the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 28 Jul 2026
CMIP6-MedPlus dataset: climate projections for the Mediterranean region using statistical downscaling
Abstract. Global climate-model projections are often too coarse to represent the spatial variability required for regional assessments. This study presents CMIP6-MedPlus, an open-access climate projection dataset for an extended Mediterranean region, available at https://doi.org/10.5281/zenodo.17898529 (Todaro et al., 2025). It provides daily precipitation and near-surface air temperature at 0.25° spatial resolution for the period 1985–2100, based on five Coupled Model Intercomparison Project Phase 6 (CMIP6) Global Climate Models (GCMs) and two Shared Socioeconomic Pathways (SSP1-2.6 and SSP3-7.0). CMIP6-MedPlus was generated using a statistical downscaling framework that combines deep learning–based spatial refinement with bias correction. First, a Convolutional Neural Network (CNN)-based method was applied to enhance spatial resolution: ERA5 reanalysis fields were upscaled to the GCM resolution and used to train the CNN to learn the mapping between coarse- and fine-scale representations of the same variable. The trained CNN was then applied to GCM outputs to generate more spatially detailed fields. Second, quantile delta mapping was applied to the CNN-refined fields to correct systematic biases relative to ERA5 while preserving climate change signals. The resulting dataset provides spatially consistent regional climate information suitable for climate-impact across Mediterranean and neighbouring areas and serves as a basis for subsequent refinement at basin and local scales.
- Preprint
(4208 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 27 Sep 2026)
-
CC1: 'Comment on essd-2026-337', Rasmus Benestad, 04 Aug 2026
reply
-
AC1: 'Reply on CC1', Valeria Todaro, 21 Aug 2026
reply
Dear Dr Benestad,
Thank you for initiating this discussion and for pointing us to additional perspectives within the empirical-statistical downscaling (ESD) literature. The ESD literature is extensive and encompasses a broad range of methodological traditions, classifications, and applications. Our Introduction was therefore necessarily selective and intended to provide the methodological background needed to introduce CMIP6-MedPlus, rather than a comprehensive review of the ESD field. Nevertheless, we agree that some formulations are too restrictive and that the broader methodological landscape can be represented more clearly.
In particular, the distinction among Perfect Prognosis (PP), Model Output Statistics (MOS), and Weather Generators should not be presented as exhaustive or mutually exclusive. The CORDEX ESD white paper describes a wider range of approaches, including PP, MOS, hybrid PP–MOS approaches, bias adjustment, and weather generators, which may be applied individually or in combination (Gutiérrez et al., 2022). We will reflect this broader perspective in the revised manuscript.
The distinction between “downscaling weather” and “downscaling climate” also provides a useful perspective. In the terminology used by Benestad et al. (2025) and Gutiérrez et al. (2022), the former refers broadly to downscaling sequences of atmospheric states, whereas the latter refers to directly downscaling statistical properties such as parameters of probability distributions. Using this terminology, CMIP6-MedPlus is primarily a time-series or “weather-downscaling” product because its main output consists of continuous daily gridded fields, although QDM operates on their statistical distributions. Directly targeting distributional parameters can be advantageous when probabilities, return values, or extremes are the quantities of primary interest, whereas daily time-series products are particularly useful for applications in which temporal sequencing and persistence are relevant, for example hydrological and other process-based applications. We therefore regard the two strategies as complementary and application-dependent. In the case of CMIP6-MedPlus, the objective is to provide a daily dataset that can support a broad range of downstream applications rather than to target a specific set of future statistical indicators.
We also agree that the concept of minimum skillful scale is relevant when interpreting climate-model information. As discussed by Takayabu et al. (2016), nominal grid spacing does not necessarily correspond to the smallest spatial scale at which a GCM or RCM provides skillful information. Accordingly, the 0.25° target grid of CMIP6-MedPlus should not be interpreted as implying that the driving GCMs contain dynamically resolved information at 0.25°. Rather, it represents the target grid on which statistically informed fine-scale spatial variability is reconstructed from the coarse-scale information provided by the driving model. The procedure does not create dynamically resolved GCM information below the scales at which the driving model is skillful.
Regarding AI/ML, we agree that non-stationarity, out-of-distribution generalization, representation of extremes, dependence on training data, and evaluation under future climatic conditions are important concerns (Rampal et al., 2024; Benestad, 2026). However, we do not think that these concerns support a general conclusion that AI/ML is unsuitable for downscaling climate projections. We interpret the cited critical review primarily as an argument for caution, appropriate benchmarking, and purpose-specific evaluation; it also discusses potentially useful applications of AI/ML to downscaling (Benestad, 2026). Other studies continue to investigate deep-learning approaches and their behaviour under future climatic conditions (Baño-Medina et al., 2021; Soares et al., 2024; Wang and Tian, 2024). We therefore consider this an active methodological question rather than a settled one.
In our framework, the CNN is part of the downscaling procedure but is used specifically as a spatial-refinement operator rather than being trained to infer the future climate-change response. It learns a coarse-to-fine transformation from paired coarse- and fine-resolution ERA5 fields and is then used to remap the GCM simulations onto the reference grid, replacing a conventional coarse-to-fine interpolation step. The large-scale climate evolution remains inherited from the driving GCM. This remapping is performed before grid-cell-wise bias adjustment, allowing modeled and reference distributions to be compared at corresponding spatial locations. QDM is then applied separately to adjust systematic distributional biases while retaining the modeled quantile changes. This does not eliminate the issue of future transferability, and historical performance alone cannot demonstrate that the learned relationships remain unchanged under future climate conditions.
Regarding MOS, we agree that an event-wise formulation requiring correspondence between simulated and observed weather events is not applicable to free-running coupled GCM simulations, whose internal variability is not synchronized with observations (Takayabu et al., 2016). In climate downscaling, however, the term has also been used in a broader sense to include model-specific statistical adjustment and bias-correction methods applied to GCM or RCM output (e.g. Gutiérrez et al., 2019). QDM does not require event-by-event correspondence between simulated and observed weather and is therefore more precisely described in our framework as a distribution-based bias-adjustment method.
Regarding VALUE, we distinguish between the usefulness of historical evaluation diagnostics and the ability of historical validation to demonstrate future transferability. Historical validation cannot by itself establish that relationships calibrated under present-day conditions will remain valid under changed forcing, and this limitation applies to ESD methods more generally, not only to AI/ML. At the same time, we do not think this makes VALUE-based diagnostics unsuitable for their intended purpose in our study. We use a subset of VALUE metrics to assess historical statistical properties and facilitate comparison with previous studies (Maraun et al., 2015; Gutiérrez et al., 2019), not as evidence that the method is guaranteed to remain valid under future climatic conditions.
Finally, we do not share the characterization of bias adjustment as a “red flag” in the present framework. As outlined above, the CNN and QDM are deliberately designed to address different aspects of the problem. The CNN learns spatial refinement rather than the model-specific climatological biases of individual GCMs. The persistence of such biases after CNN refinement is consequently expected and is not, by itself, evidence of a failure of the CNN. QDM subsequently addresses systematic differences in local marginal distributions, and this separation is reflected in our evaluation: the CNN is assessed independently for its spatial-refinement capability, while QDM is largely responsible for the improvement in historical distribution-based bias diagnostics. Similar multi-stage approaches have also been explored in recent climate-downscaling studies (e.g. Wang and Tian, 2024). Bias adjustment certainly introduces assumptions and uncertainties, particularly concerning the evolution of model biases under future climate conditions, and these should be explicitly recognized and evaluated. However, the CORDEX ESD framework itself (Gutierrez et al., 2022) identifies bias adjustment as an active area of research with numerous applications to GCM and RCM output.
Thank you again for raising these points and for contributing to the discussion. We believe that the suggested references and conceptual distinctions will help us clarify the terminology, better position CMIP6-MedPlus within the broader ESD literature, and more explicitly communicate the scope and limitations of the framework.
Best regards,
The authors
References
Baño-Medina, J., Manzanas, R., and Gutiérrez, J. M.: On the suitability of deep convolutional neural networks for continental-wide downscaling of climate change projections, Clim. Dyn., 57, 2941–2951, https://doi.org/10.1007/s00382-021-05847-0, 2021.
Benestad, R. E.: Artificial intelligence and downscaling global climate model future projections, https://doi.org/10.48550/arXiv.2601.00629, 2026.
Benestad, R. E., Parding, K. M., and Dobler, A.: Downscaling the probability of heavy rainfall over the Nordic countries, Hydrol. Earth Syst. Sci., 29, 45–65, https://doi.org/10.5194/hess-29-45-2025, 2025.
Gutiérrez, J. M., Cavazos, T., Evans, J., Nikulin, G., Somot, S., Maraun, D., Benestad, R. E., Hewitson, B., and Bettolli, M. L.: The Future Scientific Challenges for CORDEX: Empirical Statistical Downscaling (ESD), CORDEX White Paper, https://cordex.org/wp-content/uploads/2022/08/White-Paper-ESD.pdf, 2022.
Gutiérrez, J. M., Maraun, D., Widmann, M., Huth, R., et al.: An intercomparison of a large ensemble of statistical downscaling methods over Europe: Results from the VALUE perfect predictor cross-validation experiment, Int. J. Climatol., 39, 3750–3785, https://doi.org/10.1002/joc.5462, 2019.
Maraun, D., Widmann, M., Gutiérrez, J. M., Kotlarski, S., Chandler, R. E., Hertig, E., Wibig, J., Huth, R., and Wilcke, R. A. I.: VALUE: A framework to validate downscaling approaches for climate change studies, Earth's Future, 3, 1–14, https://doi.org/10.1002/2014EF000259, 2015.
Rampal, N., Hobeichi, S., Gibson, P. B., Baño-Medina, J., Abramowitz, G., Beucler, T., González-Abad, J., Chapman, W., Harder, P., and Gutiérrez, J. M.: Enhancing regional climate downscaling through advances in machine learning, Artif. Intell. Earth Syst., 3, e230066, https://doi.org/10.1175/AIES-D-23-0066.1, 2024.
Soares, P. M. M., Johannsen, F., Lima, D. C. A., Lemos, G., Bento, V. A., and Bushenkova, A.: High-resolution downscaling of CMIP6 Earth system and global climate models using deep learning for Iberia, Geosci. Model Dev., 17, 229–259, https://doi.org/10.5194/gmd-17-229-2024, 2024.
Takayabu, I., Kanamaru, H., Dairaku, K., Benestad, R., von Storch, H., and Christensen, J. H.: Reconsidering the quality and utility of downscaling, J. Meteorol. Soc. Jpn., 94A, 31–45, https://doi.org/10.2151/jmsj.2015-042, 2016.
Wang, F. and Tian, D.: Multivariate bias correction and downscaling of climate models with trend-preserving deep learning, Clim. Dyn., 62, 9651–9672, https://doi.org/10.1007/s00382-024-07406-9, 2024.
Citation: https://doi.org/10.5194/essd-2026-337-AC1
-
AC1: 'Reply on CC1', Valeria Todaro, 21 Aug 2026
reply
-
RC1: 'Comment on essd-2026-337', Anonymous Referee #1, 07 Sep 2026
reply
The manuscript presents CMIP6-MedPlus, a statistically downscaled climate projection dataset for the extended Mediterranean region. It provides daily precipitation and mean, minimum, and maximum temperature at 0.25° grid spacing for 1985–2100 from five CMIP6 GCMs under SSP1-2.6 and SSP3-7.0. The framework combines CNN-based spatial refinement with Quantile Delta Mapping (QDM).
The dataset is potentially useful for regional climate-impact applications, and the open availability of both data and code is a clear strength. The manuscript is generally well structured and the proposed framework is relevant to current developments in statistical downscaling.
In my view, the manuscript could be suitable for publication after the authors take into account the following issues, mainly concerning the added value of the CNN component and the preservation of the parent GCM climate-change signal.
Major comments
1. Added value of CNN relative to a simpler interpolation + QDM approach
The CNN is compared with nearest-neighbor interpolation during the ERA5 reconstruction experiment, but the final CNN+QDM product is not quantitatively compared with a simpler interpolation+QDM baseline. Sect. 3.2.3 provides only a qualitative comparison based on one precipitation event in Fig. 9.
It would strengthen the manuscript to include a more systematic comparison between interpolation/regridding + QDM and CNN + QDM, using quantitative spatial diagnostics in addition to visual examples. This would help demonstrate more clearly the added value of the CNN in the final product.
2. Preservation of the parent GCM climate-change signal
The manuscript states that the framework aims to preserve projected climate-change signals, and QDM is selected partly for this reason. The conservative constraint used in the CNN is relevant in this context: because it conserves the coarse-grid value for each day, it should ensure preservation of the parent GCM signal through the CNN step when considered at the coarse spatial scale.
However, this constraint applies to the CNN output before bias correction. QDM is subsequently applied independently to each 0.25° grid-cell series. Moreover, QDM is designed to preserve the projected quantile changes of the series provided as its input, which in this framework are the CNN-generated series. Therefore, the CNN conservation constraint alone does not demonstrate that the climate-change signal of the original parent GCM is preserved in the final CNN+QDM dataset.
The manuscript evaluates historical biases and presents future projections from the final downscaled product, but it does not directly compare the projected changes with those of the parent GCMs. I therefore suggest explicitly assessing climate-change signal preservation through the processing chain (parent GCM, CNN output, and final CNN+QDM product) for representative variables and statistics.
Minor comments
-
The term “regridding” may be confusing for the CNN step.
-
The manuscript would also benefit from a careful language check for minor grammatical and typographical issues.
Citation: https://doi.org/10.5194/essd-2026-337-RC1 -
Data sets
CMIP6-MedPlus: downscaled climate projections for the Mediterranean region Valeria Todaro, Daniele Secci, Marco D’Oria, and Maria Giovanna Tanda https://doi.org/10.5281/zenodo.17898529
Model code and software
CMIP6-MedPlus Valeria Todaro and Daniele Secci https://github.com/ValeriaTodaro/CMIP6-MedPlus
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 197 | 88 | 36 | 321 | 33 | 42 |
- HTML: 197
- PDF: 88
- XML: 36
- Total: 321
- BibTeX: 33
- EndNote: 42
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
It's nice to see new papers on empirical-statistical downscaling, and I hope that my comments on it can lead to a discussion that further benefits the writing process.
Judging from the introduction of the paper, it doesn't span all the knowledge within the field of downscaling, and I think that it can be improved by a broader representation of past studies and experiences. One trick is to ask some LLM if the manuscript captures the broad field of downscaling. For instance, I've spent 25 years developing methods within empirical-statistical downscaling, but see that most of my efforts have been neglected by my peers. This is unfortunate, since I think we all have lots to learn from each other, from discussions, and from past experiences. There is less progress when part of the community is being ignored, and there are also some indications that studies end up with wrong conclusions if not all relevant work is accounted for (https://doi.org/10.1007/s00704-015-1597-5).
One comment about the three main frameworks: Perfect Prognosis (PP), Model Output Statistics (MOS) and Weather Generators (WGs) - there is also a fourth: the hybrid PP-MOS, as outlined in https://cordex.org/wp-content/uploads/2022/08/White-Paper-ESD.pdf. In addition, there are two main approaches: the traditional downscaling of data points (e.g. days, months) which can be described as "downscaling weather" and parameters describing the shape of the distribution curves (e.g. pdf or probabilities) which can be described as "downscaling climate" (these concepts are described in one of the appendices of https://doi.org/10.5194/hess-29-45-2025).
I find it useful to introduce the concept of models' minimum skillful scale in order to explain why we use the large-scale information from global climate models to infer changes in local small-scale conditions, which is explained in https://doi.org/10.2151/jmsj.2015-042. The minimum skillful scale also applies to regional climate models used in dynamical downscaling.
I will argue that Deep learning, artificial intelligence and machine learning (AI/ML) is not appropriate for downscaling climate projections (https://arxiv.org/abs/2601.00629), and that downscaling the statistical parameters describing the shape of statistical curves is a better way to capture changes in the statistics as well as extremes. Most AI/ML studies only compare their results to studies involving downscaling of weather and not downscaling of climate.
I don't see that MOS can be applied to coupled GCMs unless they are initialised to make predictions, but I think that the hybrid PP-MOS (which also is simple and elegant) is suitable for coupled GCM projections.
For AI/ML, the VALUE protocol for validation no longer is as suitable, since climate change will change how variables relate to each other internally. Also, bias correction should not be necessary when it comes to empirical-statistical downscaling, should it? To me, it's a red flag.