Source-linked AI summary
YingLan: A Multimodal Ocean Foundation Model for Unified Representation Learning from Heterogeneous Observations
Sen Jia, Siqi Wei, Junyu Wang, Zijun Lv, Pan Liu, Meng Xu, Xiyou Fu, Yuxi Sun, Qiang Liu, Shuyu Zhang, Lin Mu
TL;DR
Heterogeneous ocean observations pose challenges for joint modeling and transferable representations. YingLan combines adaptive encoding, multiscale spatiotemporal interaction, and evolution-aware pretraining; in a seven-day Shenzhen coastal SSS forecast, it reports the lowest average RMSE among evaluated methods.
Problem
Differences in ocean observations’ data structures, physical semantics, and sampling characteristics make unified representation learning and knowledge transfer challenging.
Source: §1 · p. 2Method
YingLan combines physics-attribute-conditioned dynamic encoding, multiscale STIC interactions, and pretraining with masked reconstruction and temporal-evolution objectives.
Source: §1 · pp. 2–3Results
0.687508 PSU average RMSE in seven-day Shenzhen coastal SSS forecasting, the lowest reported error among evaluated methods.
Source: §4.3 · p. 11, §5 · p. 13Takeaways & Limitations
YingLan provides an initial evaluation of a unified multimodal framework for regional ocean forecasting using heterogeneous observations.
Source: §1 · p. 3, §5 · p. 13The temporal consistency constraint does not require ocean-state evolution to strictly satisfy a prescribed physical governing equation.
Source: §3.4 · p. 10
Abstract
from arXiv · showhide
Multimodal ocean modeling remains challenging due to substantial heterogeneity in data structures, physical semantics, and spatiotemporal characteristics across satellite imagery, gridded oceanographic fields, and buoy observations. Existing approaches largely focus on task-specific ocean modeling, while learning transferable representations from heterogeneous observations remains an open problem. To address this challenge, we propose YingLan, a multimodal ocean foundation model that integrates physics-attribute-conditioned dynamic encoding, Spatiotemporal Increment Coupling (STIC), and ocean-state evolution-aware pretraining. Specifically, the dynamic encoding mechanism generates adaptive input projections conditioned on physical semantics and observation characteristics, enabling flexible representation of heterogeneous ocean variables. STIC models multiscale spatiotemporal dependencies through cross-scale interactions of latent temporal increments. The pretraining strategy further combines masked observation reconstruction with temporal evolution objectives to facilitate the learning of coherent ocean-state representations across different prediction horizons. YingLan is implemented using MindSpore and trained on Huawei Ascend 910B NPUs, demonstrating its compatibility with domestic AI computing infrastructure. In a seven-day sea surface salinity (SSS) forecasting task conducted in the Shenzhen coastal waters, YingLan achieves an average RMSE of 0.687508 PSU, corresponding to relative error reductions of 0.58%, 5.31%, and 12.42% compared with LangYa, XiHe, and GloNet2, respectively. These results provide initial evidence of the model's effectiveness in regional ocean forecasting, while broader cross-task and cross-regional evaluations remain necessary to assess its generalization capabilities.
1 Introduction
Heterogeneous ocean observations differ in structure, physical semantics, and spatiotemporal scales, making unified representation learning and transfer across tasks challenging. YingLan combines adaptive encoding, multiscale spatiotemporal interaction, and evolution-aware pretraining, and evaluates regional transferability using Shenzhen coastal waters as a case study.
1 Introduction
- Existing work leaves unified representation learning, heterogeneous observation encoding, multiscale dependency modeling, and ocean-state evolution learning as open questions.Prior studies span ocean forecasting, shared-representation pretraining, and multimodal remote-sensing pretraining.Source: §1 · p. 2
- Heterogeneous observations differ in data structure, variable semantics, sampling, and spatiotemporal scales, challenging joint modeling and knowledge transfer.Satellite imagery, oceanographic fields, and buoy observations capture complementary but distinct aspects of ocean conditions.Source: §1 · p. 2
- YingLan integrates satellite imagery, gridded oceanographic fields, and buoy time series in a unified framework for transferable representations and downstream adaptation.Its stated evaluation examines regional applications, including forecasting and, where supported by the design, few-shot adaptation, cross-temporal evaluation, and cross-region transfer.Source: §1 · p. 3
- Physics-attribute-conditioned dynamic encoding generates input projections using physical semantics and observation conditions to adaptively encode heterogeneous observations.The mechanism distinguishes variable-specific physical attributes from observation-specific characteristics.Source: §1 · p. 3
- STIC exchanges information across scales through latent temporal increments, linking localized ocean structures with large-scale environmental contexts.Evolution-aware pretraining jointly optimizes masked observation reconstruction, horizon-conditioned latent increment prediction, and temporal compositional consistency.Source: §1 · p. 3
2 Related Work
Prior ocean modeling and foundation-model work explores physical-process modeling, reconstruction, forecasting, and modality-specific pretraining. The related work identifies unified representations of heterogeneous ocean observations as a remaining direction.
2.1 Ocean Foundation Models and Pretraining Paradigms
- Ocean foundation-model approaches include task-specific physical modeling, reconstruction, and forecasting, while pretrained representations are proposed for transfer across ocean applications.Examples include recurrent-network correction of wave evolution, partial-convolution ocean-state assimilation, and WenHai's multivariate forecasting.Source: §2.1 · p. 3
- Ocean pretraining spans physical-field, SAR-image, and language-oriented approaches, but their frameworks address different modalities and task requirements.Examples include AI-GOMS, WV-Net, OceanGPT [18], and OceanPile [19].Source: §2.1 · pp. 3–4
2.2 Heterogeneous Ocean Representation and Multiscale Spatiotemporal Learning
- Multisource ocean-data fusion has been studied through spatial downscaling, field reconstruction, and multimodal spatiotemporal learning.Examples combine SSH and SST [20], fuse in situ, satellite, and model data, use reanalysis with satellite and buoy measurements, and study sparse-observation fusion.Source: §2.2 · p. 4
- Ocean spatiotemporal models use hierarchical extraction, attention, and multiscale fusion to represent dependencies across spatial scales and temporal horizons.Examples address SST forecasting and ENSO prediction.Source: §2.2 · p. 4
3 Method
YingLan unifies satellite imagery, gridded oceanographic fields, and buoy time series through dynamic encoding, multiscale spatiotemporal interaction, and evolution-aware pretraining. These components condition representations on observation attributes, couple temporal increments across scales, and train for coherent latent transitions across prediction horizons.
3.1 Overall Architecture
- YingLan combines satellite imagery, gridded oceanographic fields, and buoy time series in a shared framework with three components: dynamic encoding, STIC, and evolution-aware pretraining.After pretraining, its shared encoder is adapted to regional monitoring and prediction through task-specific heads and fine-tuning.Source: §3.1 · pp. 4–5
3.2 Physics-Attribute-Conditioned Dynamic Encoding
- Dynamic encoding generates input projections conditioned jointly on physical semantics and observation conditions, distinguishing variable meaning from measurement circumstances.Physical-semantic attributes describe a variable and its properties; observation conditions include sensor configuration, depth, and spatial resolution where applicable.Source: §3.2 · pp. 5–6
3.3 Spatiotemporal Increment Coupling (STIC) Module
- STIC transfers latent temporal increments between spatial scales and couples them with target-scale historical states to represent interactions between local structures and large-scale contexts.It reconstructs scale-specific dynamic features at a shared resolution while retaining the input representation through a residual connection.Source: §3.3 · pp. 7–9
3.4 Ocean-State Evolution-Aware Pretraining
- Evolution-aware pretraining combines masked reconstruction, horizon-conditioned latent-state prediction, and temporal compositional consistency to learn representations of ocean-state changes across horizons.The consistency constraint encourages a direct long-horizon increment to agree approximately with cumulative increments from successive shorter predictions, without requiring a prescribed physical governing equation.Source: §3.4 · pp. 9–10
4.1 Datasets
YingLan’s pretraining repository combines diverse in situ, satellite, and reanalysis sources, while downstream datasets cover heterogeneous coastal observations and applications in Shenzhen and adjacent waters.
4.1.1 Pretraining Datasets
- The pretraining repository integrates Argo, WOD, and EN4 observations; MODIS, AMSR-2, and GHRSST satellite products; and GLORYS, CORA, and NCEP-NCAR reanalysis datasets.The sources include temperature and salinity observations and gridded products, satellite-derived sea surface temperature, and oceanic and atmospheric reanalysis.Source: §4.1.1 · p. 10
- The repository contains more than 6 TB of data spanning multiple observation types, extended temporal periods, and broad geographic coverage.Its data include profiling-float sequences, multitemporal satellite imagery, and temporally resolved gridded fields.Source: §4.1.1 · p. 11
4.1.2 Downstream Evaluation Datasets
- Downstream datasets from Shenzhen and adjacent coastal waters cover forecasting, typhoon prediction, ecosystem monitoring, and marine hazard early warning.They include gridded fields, buoy and hydrometeorological time series, satellite and UAV imagery, and in situ survey records.Source: §4.1.2 · p. 11
4.2 Experimental Setup
The experiments compare YingLan with three ocean forecasting models using task-specific metrics and consistent evaluation data; implementation uses MindSpore on Huawei Ascend 910B NPUs.
4.2.1 Baselines and Evaluation Metrics
- For seven-day SSS forecasting, average RMSE across all seven lead times is reported in PSU and compared with LangYa, XiHe, and GloNet2.Methods use consistent test samples, spatial grids, and valid-ocean masks; other tasks use metrics selected for their targets and characteristics.Source: §4.2.1 · p. 11
4.2.2 Implementation Details
- YingLan is implemented with MindSpore and trained on Huawei Ascend 910B NPUs.Training uses MindSpore automatic differentiation and tensor computation within the Ascend environment.Source: §4.2.2 · p. 11
4.3 Downstream Task Performance
On seven-day SSS forecasting in Shenzhen coastal waters, YingLan has the lowest reported average RMSE among the compared models.
4.3 Downstream Task Performance
- Seven-day SSS forecasting in Shenzhen coastal waters is evaluated using average RMSE across the seven forecast lead days.Lower average RMSE indicates better forecasting performance.Source: §4.3 · p. 11
- 0.687508 PSU average RMSE is the lowest among the compared models, beating LangYa (0.691538 PSU), XiHe (0.726042 PSU), and GloNet2 (0.785023 PSU).The reported average is across seven forecast lead times; lower RMSE indicates higher prediction accuracy.Source: §4.3 · pp. 11–12, Figure 2 · p. 12
4.4 Results analysis of ablation experiment
The ablation setup removes each of three YingLan components individually and evaluates the variants on seven-day SST forecasting in Shenzhen coastal waters.
4.4 Results analysis of ablation experiment
- Three ablation variants each remove one component: dynamic encoding, STIC, or the ocean-state evolution-aware pretraining strategy.The study is designed to investigate the contributions of these components.Source: §4.4 · p. 12
- The variants are evaluated on seven-day SST forecasting in Shenzhen coastal waters.Source: §4.4 · p. 12
- Average RMSE across seven forecast lead times is the ablation evaluation metric, with results presented in Table 1.Source: §4.4 · p. 12
5 Conclusion
YingLan combines adaptive feature encoding, multiscale spatiotemporal integration, and joint pretraining to model heterogeneous ocean observations. In Shenzhen seven-day SSS forecasting, it reports the lowest error among evaluated methods, while broader downstream and cross-regional evaluations remain future work.
5 Conclusion
- YingLan adaptively encodes ocean variables using physical semantics and observation conditions, then integrates temporal variations across spatial hierarchies.Joint pretraining objectives target both observational structures and ocean-state dynamics.Source: §5 · pp. 12–13
- 0.687508 PSU average RMSE was reported for seven-day SSS forecasting in Shenzhen, the lowest error among evaluated methods.The conclusion identifies broader downstream-task adaptation and cross-regional generalization as future research directions.Source: §5 · p. 13
- The framework unifies satellite imagery, gridded oceanographic fields, and buoy time-series observations for transferable ocean representations.The work also reports implementation with MindSpore and training on Huawei Ascend 910B NPUs.Source: §5 · p. 13
- Future research will investigate broader downstream-task adaptation, cross-regional generalization, and robust modeling under incomplete and heterogeneous observations.Source: §5 · p. 13
A Additional Details
The appendix passage is an instruction to add derivations, experimental details, and supplemental tables, and to delete the appendix if unnecessary.
A Additional Details
- The appendix placeholder calls for derivations, extra experimental details, and supplemental tables, or deletion if unnecessary.Source: A Additional Details · p. 16