Overview: the ocean carbon sink mitigates 25% of annual anthropogenic CO2 emissions

To monitor this key climate service, air-sea CO2 fluxes across the globe must be known. Machine learning capably interpolates sparse surface ocean PCO2 data by estimating relationships in a complex state space that is largely independent of physical space and time. This spatio-temporal independence makes IT difficult to understand the impact of specific data on algorithmic skill. This project will develop approaches to quantify the value of individual PCO2 observations to ML interpolation algorithms. The largest global database of surface ocean PCO2 observations, the surface ocean CO2 atlas (SOCAT) covers only 1.6% of the global ocean (at 1X1 since the 1980S) and is biased towards the northern hemisphere. Machine learning (ML) is now widely used to interpolate these data to derive monthly estimates of spatially-resolved air-sea CO2 fluxes, and from these, estimates of the integrated ocean carbon sink. While mean fluxes and seasonality are well constrained at the large scale, data are insufficient to robustly reconstruct interannual to decadal variations. Looking to the future, IT can be expected that data coverage will always be a challenge due to the great size and variability of the ocean, difficult observing conditions, and constrained resources. Various studies have shown by trial-and-error that PCO2 data in the high latitude winter seasons are most critically needed to improve ML-based interpolations. Comparable understanding of whether there are some places where more data than necessary are being collected has not been developed. In order to develop a robust, sustainable monitoring system for the ocean carbon sink, a capability to formally quantify the value of individual PCO2 observations is needed. This valuation is a prerequisite to optimizing the distribution of limited observations so as to obtain the maximum possible skill from ML-based interpolation algorithms. Understanding the value of specific data to algorithms is an active area of ML research. This research has been driven primarily by the need for fair compensation to data providers - how to differentiate data that substantially improves a product-recommendation algorithm from data of limited marginal utility? These studies provide a valuable intellectual resource that can be applied to surface ocean PCO2 data and its interpolation. With the data shapley method (ghorbani and ZOU, 2019), this project will assign a specific value to each data point in the SOCAT database. Potential skill enhancements from alternative sampling patterns that raise the average value of the dataset without increasing its size will be assessed using the multi-model, multi-ensemble testbed developed by our team. The testbed approach is particularly useful because IT supports comprehensive evaluations of algorithmic skill through comparison to the original model fields. This project focuses on element #3 of option 1B. In addition to publications and presentations, all code will be made freely available on github, allowing the techniques developed to have broad application. An advisory committee that is well-connected to the PCO2 observing and mapping communities will further ensure that results will impact the future observing system.