ORCAst: Operational High-Resolution Current Forecasts

TL;DR

ORCAst introduces an operational, high-resolution ocean current forecasting framework trained exclusively on observational data, achieving state-of-the-art performance in extratropical regions. The model employs a multi-arm encoder–decoder architecture with per-variable encoders/decoders and a Gated Spatio-Temporal Attention translator, and it is trained in three stages that progressively incorporate nadir SSH, SWOT SSH, and drifter currents. Regional specialization further boosts accuracy, with Mediterranean, Gulf Stream, and Agulhas models outperforming the global variant, particularly for 7-day forecasts. The work demonstrates practical applicability using drifter and ship data, analyzes input-data contributions, and discusses limitations and future directions, including equatorial forecasting and probabilistic extensions.

Abstract

We present ORCAst, a multi-stage, multi-arm network for Operational high-Resolution Current forecAsts over one week. Producing real-time nowcasts and forecasts of ocean surface currents is a challenging problem due to indirect or incomplete information from satellite remote sensing data. Entirely trained on real satellite data and in situ measurements from drifters, our model learns to forecast global ocean surface currents using various sources of ground truth observations in a multi-stage learning procedure. Our multi-arm encoder-decoder model architecture allows us to first predict sea surface height and geostrophic currents from larger quantities of nadir and SWOT altimetry data, before learning to predict ocean surface currents from much more sparse in situ measurements from drifters. Training our model on specific regions improves performance. Our model achieves stronger nowcast and forecast performance in predicting ocean surface currents than various state-of-the-art methods.

Figures (15)

• Figure 1: ORCAst is an operational neural network model that achieves state-of-the-art forecasts of ocean surface currents in extratropical latitudes. It is trained exclusively on observational data, including drifters, SWOT and nadir altimetry, SST and CHL observations, in a multi-stage training process.
• Figure 2: In situ observations of currents from drifters in three key regions in 2023, used for evaluation.
• Figure 3: SWOT SSH observation, with visible cyclonic and anti-cyclonic eddies at high resolution. SWOT consists of two 50 km wide bands, separated by a 20 km gap covered by traditional nadir altimetry instruments.
• Figure 4: SWOT SSH observation, with visible internal gravitational waves. In particular, close to the Equator, it is more difficult to estimate ocean surface currents from SSH as a result of the reduced Coriolis effect.
• Figure 5: Overview of ORCAst. Our models inputs temporal sequences of multivariate satellite observations (here SSH, SST, CHL) from $t=0$ to $t=T$. Each variable is inputted to a different spatial encoder, which extracts information from each timestep separately. A spatio-temporal positional embedding is applied to inform the network of geographical coordinates and seasonal information. A spatio-temporal translator $t_\varphi$ is used to produce the forecast encodings. Finally, each timestep is decoded independently with three univariate decoders $g_\psi$, forecasting SSH, $U$ and $V$ components separately.