Comment on essd-2022-308

Abstract

Over a decade ago, oceanographers began installing oxygen sensors on Argo floats to be deployed throughout the world ocean with the express objective of better constraining trends and variability in the ocean’s inventory of oxygen. Until now, measurements from these Argo-mounted oxygen sensors have been mainly used for localized process studies on air–sea oxygen exchange, biological pump efficiency, upper ocean primary production, and oxygen minimum zone dynamics. Here we present a four-dimensional gridded product of ocean interior oxygen, derived via machine learning algorithms trained on dissolved oxygen observations from Argo-mounted sensors and discrete measurements from ship-based surveys, and applied to temperature and salinity fields constructed from the global Argo array. The data product is called GOBAI-O2 for Gridded Ocean Biogeochemistry from Artificial Intelligence – Oxygen (Sharp et al., 2022; https://doi.org/10.25921/z72m-yz67; last access: 30 Aug. 2022); it covers 86 % of the global ocean area on a 1° latitude by 1° longitude grid, spans the years 2004–2021 with monthly resolution, and extends from the ocean surface to two kilometers in depth on 58 levels. Two machine learning algorithms — random forest regressions and feed-forward neural networks — are used in the development of GOBAI-O2, and the performance of those algorithms is assessed using real observations and Earth system model output. GOBAI-O2 is evaluated through comparisons to the World Ocean Atlas and to direct observations from large-scale hydrographic research cruises. Finally, potential uses for GOBAI-O2 are demonstrated by presenting average oxygen fields on isobaric and isopycnal surfaces, average oxygen fields across vertical–meridional sections, climatological cycles of oxygen averaged over different pressure intervals, and a globally integrated oxygen inventory time series.