Reconstructing ocean subsurface salinity at high resolution using a machine learning approach

Abstract

Abstract. A gridded ocean subsurface salinity dataset with global coverage is useful for research on climate change and its variability. Here, we explore the feed-forward neural network (FFNN) approach to reconstruct a high-resolution (0.25∘ × 0.25∘) ocean subsurface (1–2000 m) salinity dataset for the period 1993–2018 by merging in situ salinity profile observations with high-resolution (0.25∘ × 0.25∘) satellite remote-sensing altimetry absolute dynamic topography (ADT), sea surface temperature (SST), sea surface wind (SSW) field data, and a coarse-resolution (1∘ × 1∘) gridded salinity product. We show that the FFNN can effectively transfer small-scale spatial variations in ADT, SST, and SSW fields into the 0.25∘ × 0.25∘ salinity field. The root-mean-square error (RMSE) can be reduced by ∼11 % on a global-average basis compared with the 1∘ × 1∘ salinity gridded field. The reduction in RMSE is much larger in the upper ocean than the deep ocean because of stronger mesoscale variations in the upper layers. In addition, the new 0.25∘ × 0.25∘ reconstruction shows more realistic spatial signals in the regions with strong mesoscale variations, e.g., the Gulf Stream, Kuroshio, and Antarctic Circumpolar Current regions, than the 1∘ × 1∘ resolution product, indicating the efficiency of the machine learning approach in bringing satellite observations together with in situ observations. The large-scale salinity patterns from 0.25∘ × 0.25∘ data are consistent with the 1∘ × 1∘ gridded salinity field, suggesting the persistence of the large-scale signals in the high-resolution reconstruction. The successful application of machine learning in this study provides an alternative approach for ocean and climate data reconstruction that can complement the existing data assimilation and objective analysis methods. The reconstructed IAP0.25∘ dataset is freely available at https://doi.org/10.57760/sciencedb.o00122.00001 (Tian et al., 2022).

An Updated Estimate of the Indonesian Throughflow Geostrophic Transport: Interannual Variability and Salinity Effect

“…To account for the ocean salinity ( S ) effect, we used two versions of the Institute of Atmospheric Physics (IAP) monthly salinity gridded datasets (Cheng et al., 2020; Tian et al., 2022). One is the 25‐member, 0.25° × 0.25° product of 1993–2018, and the other is the 26‐member, 0.50° × 0.50° product of 1984–2017.…”

“…These data are interpolated onto the same depth levels of the T profiles. In‐situ salinity observations, sea surface temperature, surface winds, altimeter‐based sea surface height, and a coarse‐resolution gridded salinity product are merged to reconstruct the monthly S fields (Tian et al., 2022). The feed‐forward neural network and the Monte Carlo dropout approach are used in member generation to ensure that the reconstructed members account for all major error sources including salinity instruments, representativeness, and methodology (Tian et al., 2022).…”

“…In‐situ salinity observations, sea surface temperature, surface winds, altimeter‐based sea surface height, and a coarse‐resolution gridded salinity product are merged to reconstruct the monthly S fields (Tian et al., 2022). The feed‐forward neural network and the Monte Carlo dropout approach are used in member generation to ensure that the reconstructed members account for all major error sources including salinity instruments, representativeness, and methodology (Tian et al., 2022). These two S datasets are interpolated onto the IX1 line to match T (Figure 1d), forming two versions of the ITF G estimate: the 1993–2018 and 1984–2017 versions (Figure S1 in Supporting Information ).…”

“…Hu and Sprintall (2016) suggested that the salinity component contributed ∼36% to the ITF interannual variability and dictated its strengthening since ∼2004. The recently available observational salinity data products (Cheng et al., 2020; Tian et al., 2022) may promote the understanding of the salinity effect.…”

Physics‐Informed Deep Operator Learning Based on Reduced‐Order Modeling for Retrieving the Ocean Interior Density From the Surface

“…These models rely on the assumption that the relationship between surface data and subsurface data can be represented by known function forms, with unknown coefficients determined through regression methods. While these models are straightforward and flexible, their simplicity may not be adequate in handling complex, nonlinear problems (L. Meng et al., 2021; Tian et al., 2022). Methods based on machine learning have recently become the most widely used class of data‐driven models, such as artificial neural network (Ali et al., 2004; Bao et al., 2019; Su et al., 2020), self‐organizing maps (Wu et al., 2012; C. Chen et al., 2018), support vector machine (Su et al., 2015, 2018), clustering neural network (W. Lu et al., 2019), and random forest (Su et al., 2018).…”