- 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
- 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.
- 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
- 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.
- 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.
- 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.
- 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
- The distinction from prior region-adaptive methods, such as ClassSR and GeoFAR, and scale-adaptive methods, such as EQSR, requires clearer explanation.
- The manuscript does not sufficiently explain how shared probe allocation accommodates potentially different refinement needs across atmospheric variables.
- Guaranteeing one probe per cell, including smooth regions, lacks sufficient justification. A comparison with fully global selection under the same budget is needed.
- 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
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}$$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}$$
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}$$