Applying neural network learning to deconvolve the δ18O – sea surface salinity relationship in the global ocean

Abstract:

A robust global δ¹⁸O_sw dataset provides an essential benchmark for multiple oceanographic and hydrologic applications. Particularly, surface δ¹⁸O_sw values are highly correlated with sea surface salinity (SSS), which allows for valuable reconstructions of past ocean salinity and changes in the atmospheric hydrologic cycle. This relationship is known to vary spatially and temporally, but the relationship has, to date, only been assessed in predetermined ocean regions and water masses, or at the local, island-scale. Here, we create a self-organizing map (SOM), which utilizes an unsupervised neural network, to assess spatial patterns of δ¹⁸O_sw and the δ¹⁸O_sw-SSS relationship across the global ocean. To accomplish this, we use an updated global δ¹⁸O_sw dataset of mixed layer (<25m) δ¹⁸O_sw and SSS data from the NASA GISS Global Seawater Oxygen-18 database. Previous work created a global gridded δ¹⁸O_sw dataset by defining interpolated regions from sea surface circulation patterns and regions likely to have distinct δ¹⁸O_sw signatures. In the new dataset, the spatial interpolation is generated with an unsupervised neural network, which has no prior imposed boundaries or parameters. By using a neural network to interpolate the sparse spatial δ¹⁸O_sw data, we create a dataset which is based solely on the spatial distribution of the data and the depth of the sample. We define a number of statistically significant δ¹⁸O_sw regions from this dataset. Building on the interpolated data, we will use a SOM to generate a δ¹⁸O_sw – SSS relationship for defined regions using seawater samples with paired δ¹⁸O_sw and SSS measurements (N = 4785). Additional spatial clustering will be done on the δ¹⁸O_sw – SSS dataset to pull out distinct regions. These regions will be assessed alongside circulation patterns and other water mass identifiers in the mixed layer, the δ¹⁸O_sw clustered regions, and regional precipitation δ¹⁸O values. Ideally, the defined δ¹⁸O_sw – SSS relationships can be used to assess spatial changes in the linear slope and intercept values as indicators of spatially varying hydroclimatic conditions and provide insights into the drivers of variability in this relationship. These relationships can also be applied in reconstructions of past ocean salinity and the associated changes in the hydrologic cycle.