Global Ocean Carbon Dioxide Flux Mapping Techniques: Evaluation, Development, and Discrepancies

• Main Author: Gloege, Lucas
• Format: Thesis
• Language: unknown
• Published: Columbia University 2020
• Subjects: Geochemistry; FOS Earth and related environmental sciences; Atmospheric carbon dioxide; Ocean-atmosphere interaction; Ocean-atmosphere interaction > Measurement; Carbon dioxide sinks; Southern Ocean
• Online Access: https://dx.doi.org/10.7916/d8-j3p1-tf92; https://academiccommons.columbia.edu/doi/10.7916/d8-j3p1-tf92

Atmospheric CO₂ is projected to increase for the foreseeable future. The amount of CO₂ that remains in the atmosphere is regulated, in large part, by the ocean. As the long-term response to the changing atmospheric pCO₂ unfolds, the ocean sink will continue to be modified on seasonal to decadal timescales by climate variability and change. The magnitude of this variability is an active area of research. Accurately quantifying this variability is a challenge given the paucity of direct in-situ observations. In order calculate the global air-sea CO₂ sink, ocean pCO₂ needs to be known, or at least accurately estimated, at all locations at regular intervals. Two approaches to estimate air-sea CO₂ flux are, 1) from simulations of the Earth system and 2) data gap-filling mapping techniques. The goals of this thesis are to 1) rigorously quantify errors in a leading pCO₂ and ocean CO₂ sink mapping technique and 2) to evaluate the efficacy of adding Earth system model based estimates of ocean pCO₂ as a first guess into machine learning based mapping techniques. To meet the first goal, we use a suite of Large Ensemble model members as a testbed to evaluate a leading pCO₂ gap-filling approach (SOM-FFN). We find that the SOM-FFN performs well when sufficient data is available, but overestimates Southern Ocean decadal variability by about 39%. To meet our second goal, we incorporate Earth system model pCO₂ output into machine learning techniques either by adding the output as an additional feature or by post-processing the model output by learning the misfit (misfit=observation-model) and correcting for it. We find that blending model output and observations using machine learning marginally improves prediction accuracy. In addition, we discuss the potential of the learned misfits as a new model diagnostic tool, which can be used to visualize spatiotemporal pCO₂ estimates. Taken together, this study has significant implications in the development of carbon monitoring systems, in turn aiding policy making and improving our understanding of the evolution of the air-sea CO₂ sink.