Reconstructing North Atlantic Ocean Heat Content Using Convolutional Neural Networks

Slowly varying ocean heat content is one of the most important variables when describing cli-
mate variability on interannual to decadal time scales. Since observation-based estimates of
ocean heat content require extensive observational coverage, incomplete observations are often
combined with numerical models via data assimilation to simulate the evolution of oceanic heat.
However, incomplete observations, particularly in the subsurface ocean, lead to large uncertain-
ties in the resulting model-based estimate. As an alternative approach, Kadow et al (2020) have
proven that artificial intelligence can successfully be utilized to reconstruct missing climate in-
formation for surface temperatures. In the following, we investigate the possibility to train their
three-dimensional convolutional neural network to reconstruct missing subsurface temperatures
to obtain ocean heat content estimates with a focus on the North Atlantic ocean.
The network is trained and tested to reconstruct a 16 member Ensemble Kalman Filter assimi-
lation ensemble constructed with the Max-Planck Institute Earth System Model for the period
from 1958 to 2020. Specifically, we examine whether the partial convolutional U-net represents
a valid alternative to the Ensemble Kalman Filter assimilation to estimate North Atlantic sub-
polar gyre ocean heat content.
The neural network is capable of reproducing the assimilation reduced to datapoints with ob-
servational coverages within its ensemble spread with a correlation coefficient of 0.93 over the
entire time period and of 0.99 over 2004 – 2020 (the Argo-Era). Additionally, the network is
able to reconstruct the observed ocean heat content directly from observations for 12 additional
months with a correlation of 0.97, essentially replacing the assimilation experiment by an extrap-
olation. When reconstructing the pre-Argo-Era, the network is only trained with assimilations
from the Argo-Era. The lower correlation in the resulting reconstruction indicates higher un-
certainties in the assimilation outside of its ensemble spread at times with low observational
density. These uncertainties are highlighted by inconsistencies in the assimilation’s represen-
tations of the North Atlantic Current at times and grid points without observations detected
by the neural network. Our results demonstrate that a neural network is not only capable of
reproducing the observed ocean heat content over the training period, but also before and after
making the neural network a suitable candidate to step-wise extend or replace data assimilation.