acceptodds
Under review as a conference paper at ICLR 2027

RainAtlas: A Multi-Continental Dataset for Precipitation Downscaling

Abstract

Extreme rainfall events are increasing in intensity and frequency as climate change accelerates. While kilometer-scale precipitation forecasts are critical for supporting local decision-making, the limited availability of high-resolution precipitation observations hinders their accuracy, especially in under-resourced regions. Machine learning models are widely used to downscale precipitation data to km-scale, but their application to unseen geographies presents challenges. First, processing raw high-resolution precipitation datasets across regions requires significant engineering and domain expertise. Second, generalization across regions remains difficult. To help overcome these barriers, we release RainAtlas, a large-scale, ML-ready and multi-continental dataset for precipitation downscaling. Covering three continents, RainAtlas harmonizes heterogeneous hourly km-scale observations to a common 2-km grid. Each regional partition contains around 210,000 aligned low- and high-resolution precipitation pairs, respectively from ERA5 reanalysis and direct observations. We benchmark state-of-the-art ML-based downscaling models across RainAtlas using a wide range of metrics. Our evaluation reveals substantial variance in out-of-domain generalization depending on the training regions. This underscores the need for cross-regional, multi-source km-scale evaluation, establishing RainAtlas as a well-positioned benchmark for precipitation downscaling research.

Primary area:
datasets and benchmarks
Keywords:
precipitation downscaling, geographical generalization, dataset, benchmark

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_eHSUOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3(+1)
Soundness
2
Read review
Paper summary

The paper introduces RainAtlas, a dataset and benchmark for precipitation downscaling and cross-regional generalization. It harmonizes hourly precipitation observations from the United States, Europe, and East Asia onto 2 km grids and pairs them with ERA5 and static geographical covariates. Comparisons of deterministic and generative methods show favorable performance for generative approaches across several metrics, while cross-regional generalization depends strongly on the training region.

Critical strengths
  1. The dataset has clear practical value. Pairing heterogeneous precipitation products across multiple regions with standardized input variables reduces the data preparation burden and provides a reusable foundation for cross-regional downscaling research.

  2. The comparison includes deterministic and several generative methods, with multiscale and probabilistic evaluation under both single-region and multi-region training. So, I believe that the benchmark offers broad experimental coverage.

Critical weaknesses
  1. Nearest-neighbor resampling and the removal of entire time steps based on anomalous values may affect precipitation totals, extremes, and spatial structures, but the main text lacks quantitative before-and-after comparisons. For a dataset contribution, these choices should be validated for their effects on key precipitation statistics and model comparisons.

  2. Regions differ simultaneously in observational products and training periods, so the reported generalization gaps reflect several distribution shifts. More targeted, matched comparisons are needed to distinguish the roles of climatic diversity, observational quality, and temporal coverage. Therefore, the interpretation of cross-regional performance remains under-supported.

Questions for authors
  1. Does the cross-regional advantage of MRMS-trained models persist after matching sample counts and seasonal distributions and stratifying by climate regime and observational quality?

  2. Can representative precipitation fields and error maps support this analysis?

3 comments

We thank the reviewer for the review, and for noting the "clear practical value" of the dataset and the "broad experimental coverage" of the benchmark. We address the weaknesses and answer the questions below and in the following comment.

W1. Effect of nearest-neighbour resampling and timestep removal

"Nearest-neighbor resampling and the removal of entire time steps based on anomalous values may affect precipitation totals, extremes, and spatial structures, but the main text lacks quantitative before-and-after comparisons."

Schemes such as bilinear interpolation or conservative (area-weighted) remapping average the ~4 source cells that fall into each 2 km cell, which smooths the field and lowers peak intensities. Nearest-neighbour keeps one source value per target cell unchanged, so observed intensities are never smoothed, at the cost of occasionally dropping source values, sometimes including the local maximum. We quantified the impact of each scheme on MRMS over 2022, on 500 random hours and on the most intense hour of the 100 most intense days, comparing the native 0.01° field (area-weighted statistics) with the released nearest-neighbour regridding and with bilinear and conservative remapping onto the same 2 km grid:

MRMS 2022NativeNearest-neighbourBilinearConservative
500 random hours
Wet fraction (> 0.1 mm/h)0.04770.04770.04930.0491
Mean (mm/h)0.08180.08180.08180.0818
p99.9 / p99.99 (mm/h)9.5 / 26.99.5 / 26.99.4 / 26.49.4 / 26.5
Domain max, median ratio to native10.9850.9370.948
Domain max, 5th percentile of ratio to native10.8770.7480.767
100 most intense hours (100 different days)
Wet fraction (> 0.1 mm/h)0.05040.05040.05230.0521
p99.9 / p99.99 (mm/h)11.8 / 36.211.7 / 36.211.6 / 35.611.7 / 35.8
Domain max, median ratio to native10.9740.9290.939
Domain max, 5th percentile of ratio to native10.7130.5260.539

Nearest-neighbour preserves the intensity distribution (wet fraction, mean and upper quantiles within 0.1 mm/h of the native values), whereas bilinear and conservative remapping slightly increase light rain (+3% wet fraction) and lower the tail statistics. We agree that this should be documented, and will add this table to Appendix A in a new paragraph on reprojection and resampling. The EA dataset is resampled from a 0.01° grid with the same nearest-neighbour scheme, so the same conclusions hold there.

Regarding the removal of entire timesteps from EURADCLIM, we quantified its effect on the resulting distribution. The filter leaves the wet fraction unchanged and slightly lowers the upper quantiles, with the largest effect at the 99.99th percentile (16.8 → 14.9 mm/h). Quantiles are computed over all cells of the timesteps considered, dry cells included. We thank the reviewer for this recommendation and will include this table in Appendix A.

EURADCLIM 2017–2022Wet fraction (> 0.1 mm/h)p99 (mm/h)p99.9 (mm/h)p99.99 (mm/h)
All timesteps (before filter)0.1101.96.216.8
Kept timesteps (after filter)0.1101.86.014.9

W2. Several distribution shifts at once

"Regions differ simultaneously in observational products and training periods, so the reported generalization gaps reflect several distribution shifts. More targeted, matched comparisons are needed to distinguish the roles of climatic diversity, observational quality, and temporal coverage."

On observational quality: a matched comparison would require a second product over the same region and period. Such products exist for parts of the US (NCEP Stage IV) and Europe (RADKLIM), but they are either coarser or built from the same radar measurements. RainAtlas provides a common framework in which such products could be added. We will clearly state this in the limitations, as an avenue for future work.

On temporal coverage: the validation (2021) and test (2022) years are the same for all three regions. Only the training windows differ, by 5 to 6 years, between MRMS and EURADCLIM / EA. Over that offset, the forced change in the precipitation distribution is negligible compared with internal climate variability (details in our response to Reviewer YDwa, concern D). How this variability affects model performance remains an open question, which we will state as future work.

On climatic diversity: we considered using the sampling strategy to test whether over-representing certain Köppen-Geiger classes improves geographical generalization, but decided it was outside the scope of this paper. We will explicitly state this direction as future work.

We still think the reader deserves hypotheses for the MRMS result. The two we find most plausible are that (1) MRMS is the most homogeneous product, built from the largest number of ground-based radars in a single unified network, and (2) it covers the most climatically diverse domain. The revised text will present these as hypotheses that the paper's results do not test, and as an invitation for further work on RainAtlas.

Q1. Does the MRMS advantage persist after matching?

"Does the cross-regional advantage of MRMS-trained models persist after matching sample counts and seasonal distributions and stratifying by climate regime and observational quality?"

Sample counts are matched by construction: every test set contains 30,000 crops and every training set 150,000 crops, for every region and combination of regions. We refer the reviewer to our response to W2 regarding stratification by climate regime and observational quality.

Q2. Representative fields and error maps

"Can representative precipitation fields and error maps support this analysis?"

Spatial error maps would be a valuable complement, but they cannot be derived from our current evaluation. Our test sets consist of 30,000 crops sampled across space and time with importance or random sampling, so they do not cover each domain completely. Per-grid-cell error maps would require new test sets spanning each full domain with overlapping crops over a fixed period, and inference for every model and training configuration, at a substantial additional compute cost. We therefore leave domain-wide error maps to future work.

Regarding precipitation fields, we thank the reviewer for the suggestion. In the camera-ready version, we will add visualizations of the predicted precipitation fields from all benchmarked models and training configurations for four intense precipitation events per domain, selected from the test set using an objective criterion (e.g., the highest domain-maximum intensities). This will let readers directly compare how each model reproduces extreme precipitation structures, both in-domain and on held-out regions.

We thank the reviewer again for their review. We believe we have answered each of the reviewer's concerns and questions, and we would welcome further discussion if any remain unaddressed.

Thank you for your thorough and patient responses. My concerns have been addressed, and I believe this work meets the standards for acceptance at ICLR. I increased my rating. Good luck!

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

This paper addresses two challenges in Precipitation Downscaling: the difficulty of harmonizing high-resolution precipitation observations and the limited geographical generalization of machine learning models (Problem). The authors argue that kilometer-scale precipitation estimates are important for extreme weather and disaster risk assessment, while high-resolution observations remain scarce and unevenly distributed, motivating the need for a unified dataset supporting cross-regional training and evaluation (Motivation). Furthermore, differences in topography, climatic conditions, and observational characteristics across regions make the statistical relationships between low- and high-resolution precipitation difficult to transfer, highlighting the need for a standardized multi-region dataset and benchmark (Insight). To this end, the authors introduce RainAtlas, a multi-continental dataset and benchmark for kilometer-scale precipitation downscaling and geographical generalization. RainAtlas integrates high-resolution precipitation observations from the United States, Europe, and East Asia. Through unified projections, quality control, and spatiotemporal alignment, these observations are standardized to a 2-km resolution and paired with 24-km ERA5 reanalysis data and static geographical covariates, resulting in over 1.2 million samples. In the experiments, the authors evaluate five machine learning models using multiple metrics to assess In-domain Performance and Out-of-domain Generalization. The results show that models generally perform better in-domain, while geographical generalization depends strongly on the source of the training data.

Critical strengths

The dataset has meaningful research value. By integrating high-resolution precipitation observations from the United States, Europe, and East Asia, the authors construct a unified 2-km precipitation downscaling dataset. It provides a standardized foundation for research while reducing the data processing and usage barriers associated with heterogeneous observational sources.

The benchmark provides a systematic evaluation. Using five machine learning models and multiple evaluation metrics, the authors investigate the impact of training data sources on model generalization through single-region and multi-region training experiments, providing a valuable benchmark for future research.

Critical weaknesses

There is room for further expansion of geographical coverage. RainAtlas currently covers three regions in the Northern Hemisphere: the United States, Europe, and East Asia, without including regions such as South America. Expanding its geographical coverage in future work could enable a more comprehensive evaluation of model generalization across diverse climatic conditions.

Questions for authors

Impact of extreme precipitation filtering on the data distribution. Why did the authors discard entire time steps containing precipitation values above 125 mm/h in EURADCLIM, affecting approximately 5% of all time steps?

Role of the Importance Sampling parameter. The paper uses different values of m for training and testing. However, according to the provided equations, this parameter appears to cancel out when the final sampling probabilities are normalized. Could the authors further clarify the role of m?

3 comments

We thank the reviewer for their thoughtful review of the paper, for noting the "meaningful research value" of the dataset, and that RainAtlas provides "a valuable benchmark for future research". We address the weakness raised by the reviewer and answer their questions below and in the following comments.


W1. Geographical coverage

"RainAtlas currently covers three regions in the Northern Hemisphere: the United States, Europe, and East Asia, without including regions such as South America. Expanding its geographical coverage in future work could enable a more comprehensive evaluation of model generalization across diverse climatic conditions."

This is certainly true. Unfortunately, we are limited by available data products that meet the criteria of km-scale hourly fields, estimates based on ground-based radar and rain gauges (possibly complemented by satellite data, as for MRMS and EA), public access, and multi-year coverage. Such products remain concentrated in the Northern Hemisphere, and this scarcity is precisely what motivates studying geographical generalisation. This is also why we designed our processing framework to be modular, so that new regions can be added with the same projection, processing, and train/val/test sampling.

To motivate the community to integrate new datasets with RainAtlas, we will cite two candidate datasets that we believe could bring major benefits for geographical generalization:

  1. the recently released RRBF-KMA radar–gauge merged dataset for South Korea (Ryu et al., 2026), and
  2. the emerging Australian BRAIN hourly analysis merging radar, satellite and gauges (Zhang et al., 2026).

For South America, unfortunately, no available product currently meets RainAtlas' criteria.


Q1. EURADCLIM filtering

"Why did the authors discard entire time steps containing precipitation values above 125 mm/h in EURADCLIM, affecting approximately 5% of all time steps?"

EURADCLIM authors report that residual radar artefacts remain present in the dataset, which "limit the applicability of EURADCLIM at the grid cell scale, especially for use in extreme value modeling". They provide several recommendations for dealing with these outliers, such as removing "entire composites that have severe artifacts over larger areas", or discarding grid cells with values above a certain threshold, while noting that fixing this threshold "is a bit arbitrary". We decided to follow a conservative and easy-to-implement approach (i.e. discarding entire timesteps containing values above a fixed threshold of 125 mm/h), since the proportion of affected timesteps remained reasonable (around 5%). This guarantees no artefact reaches the training targets.

On the choice of threshold: the highest hourly point-gauge totals on record in Europe are around 180 mm (184.6 mm in Spain in 2024, 181 mm in Italy in 2021), and such point values are diluted when averaged over a 4 km² cell, i.e. RainAtlas's resolution. Our 125 mm/h threshold is also in line with previous work on EURADCLIM, which capped 30-min intensities at 80 mm/h (Matthews et al., 2025) and hourly values at 150 mm/h (Overeem et al., 2024). The effect of the filter on the precipitation distribution is analysed in our response to Reviewer YDwa (C.2).


Q2. Role of the importance-sampling parameter mm

"The paper uses different values of m for training and testing. However, according to the provided equations, this parameter appears to cancel out when the final sampling probabilities are normalized. Could the authors further clarify the role of m?"

The reviewer is right, and we thank them for catching this. In Equation 1 the score is S=min⁡(1,mC∑ig(xi)),S = \min\left(1, \frac{m}{C}\sum_{i} g(x_i)\right), and since g(x)=1−exp⁡(−x/s)g(x) = 1 - \exp(-x/s) lies in [0,1)[0, 1), the unclipped score is at most mm. With m=0.1m = 0.1 (train/val) and m=0.2m = 0.2 (test), the clipping is never active, so mm cancels exactly in the normalisation and has no effect on the sampling distribution.

The parameter that actually differs between training and test sampling is the saturation constant ss: at s=1s = 1 the score is close to the proportion of wet grid cells of the crop, whereas at s=30s = 30 it is close to the mean rain intensity, so the test sets contain more heavy precipitation events.

We kept mm only for fidelity to the formulation of Ravuri et al. (2021), which is also used by NowcastNet (Zhang et al., 2023), from which the sampling scheme is taken. In the revision we will simplify Eq. 1 to S=1C∑ig(xi),S = \frac{1}{C}\sum_{i} g(x_i), drop mm, and explain the role of ss in Section 3.5. The choice of sampling strategy and its effect on model performance is deliberately left for future work.

We thank the reviewer again for their review. We believe we have answered each of the reviewer's concerns and questions, and we would welcome further discussion if any remain unaddressed.


References

  • Matthews, F., Borrelli, P., Panagos, P., and Bezak, N. (2025). Dynamic assessment of rainfall erosivity in Europe: evaluation of EURADCLIM ground-radar data. Hydrol. Earth Syst. Sci., 29, 5299–5313. https://doi.org/10.5194/hess-29-5299-2025
  • Overeem, A., Leijnse, H., van der Schrier, G., van den Besselaar, E., Garcia-Marti, I., and de Vos, L. W. (2024). Merging with crowdsourced rain gauge data improves pan-European radar precipitation estimates. Hydrol. Earth Syst. Sci., 28, 649–668. https://doi.org/10.5194/hess-28-649-2024
  • Ravuri, S., Lenc, K., Willson, M., et al. (2021). Skilful precipitation nowcasting using deep generative models of radar. Nature, 597, 672–677. https://doi.org/10.1038/s41586-021-03854-z
  • Ryu, S., Song, J. J., Lim, K. S., and Lee, G. (2026). RRBF-KMA: high-resolution radar–gauge merged precipitation dataset for South Korea (2016–2024). Earth Syst. Sci. Data, 18, 7679–7695. https://doi.org/10.5194/essd-18-7679-2026
  • Zhang, Y., Long, M., Chen, K., et al. (2023). Skilful nowcasting of extreme precipitation with NowcastNet. Nature, 619, 526–532. https://doi.org/10.1038/s41586-023-06184-4
  • Zhang, Y., Wang, Q. J., Frost, A. J., Pudashine, J., Trewin, B., Velasco-Forero, C., Su, C.-H., and Villani, V. (2026). Design and trial implementation of a continental-scale, kilometre-resolution hourly precipitation analysis for Australia using satellite, radar and gauges. Hydrol. Earth Syst. Sci., 30, 6701–6730. https://doi.org/10.5194/hess-30-6701-2026

I thank the authors for their thoughtful response and clarifications. Most of my concerns have been adequately addressed, and I will maintain my original decision.

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

The paper introduces RainAtlas, a multi-continental dataset and benchmark for precipitation downscaling. It combines high-resolution precipitation observations from the US, Europe, and East Asia with low-resolution ERA5 inputs, providing a new dataset and evaluation setup for ML downscaling models. The authors benchmark several deterministic and generative models in both in-domain and cross-domain settings.

Critical strengths
  1. The dataset is a substantial and potentially useful contribution. RainAtlas puts together multiple precipitation observation products into a common LR/HR matched dataset, which has not been done previously and is important considering the gaps in global data availability. The data processing and public release can help subsequent ML research in this area.
  2. The benchmark provides a broad and relevant evaluation of existing methods. The authors compare several recent deterministic and generative approaches across in-domain and cross-region settings, using several metrics, and present the results coherently.
Critical weaknesses
  1. Evaluation component does not go far enough in analyzing model behavior.
    • Each region uses a different high-resolution precipitation product, and the training periods also differ across regions. As such, the observed transfer performance may reflect differences in observational characteristics rather than climatic or geographic variability alone. The interpretation of MRMS's superior generalization is then not well supported, and there is no additional analysis which can try and separate by climate, terrain, or precipitation regime to provide additional scientific insight. It also means much of the language is overclaiming things which cannot be said from the results in the paper.
    • Related work such as RainShift report 99th percentile metrics to evaluate at extremes, but that or other approaches to evaluate at extremes is left out here. However, extremes are are a challenging and important part of precipitation related tasks.
  2. Some lack of clarity in preprocessing steps.
    • The normalization procedure for held-out regions is not clearly defined.
    • The authors discard entire EURADCLIM timesteps whenever any grid cell exceeds 125 mm/h, removing approximately 5% of the available timesteps. Couldn’t this eliminate genuine severe precipitation events as well as otherwise valid observations? Given the paper's emphasis on extremes, an analysis of the precipitation climatology and upper-tail statistics before and after filtering could help to establish that the resulting dataset remains representative.
  3. The justification used for the assumptions in Section 4.1 seems incorrect and overstated. Atmospheric predictability as commonly cited from Lorenz, refers to how rapidly errors grow in forecasts initialized from a particular atmospheric state. It does not mean that meteorological observations separated by 14 days become statistically independent for the purposes of an ML-based model. Intuitively, there might be long term trends and changes in precipitation patterns due to climate change, which motivate temporal splits in most weather-related work. Some empirical / statistical analysis would do a lot more to back up this huge assumption than the current claim.
Questions for authors
  1. Could the authors provide some more context for discarding so many of the EURADCLIM time steps, and how this affects the precipitation distribution, particularly for extremes?
  2. Could the authors clarify how normalization is done when evaluating on a held out region, in particular does the procedure use any HR precipitation statistics from the target region?
  3. Can the authors provide any additional evidence the assumption in section 4.1 makes sense and does not effect the soundness of the evaluation?
3 comments

We thank the reviewer for a careful and constructive review, and for recognising RainAtlas as "a substantial [...] contribution". The concerns raised fall into four groups: (A) confounded sources of distribution shift weaken claims of MRMS superiority with respect to geographical generalization, (B) evaluation at extremes is left out, (C) preprocessing clarity and impact (normalisation, EURADCLIM filtering), (D) the temporal-independence assumption of Section 4.1 and the validity of temporal train/val/test splits. We address each below and in the following comments.


A. Confounded sources of distribution shift

"Each region uses a different high-resolution precipitation product, and the training periods also differ across regions. [...] The interpretation of MRMS's superior generalization is then not well supported [...] much of the language is overclaiming things which cannot be said from the results in the paper."

We agree: the observed superiority of MRMS regarding geographical generalization cannot be fully explained from experiments and results presented in the paper. Fully disentangling observational quality from climate would require a second product over the same region and period.

On the training periods: the offset between the MRMS and the EURADCLIM / EA training windows is about five years, over which the forced change in the precipitation distribution is negligible compared with internal climate variability (details in our response to concern D). How that variability shapes the training distribution is an open question, which we will state as future work.

On stratification by climate type: we considered using the sampling strategy to test whether over-representing certain Köppen-Geiger classes improves geographical generalisation, and decided it was outside the scope of this paper. We will explicitly state this direction as interesting future work.

There are several possible explanations for the MRMS result. The two we find most plausible are (1) MRMS is the most homogeneous product, built on the largest number of ground-based radars from a single unified network, and (2) it covers the most climatically diverse domain. The revised text will present these as hypotheses arising from the observed results.


B. Evaluation at extremes

"Related work such as RainShift report 99th percentile metrics to evaluate at extremes, but that or other approaches to evaluate at extremes is left out here. However, extremes are a challenging and important part of precipitation related tasks."

We fully agree that extremes are of the utmost importance for precipitation downscaling. We argue that our evaluation does target them, although we should have said so explicitly:

  • LHD (logarithmic histogram distance, Appendix D, Eq. 15) compares log-binned intensity histograms of prediction and target, so it heavily penalises models that under- or over-estimate the right tail. This is what exposes the right-tail overestimation of SerpentFlow and the consistency model in Section 5.1, and the tails are shown directly in the intensity histograms of Figs. 8, 12 and 17.
  • FSS (fractions skill score, Eq. 13) is computed at thresholds of 1, 5, 10 and 20 mm/h (Figs. 9 and 13); 20 mm/h is well above the 7.6 mm/h at which the AMS Glossary of Meteorology classifies rain as "heavy".
  • Our importance-sampling strategy (Section 3.5) deliberately over-represents heavy precipitation, and the test sets use a larger saturation hyperparameter than the training sets (s = 30 vs. s = 1), which weights crops by rain amount rather than by wet area. As a result, every evaluation on the importance-subsampled test sets already weights model performance toward heavy and extreme precipitation. The table below quantifies this on the 2022 test sets, comparing importance subsampling with uniform random subsampling:
Test set (2022)MRMS importanceMRMS randomEURADCLIM importanceEURADCLIM randomEast Asia importanceEast Asia random
Cells > 10 mm/h (%)0.970.090.210.032.520.13
Crops with max > 20 mm/h (%)51.111.425.96.852.56.4
Crops with max > 50 mm/h (%)16.22.46.41.20.20.01
Mean crop maximum (mm/h)26.86.616.95.723.04.4
Crops with > 50% wet cells (%)19.41.348.411.424.11.5

In addition to this table, we will add the full set of Table 2 statistics for the test sets to Appendix B. Results on the randomly-subsampled test sets, i.e. under "normal" conditions, are also reported in Appendix E.1.2 and E.2 for all models and metrics, so both regimes are covered.

C.1. Normalisation for held-out regions

"The normalization procedure for held-out regions is not clearly defined." / "does the procedure use any HR precipitation statistics from the target region?"

No. When evaluating on a held-out region, the z-score statistics are those of the training region(s) of the evaluated configuration, exactly as in the in-domain setting. No statistics (LR or HR) of the target region are used at any point. We will rewrite the "Transformation and normalization" paragraph of Appendix C, whose current wording ("independently for each dataset") is ambiguous on this point.

C.2. Discarding EURADCLIM timesteps above 125 mm/h

"The authors discard entire EURADCLIM timesteps whenever any grid cell exceeds 125 mm/h, removing approximately 5% of the available timesteps. Couldn't this eliminate genuine severe precipitation events as well as otherwise valid observations?"

Yes. EURADCLIM authors report that residual radar artefacts remain present in the dataset, which "limit the applicability of EURADCLIM at the grid cell scale, especially for use in extreme value modeling". They provide several recommendations for dealing with these outliers, such as removing "entire composites that have severe artifacts over larger areas", or discarding grid cells with values above a certain threshold, while noting that fixing this threshold "is a bit arbitrary".

We decided to follow a conservative and easy-to-implement approach (i.e. discarding entire timesteps containing values above a fixed threshold of 125 mm/h), since the proportion of affected timesteps remained reasonable (around 5%). This guarantees no artefact reaches the training targets. However, it is correct to note that this removes valid precipitation events, even extreme events, happening elsewhere in the region, or even around the problematic grid cells. We agree that this choice should be better documented, and it should be highlighted to the reader that it is far from the optimal approach, which we believe should consist in comparing flagged extreme values with other observational products such as pure rain-gauge observations. This alternative remains difficult to implement due to the sparsity and access difficulty of such products over the entire EURADCLIM domain.

"Given the paper's emphasis on extremes, an analysis of the precipitation climatology and upper-tail statistics before and after filtering could help to establish that the resulting dataset remains representative."

Effect on the climatology: the applied filtering leaves the proportion of wet grid cells unchanged and slightly lowers the upper quantiles, with the largest effect at the 99.99th percentile (16.8 → 14.9 mm/h). Quantiles are computed over all cells of the timesteps considered, dry cells included. We thank the reviewer for this recommendation and include this table in Appendix A.

EURADCLIM 2017–2022Wet fraction (> 0.1 mm/h)p99 (mm/h)p99.9 (mm/h)p99.99 (mm/h)
All timesteps (before filter)0.1101.96.216.8
Kept timesteps (after filter)0.1101.86.014.9

D. Temporal-independence assumption

"Atmospheric predictability as commonly cited from Lorenz, refers to how rapidly errors grow in forecasts initialized from a particular atmospheric state. It does not mean that meteorological observations separated by 14 days become statistically independent for the purposes of an ML-based model."

The reviewer is right, and we will remove the Lorenz-based justification from Section 4.1 and from the caption of Figure 6. However, we argue that the particular temporal split applied to MRMS remains valid for two reasons:

  1. For precipitation, the forced and low-frequency signal is extremely small compared with weather-scale variability, unlike for temperature. In large-ensemble projections, internal variability dominates regional precipitation projections in the near term (Lehner et al., 2020).
  2. Our models receive no temporal context: no adjacent timesteps and no date or time metadata. There is therefore no input channel through which a model could exploit a slow mode or a trend.

Still, one could object that because hourly extremes scale with temperature at roughly 7% per °C of warming, following the Clausius-Clapeyron relation, MRMS may contain extreme events that would not have happened under colder climates similar to those observed from EURADCLIM and EA. However, the training periods are only 5–6 years apart (2017 to 2020 vs. 2023 to 2025). At the current warming rate of about 0.27 °C per decade globally (Forster et al., 2026), this corresponds to 0.15 °C of warming between the training periods, or roughly a 1% increase in extreme precipitation. That is minimal compared to internal climate variability. We update the justification in Section 4.1 and Figure 6 to reflect these new arguments.

We thank the reviewer again for their detailed comments, and especially for pointing out that our Lorenz argument was invalid, which significantly weakened the justification for our temporal split configuration. We believe we have answered each of the reviewer's concerns and questions, and we would welcome further discussion if any remain unaddressed.


References

  • American Meteorological Society (2012). Rain. Glossary of Meteorology. https://glossary.ametsoc.org/wiki/Rain
  • Lehner, F., Deser, C., Maher, N., Marotzke, J., Fischer, E. M., Brunner, L., Knutti, R., and Hawkins, E. (2020). Partitioning climate projection uncertainty with multiple large ensembles and CMIP5/6. Earth Syst. Dynam., 11, 491–508. https://doi.org/10.5194/esd-11-491-2020
  • Forster, P. M., Walsh, T., Smith, C., Lamb, W. F., Lamboll, R., Cassou, C., Hauser, M., Hausfather, Z., Lee, J.-Y., Palmer, M. D., von Schuckmann, K., Slangen, A. B. A., Szopa, S., Trewin, B., Yun, J., Gillett, N. P., Jenkins, S., Matthews, H. D., Raghavan, K., Ribes, A., Rogelj, J., Rosen, D., Zhang, X., Allen, M., Andrew, R. M., Atkinson, C., Betts, R. A., Bombelli, A., Burgess, S. N., Cheng, L., Claxton, H. E., Friedlingstein, P., Frölicher, T. L., Domingues, C. M., Gasser, T., Gregory, C. H., Hoesly, R. M., Huppmann, D., Ishii, M., Kadow, C., Karwat, A., Kennedy, J., Killick, R. E., Kovilakam, M. V. M., Krummel, P. B., Lan, X., Lamarque, J.-F., Liné, A., Martín-Míguez, B., Monselesan, D. P., Morice, C., Mühle, J., Mussak, P., Peters, G. P., Pirani, A., Pongratz, J., Rigby, M., Rohde, R., Savita, A., Seneviratne, S. I., Smith, S. J., Taha, G., Tassone, C., Thorne, P., Wells, C., Western, L. M., van der Werf, G. R., Wijffels, S. E., Zecchetto, M., Zhong, J., Zhang, X.-Y., Masson-Delmotte, V., and Zhai, P. (2026). Indicators of Global Climate Change 2025: annual update of key indicators of the state of the climate system and human influence. Earth Syst. Sci. Data, 18, 3889–3933. https://doi.org/10.5194/essd-18-3889-2026

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.