taylor Runyan
• Format: Text
• Language: unknown
• Published: Scholars Junction 2021
• Subjects: Earth science; geology; marine geology; machine learning; geospatial; geophysics; Methane hydrate
• Online Access: https://scholarsjunction.msstate.edu/td/5280; https://scholarsjunction.msstate.edu/context/td/article/6280/viewcontent/taylor_lee_dissertation.pdf

Seafloor properties, including total organic carbon (TOC) and the vertical thickness (isochores) of geological units, are sparsely measured on a global scale and spatial interpolation (prediction) techniques are often used as a proxy for observations. Previous geospatial interpolations of seafloor TOC exhibit gaps where little to no observed data exists. Recent machine learning techniques, based upon a suite of geophysical and geochemical properties (e.g., seafloor biomass, porosity, distance from coast) show promise in making globally complete, comprehensive, and statistically robust geospatial seafloor predictions. Here I apply a non-parametric (i.e., data-driven) machine learning (ML) algorithm, specifically k-nearest neighbors (kNN), to estimate the global distribution of seafloor TOC and marine isochores. This machine learning approach shows major advantages relative to geospatial interpolation, including results that are quantitative, easily updatable, accompanied with uncertainty estimation, and agnostic to spatial gaps in observations. Additionally. analysis of parameter space sample density provides a guide for future sampling. Resulting predictions of the global distribution of seafloor TOC and marine isochore thicknesses were used with ML workflow to predict other seafloor parameters (e.g., heat flow, temperature, salinity) in order to constrain the global distribution of the base of hydrate stability zone and methane generation for all sub-seafloor sediments. Estimating global carbon budgets is first-order dependent on accurate model input, therefore our estimate of the base of hydrate stability zone, and subsequent carbon and methane accumulation in the subseafloor yields improvement over the standard interpolation techniques used in previous global modeling analyses. By using these globally updateable machine learning parameters as the input to predictions, results provide easily updated global budgets of total carbon and methane generated. This dissertation presents valuable new global ...