A rainfall estimation ensemble method using a cGAN for SEVIRI geostationary satellite data

Janner, Selina; Polz, Julius; Glawion, Luca; Kunstmann, Harald; Chwala, Christian

doi:10.5194/egusphere-2026-4000

Preprints

https://doi.org/10.5194/egusphere-2026-4000

Preprints

19 Aug 2026

| 19 Aug 2026

A rainfall estimation ensemble method using a cGAN for SEVIRI geostationary satellite data

Selina Janner, Julius Polz, Luca Glawion, Harald Kunstmann, and Christian Chwala

Abstract. Geostationary satellites (GEOs) provide real-time information on the atmosphere and can be used for quantitative precipitation estimation (QPE) with low latency. The relation between GEO observations and precipitation is complex, though, limiting QPE performance.

Here we present a conditional generative adversarial network (cGAN) to perform QPE using data from SEVIRI infrared and water vapor channels. The model is trained with precipitation maps derived from the German radar network (RADKLIM-YW).

The model is evaluated in Germany for a separate time period and compared with the operational GEO QPE product PDIR-Now. Our main findings are: 1. The cGAN produces more realistic precipitation structures and intensities (including high precipitation rates) and has a higher location accuracy then the benchmark product; 2. Ensemble predictions are spatially under dispersed, uncertainty in precipitation intensity is captured better than uncertainty in precipitation location. Ensemble spread and other limitations, such as spurious amplified errors at extreme intensities, do not substantially reduce the overall skill improvements over the benchmark product. Future work should focus on extending the framework across a other geographic domains to assess generalization under varying viewing geometries and precipitation regimes.

Received: 06 Jul 2026 – Discussion started: 19 Aug 2026

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

Selina Janner, Julius Polz, Luca Glawion, Harald Kunstmann, and Christian Chwala

Status: final response (author comments only)

RC1:
'Comment on egusphere-2026-4000', Anonymous Referee #1, 07 Oct 2026
The manuscript presents a cGAN-based framework for precipitation retrieval from SEVIRI infrared and water-vapor observations, using RADKLIM-YW as the reference and PDIR-Now as the main benchmark. The topic is relevant to near-real-time satellite precipitation estimation, and the evaluation includes conventional accuracy metrics, spatial-structure diagnostics, and ensemble analyses. However, I have several concerns regarding the methodological novelty, attribution of the reported improvements, and interpretation and evaluation of the ensemble predictions. I therefore recommend major revision before the manuscript can be considered for publication.
The novelty of the proposed method needs to be demonstrated more convincingly. Deep-learning-based GEO infrared precipitation retrieval, including cGAN-based approaches, has already been explored, whereas the current evaluation mainly compares the proposed model with PDIR-Now. Since PDIR-Now is an operational product rather than a state-of-the-art deep-learning baseline, this comparison alone is insufficient to establish a clear methodological advance. I suggest including at least one representative deep-learning retrieval method under comparable conditions and better positioning the study relative to recent AI-based and probabilistic GEO-QPE approaches.

The specific contribution of adversarial training should be isolated. The generator contains several architectural components, so the current results cannot demonstrate that the reported improvements are specifically due to the cGAN formulation. A controlled ablation using the same generator, inputs, training data, and preprocessing but without the adversarial loss would provide direct evidence of whether adversarial training improves spatial structure, precipitation intensity, and overall retrieval skill.

The manuscript should distinguish more clearly between the cGAN formulation and the source of ensemble stochasticity. The ensemble is generated through Monte Carlo dropout during inference rather than through an explicit stochastic latent variable in the GAN. The resulting spread should therefore be interpreted more carefully as an approximate uncertainty estimate associated with model perturbations, rather than evidence that the full conditional precipitation distribution has been learned.

If ensemble uncertainty is presented as an important contribution, its probabilistic verification should be strengthened. The current evaluation relies mainly on case studies, member ranges, FSS distributions, and transects. A small set of standard diagnostics, such as CRPS, a rank/PIT histogram or reliability diagram, and a spread-skill or interval-coverage assessment, would provide a more quantitative evaluation of calibration and underdispersion. The manuscript should also clearly state whether the long-period cGAN results are based on a fixed ensemble member, the ensemble mean/median, or another statistic.

The added value of using two consecutive SEVIRI time steps is not directly demonstrated. A simple T=1 versus T=2 comparison would help determine whether the previous image provides meaningful additional information. If such an experiment is not feasible, the claims that the model demonstrably exploits cloud-development information and achieves temporal consistency should be made more cautious.

Several important training details should be reported more clearly for reproducibility, particularly the numerical weights of the adversarial and MSE losses, dropout rate, optimizer, learning rate, number of training epochs, and other key optimization settings.

The interpretation of extreme precipitation should be more cautious. The cGAN improves the representation of moderate-to-high precipitation, but the rarest tail appears strongly amplified relative to the reference. Larger generated maxima should therefore not automatically be interpreted as improved extreme-precipitation estimation.

The treatment of FSS should be clarified. Since FSS_uniform depends on precipitation coverage and therefore varies with the precipitation threshold, the use of a fixed FSS > 0.5 criterion for all thresholds needs further justification, or the corresponding threshold-dependent reference levels should be shown.

Several presentation and terminology issues should also be addressed. April 2021 is more appropriately described as a validation or model-selection period rather than a test period; the rationale for using 200 ensemble members should be briefly explained; Figure 4(a) would benefit from a direct model-color legend; and the manuscript requires a careful language check because a noticeable number of grammatical and spelling errors remain.
Citation: https://doi.org/10.5194/egusphere-2026-4000-RC1
RC2:
'Comment on egusphere-2026-4000', Anonymous Referee #2, 15 Oct 2026
General assessment
The manuscript addresses a relevant problem in satellite-based quantitative precipitation estimation and presents encouraging improvements over an established operational retrieval. The use of high-resolution radar observations and the exploration of ensemble generation are valuable aspects of the study. However, several aspects require clarification to establish the study’s contribution and support the interpretation of its results. These concern the comparative experimental design, temporal aggregation, spatial and high-intensity precipitation verification, and ensemble uncertainty evaluation.
Major Comments1. Scientific contribution and discussion
Hayatbini et al. (2019) apparently demonstrated cGAN-based precipitation retrieval from geostationary imagery. Please distinguish the methodological advances and new findings here from extensions involving SEVIRI, verification, and ensemble generation. The manuscript also needs a more thorough discussion of the literature: many results are presented without comparison with previous studies or the state of the art. Example, Section 4.2, are these properties new findings of this study or established features of the architecture? Such comparisons would help contextualize the findings and clarify the study’s contribution.
2. Dataset representativeness (Section 2.1.1)
The training and validation periods appear short for this type of model, particularly for potential operational use. Please justify whether they provide sufficient data and adequately represent relevant precipitation regimes and intensities, drawing comparisons with similar AI-based retrieval studies. Clarify whether computational constraints limited the dataset size and whether longer training or validation periods were tested without improving performance. The use of a single month for model selection also requires discussion of its representativeness.
3. Model configuration and selection (Sections 2.2.1–2.2.2)
Please report the final model configuration, the hyperparameters and ranges tested, and its performance relative to alternative configurations, clarifying whether the differences were substantial. Ablation and sensitivity experiments would help establish the added value of the WV channel and whether two consecutive images provide sufficient temporal context to capture relevant precipitation evolution. This information would strengthen reproducibility and the justification of the design choices.
The suitability of mean FSS for checkpoint selection also requires justification. Its use in the cited spatiotemporal downscaling study does not necessarily establish its suitability as the sole selection criterion for retrieval, where rainfall-amount accuracy is also essential. High mean FSS may conceal poor intense-rainfall performance, particularly given the modest quantitative skill reported. Please clarify whether alternative selection metrics were considered and compare their effects on model performance with those of FSS-based selection.
4. Baseline comparison (Section 2.1.4)
Please justify using PDIR-Now as the sole baseline and clarify whether established products such as ERR/SCaMPR or Hydro-Estimator were considered, discussing their availability, coverage, suitability, and any access limitations. Operational benchmarks assess comparative performance but do not isolate the contribution of adversarial training. Attributing improvements specifically to this component requires a controlled comparison with an otherwise equivalent model trained without the adversarial loss.
5. Intensity-dependent and absolute performance (Sections 3.1–3.2.3 and Tables 1 and 2)
Please report categorical skill across precipitation thresholds and rainfall errors conditional on observed precipitation and stratified by intensity (e.g. 0.1, 1, 5, 10 mm/h or percentiles). Aggregate metrics may conceal poor intense-rainfall performance or be influenced by dry conditions.
Minor Comments
Section 3.2: I may have misunderstood the procedure, but the construction of hourly totals is unclear. Four 5-minute cGAN predictions sampled at 15-minute intervals appear to represent only 20 minutes of accumulation per hour. Please explain how the intervening periods were accounted for, whether radar hourly totals include all twelve 5-minute intervals, and how temporal coverage was made consistent across products.

Section 3.2.3: Higher FSS than PDIR-Now does not necessarily establish satisfactory absolute skill. At 10 mm/h, FSS exceeds 0.5 only at approximately 53 × 99 km. Please discuss this limitation and its implications for applications at finer spatial scales.

Section 3.2.4: Similar RAPSD indicates agreement in spatial variability across scales, but the power spectrum does not retain phase information. Fields can therefore have similar RAPSD despite different precipitation locations. Please acknowledge this limitation and interpret location accuracy using complementary spatial or pixelwise metrics.

Section 3.3.1: I may have misunderstood the scope of the analysis, but ensemble evaluation appears limited to two selected events. These illustrate member variability but do not establish reliable uncertainty estimates across the test period. Please explain why ensemble reliability and calibration were not evaluated against RADKLIM-YW over the full test dataset, including any limitations preventing this assessment.

Line 47, Section 2.2.3, and Table 1: Define MSE, FSS, and all other metrics and contingency-table quantities at their first occurrence in the main text, before their use. Write terms in full before introducing abbreviations and make the Table 1 caption understandable independently.

Dataset terminology, Lines 72–75: Clarify the distinction between “testing” and “evaluation.” Consider using “training” for parameter fitting, “validation” for model selection and tuning, and “test” for independent performance assessment.

RADKLIM-YW description, Section 2.1.3: Please describe the underlying radar product and its processing into precipitation accumulations, including whether PPI or CAPPI products are involved. Report grid dimensions and summarize quality control, climatological correction, and gauge adjustment with supporting references.

Line 102: Please clarify whether the interpolation method used for RADKLIM-YW is commonly adopted in satellite precipitation studies and cite previous applications.

Wet-pixel definition for Tables 1: Please clarify whether Table 1 also uses the 0.1 mm/h wet-pixel threshold. State the applicable threshold before each table rather than introducing it only later.

Figure 3: Please specify in the caption the panel letters corresponding to each metric: PCC, MAE, POD, FA Ratio, and Bias.

Lines 213-214: Please discuss whether the absence of orography-related patterns in PCC is expected or could reflect smoothing in the cGAN predictions. Support the interpretation with previous studies and distinguish plausible explanations from demonstrated effects.

Ensemble size and displayed members, Line 225, Figure 5, and Section 3.3.1: Please justify using 200 members and explain how Members 0–19 were selected for display and whether they represent the full ensemble. State whether sensitivity to ensemble size was assessed. The claim of limited spatial spread should be supported by the full ensemble or restricted to the three illustrated members (Line 263)..

Line 286-287 : Please clarify whether collocation accounts for the actual SEVIRI acquisition time over Germany and the radar accumulation interval, as these may affect the reported 15-minute delay.

Lines 360-361: Please justify describing PDIR-Now as a “physics-informed operational retrieval.” Empirically calibrated infrared–rainfall relationships do not necessarily imply explicit physical constraints.

Please verify the consistency of the bias definition, units, and sign convention in Equation C4, Tables 1-2, and the accompanying interpretation.
Citation: https://doi.org/10.5194/egusphere-2026-4000-RC2

Selina Janner, Julius Polz, Luca Glawion, Harald Kunstmann, and Christian Chwala

Model code and software

prob_cGAN_for_GEO_QPE: Initial Release v1.0.0 Selina Janner https://doi.org/10.5281/zenodo.20624016

Video supplement

Supplemental video for the paper "A rainfall estimation ensemble method using a cGAN for SEVIRI geostationary satellite data" Selina Janner https://doi.org/10.5446/73585

Selina Janner, Julius Polz, Luca Glawion, Harald Kunstmann, and Christian Chwala

Viewed

Total article views: 427 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
199	134	94	427	96	94

HTML: 199
PDF: 134
XML: 94
Total: 427
BibTeX: 96
EndNote: 94

Views and downloads (calculated since 19 Aug 2026)

Month	HTML	PDF	XML	Total
Aug 2026	20	13	19	52
Sep 2026	79	43	58	180
Oct 2026	100	78	17	195

Cumulative views and downloads (calculated since 19 Aug 2026)

Month	HTML	PDF	XML	Total
Aug 2026	20	13	19	52
Sep 2026	79	43	58	180
Oct 2026	100	78	17	195

Viewed (geographical distribution)

Total article views: 415 (including HTML, PDF, and XML) Thereof 415 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 01 Nov 2026

Short summary

To improve rainfall monitoring from geostationary satellite observations, we developed and artificial intelligence model that estimates rainfall from geostationary satellite infrared images. The model produced more realistic rainfall patterns and intensities at more accurate locations than a widely used operational product. However the model still struggles for example with overestimating rare extreme intensities. Future work should test this approach across more geographically diverse regions.


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