A satellite model forecasts ocean eddies as fields and paths
HaiRuo: Forecasting Ocean Eddies as Fields and Objects with a Satellite-Altimetry Foundation Model · arXiv:2611.14136
In plain words
Ocean eddies are swirling currents that can travel for weeks. HaiRuo is a system that uses 30 days of satellite maps to predict ocean conditions for the next 15 days, along with the paths of eddies already present.
One part learns ocean patterns from unlabeled satellite observations; another follows each predicted eddy through the forecast and adjusts its path.
In a global test, HaiRuo’s average day-15 centre error was lower than PredRNN-V2’s, though that result comes from a specific test period and set of regions.
The research
Knowing where an eddy will move matters because these large ocean currents carry heat, salt and nutrients across basins.
A forecast that predicts both ocean conditions and eddy paths could change how much confidence to place in locations inferred afterward from field forecasts, especially where labelled training data are sparse.
HaiRuo combines a satellite-trained ocean model with an eddy-path decoder, forecasting ocean fields and individual eddy trajectories from the same maps.
The system was tested on an independent global evaluation from 2020 to 2022, with models trained under a shared protocol and scored once. HaiRuo’s mean day-15 centre error was 25.20 km, compared with 30.58 km for PredRNN-V2, averaged across three seeds.
The same direction appeared outside the four labelled training regions, where HaiRuo’s mean error was lower than TAU’s. Separate tests of its backbone and decoder help explain the result, though they used different settings.
The global comparison used delayed-time satellite maps and a test period from 2020 to 2022. The forecast-field comparison with WenHai and XiHe used a fixed tracker on 48 test dates; the reported experiments did not use the near-real-time data stream described for operational forecasting.
What the evidence shows
- Forecasting fields and eddies
HaiRuo is a forecasting system that predicts ocean fields and eddy paths together from satellite-altimetry maps. Its global evaluation compared it with field-forecasting methods under a shared protocol, using the same 1,024 labelled windows and training budget.
The central question is whether predicting eddies as objects within the forecast improves their predicted locations, as measured by distance from observed centres after 15 days.
- Lower error in the global test
For atlas eddies present at both the start and day 15, D15 measures the distance between predicted and observed centres on day 15. The three-seed mean was 25.20 km for HaiRuo and 30.58 km for PredRNN-V2, supporting lower error in this global test.
- The result beyond training regions
The same day-15 centre-distance measure was lower for HaiRuo outside the four regions used for labelled training: its mean was 27.64 km, compared with 33.52 km for TAU.
This regional comparison suggests the result extends beyond labelled training areas, but does not show that every ocean region behaves alike.
- What the components contribute
On Dev48, adding frozen HaiRuo-1B to the same XuanJi readout lowered day-15 centre error, supporting a contribution from the pretrained representation in that setup. Separately, adding two-round XuanJi object attention lowered the error for each of six video-prediction backbones.
These component tests support contributions from both parts, but their settings differ from the global evaluation.
- A common tracker comparison
Passing each forecast field through the same fixed tracker tests field quality separately from directly scoring a system’s predicted trajectories. On the 48 Test48 dates, HaiRuo’s field yielded 29.5 km centre error, compared with 47.1 km for WenHai and 51.6 km for XiHe.
Scoring HaiRuo’s predicted trajectories directly gave 10.1 km on Test48 and 11.2 km globally, 77–78% below WenHai and XiHe in the global comparison. The field and trajectory results use different scoring approaches.
Deep
- Original paper authors
- Xukun Qin, Haochen Sun, Yuyang Tian, Ge Chen
- Submitted to arXiv
- License
- CC BY 4.0