acceptodds
Under review as a conference paper at ICLR 2027

Learning Field Reconstruction from Incomplete Data by Globally Correcting Local Estimates

Abstract

Reconstructing physical fields from training samples that are always incomplete requires learning spatial structure from fragmented observations.Existing context–query work establishes how held-out observations provide valid training targets, but this does not make the complete-field distribution identifiable when every training field is incomplete.With finite data, weak evidence of sharp transitions and localized variations can further favor averaged predictions that attenuate local detail.A structural prior is therefore needed to favor plausible completions; local spatial relationships offer one grounded in the observations.We propose a locally constructed, globally revisable estimator that explicitly learns local field estimates and subsequently corrects them using full-domain observations.A shared coordinate-conditioned predictor learns from incomplete patches, allowing relatively well-observed neighborhoods to provide direct supervision of local structure.Its overlapping predictions are reconciled into an observation-conditioned consensus field.A full-domain estimator retains the original observations and learns a residual correction around this frozen field estimate, allowing locally constructed structure to be revised by broader evidence.The local estimate serves as both an explicit input, accompanied by its discrepancies with the observations, and a prediction starting point that the global model can revise.On three real-world ocean datasets with authentic observation gaps, our estimator achieves the lowest MSE and highest PSNR on withheld source-supported values, reducing MSE by 28.9%–34.5% against the strongest external baseline.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Incomplete Data, Prior, Diffusion Models

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_Jjm5OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses point reconstruction of physical fields when every training sample is itself incomplete, following the context–query supervision of IDCD and OAMP: available observations are split into a supplied context and withheld queries, and only the queries are supervised. The proposed estimator has two stages.

  1. Local prior. A coordinate-conditioned patch predictor is trained with the query loss restricted to patches, then frozen. At reconstruction time, it is evaluated on every patch of an overlapping covering, and the predictions are averaged into a consensus field Πt\Pi_t, called the local prior.
  2. Global residual correction. A full-domain corrector receives the context, the mask, sg(Πt)\mathrm{sg}(\Pi_t), the overlap count WW, and two discrepancy maps between the prior and the observations (SctxS_{ctx}, S∂S_\partial). It outputs a zero-initialized residual, so the reconstruction is Πt+rψ\Pi_t + r_\psi, trained with the same query loss.

Inference follows IDCD's single-step prediction at near-zero diffusion time. Ocean data use the full supplied mask; PDE data use an ensemble of K=8K=8 subcontexts.

Experiments.

  • Three CMEMS ocean products (Black Sea CHL, Baltic Sea NANO, Global Ocean SSS) with authentic gaps, at 64×6464\times64 and 128×128128\times128, scored only on withheld source-supported values through a cached real-mask overlay protocol shared by all methods.
  • Shallow Water and Navier–Stokes under pixel and block masks.

Results.

  • On the ocean products, the method has the lowest MSE and highest PSNR everywhere, with 28.9–34.5% lower MSE than the strongest external baseline at 64×6464\times64.
  • Controls against a parameter-matched no-prior model (Full Direct) and Prior Direct attribute 8.3–9.0% and a further 9.8–16.2% MSE reduction to prior conditioning and to the residual parameterization, respectively.
  • Replacing the learned prior with OAMP or DINEOF outputs does worse.
  • Ablations cover coordinates, input maps, fusion rules, and patch size.
  • An intervention study shows that the corrector responds to observations outside the prior's support, with dataset-dependent effects on accuracy.

Theory. The appendix includes a non-identifiability result for the complete-field distribution and the observation that unrestricted direct and residual predictors share the same optimum. The authors are explicit that the claimed benefit concerns finite-model learning rather than representational power.

Critical strengths

1. Consistent gains under a careful and fair evaluation protocol. The ocean evaluation is the most convincing part of the paper:

  • splits are chronological;
  • input and scoring supports come from authentic masks of other days and are cached and shared across all methods;
  • only source-supported values are scored, and normalization statistics are fitted on training observations only.

Within this protocol, the method wins on all three products, at both resolutions, and across seeds with small standard deviations. The margin also holds against the parameter-matched Full Direct control, which received comparable training time: about 18–23% lower MSE at 64×6464\times64, computed from Table 2. Because these numbers come from real cloud and swath gaps rather than synthetic masks, I take them as practically meaningful for satellite gap filling.

2. The design choices are tested with controls, and the claims are scoped honestly. The paper separates three effects:

  • prior conditioning (Full Direct vs. Prior Direct);
  • the residual output form (Prior Direct vs. Prior Residual);
  • the choice of reference (Global2Global, Low-rank2Global).

It then checks the corrector inputs and the fusion rule. The remote-evidence intervention in Sec. 4.5 is a nice touch, especially since its mixed result on SSS is reported rather than hidden: removing far observations reduces error there. The theory is also stated with the right caveats. Appendix D.4 openly says that direct and residual predictors have the same unrestricted optimum, so the claim is about how finite models learn. This makes me trust the empirical claims more than usual.

Critical weaknesses

1. The central hypothesis, that local learning helps, is not cleanly separated from simply stacking two networks.

The paper's story is that organizing supervision around small patch problems lets a finite model learn local structure better. The final estimator, however, also differs from Full Direct in having two networks in sequence, a frozen first-stage output, extra hand-designed input maps, and a residual skip. None of the reported controls varies only locality:

  • Global2Global uses the external OAMP output as its prior, which has a different backbone and capacity, and its prior is already weaker than Full Direct (Table 1 vs. Table 2).
  • It also performs oddly: 41.695 on SSS, worse than Full Direct alone at 32.895.

The clean test would be a global-first cascade with the same backbone and parameter count as the patch predictor, trained full-domain, frozen, and fed to the same corrector with the same cue maps. Without it, I cannot tell whether the gain comes from locality or from a generic two-stage refinement, which is a well-known recipe (coarse-to-fine inpainting, PatchRefiner, CRITER, all listed in Table 4). The theoretical sections do not help decide this either: D.1 is a standard parity construction that the method does not use, D.3 is immediate from the construction, and D.4 shows there is no representational advantage. As it stands, the contribution is an effective recipe whose mechanism is asserted rather than demonstrated.

2. The headline margin is overstated, and some setup details needed to judge the comparison are missing.

  • The headline compares against a baseline weaker than the authors' own control. The 28.9–34.5% figure is measured against external baselines. Yet the parameter-matched OAMP variant (Full Direct) is substantially better than the OAMP entry in Table 1 on every product (e.g., 32.895 vs. 44.082 on SSS). Against Full Direct, the gains are roughly 23%, 18%, and 19%. These are still good, but the abstract should report them.
  • Architecture and compute are not described. I could not find the backbone of the patch predictor, the corrector, or Full Direct in the main text, nor the parameter counts and training compute. This makes the parameter-matched claim hard to verify.
  • The ocean setting is narrow relative to how gap filling is done in practice. Fields are resampled from about 1 km to 64×6464\times64, each day is treated independently, and temporal context is ignored, although DINEOF and deep methods such as DINCAE exploit it. It is not stated how DINEOF was configured here (spatio-temporal stacking or per-field).
  • The PDE gains look like an outlier. On Shallow Water Pixel 30%, the method's MSE is about 25x lower than every baseline. That subset has only 81 distinct radial dam-break initial states (57 for training) and the model receives absolute coordinates, so I suspect much of this gain reflects the simplicity of the dataset rather than the method.
  • Some results are thinly supported. Block 8/9 and 128×128128\times128 are single runs.
Questions for authors
  1. Locality versus stacking. Can you add a global-first cascade? Concretely: a full-domain predictor with the same architecture family and parameter count as the patch predictor, trained with the same query loss and coordinates, frozen, and used as Πt\Pi_t for the same residual corrector and cue maps. If Prior Residual still clearly wins, the locality claim is supported. If not, the contribution reduces to a two-stage refinement recipe.

  2. Architecture and compute (clarification). What are the backbones and parameter counts of the patch predictor, the corrector, Full Direct, and the external OAMP? What is the training compute of each stage? Why does the parameter-matched Full Direct beat the published-size OAMP by up to 25%? Is this capacity alone, or other changes as well? Would you be willing to report the main improvement relative to Full Direct?

  3. Prior Direct vs. Prior Residual. The 9.8–16.2% gap is attributed to the residual form together with its initialization. Does the gap persist if Prior Direct is trained longer, or if it is given an explicit skip that starts at Πt\Pi_t but is free to drop it? This would show whether the effect is about optimization speed or about the final solution.

  4. Shallow Water. How much of the roughly 25x gain on Pixel 30% remains without absolute coordinates, or with a split that holds out initial-state radii outside the training range? I want to understand whether the model is memorizing the radial geometry.

Smaller clarifications

  1. How was DINEOF run, per field or with temporal stacking? Did you consider a temporal-context baseline such as DINCAE, given that the products are daily?
6 comments

W1 - problem setting and relation to staged reconstruction

We agree that staged prediction and residual refinement are established ideas. Our contribution concerns how to organize learning when complete target fields are never available and local structure receives only fragmented supervision. This distinction motivates the local-first design and the controls discussed in our next comment.

Why the supervision constraint changes the design question. In our setting, observations serve as both the reconstruction input and the only available training targets. Holding out observed entries creates valid query supervision, but does not reveal values that were never observed. With finite data, evidence for sharp transitions and localized variation can therefore remain weak, favoring averaged predictions. Increasing network capacity alone cannot supply those missing targets.

We deliberately give local reconstruction its own training stage. A shared coordinate-conditioned predictor pools context-query relationships across incomplete patches, allowing relatively well-observed neighborhoods to supervise recurring local structure. After freezing the local predictor, we fuse its overlapping predictions into a reference field for each input. The full-domain model then learns to revise this starting estimate using the original observations and discrepancy cues. Cropping reorganizes existing evidence. This gives the two stages distinct learning roles: learning local structure and revising it with full-domain evidence.

Concrete comparison with the methods named in W1. Building on Table 4 and Appendix A.3, the table below separates the tasks, supervision, motivation, and learning organization of these methods.

MethodTaskTraining supervisionMain challengeInitial referenceRefinement scopeTraining organizationShared elements
Contextual Attention (coarse-to-fine)Image inpaintingComplete images with artificial holesCoherent structure and texture across large holesCoarse image completionImage refinement using distant contextual featuresReconstruction loss; refinement also uses GAN lossesInitial prediction followed by learned refinement
PatchRefinerHigh-resolution monocular depthSynthetic depth, real depth labels and pseudo-labelsMemory/computation; depth scale and fine detailGlobal coarse depth from downsampled RGBPatchwise residual depth from high-resolution RGBCoarse predictor frozen before refiner trainingFrozen reference; additive residual refinement
CRITERSatellite SST reconstructionIncomplete sequences; loss on available measurementsFine spatial variability under sparse observationsGlobal low-frequency estimate from consecutive fieldsIterative full-field residuals and uncertaintyCoarse transformer frozen before refinement trainingIncomplete-data learning; frozen reference; residual correction
OursSpatial field imputationHeld-out observed entries of never-complete fieldsFragmented supervision of local structureConsensus of overlapping local predictionsFull-domain residual correction using original observationsLocal predictor frozen before corrector trainingStaged prediction and residual learning

The motivation thus differs from decomposing high-resolution visual prediction primarily to manage computation, receptive fields, or global-scale/detail trade-offs. CRITER shares our incomplete-data setting; the key distinction there is the local-first organization of learning.

Different supervision and observation geometry can favor different decompositions. Our work investigates which organization is effective under never-complete field supervision through a local-first construction and controlled comparisons. These results establish the value of giving local structure a dedicated learning stage before full-domain correction.

W1 / Q1 - evidence separating local learning from stacking

1. The global-first cascade tests the organization of the first stage. Global2Global in Section 4.3, Table 2 uses standard OAMP as its first stage. Appendix E.4 now makes explicit that it shares the local predictor's Karras U-Net family, depth, channel-width schedule, query loss, and coordinate conditioning, but is trained full-domain. It is frozen before training the same residual corrector architecture, preserving the input channels and recomputing the discrepancy maps. Its first stage has approximately 431M parameters versus 409M for the local predictor (Table 8), providing the requested global-first control with comparable capacity.

The relevant Table 2 results are reproduced below. Entries are MSE in units of 10−310^{-3}, averaged over independent training seeds; the PDF also reports standard deviations.

EstimatorCHLNANOSSS
Full Direct9.35023.55932.895
Global2Global9.17525.09741.695
Prior Direct8.54621.44530.151
Prior Residual (Ours)7.16419.34526.798

The local-first estimator reduces MSE by 21.9%-35.7% relative to Global2Global, while retaining the same correction design. On SSS, the cascade improves its own OAMP reference from 44.082 to 41.695 but remains worse than Full Direct (32.895): adding a second network is not sufficient to reproduce our result.

The comparison also reveals the value of the intermediate representation. On CHL and NANO, the Local Prior alone is less accurate than standalone OAMP (15.227 vs. 10.263 and 30.576 vs. 27.220), yet its corrected result is better than Global2Global. A local reference can therefore be more useful for subsequent correction even when its standalone MSE is higher.

2. Capacity-matched and crop-training controls support the division of learning. Full Direct allocates approximately 841M parameters to one full-domain model, versus 840M across our two stages, and receives comparable measured training cost (Appendix E.5). Our MSE reduction remains 17.9%-23.4%. Separately, the single OAMP model trained on random crops with the same query loss gives CHL/SSS MSE of 15.342/38.338, versus 7.164/26.798 for Ours (Table 24). This comparison supports constructing and then correcting a local field estimate over crop exposure alone under this protocol.

3. Prior conditioning and the residual starting point have distinguishable roles. Prior Direct improves on Full Direct by 8.3%-9.0% before introducing residual output prediction. Prior Residual gives a further 9.8%-16.2% improvement. The explicit-skip control in Table 26 nearly closes the CHL gap (7.210 vs. 7.164), while SSS retains a larger difference (29.882 vs. 26.798). These results identify a contribution from the starting prediction and parameterization. Table 14 separately tests the corrector's cue maps, whose removal raises error, supporting their role in connecting the learned local estimate to the observations.

4. The evidence extends beyond the diffusion objective and aggregate MSE. Keeping the two-stage organization with plain regression yields CHL/SSS MSE of 7.398/28.102 (Table 24), retaining much of the gain. Moreover, Table 22 reports lower interior-gradient and frequency-coefficient errors than parameter-matched Full Direct on all four tested PDE settings. This supports the intended recovery of spatial structure as well as improved pointwise prediction.

What the controls establish. These controls support a local-first division of learning under incomplete supervision: construct a useful local reference, then revise it using full-domain observations. The theory in Appendices D.1, D.3, and D.4 defines the identifiability limits, local-support boundary, and shared unrestricted prediction target.

remaining questions and clarifications

W2 / Q2: architecture, compute, and the headline comparison.

Full Direct widens standard OAMP while retaining its depth, no-prior interface, and query objective. The capacity increase distinguishes it from the published-size baseline. Appendix E.5 now gives the backbones, component and total parameter counts, training settings, and measured training costs.

The abstract now reports the 17.9%-23.4% MSE reduction against parameter-matched Full Direct, alongside the separately identified external-baseline comparison. This makes the stronger comparator central to the headline claim.

Q3: longer training and an explicit skip.

Prior Direct had reached a validation plateau under the shared early-stopping procedure. We have not added a separate extended-training experiment. The requested explicit-skip control is now in Appendix F.9, Table 26. It nearly closes the CHL gap, while Prior Residual retains an advantage on SSS. The remaining effect therefore depends on the dataset. Because this control changes both the starting prediction and the trainable skip pathway, it informs their combined role in finite-model learning.

Q4: Shallow Water coordinates and initial radii.

Both requested controls are in Appendix F.8. The large gain persists after removing Stage-1 absolute coordinates. It also persists on radii outside the training range, including against Full Direct, with trajectories kept intact across the split. Thus the benefit does not depend solely on the local coordinate input or interpolation among training initial states. Recurring dam-break structure and nearby observations nevertheless make Pixel 30% particularly favorable to local learning, explaining why its gain is especially large.

Q5 / W2: ocean setting and temporal baselines.

DINEOF uses temporal stacking, as clarified in Appendix E.4. We considered DINCAE in selecting the baselines: Barth et al. (2024), Section 5, Tables 2-3 report lower ensemble-mean RMSE for DINDiff on the Black Sea development and test sets. We therefore included DINDiff; we have not added a DINCAE experiment.

Reviewer_QrTZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper is about filling gaps in satellite ocean data when you never have a complete map to train on. It follows the IDCD and OAMP idea of hiding some of the observed values and learning to predict them. The method works in two steps. First, a small model learns to fill in hidden values inside small patches, and these patch predictions are averaged into a first guess for the whole map. Then a second model looks at the full map, the first guess, and where that guess disagrees with the data, and learns how to correct it. It gets the best results on three real ocean datasets and two PDE datasets.

Critical strengths

1- The results are solid. The method beats all baselines on the three ocean datasets and the numbers are averaged over several seeds. I also liked the controls. A baseline with the same number of parameters still does worse so the gain isn’t just from a bigger model. Replacing the first guess with OAMP or DINEOF also does worse so the local prior really matters.

2- The ablations are thorough and back up most design choices. The test showing the second model actually uses far-away observations is a nice touch.

Critical weaknesses

1- The idea itself isn’t very new. Making a first guess and then learning a correction is a common pattern in image inpainting and coarse to fine methods. The theory part doesn’t really explain why it helps, and the paper even admits the direct and residual versions have the same optimum in theory. So I’m still not sure what’s behind the gain. It could be the patch training acting like augmentation, having two models stacked, or the residual setup itself.

2- Some comparisons are hard to read. At test time the method is basically one prediction trained to minimize MSE, while several baselines are generative models that sample which naturally gives them higher MSE. I also missed simple baselines like interpolation or kriging which people commonly use for this kind of data. This matters most for Shallow Water Pixel 30% where the method is about 25x better than everything else which is surprised me. And honestly, the paper is very dense and took a lot of effort to follow.

Questions for authors

Clarifications:

1- How does your method do against simple interpolation or kriging, especially on the Pixel 30% settings? 2- Why is the gap on Shallow Water Pixel 30% so large, even against your own parameter-matched baseline? 3- If you train a single full-map model on random crops with the same loss, how close does it get? That would show how much comes from the patch training. 4- Since you only use one near-zero noise step at test time, do you really need the diffusion setup? Would plain regression with the same loss work just as well?

Suggestions:

1- It would be good to see results at 256×256, like OAMP, to check if the method scales. 2- A better presentation for the main text and ideas would make the paper much easier to follow.

3 comments

Critical Weakness 1: contribution and sources of improvement.

Our contribution establishes how local learning and global correction can be organized to improve reconstruction from never-complete training fields. A first estimate followed by correction is a familiar pattern; the substantive question here is how to learn the estimate, reconcile overlapping predictions, and connect it to a corrector when complete supervision is unavailable.

Table 2 addresses the stacking and residual explanations: Global2Global retains the correction design but learns its prior full-domain and is less accurate than the local-first model, and Prior Direct already improves on Full Direct before adding residual prediction. Section 4.4 and Tables 14-16 further test the prior connection and local covering. Together with the added crop-only, regression, and skip controls, these experiments establish which choices make the combination effective.

Appendix D.4 identifies the shared unrestricted prediction target. The estimator and skip controls examine how the representation and initialization affect learning in finite networks, as discussed in Sections 4.3-4.4.

Critical Weakness 2: baseline objectives.

The comparison against Full Direct uses the same query-supervised point-estimation setting, and the inference comparisons are given in Tables 2 and 20. The gain therefore persists when the comparator also predicts a point estimate. It is also visible in the fields: Ours preserves the vorticity structure smoothed by parameter-matched IDCD in Figure 4, and clearer local variation than parameter-matched OAMP in Figure 1. The gradient and spectral analysis in Appendix F.6 supports these observations. This evidence ties the benefit to recovered spatial features, beyond the distinction between scoring samples and scoring point estimates.

Q1: interpolation and kriging.

The ordinary-kriging comparison in Appendix F.7, Table 23 gives the following results on complete-reference 64×6464\times64 PDE fields. MSE is in units of 10−310^{-3} and PSNR is in dB; ± denotes standard deviation.

SettingKriging MSE ↓Ours MSE ↓Kriging PSNR ↑Ours PSNR ↑
Shallow Water Pixel 30%1.2580.035 ± 0.00738.9751.63 ± 0.96
Navier-Stokes Pixel 30%16.9054.623 ± 0.04624.1529.60 ± 0.04
Shallow Water Block 8/914.9173.48131.5545.04
Navier-Stokes Block 8/936.86525.32620.9422.59

Ours improves both metrics in all four settings. Under Pixel 30%, MSE is reduced by 97.2% on Shallow Water and 72.7% on Navier-Stokes relative to kriging. These results show the benefit of learning recurring spatial relationships across incomplete fields over spatial interpolation from the supplied observations.

Q2: the large Shallow Water gain.

Recurring dam-break structure and nearby observations favor local learning under Pixel 30%: most missing sites still have an observed immediate neighbor. This explains why this benchmark is especially favorable to the proposed division of learning. Appendix F.8 now tests the coordinate and initial-state explanations directly. The large gain persists after removing the local predictor's absolute coordinates. With initial radii outside the training range, it also persists against parameter-matched Full Direct. These results support reusable reconstruction patterns across the tested coordinate and initial-radius settings.

Q3: a full-map model trained on random crops.

We have added this control with the same predictor and query loss in Appendix F.7, Table 24. The two-stage model is substantially more accurate, showing the benefit of constructing a local field estimate and learning its full-domain revision over crop exposure alone.

Q4: is diffusion needed?

Plain regression works well with the same two-stage organization; diffusion adds a further improvement (Table 24). Diffusion training exposes the model to reconstruction across noise levels. The core reconstruction organization is therefore effective without diffusion, while noise-level training improves its accuracy.

Resolution and presentation.

The existing 128×128128\times128 results use separately trained models; evaluation at 256×256256\times256 remains future work. The revised main discussion states what each control establishes, while detailed new protocols and results are collected in Appendices F.7-F.9 to make the argument easier to follow.

Dear Reviewer QrTZ,

Thank you again for your thorough review. As the discussion period draws to a close, we would be grateful to know whether our response and the revised paper have addressed your concerns. If any points remain unclear, we would be happy to clarify them.

Thanks for the reply and the new experiments. They answer most of my questions. I’ll maintain my score.

Reviewer_jKwROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies physical-field reconstruction when every training sample is incomplete. It proposes a two-stage estimator that constructs local predictions and subsequently revises them through full-domain residual correction. A shared, coordinate-conditioned patch predictor learns local spatial relationships through context–query partitions of available observations. Its overlapping predictions are fused into a consensus field. A full-domain network then receives this frozen estimate, the retained observations, and discrepancy features to predict a residual correction. Experiments cover three ocean datasets with authentic observation gaps and complete-reference Shallow Water and Navier–Stokes datasets. The method consistently improves reconstruction accuracy in the evaluated settings, with estimator comparisons, component ablations, and remote-observation interventions examining the roles of its components.

Critical strengths

The empirical improvements are consistent and supported by meaningful controls. In Table 1, the method achieves the lowest MSE on all three ocean datasets, improving over the strongest external baseline for each dataset by 28.9%–34.5%. More importantly, Table 2 includes parameter-matched Full Direct, Prior Direct, Prior Residual, and two alternative priors. These comparisons examine the value of prior conditioning, residual prediction, and the source of the reference estimate. The PDE experiments and higher-resolution results provide complementary validation. Together, these controls make the empirical evidence more convincing than benchmark improvements alone.

The division of labor between local prediction and full-domain correction is coherent, and the scope of the claims is carefully stated. The paper explicitly recognizes that the local prior is a deterministic transformation of the retained context, with its value lying in finite-model learning rather than additional information or a resolution of distributional non-identifiability. The remote-observation interventions further distinguish responsiveness to nonlocal information from improvements in accuracy, including the mixed findings on SSS. This analysis helps clarify both the contribution and its limitations.

Critical weaknesses
  1. Essential implementation and statistical details are missing, limiting reproducibility and assessment of the controlled comparisons. Although the appendices describe the algorithmic interfaces and evaluation protocols, I could not locate the concrete local/global network architectures, numerical parameter counts, optimizer, learning rate, batch size, training steps, or number of independent training seeds. Appendix E.4 states that Full Direct is parameter matched and has comparable training time, but provides neither numerical parameter counts nor training times. Tables 16–17 also leave the confidence-interval construction and treatment of dependence among blocks from the same field unspecified. These omissions matter because the central empirical argument relies on comparisons under finite model capacity and training budgets.

  2. The quantitative evidence for improved fine-scale structure is less direct than the evidence for lower pointwise error. Reducing oversmoothing and preserving local detail are important motivations, yet all reported accuracy metrics are MSE and PSNR, while improvements in fine-scale structure are primarily illustrated through visual examples. Lower pointwise error alone does not establish more accurate sharp transitions or high-frequency spatial variation, particularly when ocean metrics are computed in a transformed, normalized, and clipped space. The complete-reference PDE datasets would support a direct assessment using gradient errors or frequency-resolved spectral errors. This limitation does not invalidate the MSE improvements, but weakens the evidence for the claimed structural advantage.

Questions for authors

Please provide the concrete architectures, component-level and total parameter counts, training steps, measured training times, main optimization hyperparameters, and number of independent training seeds for the two-stage model and Full Direct. What rules were used for hyperparameter and checkpoint selection across methods?

How were the 95% confidence intervals in Tables 16–17 constructed? How does the procedure account for dependence among blocks within the same field, repeated evaluations of the same sample across training seeds, and potentially correlated samples across time?

On the complete-reference PDE datasets, do improvements over parameter-matched Full Direct also hold under gradient-error or spatial-spectrum metrics? This would directly affect my assessment of whether local learning improves structural fidelity.

3 comments

structural fidelity and implementation

Critical Weakness 2 / Q3: structural fidelity.

The visualizations in Figures 1 and 4 show that our method substantially outperforms the parameter-matched baselines in recovering local structure. In Figure 4, Ours recovers the elongated vorticity feature inside the held-out Navier-Stokes block, preserving its shape, orientation, and local contrast; parameter-matched IDCD largely smooths this feature away. In Figure 1, Ours preserves distinct local variations and transitions in Baltic Sea NANO that parameter-matched OAMP visibly flattens. The recovered structures are much closer to the reference, making the improvement directly visible in the reconstructed fields.

The requested complete-reference PDE evaluation is now in Appendix F.6, Table 22. Improvements over parameter-matched Full Direct hold for both interior-gradient and frequency-coefficient errors in the tested Pixel 30% and Block 8/9 settings. Interior edges assess variation wholly inside the reconstructed region; the spectral comparison retains coefficient phase or sign to assess spatial alignment as well as variation strength. The improvements are largest under pixel observations and smaller under block missingness, showing how the structural benefit varies with observation geometry.

The sharp-transition analysis also shows better recovery where reference gradients are strongest. Under Pixel 30%, Ours recovers more of the high-frequency variation suppressed by Full Direct on Navier-Stokes, while reducing excess variation on Shallow Water. Both changes improve fidelity to the reference.

Regarding the transformed evaluation space, the additional analyses include affine denormalization and evaluation of unchanged predictions against unclipped references; the model ordering is preserved. The Shallow Water frequency improvement also persists under the alternative windowed-FFT evaluation. Together, the visual comparisons and these tests demonstrate improved recovery of spatial features across these evaluation settings.

Critical Weakness 1 / Q1: implementation and controlled comparisons.

Appendix E.5 now specifies the architectures, component and total parameter counts, optimization settings, configured training schedule, independent seeds, and measured training costs. The Full Direct comparison allocates approximately the same total model capacity to one full-domain predictor as our complete two-stage estimator.

For our stages and Full Direct, checkpoint selection uses the lowest validation MSE with early stopping. Reconstruction mode, patch scale, and fusion settings are selected on validation data, as explained in Appendices B.4 and F.3. The new training-cost table reports the combined two-stage cost to the selected checkpoints, including frozen-prior construction during corrector training; it is compared with Full Direct's training cost.

Q2 - confidence intervals, purpose, and dependence

Appendix F.4 now specifies the interval construction (p. 36). The original Tables 16-17 are Tables 18-19 in the revised PDF. These experiments ask two separate questions: does the global corrector use observations unavailable to the local prior, and does that use improve reconstruction accuracy? The confidence intervals quantify uncertainty in the mean diagnostic effects.

What is measured, and why. For a scored block BB, let U(B)U(B) be the union of the patch supports that determine its local prior. Table 18 removes the entire farthest third of supplied observations outside U(B)U(B). The prior on BB remains unchanged for every eligible evaluation, while the final prediction changes. This tests whether the corrector actually uses evidence beyond local support.

Table 19 measures the paired block-error change, Δa(B)=MSEB(x^a)−MSEB(x^full)\Delta_a(B)=\mathrm{MSE}_B(\widehat{x}^{a})-\mathrm{MSE}_B(\widehat{x}^{\mathrm{full}}), after deleting observations from Remote Near, Remote Mid, Remote Far, or Local Near. Each arm removes the same number of observations for a given block. The checkpoint, original supplied input, near-zero perturbation, and scoring block are shared across arms (Table 17). Pairing controls evaluation-item and model differences; matching the deletion count keeps the removal dose equal when comparing bands.

Construction and rationale. The revised PDF specifies the following procedure:

  1. Compute each block-level displacement or paired error difference before resampling.
  2. In each of 10,000 bootstrap replicates, independently sample, with replacement, the original number of sample_id x training-seed units. Retain all eligible blocks within every selected unit.
  3. Recompute the pooled mean of the block-level measurements. Use its 2.5th and 97.5th bootstrap percentiles as the two-sided 95% interval endpoints.

Keeping a field-seed unit's blocks together preserves their within-unit dependence instead of treating each block as an independent observation. Recomputing the pooled block mean matches the reported statistic: eligible blocks have equal weight, so units with more eligible blocks contribute more measurements. The percentile construction allows an asymmetric interval without imposing a Gaussian approximation. The 95% level is the nominal coverage level, and the 10,000 replicates approximate the corresponding bootstrap quantiles.

What the intervals help establish. Tables 18-19 report the following mean effects [95% CI], in units of 10−510^{-5}.

DatasetFinal-output displacement: entire Remote Far tercile (Table 18)Change in block MSE: matched-count Remote Far deletion (Table 19)
Baltic Sea NANO6.46 [4.71, 8.61]5.25 [1.14, 9.56]
Black Sea CHL5.80 [4.71, 7.09]3.72 [0.65, 6.75]
Global Ocean SSS62.30 [56.96, 68.02]-12.85 [-18.64, -7.09]

The corrector responds beyond local support on all three products. Under matched-count deletion, the positive error changes on NANO and CHL support an average benefit from the removed far observations, whereas the negative change on SSS shows the opposite. Together, the diagnostics show how the accuracy effect of remote evidence varies across products.

Dependence. Evaluations of the same field under different seeds are sampled separately; their shared-field dependence is not explicitly modeled. We did not resample contiguous time blocks, date groups, or trajectories, so temporal dependence is also not explicitly modeled.

Upon acceptance, we will fully open-source the experimental code, including the statistical and evaluation routines, training configurations, and preprocessing.

Dear Reviewer jKwR,

Thank you again for your thorough review. As the discussion period draws to a close, we would be grateful to know whether our response and the revised paper have addressed your concerns. If any points remain unclear, we would be happy to clarify them.

Comments on the paper

1 comment

General Response

Our contribution is an organization of reconstruction learning grounded in never-complete observations. Local patch learning pools observable spatial relationships into an explicit field estimate; full-domain learning then revises it using broader evidence. This gives local structure a dedicated learning role when complete-field supervision is unavailable. The main issue raised across the reviews is which aspects of this combination produce the benefit. The existing controls and added experiments address that question by examining the prior, its connection to the corrector, crop exposure, the training objective, and the initial prediction.

The results support a deliberate division of learning. Crop training alone does not reproduce the full method's benefit; direct regression retains much of it, with diffusion training contributing a further improvement. The explicit-skip control recovers much of the CHL gain and leaves a larger difference on SSS, helping distinguish what the residual formulation contributes in finite models. The coordinate and initial-radius controls also retain the Shallow Water advantage. These findings sharpen the explanation of the method rather than relying on benchmark gains alone.

The spatial benefit is visible in Figures 1 and 4, where parameter-matched baselines smooth features retained by Ours. The added complete-reference PDE analysis connects those observations to gradient, spectral, and sharp-transition errors. This supports the claimed structural improvement beyond aggregate pixelwise scores.

Revisions incorporated into the paper.

  • The abstract and Section 4.3 explicitly distinguish the comparison against parameter-matched Full Direct from external-baseline comparisons. Appendix E.4 clarifies that Global2Global implements the global-first design, with the same backbone family and comparable first-stage capacity, and that DINEOF uses temporal stacking.
  • Appendix E.5 adds concrete implementation, training, model-selection, capacity, and computational-cost details. Appendix F.4 specifies the confidence-interval construction; our response to jKwR directly explains how blocks, repeated-seed evaluations, and dates are treated.
  • Appendix F.6 adds the gradient, spectral, sharp-transition, and evaluation-space analyses.
  • Appendices F.7-F.9 add kriging, random-crop full-map training, two-stage regression, local-coordinate removal, initial-radius extrapolation, and explicit-skip controls. The main experimental discussion now states what these comparisons establish.

Reproducibility. The revised reproducibility statement points to the model interfaces and algorithms, data provenance and preprocessing, implementation and training settings, and statistical procedure. Upon acceptance, we will fully open-source the experimental code, including model implementations, training and evaluation procedures, configurations, preprocessing, and statistical diagnostics. The datasets remain available from their original providers under the providers' access and licensing terms.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.