Estimating daily subsurface thermohaline structure from satellite data: A deep network with embedded empirical orthogonal functions

Abstract

Estimating subsurface thermohaline structure from concurrent satellite data is a meaningful way to enrich internal oceanic observations. As a powerful tool for data mining, many studies have used machine learning in subsurface reconstruction, but most conventional applications have been purely black-box in nature without further consideration of oceanic characteristics. Instead, proposed here for the first time is a semi-explicit deep network for reconstructing the oceanic interior from surface data. Named EEFFNN, the method embeds empirical orthogonal functions extracted from reanalysis data (the EE part of the name) into the inner framework of a feed-forward neural network (the FFNN part of the name). Comparison with Argo profiles and reanalysis data shows that EEFFNN can significantly outperform conventional machine-learning algorithms in estimating subsurface thermohaline structures and especially subsurface-intensified eddies. Also, EEFFNN can perform thermohaline reconstruction in one pass, making it more lightweight than “shallow” machine-learning algorithms such as random forest. Overall, EEFFNN shows promise for being applied to operational thermohaline reconstruction in the near future.