Pre-training for Deep Statistical Climate Downscaling: A case study within the Spanish National Adaptation Plan (PNACC)

González-Abad, Jose; Iturbide, Maialen; Hernanz, Alfonso; Gutiérrez, José Manuel

doi:10.5194/egusphere-2025-3754

Preprints

https://doi.org/10.5194/egusphere-2025-3754

Preprints

18 Sep 2025

| 18 Sep 2025

Pre-training for Deep Statistical Climate Downscaling: A case study within the Spanish National Adaptation Plan (PNACC)

Jose González-Abad, Maialen Iturbide, Alfonso Hernanz, and José Manuel Gutiérrez

Abstract. Deep Learning (DL) has recently emerged as a promising approach for statistical climate downscaling. In this study, we investigate the use of pre-training in this context, building on the DeepESD model developed for the Spanish National Adaptation Plan (PNACC), which uses ERA5 predictors and the 5 km ROCIO-IBEB national gridded predictand dataset. We evaluate the effectiveness of different fine-tuning strategies to adapt this pre-trained model to alternative regional predictand datasets, specifically a point-based station dataset. The objective is to develop downstream downscaling methods that maintain consistency with the original national-scale model while capturing the specific characteristics of regional and local datasets.

We analyze the benefits of fine-tuning in terms of faster convergence, improved generalization, and greater consistency. Using eXplainable Artificial Intelligence (XAI) techniques, we examine the relationships learned by the models and compare the resulting climate change signals. Our results demonstrate that pre-training provides a robust foundation for statistical downscaling, particularly in cases with limited spatial and/or temporal data availability (e.g., local high-resolution datasets available only for short periods), thereby reducing epistemic uncertainty and improving the reliability of future climate projections. Overall, this approach represents a step toward standardizing DL-based downscaling models to ensure more coherent and consistent climate projections across national and regional scales.

Received: 01 Aug 2025 – Discussion started: 18 Sep 2025

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Preprint (PDF, 4913 KB)

Notice on discussion status
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
Preprint (4913 KB)

Download & links

The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.

Journal article(s) based on this preprint

01 Jul 2026

Pre-training for deep statistical climate downscaling: enhancing consistency and robustness across regional datasets

Jose González-Abad, Maialen Iturbide, Alfonso Hernanz, and José Manuel Gutiérrez

Geosci. Model Dev., 19, 5781–5804, https://doi.org/10.5194/gmd-19-5781-2026,https://doi.org/10.5194/gmd-19-5781-2026, 2026

Short summary

Jose González-Abad, Maialen Iturbide, Alfonso Hernanz, and José Manuel Gutiérrez

Interactive discussion

Status: closed

CEC1:
'Comment on egusphere-2025-3754 - No compliance with the policy of the journal', Juan Antonio Añel, 11 Oct 2025

Dear authors,
Unfortunately, after checking your manuscript, it has come to our attention that it does not comply with our "Code and Data Policy".

https://www.geoscientific-model-development.net/policies/code_and_data_policy.html
To access part of the data used in your work, you have linked web pages from AEMET. However, these are not suitable repositories for scientific publication, and we can not accept them. Therefore, the current situation with your manuscript is irregular. Please, publish your data in one of the appropriate repositories and reply to this comment with the relevant information (link and a permanent identifier for it (e.g. DOI)) as soon as possible, as we can not accept manuscripts in Discussions that do not comply with our policy.
Also, you must include a modified 'Code and Data Availability' section in a potentially reviewed manuscript, containing the information of the new repositories.
I must note that if you do not fix this problem, we cannot accept your manuscript for publication in our journal.
Juan A. Añel

Geosci. Model Dev. Executive Editor

Citation: https://doi.org/10.5194/egusphere-2025-3754-CEC1
- AC1:
  'Reply on CEC1', José González-Abad, 13 Oct 2025
  
  Dear Juan A. Añel,
  We apologize for the oversight. All data previously linked through the AEMET URL, corresponding to the predictands used to train the model, have now been made available via the following Zenodo repository: https://zenodo.org/records/17338349 (DOI: 10.5281/zenodo.17338349). We commit to updating the corresponding section to include this modification in the next review of the manuscript.
  Best,
  Jose
  
  Citation: https://doi.org/10.5194/egusphere-2025-3754-AC1
  - CEC2: 'Reply on AC1', Juan Antonio Añel, 13 Oct 2025
    
    Dear authors,
    Many thanks for the quick reply and for solving the pending issues regarding the data.
    We can consider now the current version of your manuscript in compliance with the policy of the journal.
    Juan A. Añel
    Geosci. Model Dev. Executive Editor
    
    Citation: https://doi.org/10.5194/egusphere-2025-3754-CEC2
RC1:
'Comment on egusphere-2025-3754', Anonymous Referee #1, 23 Oct 2025
This study presents an important exploration of using pre-trained deep learning (DL) models for climate downscaling, aiming to maintain physical consistency between large-scale predictors and localized datasets. By systematically testing multiple training strategies (pre-training, partial fine-tuning, full fine-tuning, and full training), the authors demonstrate the robustness and efficiency of applying pre-trained models on the station-based dataset. However, as the authors note in the discussion, “this benefit does not necessarily translate into improved accuracy on STATIONS-IBEB, likely due to the presence of higher and more localized extreme values, which are more challenging to model than their smoothed counterparts in the interpolated ROCIO-IBEB gridded dataset.” This observation raises a critical issue: while pre-training improves efficiency and generalization, it may limit the model’s ability to capture localized extremes that define station-based observations. Clarifying this trade-off would deepen the study’s insight into how pre-trained DL models balance physical consistency and predictive reliability in downscaling applications. The following comments aim to clarify and deepen several aspects of this discussion.
The key distinction between ROCIO-IBEB and STATION-IBEB lies in their treatment of local extremes. Since station-based datasets inherently preserve localized weather phenomena, it would be helpful to elaborate on the rationale for using STATION-IBEB as the downscaling target and to contrast its statistical characteristics—particularly the distribution tails representing extreme events—with those of ROCIO-IBEB. This clarification would highlight the physical implications of transferring knowledge between datasets with distinct spatial and statistical properties.

Although fine-tuned models converge faster and achieve comparable performance to fully-trained models in terms of RMSE and mean bias, Figure 4 (right column) suggests that fully-trained models perform slightly better for extreme metrics such as TXx and TNn. This raises an important question about the ability of pre-trained models to represent localized extremes, which are critical for reliable high-impact weather downscaling. A focused evaluation of model skill over the extreme subsets of both ROCIO-IBEB and STATION-IBEB would help determine whether performance limitations stem from the coarse representation of extremes in the pre-training data or from the fine-tuning process itself, which may not fully adapt to station-scale variability.

The aggregated saliency map results reveal differences between full-training and pre-trained models, yet it is unclear whether these reflect meaningful large-scale dependencies capable of inferring local extremes or potential overfitting to dominant features. Providing examples of regional or event-specific saliency maps, rather than only aggregated values, would clarify whether the learned features correspond to physically interpretable meteorological patterns or spurious correlations introduced during training.

In Section 4.4, fine-tuned models trained with datasets containing varying fractions of missing data show lower RMSE values, attributed to pre-learned representations improving generalization. However, such pre-training could potentially smooth out localized extremes in unseen data. Evaluating performance specifically under extreme conditions in these incomplete datasets would strengthen the interpretation of how pre-training affects robustness and physical fidelity when data coverage is limited.

Overall, this study makes a valuable contribution to understanding how pre-trained DL models can be adapted for regional climate applications. Further analysis focusing on extreme events and saliency-based interpretability would enhance confidence in the approach and clarify the trade-offs between maintaining physical consistency and capturing localized, high-impact weather phenomena.
Citation: https://doi.org/10.5194/egusphere-2025-3754-RC1
- AC3: 'Reply on RC1', José González-Abad, 22 Dec 2025
  
  The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2025/egusphere-2025-3754/egusphere-2025-3754-AC3-supplement.pdf
  
  Citation: https://doi.org/10.5194/egusphere-2025-3754-AC3
RC2:
'Comment on egusphere-2025-3754', Anonymous Referee #2, 24 Oct 2025

In this study, the authors aim to explore the benefits of transfer learning in the context of ML-based statistical downscaling for climate applications. Specifically, the goal is to understand the benefits of pre-training the latent representations of ML-based downscaling models on a core dataset, to achieve improved skill or higher consistency when the downscaling model is fine-tuned for a different downscaling task.
This topic is certainly of practical and scientific interest for climate modelers, and a good fit for the journal. However, I find the implementation of this study to be largely uninformative regarding the questions posed by the authors in the abstract and introduction. In my view, this is due to the choices made by the authors for the pre-training and fine-tuning datasets, as well as the forms of fine-tuning explored for the DeepESD model. For these reasons, which I detail below, I find this manuscript in its current form unsuitable for publication. I encourage the authors to find a more effective lens through which the questions that they set out to study can be answered.
Major comments:
- Transfer learning is a well-established way to fine-tune large ML models, pre-trained on an extensive pre-existing dataset, on a smaller dataset that is more representative of some final task. The manuscript instead explores pre-training roughly 17,000 parameters of the final DeepESD models, out of a total of ~4.4M parameters for temperature and ~7.5M parameters for precipitation, respectively. This can hardly be called fine-tuning, when the pre-trained parameters represent less than 1% of the total number of parameters in both cases. This partly explains why all the variants yield statistically equivalent results (Figures 4-7), and why no conclusions can be drawn from this experimental setup.
- The final target dataset of interest, STATIONS-IBEB, is used to construct the pre-training dataset, ROCIO-IBEB. This setup omits the most important practical aspect of transfer learning: can we train on one dataset to improve predictive skill on a different dataset? This is an important question for some of the applications cited by the authors: pre-train on a national-level dataset, and fine-tune on a local dataset with different statistics (Taboada et al, 2024). The current setup is too idealized and not representative of the situations where transfer learning may actually be useful, in my opinion.
- Figure 3 shows that pre-training does not yield improved results, only faster training for a relatively inexpensive model where training cost is not really an issue. The rest of the results also fail to show any positive effects of pre-training. I think this is because at the level of the high-level representations of the data learned by the convolutional layers, the datasets ROCIO-IBEB and STATIONS-IBEB are largely indistinguishable (since they share the same data sources). This leads me to believe that the improved skill of fine-tuned models when the STATIONS-IBEB dataset is artificially shrunk (Fig 9) is due to the fact that you are actually showing a very similar version of the omitted samples to the model through ROCIO-IBEB. This is another example of why useful conclusions cannot be drawn given the similarity of the two datasets considered.
Minor comments:
- L38-40: "Diverging outcomes, which may confuse users". Improving consistency at the expense of capturing the true uncertainty of regional climate projections is actually a disservice to the downstream users, because it leads to biased estimates of risk.
- L82: Calling a 1km resolution downscaled dataset spanning thousands of years and supported by extremely sparse observations is a stretch (Karger et al, 2023).
- L123: "the most widely used in the downscaling literature". I don't think this method is that well established (60 studies reference it), so this needs to be toned down. There are studies from 2024 on downscaling with more references (e.g., CorrDiff), and deterministic ML-based downscaling models are not representative of the state of the art anymore.
- The version of DeepESD used in this paper is different than the one introduced by Bano-Medina et al (2022) for temperature. The MSE loss assumes a homogeneous uncertainty estimate, unlike the original Gaussian log-likelihood where the variance is explicitly modeled. I would also say that the deterministic MSE is no longer a "widely adopted" loss in downscaling due to its tendency to smooth fields in space and underestimate extremes.
- Fig 2: The legend reads "full-tuning" for the right column It should be full fine-tuning.
- Fig 3: Is this the training loss or the validation loss? If the former, please change to the validation loss, which is more representative of operational skill. Otherwise, please show both.
- L212: "take about half the number of epochs": Certainly not to reach the best final skill, since the fully trained model is better. How are you defining a common final time for all models to assert this?
- The results shown in Figures 4 and 5 for the ROCIO-IBEB dataset are not comparable to those in the STATIONS-IBEB dataset, since the former is a smoothed interpolation of the latter. Errors on the ROCIO-IBEB dataset will always be lower.
- FIgure 9: I think the legend should refer to different versions of STATIONS-IBEB, not ROCIO-IBEB.
- Discussion: I do not know where the study demonstrates "the potential of pre-training" (L303), or the affirmation that "fine-tuning the extractor appears to be beneficial". Beyond Figure 9, which has some issues I raised before, the other results are largely equivalent for all variants.

Citation: https://doi.org/10.5194/egusphere-2025-3754-RC2
- AC2: 'Reply on RC2', José González-Abad, 22 Dec 2025
  
  The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2025/egusphere-2025-3754/egusphere-2025-3754-AC2-supplement.pdf
  
  Citation: https://doi.org/10.5194/egusphere-2025-3754-AC2

Peer review completion

AR – Author's response | RR – Referee report | ED – Editor decision | EF – Editorial file upload

AR by José González-Abad on behalf of the Authors (03 Feb 2026) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (09 Feb 2026) by Po-Lun Ma

RR by Anonymous Referee #2 (12 Feb 2026)

Suggestions for revision or reasons for rejection

I would like to thank the authors for including results for the additional dataset, STATIONS-CAT. The revised manuscript addresses most of my initial concerns, and I am happy to recommend it for publication pending a few minor remaining points. Please note that all line numbers in the following comments refer to the manuscript with tracked changes.

1. In the discussion (L433), the authors conclude that "models with limited data availability may fit noise and learn relationships that are less aligned with known dynamics." If this were the case for STATIONS-CAT, the fully trained model would likely exhibit higher RMSEs or biases in the test set compared to the pre-trained models. It would strengthen the manuscript if the authors could support this hypothesis with a result for STATIONS-CAT similar to Figures 3, 4, or 8 (adding just one of these would suffice).

2. In Figure 6, it is currently difficult to extract meaningful information from the temperature fields because the color patterns across all models appear nearly identical. This is problematic, as Figure 6 provides crucial insight into whether pretraining helps constrain climate change signals more consistently. I recommend reducing the temperature range in the colorbars for TNn and TXx to make the differences between models more discernible. For instance, ranges of 1 to 4 °C for TNn and 4 to 8 °C for TXx seem to capture most of the variability and would likely provide much better contrast.

3. Some qualitative statements in the text would be more compelling if backed by quantitative analysis. For example, regarding the reduced extreme precipitation in northern Spain or the changes in the Duero River (L324): could relative errors or biases be provided to support these observations?

Minor comments:

L104: "as well as global high-res grids" -> This phrase should probably be deleted since the corresponding reference has been removed.

L386: I believe the reference here should be to fine-tuning on STATIONS-CAT, not STATIONS-IBEB.

L392: Should "Fig 5" be "Fig 10"?

Hide

RR by Anonymous Referee #1 (26 Feb 2026)

Suggestions for revision or reasons for rejection

While the previous version emphasized efficiency and generalization of pre-training for station downscaling, the revision shifts the focus to improving cross-dataset consistency and the robustness of learned relationships across national and regional products, rather than maximizing predictive skill for individual station datasets. The introduction of the independent STATIONS-CAT dataset and the expanded interpretability analysis reinforce this emphasis on transferable representations and coherent climate-change signals across diverse observational products. This repositioning clarifies the intended scope of the proposed framework and enhances its relevance for multi-dataset downscaling consistency.
However, this reframing raises an important concern regarding the interpretation of robustness and the practical motivation for transferring a model trained on gridded data to station observations. In statistical downscaling, adapting a model pre-trained on a spatially smoothed gridded dataset is typically motivated by the need to recover localized variability and extremes attenuated by interpolation. While the manuscript argues that fine-tuning improves cross-dataset consistency without degrading overall performance, evaluation of effectiveness for station-scale extremes remains limited, and predictive skill is explicitly de-emphasized. It therefore remains unclear whether the transfer-learning framework preserves or improves the ability to infer local extreme phenomena relative to models trained directly on station data. Clarifying the relationship between representational consistency and predictive effectiveness at the station scale would strengthen the practical relevance of the approach.
If cross-dataset consistency and transferable physical relationships are the central goals, the reliability of the pre-trained model and the physical validity of the learned relationships become critical assumptions. In this respect, the interpretation of the Aggregated Saliency Map (ASM) diagnostics appears preliminary and does not yet demonstrate that the pre-trained model captures the dominant geographic controls governing station-scale climate variability. Minimum temperature and precipitation are strongly modulated by topography, elevation, coastal proximity, and land–sea contrast, yet the ASM discussion largely focuses on large-scale atmospheric predictors without examining how such geographically mediated effects are represented when training on the smoothed ROCIO-IBEB grid. Because spatial interpolation reduces local gradients and extremes, it remains uncertain whether the pre-trained relationships adequately encode the geographic conditioning present in station observations.
This uncertainty also affects the claimed transferability of the learned relationships. The manuscript implicitly assumes that consistency of large-scale sensitivities in the gridded domain implies feasibility of transfer to station-scale variability and extremes, but this assumption is not yet substantiated by the ASM analysis. Further examination of how geographic controls are expressed in the learned relationships—particularly across regions with strong orographic or coastal gradients—would strengthen the physical basis of the proposed framework. The ASM results in Figures 5 and 9 provide useful initial indications of regime-dependent sensitivities, yet deeper analysis is needed. For example, stratifying ASM by geographic station subsets (elevation, coastal distance, geographic region) could reveal whether the learned relationships remain physically consistent across heterogeneous environments. In addition, Figure 10 shows that full training and full fine-tuning converge to distinct large-scale sensitivities despite identical architectures, suggesting different local minima in the loss landscape. It would therefore be informative to assess whether these models exhibit systematically different behavior for individual events or regional station clusters, particularly for extremes. Such differences could indicate regime-dependent overfitting to localized processes or alternative physically meaningful relationships.
Concretely, the authors could (i) compute ASM separately for stations stratified by elevation, coastal distance, or physiographic region; (ii) compare case-wise errors and saliency patterns between full training and full fine-tuned models for selected extreme events or regional subsets; and (iii) evaluate whether differences between regimes concentrate in stations or periods associated with known local extreme processes (e.g., orographic precipitation, coastal convection, cold-air pooling). These analyses would clarify whether divergence between training strategies reflects physically interpretable regime dependence or optimization variability, and would provide stronger evidence that the pre-trained relationships are both physically grounded and transferable to station-scale downscaling.
Overall, demonstrating that the pre-trained relationships remain physically interpretable across geographically stratified station subsets and that fine-tuned models preserve or improve skill for station-scale variability and extremes is essential to establish that the proposed framework supports effective station-based downscaling beyond representational consistency. Such evidence would be necessary to align the study with typical expectations for methodological contributions in statistical climate downscaling and to substantiate the current title and scope. Given that these aspects are not yet adequately addressed in the present revision, I recommend rejection of the manuscript in its current form under the stated framing. However, the proposed approach has clear potential, and I strongly encourage the authors to undertake the suggested station-scale performance and physical interpretability analyses and resubmit a more comprehensive and fully validated manuscript. Establishing this link between cross-dataset consistency and reliable inference of localized extremes would substantially strengthen both the scientific credibility and practical relevance of the work for publication in GMD.

Referee Report: PDF

Hide

ED: Reconsider after major revisions (27 Feb 2026) by Po-Lun Ma

Dear José González-Abad,

Thank you for submitting the revised manuscript. I now have received reports from two reviewers. While reviewers acknowledged the effort revising the manuscript, they still raised significant concern regarding the study. Please review their comments and address them carefully when you submit the revision. The reviewers' comments are attached below.

Regards,

Po-Lun

***
Reviewer 1:

While the previous version emphasized efficiency and generalization of pre-training for station downscaling, the revision shifts the focus to improving cross-dataset consistency and the robustness of learned relationships across national and regional products, rather than maximizing predictive skill for individual station datasets. The introduction of the independent STATIONS-CAT dataset and the expanded interpretability analysis reinforce this emphasis on transferable representations and coherent climate-change signals across diverse observational products. This repositioning clarifies the intended scope of the proposed framework and enhances its relevance for multi-dataset downscaling consistency.

However, this reframing raises an important concern regarding the interpretation of robustness and the practical motivation for transferring a model trained on gridded data to station observations. In statistical downscaling, adapting a model pre-trained on a spatially smoothed gridded dataset is typically motivated by the need to recover localized variability and extremes attenuated by interpolation. While the manuscript argues that fine-tuning improves cross-dataset consistency without degrading overall performance, evaluation of effectiveness for station-scale extremes remains limited, and predictive skill is explicitly de-emphasized. It therefore remains unclear whether the transfer-learning framework preserves or improves the ability to infer local extreme phenomena relative to models trained directly on station data. Clarifying the relationship between representational consistency and predictive effectiveness at the station scale would strengthen the practical relevance of the approach.

If cross-dataset consistency and transferable physical relationships are the central goals, the reliability of the pre-trained model and the physical validity of the learned relationships become critical assumptions. In this respect, the interpretation of the Aggregated Saliency Map (ASM) diagnostics appears preliminary and does not yet demonstrate that the pre-trained model captures the dominant geographic controls governing station-scale climate variability. Minimum temperature and precipitation are strongly modulated by topography, elevation, coastal proximity, and land–sea contrast, yet the ASM discussion largely focuses on large-scale atmospheric predictors without examining how such geographically mediated effects are represented when training on the smoothed ROCIO-IBEB grid. Because spatial interpolation reduces local gradients and extremes, it remains uncertain whether the pre-trained relationships adequately encode the geographic conditioning present in station observations.

This uncertainty also affects the claimed transferability of the learned relationships. The manuscript implicitly assumes that consistency of large-scale sensitivities in the gridded domain implies feasibility of transfer to station-scale variability and extremes, but this assumption is not yet substantiated by the ASM analysis. Further examination of how geographic controls are expressed in the learned relationships—particularly across regions with strong orographic or coastal gradients—would strengthen the physical basis of the proposed framework. The ASM results in Figures 5 and 9 provide useful initial indications of regime-dependent sensitivities, yet deeper analysis is needed. For example, stratifying ASM by geographic station subsets (elevation, coastal distance, geographic region) could reveal whether the learned relationships remain physically consistent across heterogeneous environments. In addition, Figure 10 shows that full training and full fine-tuning converge to distinct large-scale sensitivities despite identical architectures, suggesting different local minima in the loss landscape. It would therefore be informative to assess whether these models exhibit systematically different behavior for individual events or regional station clusters, particularly for extremes. Such differences could indicate regime-dependent overfitting to localized processes or alternative physically meaningful relationships.

Concretely, the authors could (i) compute ASM separately for stations stratified by elevation, coastal distance, or physiographic region; (ii) compare case-wise errors and saliency patterns between full training and full fine-tuned models for selected extreme events or regional subsets; and (iii) evaluate whether differences between regimes concentrate in stations or periods associated with known local extreme processes (e.g., orographic precipitation, coastal convection, cold-air pooling). These analyses would clarify whether divergence between training strategies reflects physically interpretable regime dependence or optimization variability, and would provide stronger evidence that the pre-trained relationships are both physically grounded and transferable to station-scale downscaling.

Overall, demonstrating that the pre-trained relationships remain physically interpretable across geographically stratified station subsets and that fine-tuned models preserve or improve skill for station-scale variability and extremes is essential to establish that the proposed framework supports effective station-based downscaling beyond representational consistency. Such evidence would be necessary to align the study with typical expectations for methodological contributions in statistical climate downscaling and to substantiate the current title and scope. Given that these aspects are not yet adequately addressed in the present revision, I recommend rejection of the manuscript in its current form under the stated framing. However, the proposed approach has clear potential, and I strongly encourage the authors to undertake the suggested station-scale performance and physical interpretability analyses and resubmit a more comprehensive and fully validated manuscript. Establishing this link between cross-dataset consistency and reliable inference of localized extremes would substantially strengthen both the scientific credibility and practical relevance of the work for publication in GMD.

Reviewer 2:

I would like to thank the authors for including results for the additional dataset, STATIONS-CAT. The revised manuscript addresses most of my initial concerns, and I am happy to recommend it for publication pending a few minor remaining points. Please note that all line numbers in the following comments refer to the manuscript with tracked changes.

1. In the discussion (L433), the authors conclude that "models with limited data availability may fit noise and learn relationships that are less aligned with known dynamics." If this were the case for STATIONS-CAT, the fully trained model would likely exhibit higher RMSEs or biases in the test set compared to the pre-trained models. It would strengthen the manuscript if the authors could support this hypothesis with a result for STATIONS-CAT similar to Figures 3, 4, or 8 (adding just one of these would suffice).

2. In Figure 6, it is currently difficult to extract meaningful information from the temperature fields because the color patterns across all models appear nearly identical. This is problematic, as Figure 6 provides crucial insight into whether pretraining helps constrain climate change signals more consistently. I recommend reducing the temperature range in the colorbars for TNn and TXx to make the differences between models more discernible. For instance, ranges of 1 to 4 °C for TNn and 4 to 8 °C for TXx seem to capture most of the variability and would likely provide much better contrast.

3. Some qualitative statements in the text would be more compelling if backed by quantitative analysis. For example, regarding the reduced extreme precipitation in northern Spain or the changes in the Duero River (L324): could relative errors or biases be provided to support these observations?

Minor comments:

L104: "as well as global high-res grids" -> This phrase should probably be deleted since the corresponding reference has been removed.

L386: I believe the reference here should be to fine-tuning on STATIONS-CAT, not STATIONS-IBEB.

L392: Should "Fig 5" be "Fig 10"?

Hide

AR by José González-Abad on behalf of the Authors (24 Apr 2026) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (04 May 2026) by Po-Lun Ma

RR by Anonymous Referee #1 (16 May 2026)

Suggestions for revision or reasons for rejection

The revised manuscript has been substantially improved, and I would like to thank the authors for carefully addressing my previous comments. The revision more convincingly justifies the effectiveness of the fine-tuned models by examining saliency maps in greater detail and assessing their physical consistency, rather than relying solely on predictive performance metrics and aggregated saliency. The added figures and associated physical interpretations greatly strengthen the manuscript. In particular, Figures 12 and 13 provide more convincing evidence that the saliency-map patterns are physically meaningful and that the pre-trained models retain robust and transferable relationships after fine-tuning. Figure 9 helps demonstrate that fine-tuning maintains station-scale performance for precipitation extremes.
As the authors note in their response, these analyses “approach the practical limits of current XAI techniques: saliency-based methods quantify input sensitivity but do not provide causal explanations.” I therefore believe that examining saliency maps for specific cases, and assessing their alignment with the conceptual model of the relevant local weather processes, is a necessary and valuable step toward improving the explainability of AI-based downscaling models. The revised manuscript demonstrates this approach well and provides stronger support for the robustness of the pre-trained models in transferring learned relationships from a gridded dataset to station-based datasets.
My only remaining suggestion is that the authors add a brief discussion in the Introduction, perhaps around Line 77, on the importance of physical alignment between saliency patterns and the synoptic factors controlling local extreme weather. This would further highlight the authors’ effort to push saliency-based XAI beyond purely diagnostic sensitivity analysis and toward physically interpretable evaluation of learned relationships. It would also help clarify why the added case-wise and physically grounded saliency analyses are critical to the manuscript’s contribution.
Overall, the revised manuscript addresses my initial concerns, and I am happy to recommend it for publication after this final minor revision is considered.

Referee Report: PDF

Hide

ED: Publish subject to minor revisions (review by editor) (21 May 2026) by Po-Lun Ma

AR by José González-Abad on behalf of the Authors (03 Jun 2026) Author's response Author's tracked changes Manuscript

ED: Publish as is (03 Jun 2026) by Po-Lun Ma

AR by José González-Abad on behalf of the Authors (05 Jun 2026)

Interactive discussion

Status: closed

CEC1:
'Comment on egusphere-2025-3754 - No compliance with the policy of the journal', Juan Antonio Añel, 11 Oct 2025

Dear authors,
Unfortunately, after checking your manuscript, it has come to our attention that it does not comply with our "Code and Data Policy".

https://www.geoscientific-model-development.net/policies/code_and_data_policy.html
To access part of the data used in your work, you have linked web pages from AEMET. However, these are not suitable repositories for scientific publication, and we can not accept them. Therefore, the current situation with your manuscript is irregular. Please, publish your data in one of the appropriate repositories and reply to this comment with the relevant information (link and a permanent identifier for it (e.g. DOI)) as soon as possible, as we can not accept manuscripts in Discussions that do not comply with our policy.
Also, you must include a modified 'Code and Data Availability' section in a potentially reviewed manuscript, containing the information of the new repositories.
I must note that if you do not fix this problem, we cannot accept your manuscript for publication in our journal.
Juan A. Añel

Geosci. Model Dev. Executive Editor

Citation: https://doi.org/10.5194/egusphere-2025-3754-CEC1
- AC1:
  'Reply on CEC1', José González-Abad, 13 Oct 2025
  
  Dear Juan A. Añel,
  We apologize for the oversight. All data previously linked through the AEMET URL, corresponding to the predictands used to train the model, have now been made available via the following Zenodo repository: https://zenodo.org/records/17338349 (DOI: 10.5281/zenodo.17338349). We commit to updating the corresponding section to include this modification in the next review of the manuscript.
  Best,
  Jose
  
  Citation: https://doi.org/10.5194/egusphere-2025-3754-AC1
  - CEC2: 'Reply on AC1', Juan Antonio Añel, 13 Oct 2025
    
    Dear authors,
    Many thanks for the quick reply and for solving the pending issues regarding the data.
    We can consider now the current version of your manuscript in compliance with the policy of the journal.
    Juan A. Añel
    Geosci. Model Dev. Executive Editor
    
    Citation: https://doi.org/10.5194/egusphere-2025-3754-CEC2
RC1:
'Comment on egusphere-2025-3754', Anonymous Referee #1, 23 Oct 2025
This study presents an important exploration of using pre-trained deep learning (DL) models for climate downscaling, aiming to maintain physical consistency between large-scale predictors and localized datasets. By systematically testing multiple training strategies (pre-training, partial fine-tuning, full fine-tuning, and full training), the authors demonstrate the robustness and efficiency of applying pre-trained models on the station-based dataset. However, as the authors note in the discussion, “this benefit does not necessarily translate into improved accuracy on STATIONS-IBEB, likely due to the presence of higher and more localized extreme values, which are more challenging to model than their smoothed counterparts in the interpolated ROCIO-IBEB gridded dataset.” This observation raises a critical issue: while pre-training improves efficiency and generalization, it may limit the model’s ability to capture localized extremes that define station-based observations. Clarifying this trade-off would deepen the study’s insight into how pre-trained DL models balance physical consistency and predictive reliability in downscaling applications. The following comments aim to clarify and deepen several aspects of this discussion.
The key distinction between ROCIO-IBEB and STATION-IBEB lies in their treatment of local extremes. Since station-based datasets inherently preserve localized weather phenomena, it would be helpful to elaborate on the rationale for using STATION-IBEB as the downscaling target and to contrast its statistical characteristics—particularly the distribution tails representing extreme events—with those of ROCIO-IBEB. This clarification would highlight the physical implications of transferring knowledge between datasets with distinct spatial and statistical properties.

Although fine-tuned models converge faster and achieve comparable performance to fully-trained models in terms of RMSE and mean bias, Figure 4 (right column) suggests that fully-trained models perform slightly better for extreme metrics such as TXx and TNn. This raises an important question about the ability of pre-trained models to represent localized extremes, which are critical for reliable high-impact weather downscaling. A focused evaluation of model skill over the extreme subsets of both ROCIO-IBEB and STATION-IBEB would help determine whether performance limitations stem from the coarse representation of extremes in the pre-training data or from the fine-tuning process itself, which may not fully adapt to station-scale variability.

The aggregated saliency map results reveal differences between full-training and pre-trained models, yet it is unclear whether these reflect meaningful large-scale dependencies capable of inferring local extremes or potential overfitting to dominant features. Providing examples of regional or event-specific saliency maps, rather than only aggregated values, would clarify whether the learned features correspond to physically interpretable meteorological patterns or spurious correlations introduced during training.

In Section 4.4, fine-tuned models trained with datasets containing varying fractions of missing data show lower RMSE values, attributed to pre-learned representations improving generalization. However, such pre-training could potentially smooth out localized extremes in unseen data. Evaluating performance specifically under extreme conditions in these incomplete datasets would strengthen the interpretation of how pre-training affects robustness and physical fidelity when data coverage is limited.

Overall, this study makes a valuable contribution to understanding how pre-trained DL models can be adapted for regional climate applications. Further analysis focusing on extreme events and saliency-based interpretability would enhance confidence in the approach and clarify the trade-offs between maintaining physical consistency and capturing localized, high-impact weather phenomena.
Citation: https://doi.org/10.5194/egusphere-2025-3754-RC1
- AC3: 'Reply on RC1', José González-Abad, 22 Dec 2025
  
  The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2025/egusphere-2025-3754/egusphere-2025-3754-AC3-supplement.pdf
  
  Citation: https://doi.org/10.5194/egusphere-2025-3754-AC3
RC2:
'Comment on egusphere-2025-3754', Anonymous Referee #2, 24 Oct 2025

In this study, the authors aim to explore the benefits of transfer learning in the context of ML-based statistical downscaling for climate applications. Specifically, the goal is to understand the benefits of pre-training the latent representations of ML-based downscaling models on a core dataset, to achieve improved skill or higher consistency when the downscaling model is fine-tuned for a different downscaling task.
This topic is certainly of practical and scientific interest for climate modelers, and a good fit for the journal. However, I find the implementation of this study to be largely uninformative regarding the questions posed by the authors in the abstract and introduction. In my view, this is due to the choices made by the authors for the pre-training and fine-tuning datasets, as well as the forms of fine-tuning explored for the DeepESD model. For these reasons, which I detail below, I find this manuscript in its current form unsuitable for publication. I encourage the authors to find a more effective lens through which the questions that they set out to study can be answered.
Major comments:
- Transfer learning is a well-established way to fine-tune large ML models, pre-trained on an extensive pre-existing dataset, on a smaller dataset that is more representative of some final task. The manuscript instead explores pre-training roughly 17,000 parameters of the final DeepESD models, out of a total of ~4.4M parameters for temperature and ~7.5M parameters for precipitation, respectively. This can hardly be called fine-tuning, when the pre-trained parameters represent less than 1% of the total number of parameters in both cases. This partly explains why all the variants yield statistically equivalent results (Figures 4-7), and why no conclusions can be drawn from this experimental setup.
- The final target dataset of interest, STATIONS-IBEB, is used to construct the pre-training dataset, ROCIO-IBEB. This setup omits the most important practical aspect of transfer learning: can we train on one dataset to improve predictive skill on a different dataset? This is an important question for some of the applications cited by the authors: pre-train on a national-level dataset, and fine-tune on a local dataset with different statistics (Taboada et al, 2024). The current setup is too idealized and not representative of the situations where transfer learning may actually be useful, in my opinion.
- Figure 3 shows that pre-training does not yield improved results, only faster training for a relatively inexpensive model where training cost is not really an issue. The rest of the results also fail to show any positive effects of pre-training. I think this is because at the level of the high-level representations of the data learned by the convolutional layers, the datasets ROCIO-IBEB and STATIONS-IBEB are largely indistinguishable (since they share the same data sources). This leads me to believe that the improved skill of fine-tuned models when the STATIONS-IBEB dataset is artificially shrunk (Fig 9) is due to the fact that you are actually showing a very similar version of the omitted samples to the model through ROCIO-IBEB. This is another example of why useful conclusions cannot be drawn given the similarity of the two datasets considered.
Minor comments:
- L38-40: "Diverging outcomes, which may confuse users". Improving consistency at the expense of capturing the true uncertainty of regional climate projections is actually a disservice to the downstream users, because it leads to biased estimates of risk.
- L82: Calling a 1km resolution downscaled dataset spanning thousands of years and supported by extremely sparse observations is a stretch (Karger et al, 2023).
- L123: "the most widely used in the downscaling literature". I don't think this method is that well established (60 studies reference it), so this needs to be toned down. There are studies from 2024 on downscaling with more references (e.g., CorrDiff), and deterministic ML-based downscaling models are not representative of the state of the art anymore.
- The version of DeepESD used in this paper is different than the one introduced by Bano-Medina et al (2022) for temperature. The MSE loss assumes a homogeneous uncertainty estimate, unlike the original Gaussian log-likelihood where the variance is explicitly modeled. I would also say that the deterministic MSE is no longer a "widely adopted" loss in downscaling due to its tendency to smooth fields in space and underestimate extremes.
- Fig 2: The legend reads "full-tuning" for the right column It should be full fine-tuning.
- Fig 3: Is this the training loss or the validation loss? If the former, please change to the validation loss, which is more representative of operational skill. Otherwise, please show both.
- L212: "take about half the number of epochs": Certainly not to reach the best final skill, since the fully trained model is better. How are you defining a common final time for all models to assert this?
- The results shown in Figures 4 and 5 for the ROCIO-IBEB dataset are not comparable to those in the STATIONS-IBEB dataset, since the former is a smoothed interpolation of the latter. Errors on the ROCIO-IBEB dataset will always be lower.
- FIgure 9: I think the legend should refer to different versions of STATIONS-IBEB, not ROCIO-IBEB.
- Discussion: I do not know where the study demonstrates "the potential of pre-training" (L303), or the affirmation that "fine-tuning the extractor appears to be beneficial". Beyond Figure 9, which has some issues I raised before, the other results are largely equivalent for all variants.

Citation: https://doi.org/10.5194/egusphere-2025-3754-RC2
- AC2: 'Reply on RC2', José González-Abad, 22 Dec 2025
  
  The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2025/egusphere-2025-3754/egusphere-2025-3754-AC2-supplement.pdf
  
  Citation: https://doi.org/10.5194/egusphere-2025-3754-AC2

Peer review completion

AR – Author's response | RR – Referee report | ED – Editor decision | EF – Editorial file upload

AR by José González-Abad on behalf of the Authors (03 Feb 2026) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (09 Feb 2026) by Po-Lun Ma

RR by Anonymous Referee #2 (12 Feb 2026)

Suggestions for revision or reasons for rejection

Hide

RR by Anonymous Referee #1 (26 Feb 2026)

Suggestions for revision or reasons for rejection

Referee Report: PDF

Hide

ED: Reconsider after major revisions (27 Feb 2026) by Po-Lun Ma

Hide

AR by José González-Abad on behalf of the Authors (24 Apr 2026) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (04 May 2026) by Po-Lun Ma

RR by Anonymous Referee #1 (16 May 2026)

Suggestions for revision or reasons for rejection

Referee Report: PDF

Hide

ED: Publish subject to minor revisions (review by editor) (21 May 2026) by Po-Lun Ma

AR by José González-Abad on behalf of the Authors (03 Jun 2026) Author's response Author's tracked changes Manuscript

ED: Publish as is (03 Jun 2026) by Po-Lun Ma

AR by José González-Abad on behalf of the Authors (05 Jun 2026)

Journal article(s) based on this preprint

01 Jul 2026

Pre-training for deep statistical climate downscaling: enhancing consistency and robustness across regional datasets

Jose González-Abad, Maialen Iturbide, Alfonso Hernanz, and José Manuel Gutiérrez

Geosci. Model Dev., 19, 5781–5804, https://doi.org/10.5194/gmd-19-5781-2026,https://doi.org/10.5194/gmd-19-5781-2026, 2026

Short summary

Jose González-Abad, Maialen Iturbide, Alfonso Hernanz, and José Manuel Gutiérrez

Viewed

Total article views: 8,275 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
6,263	1,805	207	8,275	159	171

HTML: 6,263
PDF: 1,805
XML: 207
Total: 8,275
BibTeX: 159
EndNote: 171

Views and downloads (calculated since 18 Sep 2025)

Month	HTML	PDF	XML	Total
Sep 2025	4,162	70	35	4,267
Oct 2025	815	325	80	1,220
Nov 2025	246	270	25	541
Dec 2025	198	245	40	483
Jan 2026	289	302	10	601
Feb 2026	150	205	5	360
Mar 2026	265	220	8	493
Apr 2026	63	40	1	104
May 2026	48	93	1	142
Jun 2026	4	8	0	12
Jul 2026	15	23	1	39
Aug 2026	8	4	1	13

Cumulative views and downloads (calculated since 18 Sep 2025)

Month	HTML	PDF	XML	Total
Sep 2025	4,162	70	35	4,267
Oct 2025	815	325	80	1,220
Nov 2025	246	270	25	541
Dec 2025	198	245	40	483
Jan 2026	289	302	10	601
Feb 2026	150	205	5	360
Mar 2026	265	220	8	493
Apr 2026	63	40	1	104
May 2026	48	93	1	142
Jun 2026	4	8	0	12
Jul 2026	15	23	1	39
Aug 2026	8	4	1	13

Viewed (geographical distribution)

Total article views: 8,210 (including HTML, PDF, and XML) Thereof 8,210 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 15 Aug 2026

Download

The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.

Preprint (4913 KB)
Metadata XML

Short summary

We explore how deep learning can improve local climate projections by adapting a national model to regional data. By relying on a paradigm called pre-training, we showed that models can learn faster, generalize better, and produce more consistent results, even when data is limited. This helps make future climate projections more reliable and supports better planning at both national and local levels.


Total:	0
HTML:	0
PDF:	0
XML:	0