the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 06 Mar 2026
High-resolution mapping of air quality across Europe: an ensemble machine and deep learning framework integrating multi-scale spatial predictors (CHROMAP v1.0)
Abstract. This article presents a model for mapping air quality at high-resolution (called CHROMAP) based on the fusion of data from deterministic models, in-situ and satellite observations, and spatial proxies using an ensemble of ML and DL algorithms. Annual estimates of the SOMO35 indicator and the average concentrations of NO2, PM2.5, PM10, and O3 are produced and evaluated for the 2013–2023 period at a spatial resolution of 500 meters over the European domain. The methodology maintains consistency across all pollutant indicators while ensuring flexibility and transferability.
By including interpretable AI diagnostics, CHROMAP provides a quantitative assessment of the importance of the 26 features over 11 years for each air quality indicator. Integrating all types of stations into the regressions, the evaluation carried out reveals that the performance scores have been significantly improved compared to CAMS reanalyses (~10 km resolution) used for downscaling; with a reduction in RRMSE on average over the period of about -33 % for NO2, -21 % for O3, -10 % for SOMO35, -22 % for PM2.5 and -37 % for PM10, and an increase in R2 of 28 %, 34 %, 18 %, 14 % and 36 %, respectively. In addition, a sensitivity analysis carried out on the static exposure of the population shows that significant differences can be found with values at high resolution, especially for NO2, thus impacting the calculation of the health impact.
By ensuring sufficient availability of in-situ observations and concentration fields from CTMs for downscaling, this methodology could be extended to additional air quality indicators and applied at higher temporal frequency, opening new opportunities for comprehensive air quality assessment.
- Preprint
(1797 KB) - Metadata XML
-
Supplement
(1572 KB) - BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-1109', Anonymous Referee #1, 26 Mar 2026
-
CEC1: 'Comment on egusphere-2026-1109 - No compliance with the policy of the journal', Juan Antonio Añel, 28 Mar 2026
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.htmlIn your "Code and Data Availability" statement you say that the data that you use for your work is available upon request. I am sorry but we can not accept this. It is forbidden by our policy, and your manuscript should have never been accepted for Discussions given such violation of it. Our policy clearly states that all the code and data necessary to replicate a manuscript must be published openly and freely to anyone before submission.
The GMD review and publication process depends on reviewers and community commentators being able to access, during the discussion phase, the code and data on which a manuscript depends, and on ensuring the provenance of replicability of the published papers for years after their publication. Please, therefore, 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. We cannot have manuscripts under discussion that do not comply with our policy.
The 'Code and Data Availability’ section must also be modified to cite the new repository locations, and corresponding references added to the bibliography.
I must note that if you do not fix this problem, we cannot continue with the peer-review process or accept your manuscript for publication in GMD.
Juan A. Añel
Geosci. Model Dev. Executive EditorCitation: https://doi.org/10.5194/egusphere-2026-1109-CEC1 -
AC1: 'Reply on CEC1', Antoine Guion, 01 Apr 2026
Dear Juan A. Añel,
Thank you for your comment
In order to comply with your "Code and Data Policy", we plan to upload the data on which the manuscript depends into a "Zenodo" repository.
However, we are facing a technical problem regarding the upload of large datasets. We are in discussion with the Zenodo support team to find a solution as soon as possible.
In the meantime, the datasets are temporarily available for free access through the following FileZender link: https://filesender.renater.fr/?s=download&token=6234a3d2-346d-4f1b-8eaf-1e5b9c70aacd
Therefore, the "Code and data availability" section in the manuscript will be modified to:
“The CHROMAPv1.0 code is available at https://zenodo.org/records/18846210 (Guion, 2026). Gridded files of annual concentrations of NO2, PM2.5, PM10, O3, and the SOMO35 indicator, as well as their associated standard deviation, produced with CHROMAPv1.0 for the period 2013-2023 are available at xxx (Zenodo repository).”
We hope that this meets the requirements of the GMD journal.
Sincerely,
Antoine Guion
Citation: https://doi.org/10.5194/egusphere-2026-1109-AC1 -
CEC2: 'Reply on AC1', Juan Antonio Añel, 01 Apr 2026
Dear authors,
Thanks for the update. Please, post a reply to this comment when you have the data deposited in the repository. However, I must not that this does not change the situation with your submission, which currently we can not consider for publication in the journal, as the link provided does not comply with the requirements of the policy.
Juan A. Añel
Geosci. Model Dev. Executive Editor
Citation: https://doi.org/10.5194/egusphere-2026-1109-CEC2 -
AC2: 'Reply on CEC2', Antoine Guion, 03 Apr 2026
Dear Juan A. Añel,
The datasets are now published on a Zenodo repository, at the following permanent link: https://zenodo.org/records/19369200
Should this already be included in a new version of the revised manuscript at this point, or can it be added when submitting the revised manuscript in response to the reviewers' comments?
Best regards,
Antoine Guion
Citation: https://doi.org/10.5194/egusphere-2026-1109-AC2 -
CEC3: 'Reply on AC2', Juan Antonio Añel, 03 Apr 2026
Dear authors,
Thanks for addressing this issue so quickly. I have checked the repositories and we can consider now the current version of your manuscript in compliance with the code policy of the journal. Also, it is enough that you include the new version of the Code and Data Availability section containing the information on the new repository in a potentially requested new version by the Topical Editor, or if no additional reviews are necessary, during the proofs process for the manuscript.
Juan A. Añel
Geosci. Model Dev. Executive Editor
Citation: https://doi.org/10.5194/egusphere-2026-1109-CEC3
-
CEC3: 'Reply on AC2', Juan Antonio Añel, 03 Apr 2026
-
AC2: 'Reply on CEC2', Antoine Guion, 03 Apr 2026
-
CEC2: 'Reply on AC1', Juan Antonio Añel, 01 Apr 2026
-
AC1: 'Reply on CEC1', Antoine Guion, 01 Apr 2026
-
RC2: 'Comment on egusphere-2026-1109', Anonymous Referee #2, 21 Jul 2026
This manuscript describes a well-motivated and documented hybrid ML/DL downscaling framework (CHROMAP) that fuses CTM reanalysis, satellite, meteorological, and geographic features to produce annual average 500m air quality maps over Europe for 2013-2023. The workflow is clear, the evaluation is reasonably thorough, and the paper is a solid contribution to GMD. I recommend a minor to moderate revision. The methodology and results are sound and well presented, but the paper would benefit from tightening the cross-validation design, clarifying feature usage per pollutant, and reconciling a couple of internally inconsistent statistics. My detailed comments are as below.
Major comments:
- Risk of spatial autocorrelation leakage in cross-validation.
The stratified K-fold CV used to evaluate CHROMAP splits individual stations into folds, but many of the features vary smoothly in space. As a result, a held-out station in one-fold can be often surrounded by training stations located just a few km away with nearly identical feature values and similar target concentrations. The model may partly succeed through local spatial interpolation rather than genuine feature-to-concentration generalization, which likely inflates the reported metrics relative to the true model skill at locations with no nearby training data. The current design mainly tells us how well CHROMAP predicts at a station drawn from the same well-monitored regions (e.g. Western/Central Europe) as the training data, rather than how it performs in the sparsely monitored regions where the vast majority of the domain grid cells lie.
I recommend the authors report at least one additional evaluation using spatially blocked cross-validation, e.g. leave one country/region out from the training stations for one or two pollutants; to quantify how much performance degrades under genuine spatial extrapolation and to better support the objective about mapping unmonitored areas.
- Clarify the chemical features feed each pollutant-specific model.
Table 1 lists five CHEM_CTM_$poll field, and authors state in lines 304-305 that “Each model undergoes supervised learning using identical features (see Sect. 2.2) (except for NO2 and O3 that also includes “CHEM_Sat_no2” from 2019)”. It is not clear, for example, whether CHEM_CTM_no2, CHEM_CTM_pm25 etc. are used in the O3 model. Given the results shown in sect. 3.2, it seems only the matching chemical field is used. But it would be necessary to state it explicitly, since it affects both interpretability and the “consistency across indicators” claim.
- Reconcile mean and median bias for NO2.
In lines 333-334, the average relative difference in mean value is less than 1% for each species, but the average relative difference in median value is 8.3% for NO2. This may suggest a skewed distribution and deserves a sentence of explanation.
- Disentangle bias correction from resolution effects in the exposure analysis (sect. 3.4).
The reported differences in population exposure between CHROMAP and CAMS_ens are attributed to both bias correction and higher spatial resolution. In current format, these two contributions are conflated. A cleaner test would be to compare exposure computed from 1) raw CAMS_ens, 2) a bias-corrected but same 10-km resolution field (e.g. debiased using the same station network), and 3) the full CHROMAP. It would allow a clearer view of the contributions from these two parts.
- Ensemble justification/ablation.
The choices of ML models: RI, RF, XGB, and MLP with median aggregation is reasonable, but there is no per-model performance shown. Since the XGB often dominate on tabular geospatial data, it would be more transparent to show individual model performance to justify the ensemble median outperforms the best single learner.
- Station representativeness vs. 500 m grid support, especially for traffic stations.
R² at traffic stations (0.42–0.55 for NO2, lower still for O3/SOMO35) is markedly worse than at background stations. Part of this gap is presumably a scale mismatch: a traffic monitor represents a footprint of tens of meters, not 500 m. This is worth a sentence acknowledging that CHROMAP, like other LUR-type approaches, cannot fully resolve within-cell hotspot gradients without a local dispersion component (e.g., as in uEMEP), rather than implying the ML models alone are underperforming there.
Minor comments:
- Units for SOMO35. SOMO35 is described as “yearly sum of the maximum daily 8 hour running mean concentrations greater than 35 ppb” (lines 90-91), but the units used in Table 1 is mg/m3. Please clarify what T/P is used for the unit conversion.
- Table 1, GEO_RoadNet row. “Spatial resolution: Original road width (linear vector)” is confusing. I suppose the feature is a road length density, not width.
- Grid cell physical size. The 0.0064° grid size corresponds to 500m only at 45.5° Given the domain spans 30-72°N, the actual cell width varies substantially with latitude. It would worth a clarifying sentence of the equivalent km range.
- Temporal interpolation of GHS-POP/GHS-SMOD (only available for 2015 and 2020, nearest-neighbor-interpolated to fill 2013–2023): nearest-neighbor is a coarse choice for a variable that changes gradually; a linear interpolation (with extrapolation flagged for the tail years 2021–2023) seems more defensible and should at least be justified.
Citation: https://doi.org/10.5194/egusphere-2026-1109-RC2
Data sets
CHROMAPv1.0 Antoine Guion https://zenodo.org/records/18846210
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 1,102 | 925 | 99 | 2,126 | 234 | 126 | 137 |
- HTML: 1,102
- PDF: 925
- XML: 99
- Total: 2,126
- Supplement: 234
- BibTeX: 126
- EndNote: 137
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
please see attached review document