Medium-Range Trajectory Prediction Network Compliant to Physical Constraint for Oceanic Eddy

Ge, Linyao; Huang, Baoxiang; Chen, X.; Chen, Ge

doi:10.1109/tgrs.2023.3298020

Cited by 4 publications

(1 citation statement)

References 43 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The aforementioned methodology of developing a mathematical model for ocean currents based on oceanographic physical mechanisms can be classified as a model-driven approach. As the scale of data expands, data-driven methods, particularly deep learning, are increasingly utilized to dramatically improve the state-of-the-art in liquid computing (Guan et al, 2022), atmosphere forecasting (Lam et al, 2023;Bi et al, 2023), and oceanography predicting (Ge et al, 2023). Deep learning allows computational models that are composed of multiple processing layers, enabling the learning of data representations with varied levels of abstraction (LeCun et al, 2015).…”

Section: Introductionmentioning

confidence: 99%

GEST: Accurate global ocean surface current reconstruction withmulti-scale dynamics-informed neural network

Ge,

Wang,

Huang

et al. 2024

Preprint

View full text Add to dashboard Cite

Abstract. Exceptional precision and excellent resolution reconstruction of sea surface currents are beneficial for exploring complex oceanic dynamic processes. Normally, this required physical inversion models for global or regional oceans are constructed to reconstruct oceanic currents. These models are based on the analysis of sea surface geostrophic and Ekman currents derived from satellite observations of sea level and wind stress fields. Nevertheless, the presence of various typical dynamic processes in marine environments, such as mesoscale eddies and small-scale waves, continues to pose challenges in accurately reconstructing oceanic currents. Meanwhile, any product of surface current that neglects the contribution of wave motion would, at best, be incomplete. Therefore, in this paper, we introduce an accurate sea surface current product at a depth of 15 m, named GEST (Geostrophic-Ekman-Stokes-Tide). This product is produced by a multi-scale dynamics-informed neural network that learns the intricate representation of concealed characteristics in Ekman, geostrophic currents, wave-induced Stokes drift, and TPXO9 tidal currents. Its structure design is predicated upon the intricate coupling relationships between various ocean surface components and the veritable currents discerned by the deployment of drift buoys, with each ocean surface component correlating to discrete physical processes. Compared with the prevailing product, the GEST confers an elevation in precision by approximately 9.2 cm/s over the traditional multinomial fitting method, 10.4 cm/s beyond the OSCAR, and 8.81 cm/s surpassing GlobCurrent.

show abstract