acceptodds
Under review as a conference paper at ICLR 2027

Physics-Inspired Dual-Graph Reasoning for Submesoscale Ocean Super-Resolution

Abstract

Accurately monitoring submesoscale (kilometer-scale) ocean dynamics is essential for understanding fine-grained energy transport and exchange, yet the high cost of in-situ observations results in limited data availability. Advances in visual super-resolution (SR) enable high-resolution (HR) predictions by refining abundant low-resolution (LR) remote sensing data. However, recent generative SR approaches can introduce visually rich but unsupported details, reducing reconstruction fidelity on fine-scale structures. To address this problem, we emphasize inductive biases inspired by multiscale ocean dynamics and propose a dual-graph reasoning network, namely DRONEThe anonymous code repository can be found at https://anonymous.4open.science/r/DRONE-1E41https://anonymous.4open.science/r/DRONE-1E41.. DRONE combines a heterogeneous channel graph with a spectral-aware pixel graph to improve reconstruction fidelity without imposing explicit physical equations or conservation constraints. The channel graph couples temporal observations with spatial-frequency representations, while the pixel graph allocates connectivity according to local spectral complexity. Extensive experiments on an HR MITgcm ocean simulation dataset demonstrate that DRONE surpasses Transformer-, diffusion-, and GNN-based SR methods on standard fidelity metrics, frontal-structure measurements, and dataset-wide spectral evaluations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Submesoscale Ocean Dynamics, Graph Reasoning, Super-Resolution

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_nPNaOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a model called DRONE for super-resolving ocean atmospheric data. To enhance the reconstruction fidelity on fine-scale structures, DRONE combines a heterogeneous channel graph with a spectral-aware pixel graph to improve reconstruction fidelity without imposing explicit physical constraints. Experiments on the MITgcm ocean simulation dataset demonstrate the effectiveness of each design.

Critical strengths
  • This paper proposes a new method and attempts to address an attractive question, i.e., how to enhance the fine-scale fidelity when super-resolving ocean data.
  • Introducing specific neural designs regarding spatiotemporal relationship, and spectral-aware modeling are promising ways to enhance the fidelity of the results in the corresponding domain.
  • The paper is well-structured and easy to follow. The experiment section is formatted as several research questions to answer and then each subsection explains the results based on each question, this makes the results very clear to follow.
Critical weaknesses
  • The model design mainly focuses on modeling spatial and temporal relationships and enhancing the representation of specific frequency information, rather than using physical information or constraints, therefore the connection between the model design and its claimed physical inspiration is weak.
  • Another argument of the introduction is the lack of inductive bias in current super-resolution models for geophysical data, in response, the proposed model is claimed to introduce the inductive bias from the geophysical data. However, the definition of inductive bias is vague and somewhat exaggerated, as there is no explicit demonstration that any kind of inductive bias has been leveraged or improved in the data.
  • It is not clear if the dataset is representative enough for evaluation, for example, what are the training, testing, and validation years? Also, the simulation-based dataset covers only the Northwest Pacific region, which is not sufficient to show the model’s ability in handling different ocean conditions in real observations.
  • As a general SR architecture, the method is validated on a single variable (SST), and a single downscaling ratio (16x). This provides very limited demonstration of the model’s ability in super-resolving different ocean climate variables in different scenarios, for example, the wind and sea level pressure, which have quite different physical properties but are important for some applications like the tracking of tropical cyclones.
  • Some related works on atmospheric downscaling are missing in the literature review, for example, the DSFNO [1] proposes a neural operator-based model for climate data downscaling, and Corrdiff [2] introduces a diffusion-based model to better generate high-frequency components.

[1] Fourier Neural Operators for Arbitrary Resolution Climate Data Downscaling, JMLR, 2024. [2] CorrDiff: A Two-Step Diffusion Model for High-Resolution Weather Forecasting, Communications Earth & Environment, 2025.

Questions for authors

None

Reviewer_zsdnOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses the difficulty of preserving fine-scale structures and spectral fidelity in high-factor super-resolution of submesoscale ocean fields. It proposes DRONE, a dual-graph reasoning network that combines observations from multiple time steps with multiscale features extracted using spatial kernels. A spatiotemporal channel graph models cross-channel interactions, while a pixel-level spatial graph assigns node degrees inversely to local spectral complexity to limit excessive smoothing in high-frequency regions. A cascaded pipeline supports 16× upscaling. On MITgcm ocean simulation data, the method outperforms the evaluated baselines on standard reconstruction metrics, frontal-structure fidelity, and spectral measures. Overall, the motivation and proposed solution are appealing, and the reported experimental results are solid.

Critical strengths
  1. The architecture is coherent, with its components organized around exploiting temporal context and preserving multiscale spatial structure. The spatiotemporal channel graph connects temporal observations with spatial-frequency features, while the spectral-aware pixel graph uses sparser connectivity around textured regions and edges and broader aggregation in smooth regions. These choices are well aligned with the goal of preserving fine-scale structures in ocean super-resolution.

  2. The paper is generally clearly written and supports the proposed method with a substantial set of experiments and analyses, including comparisons across several model families, component ablations, frontal-structure visualizations, and spectral evaluation.

Critical weaknesses
  1. Although the authors explicitly state that the method does not impose physical constraints, they characterize it as “physics-inspired.” My concern is that modeling spatial and temporal components separately does not clearly distinguish the framework from conventional architectural design, while characterizing it as “physics-inspired” may be too strong. In addition, the adaptive spatial aggregation appears similar in spirit to deformable convolution or deformable attention. It remains unclear whether the proposed mechanism offers additional benefits over these established alternatives.

  2. The method relies on additional observations across time. In practical satellite-observation settings, this requirement may entail additional acquisition and data-transmission costs when suitable temporal observations are not already available, potentially limiting its applicability. Nevertheless, in light of the discussion in Remark 3.1, I consider the proposed pipeline design reasonable. However, because DRONE exploits information from multiple frames, the main comparison against single-image super-resolution models does not fully establish its advantage under matched information availability. Appendix H.2 provides useful matched-input comparisons with SwinIR-Stack and PFT-Stack, but these baselines incorporate temporal information through channel stacking. Comparisons with appropriately adapted video super-resolution methods would provide a more informative assessment of the proposed temporal modeling.

  3. There are inconsistencies in the presentation. At the beginning of Section 4, RQ1 is framed in terms of temporal generalizability, whereas Section 4.2, which is labeled RQ1, primarily presents an overall comparison with other methods. In addition, the caption in Figure.5 (b) and (c) appear to be interchanged.

Questions for authors

Addressing the following points would help reviewer reassess the paper.

  1. The current ablation studies do not sufficiently justify the proposed design. For example, DRONE_T only tests a variant with the spatial graph removed, for which a performance drop is largely expected. A more informative comparison would use a fully connected spatial graph or dense connectivity for all nodes, to verify the claim in Section 3.4 that sparse connections at edge pixels can effectively suppress low-pass smoothing.

  2. Please report comparisons with video super-resolution models adapted to this setting.

  3. Please address the presentation issues raised in W3.

Reviewer_8SwUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DRONE is a dual-graph architecture for ocean-field super-resolution that combines temporal and spatial-feature interactions with complexity-dependent pixel connectivity. It achieves the best reported reconstruction scores, but mathematical inconsistencies and limitations in the proposed metrics weaken the explanations for why it works, particularly the claims about spectral preservation and suppression of structured artifacts.

Critical strengths
  1. Recovering fine-scale ocean structure from coarse observations is practically important, and combining temporal context, multiscale features, and spatially adaptive aggregation is a reasonable approach.
  2. DRONE outperforms all compared methods on reconstruction, frontal, and spectral metrics, and the appendix adds substantive controls: matched temporal inputs, alternative degree policies, unseen degradations, and FIU-MAE sensitivity. The SST experiment shows applicability beyond salinity.
  3. The abstract and limitations correctly state that the method is physics-inspired rather than physics-constrained and does not enforce governing equations or conservation laws.
Critical weaknesses
  1. The degree-policy justification is inconsistent. Equations (34) and (35) are called equivalent, but a clipped reciprocal and a linearly decreasing normalized mapping yield different degree allocations; the paper should state which is implemented. Appendix D.4 also argues that larger degrees smooth more, while Equation (36) implies weaker suppression at larger degrees for fixed eigenvalue and hop count, and the “quantitative bound” has no derivation. Since this is a central contribution, the authors should reconcile the formulations or present spectral preservation as empirical motivation rather than an established property.
  2. FIU-MAE does not reliably penalize spurious fronts. Algorithm 1 weights gradient errors by ground-truth gradient magnitude, so a false front where the true gradient is zero receives zero weight even when inside the union mask, and only a weak penalty where the true gradient is small. Threshold sensitivity (Table 7) does not address this. The metric should be validated on controlled added/missing/displaced fronts and complemented by front precision/recall or symmetric weighting; the zero-denominator case also needs specification.
Questions for authors

see above

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.