acceptodds
Under review as a conference paper at ICLR 2027

Anchored Adaptive Probes with Scale-Aware Rendering for Atmospheric Super-Resolution

Abstract

Arbitrary-scale spatial super-resolution of atmospheric fields aims to alleviate limitations in acquiring high-resolution data by enabling a single model to meet resolution requirements, including those at unseen scales, without scale-specific training. However, reconstruction performance may be constrained if models do not adequately account for region-specific detail recovery requirements and variations in the target scale. Jointly modeling the spatial allocation and scale-adaptive use of reference information can align the distribution of feature support with regional refinement needs while adapting its spatial range of influence to the target scale. We propose APSR, a reconstruction model that combines anchored adaptive probes with scale-aware rendering. The model treats input grid nodes as stable anchors that preserve basic information and generates probes carrying auxiliary information within grid cells on demand, forming the reference information needed for reconstruction. Scale-aware rendering then adaptively regulates the influence of this reference information to enable continuous-scale reconstruction. Experiments on ERA5 show that APSR performs well within the training scale range of – and outperforms all evaluated arbitrary-scale methods at the tested extrapolation scales, including , demonstrating good generalization across scales.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
climate downscaling, image super-resolution

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_dQgYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes A²PSR, which combines anchored adaptive probes with scale-aware rendering for arbitrary-scale atmospheric super-resolution, and evaluates it on ERA5, demonstrating good performance extrapolation scales, such as 22.5×.

Critical strengths
  1. The method achieves good reconstruction performance at large extrapolation scales, demonstrating, to some extent, its generalization capability to unseen scales and indirectly supporting the rationale behind the proposed components.
  2. The authors evaluate the method using the real-world case of Hurricane Teddy and obtain promising results, which provide some evidence of its practical applicability to real-world scenarios.
Critical weaknesses
  1. The methodological novelty is insufficiently established. Prior studies, including ClassSR and GeoFAR, have explored reconstruction that accounts for regional differences, while EQSR (CVPR 2023) has investigated scale-adaptive reconstruction. The manuscript does not clearly articulate the distinct capabilities or methodological advances offered by the proposed approach relative to these works.
  2. The suitability of shared probe allocation across variables is insufficiently examined. Although region-specific detail recovery is a central motivation, the model uses a shared allocation mechanism for all three atmospheric variables. The manuscript provides limited analysis of their potentially different spatial refinement requirements and whether the shared mechanism adequately balances these demands.
  3. The one-probe-per-cell guarantee lacks sufficient justification. This strategy reserves probes for every cell, including smooth regions, before allocating the remaining budget through cross-region competition. Without a comparison against fully global candidate selection under the same budget, the benefit of enforcing this coverage constraint remains unclear.
  4. The ablations and accompanying analysis do not sufficiently establish the contributions of individual components. For example, comparing a preset radius with a learned scale-dependent radius conflates radius learning with scale conditioning. The discussion is also largely descriptive, providing limited explanation of component interactions or why their effects vary across variables and scales.
Questions for authors
  1. The distinction from prior region-adaptive methods, such as ClassSR and GeoFAR, and scale-adaptive methods, such as EQSR, requires clearer explanation.
  2. The manuscript does not sufficiently explain how shared probe allocation accommodates potentially different refinement needs across atmospheric variables.
  3. Guaranteeing one probe per cell, including smooth regions, lacks sufficient justification. A comparison with fully global selection under the same budget is needed.
  4. The ablations do not fully isolate component contributions. For example, the radius comparison conflates radius learning with scale conditioning. Component interactions and their varying effects across variables and scales also receive limited analysis.
3 comments
AuthorsOpenReview ↗

Response to Reviewer 3 (Part 1/3)

W1&Q1:

The innovation of our method in region-adaptive reconstruction lies in supplementing spatial representations through adaptive probes according to the potential detail recovery needs of different regions, so that the spatial distribution of feature support adapts to regional needs. In comparison, ClassSR adapts to regional differences by selecting network branches with different capacities for different regions; GeoFAR introduces geographic information such as terrain to adjust the frequency features of each region and enhance regional detail recovery.

The innovation of our method in scale-adaptive reconstruction lies in explicitly learning a unified rendering range for the two types of representations, anchors and probes, conditioned on the target scale. With the spatial support provided by auxiliary representations, selection and aggregation centered on the target location are no longer limited to the original grid locations. In comparison, EQSR performs scale-aware aggregation within a predefined 3×3 grid neighborhood, and the neighborhood structure itself is not learned and adjusted according to the target scale.

W2&Q2:

Probe allocation in our method is based on joint multivariable representations in the latent space. Considering the relationships among atmospheric variables, we learn shared allocation through joint encoding and a joint reconstruction objective, enabling probes to integrate information from different variables and provide common spatial support for their detail recovery. If independent probes of the same size were allocated to each variable, the total number of probes and the corresponding computational and storage costs would increase with the number of variables, and spatial support might also be allocated redundantly in regions where the needs of different variables overlap. Therefore, shared allocation also helps control computational cost and supports extension to more variables.

To examine whether shared allocation can accommodate the detail needs of each variable, we added a per-variable reconstruction comparison between adaptive allocation and random allocation at the 22.5× scale under the same probe budget (2,500 for both). The calculation of detail demand and the grouping of regions are both consistent with Appendix Table 9: the regional area-weighted interpolation mean squared error of each variable is divided by the corresponding training-set variance, and the results are averaged to obtain joint demand. Within each sample, Q1–Q4 are formed in ascending order of demand, and both methods and all three variables use the same grouping. As shown in the table, adaptive allocation reduces the reconstruction errors of all three variables in every group, and the relative improvement increases with demand. This indicates that shared allocation can bring reconstruction benefits to each variable in regions with different levels of demand, providing support for its ability to accommodate the detail recovery needs of multiple variables.

\begin{array}{cccc} \hline \text{Demand} & \text{T2M (K)} & \text{U10 (m/s)} & \text{V10 (m/s)} \\\\ \begin{array}{c} \text{group} \\\\ \hline \text{Q1} \\\\ \text{Q2} \\\\ \text{Q3} \\\\ \text{Q4} \end{array} & \begin{array}{ccc} \text{Random} & \text{Adaptive} & \text{Reduction (\\%)} \\\\ \hline 0.7288 & \mathbf{0.6603} & 9.40\\% \\\\ 1.2826 & \mathbf{1.1600} & 9.56\\% \\\\ 1.7397 & \mathbf{1.5679} & 9.87\\% \\\\ 1.7002 & \mathbf{1.5230} & 10.42\\% \end{array} & \begin{array}{ccc} \text{Random} & \text{Adaptive} & \text{Reduction (\\%)} \\\\ \hline 0.6889 & \mathbf{0.6546} & 4.99\\% \\\\ 1.0149 & \mathbf{0.9532} & 6.08\\% \\\\ 1.3401 & \mathbf{1.2516} & 6.60\\% \\\\ 1.9560 & \mathbf{1.8225} & 6.82\\% \end{array} & \begin{array}{ccc} \text{Random} & \text{Adaptive} & \text{Reduction (\\%)} \\\\ \hline 0.7007 & \mathbf{0.6827} & 2.56\\% \\\\ 0.9808 & \mathbf{0.9315} & 5.02\\% \\\\ 1.2744 & \mathbf{1.1968} & 6.08\\% \\\\ 1.8565 & \mathbf{1.7279} & 6.92\\% \end{array} \\\\ \hline \end{array}$$
AuthorsOpenReview ↗

Response to Reviewer 3 (Part 2/3)

W3&Q3:

In video processing, smoothness usually refers to small changes in color and texture; in atmospheric reconstruction, visually inconspicuous changes in physical quantities still need to be recovered accurately. Therefore, we provide basic supplementary support within each cell through one probe per cell and then use the remaining budget for focused refinement. We added an experiment with fully global selection under the same budget: the original anchors are retained, the per-cell guarantee is removed, and all 2,500 probe slots are allocated globally through learned candidate scores. As shown in the table below, removing the guarantee increases the reconstruction errors of all three variables at all four evaluation scales, supporting the effectiveness of combining basic coverage within each cell with adaptive allocation of the remaining budget.

\begin{array}{cccc} \hline \text{Probe allocation}&\text{T2M (K)}&\text{U10 (m/s)}&\text{V10 (m/s)}\\\\ \begin{array}{c} \text{strategy}\\\\ \hline \begin{array}{c}\text{Fully global selection}\\\\\text{(additional experiment)}\end{array}\\\\ \hline \begin{array}{c}\text{Full model}\\\\\text{(Table 2 in the original manuscript)}\end{array} \end{array} & \begin{array}{cccc} 2\times&4\times&8\times&22.5\times\\\\ \hline \vphantom{\begin{array}{c}1\\\\1\end{array}}0.521&0.862&1.112&1.332\\\\ \hline \vphantom{\begin{array}{c}1\\\\1\end{array}}\mathbf{0.495}&\mathbf{0.814}&\mathbf{1.045}&\mathbf{1.248} \end{array} & \begin{array}{cccc} 2\times&4\times&8\times&22.5\times\\\\ \hline \vphantom{\begin{array}{c}1\\\\1\end{array}}0.633&0.924&1.121&1.272\\\\ \hline \vphantom{\begin{array}{c}1\\\\1\end{array}}\mathbf{0.611}&\mathbf{0.884}&\mathbf{1.064}&\mathbf{1.196} \end{array} & \begin{array}{cccc} 2\times&4\times&8\times&22.5\times\\\\ \hline \vphantom{\begin{array}{c}1\\\\1\end{array}}0.596&0.881&1.067&1.208\\\\ \hline \vphantom{\begin{array}{c}1\\\\1\end{array}}\mathbf{0.574}&\mathbf{0.844}&\mathbf{1.019}&\mathbf{1.153} \end{array}\\\\ \hline \end{array}$$
AuthorsOpenReview ↗

Response to Reviewer 3 (Part 3/3)

W4&Q4:

We added an experiment with a learnable radius that does not depend on the target scale to distinguish the contributions of radius learning and scale conditioning. The comparison between a fixed non-learnable radius and a scale-independent learnable radius reflects the role of radius learning itself; when the radius is also learnable, the comparison with a scale-aware radius reflects the additional contribution of scale conditioning. The additional results show that both help improve reconstruction performance.

Regarding interactions among components, adaptive probes provide auxiliary spatial representations, while scale-aware rendering adjusts the range over which these representations are aggregated. Table 2 in the original manuscript shows that, at 22.5×, changing from a predefined fixed radius to a scale-aware radius increases the errors of all three variables when only anchors are used, but decreases all of them after adaptive probes are added. This indicates that the effect of adjusting the range depends on the available spatial support, and that the two work together. Under the same probe budget, adaptive allocation outperforms random allocation, also indicating that the allocation of auxiliary representations affects their value for reconstruction. The benefits also differ across variables and scales. The table below shows that the improvement brought by scale conditioning does not consistently increase with scale. This is related to differences in the local variations and spatial correlations of the variables: the aggregation range needs both to preserve local detail and to use surrounding information, while different variables have different needs for the two at different reconstruction scales and therefore respond differently to adjustments of the range.

\begin{array}{cccc} \hline \text{Radius}&\text{T2M (K)}&\text{U10 (m/s)}&\text{V10 (m/s)}\\\\ \begin{array}{c} \text{setting}\\\\ \hline \begin{array}{c}\text{Fixed non-learnable}\\\\\text{(Table 2 in the original manuscript)}\end{array}\\\\ \begin{array}{c}\text{Fixed learnable}\\\\\text{(additional experiment)}\end{array}\\\\ \hline \begin{array}{c}\text{Scale-aware learnable}\\\\\text{(Table 2 in the original manuscript)}\end{array} \end{array} & \begin{array}{cccc} 2\times&4\times&8\times&22.5\times\\\\ \hline \vphantom{\begin{array}{c}1\\\\1\end{array}}0.508&0.850&1.099&1.316\\\\ \vphantom{\begin{array}{c}1\\\\1\end{array}}0.502&0.839&1.072&1.269\\\\ \hline \vphantom{\begin{array}{c}1\\\\1\end{array}}\mathbf{0.495}&\mathbf{0.814}&\mathbf{1.045}&\mathbf{1.248} \end{array} & \begin{array}{cccc} 2\times&4\times&8\times&22.5\times\\\\ \hline \vphantom{\begin{array}{c}1\\\\1\end{array}}0.617&0.908&1.101&1.243\\\\ \vphantom{\begin{array}{c}1\\\\1\end{array}}0.616&0.902&1.088&1.215\\\\ \hline \vphantom{\begin{array}{c}1\\\\1\end{array}}\mathbf{0.611}&\mathbf{0.884}&\mathbf{1.064}&\mathbf{1.196} \end{array} & \begin{array}{cccc} 2\times&4\times&8\times&22.5\times\\\\ \hline \vphantom{\begin{array}{c}1\\\\1\end{array}}0.578&0.857&1.031&1.157\\\\ \vphantom{\begin{array}{c}1\\\\1\end{array}}0.577&0.856&1.029&1.154\\\\ \hline \vphantom{\begin{array}{c}1\\\\1\end{array}}\mathbf{0.574}&\mathbf{0.844}&\mathbf{1.019}&\mathbf{1.153} \end{array}\\\\ \hline \end{array}$$
Reviewer_ZU1ROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a neural network framework to tackle the task of arbitrary-scale super-resolution for atmospheric fields. This paper focuses on the challenge of recovering region-specific details and variations, and analyzes how existing coordinate-based implicit methods are constrained by a predefined grid structure and lack the flexibility to incorporate local details. To address this issue, this paper proposes a network architecture mainly consisting of two stages: (i) first constructs adaptive local probes from the coarse grid anchors, and (ii) then aggregates features through scale-aware rendering to reconstruct the atmospheric fields at the target resolution. Results show superior performance to baseline methods.

Critical strengths
  1. This paper analyzes an interesting perspective for the arbitrary-scale super-resolution task -- the region-specific detail recovery requirement is an interesting issue that existing works have not handled. The proposed framework provides a reasonable solution to this issue.
  2. The experimental results are strong -- the baseline methods are significantly outperformed in the comparison.
Critical weaknesses

Although the proposed method seems effective, I still find the overall framework structure a bit too complicated. The network architecture has many dedicated modules, and I find it hard to understand them all one by one while reading the paper. It is unclear if the model architecture can be simplified while retaining similar reconstruction quality.

Questions for authors
  1. What is the scale of the training dataset? How many are in the training set, and how many are in the test set? Given the complexity of the model architecture, I'm concerned about overfitting if the training set is not large enough.
  2. Is the proposed method specific to the atmospheric field super-resolution task? If so, which design components are dedicated for this task? If not, it may be an interesting future work to apply the principle to the image super-resolution task.
3 comments
AuthorsOpenReview ↗

Response to Reviewer 2

W1:

In terms of parameter count, our model has 7.924M parameters, fewer than the 24.013M parameters of the atmospheric super-resolution method MINet. In terms of structure, anchored adaptive probes and scale-aware rendering are organized around two clear needs: regional detail recovery and generalization across scales, with each part having a clear role.

To address the difficulty of understanding each specially designed module, we explain the roles and connections of the modules through the following process: assessing supplementary needs, generating auxiliary representations, and using these representations for reconstruction. Cross-region attention combined with scale encoding helps the model determine the supplementary needs of different regions at the target scale. Based on this, candidate probes are generated within each grid cell, and their representations are refined through information exchange among the candidates. The scoring module then determines which probes to retain, and the position and feature prediction modules generate the corresponding auxiliary spatial representations. These probes and the original grid anchors jointly provide information for reconstruction. The scale-aware radius determines the aggregation range at the target location, relative position modeling adjusts the information being passed, and confidence and distance weighting control their respective contributions. Finally, the decoder outputs the atmospheric variables at the target location.

The ablation experiments in Table 2 further show that each component contributes to reconstruction performance, and simplifying the model architecture by removing or simplifying the corresponding components leads to performance degradation to varying degrees.

Q1:

The training set uses hourly data from 2008–2018, with 96,432 samples; the validation set uses data from 2019, with 8,760 samples; and the test set uses data from 2020, with 8,784 samples. As shown in the table, as the training loss decreases, the validation errors of the three variables at 2× and 22.5× also decrease overall and tend to stabilize in the later stages of training, with only small fluctuations. There is no trend of continuously decreasing training loss accompanied by continuously worsening validation errors, indicating that no clear signs of overfitting were observed during the current training process.

\begin{array}{cccc} \hline \smash{\lower{0.6em}{\text{Epoch}}}&\text{Training}&2\times\\;\text{validation Lat-RMSE}&22.5\times\\;\text{validation Lat-RMSE} \\\\ \begin{array}{c} \phantom{\text{loss}} \\\\ \hline 1 \\\\ 2 \\\\ 3 \\\\ 4 \\\\ 5 \\\\ 6 \\\\ 7 \\\\ 8 \\\\ 9 \\\\ 10 \\\\ 11 \\\\ 12 \\\\ 13 \\\\ 14 \\\\ 15 \\\\ 16 \\\\ 17 \\\\ 18 \\\\ 19 \\\\ 20 \end{array} & \begin{array}{c} \text{loss} \\\\ \hline 0.0719 \\\\ 0.0537 \\\\ 0.0503 \\\\ 0.0482 \\\\ 0.0468 \\\\ 0.0457 \\\\ 0.0448 \\\\ 0.0441 \\\\ 0.0433 \\\\ 0.0426 \\\\ 0.0422 \\\\ 0.0417 \\\\ 0.0414 \\\\ 0.0411 \\\\ 0.0409 \\\\ 0.0406 \\\\ 0.0401 \\\\ 0.0402 \\\\ 0.0400 \\\\ 0.0398 \end{array} & \begin{array}{ccc} \text{T2M (K)}&\text{U10 (m/s)}&\text{V10 (m/s)} \\\\ \hline 0.7400&0.7324&0.6923 \\\\ 0.6600&0.6830&0.6394 \\\\ 0.6204&0.6590&0.6214 \\\\ 0.6151&0.6502&0.6117 \\\\ 0.5888&0.6425&0.6059 \\\\ 0.5788&0.6341&0.5984 \\\\ 0.5809&0.6293&0.5930 \\\\ 0.5633&0.6240&0.5898 \\\\ 0.5536&0.6199&0.5851 \\\\ 0.5468&0.6212&0.5837 \\\\ 0.5430&0.6161&0.5817 \\\\ 0.5369&0.6145&0.5802 \\\\ 0.5358&0.6150&0.5799 \\\\ 0.5334&0.6146&0.5790 \\\\ 0.5360&0.6159&0.5785 \\\\ 0.5309&0.6143&0.5780 \\\\ 0.5298&0.6138&0.5775 \\\\ 0.5299&0.6138&0.5774 \\\\ 0.5289&0.6135&0.5772 \\\\ 0.5292&0.6136&0.5773 \end{array} & \begin{array}{ccc} \text{T2M (K)}&\text{U10 (m/s)}&\text{V10 (m/s)} \\\\ \hline 1.4439&1.2840&1.2333 \\\\ 1.3784&1.2406&1.1865 \\\\ 1.3480&1.2277&1.1772 \\\\ 1.3576&1.2210&1.1746 \\\\ 1.3288&1.2284&1.1836 \\\\ 1.3125&1.2051&1.1650 \\\\ 1.3146&1.1946&1.1592 \\\\ 1.3066&1.2161&1.1778 \\\\ 1.3169&1.2058&1.1649 \\\\ 1.2994&1.2109&1.1705 \\\\ 1.2943&1.2005&1.1602 \\\\ 1.2938&1.2016&1.1655 \\\\ 1.2953&1.2037&1.1697 \\\\ 1.2978&1.2086&1.1702 \\\\ 1.2972&1.2089&1.1663 \\\\ 1.2924&1.2069&1.1677 \\\\ 1.2910&1.2074&1.1666 \\\\ 1.2869&1.2054&1.1660 \\\\ 1.2881&1.2058&1.1658 \\\\ 1.2888&1.2057&1.1662 \end{array} \\\\ \hline \end{array}$$ **Q2**: The core framework of our method is not limited to atmospheric fields. In principle, the ideas of adaptive probes and scale-aware rendering can also be applied to image super-resolution, although appropriate adjustments may be needed according to the characteristics of the task. We thank the reviewer for this suggestion and identify this application as a direction for future research.

The rebuttal has addressed most of my concerns. The roles of each proposed component within the architecture seem well justified and have been validated by the ablation study. However, I'd encourage the authors to revise the Sec. 3 of the paper to make it clearer and easier to understand. I'll retain my positive score.

We sincerely thank the reviewer for the positive assessment, thoughtful feedback, and careful consideration of our responses. We will revise Section 3 to make it clearer and easier to understand.

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

In this paper the authors propose A2SPR, for arbitrary scale atmospheric super-resolution, doing so by combining coarse-grid anchors and adaptive within-cell probes, and scale-aware feature aggregation. They use ERA5 for temperature and wind fields at 1x-2x scales for training, and reports improved reconstruction over the evaluated arbitrary-scale baselines at unseen scales up to 22.5x. Experiments investigate component contributions, probe budgets, stability, wind structures, and computational costs.

Critical strengths

First strength: the strong scale extrapolation. A single model that is only trained on 1x-2x which the outperforms evaluated arbitrary scale baselines at 4x, 8x, and 22.5x, which supports its usefulness across output resolutions. Second strength: The application-relevant validation. It is evaluated on 90 cyclone cases and shows improved spatial matching of the strong wind regions, which then provides evidence of meterological relevance beyond average reconstruction errors.

Critical weaknesses

First weakness: it is an uncertain comparison with the very closely related basline. GSSA-ViT is substantially modified, and the published performance is not reproduced, which then limits evidence of improvements over atmospheric downscaling approach.

Questions for authors

I would like to see the report of the official results from the GSSA-ViT configuration and explain why the gaussian grid and transformer depth had to be modified under the evaluation protocol?

Also, since we have the weak allocation-demand correlation, can you also report the gains of the reconstruction by detail-demand quartile against other random probes at the same budget. This can then show where adaptive allocation provides its benefit.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 1

W1&Q1:

Thank you to the reviewer for pointing out this issue. The atmospheric variables evaluated in this study differ from those in the official configuration. Limited by the computational resources available at the time, we made appropriate adjustments to the network mainly out of consideration for computational cost. After upgrading our computational resources, we conducted additional reproduction experiments with the original official configuration on the same test set and present its results alongside those of the adjusted configuration and our method in the table below. We will update the relevant experimental results and the configuration description in the appendix in the revised manuscript.

\begin{array}{lccc} \hline \text{Model configuration} & \text{T2M (K)} & \text{U10 (m/s)} & \text{V10 (m/s)} \\\\ \begin{array}{l} \phantom{2\times} \\\\ \hline \text{GSSA-ViT (official)} \\\\ \text{GSSA-ViT (adjusted)} \\\\ \hline \text{A²PSR (ours)} \end{array} & \begin{array}{cccc} 2\times & 4\times & 8\times & 22.5\times \\\\ \hline 1.017 & 1.703 & 2.074 & 2.275 \\\\ 1.364 & 1.914 & 2.180 & 2.339 \\\\ \hline \mathbf{0.495} & \mathbf{0.814} & \mathbf{1.045} & \mathbf{1.248} \end{array} & \begin{array}{cccc} 2\times & 4\times & 8\times & 22.5\times \\\\ \hline 0.958 & 1.511 & 1.815 & 1.985 \\\\ 1.156 & 1.659 & 1.886 & 2.007 \\\\ \hline \mathbf{0.611} & \mathbf{0.884} & \mathbf{1.064} & \mathbf{1.196} \end{array} & \begin{array}{cccc} 2\times & 4\times & 8\times & 22.5\times \\\\ \hline 0.911 & 1.480 & 1.804 & 1.983 \\\\ 1.104 & 1.628 & 1.885 & 2.030 \\\\ \hline \mathbf{0.574} & \mathbf{0.844} & \mathbf{1.019} & \mathbf{1.153} \end{array} \\\\ \hline \end{array}$$ **Q2**: We added a comparison of reconstruction by group between adaptive probes and random probes at the 22.5× scale under the same probe budget (2,500 for both). The calculation of detail demand and the grouping of regions are both consistent with Table 9: detail demand is measured by the regional area-weighted mean squared error between bilinear interpolation results and the ground truth. After dividing the error for each variable by its corresponding training-set variance, we average across the three variables to obtain the joint normalized detail demand. A larger value indicates more local detail that interpolation fails to recover. Within each sample, regions are divided into quartile groups Q1–Q4 in ascending order of demand, and both methods and all three variables use the same regional grouping. The results show that adaptive probes reduce the reconstruction errors of all three variables in Q1–Q4, and the relative improvement increases with the demand group. In Q4, the Lat-RMSE values of T2M, U10, and V10 decrease by 10.42%, 6.82%, and 6.92%, respectively. This indicates that the benefits of adaptive probes cover regions with different levels of demand and are more pronounced in high-demand regions. $$\small \begin{array}{cccc} \hline \text{Demand} & \text{T2M (K)} & \text{U10 (m/s)} & \text{V10 (m/s)} \\\\ \begin{array}{c} \text{group} \\\\ \hline \text{Q1} \\\\ \text{Q2} \\\\ \text{Q3} \\\\ \text{Q4} \end{array} & \begin{array}{ccc} \text{Random} & \text{Adaptive} & \text{Reduction (\\%)} \\\\ \hline 0.7288 & \mathbf{0.6603} & 9.40\\% \\\\ 1.2826 & \mathbf{1.1600} & 9.56\\% \\\\ 1.7397 & \mathbf{1.5679} & 9.87\\% \\\\ 1.7002 & \mathbf{1.5230} & 10.42\\% \end{array} & \begin{array}{ccc} \text{Random} & \text{Adaptive} & \text{Reduction (\\%)} \\\\ \hline 0.6889 & \mathbf{0.6546} & 4.99\\% \\\\ 1.0149 & \mathbf{0.9532} & 6.08\\% \\\\ 1.3401 & \mathbf{1.2516} & 6.60\\% \\\\ 1.9560 & \mathbf{1.8225} & 6.82\\% \end{array} & \begin{array}{ccc} \text{Random} & \text{Adaptive} & \text{Reduction (\\%)} \\\\ \hline 0.7007 & \mathbf{0.6827} & 2.56\\% \\\\ 0.9808 & \mathbf{0.9315} & 5.02\\% \\\\ 1.2744 & \mathbf{1.1968} & 6.08\\% \\\\ 1.8565 & \mathbf{1.7279} & 6.92\\% \end{array} \\\\ \hline \end{array}$$

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.