acceptodds
Under review as a conference paper at ICLR 2027

RainOps: Multi-Source Availability-Aware Nowcasting with Residual Probabilistic Refinement

Abstract

Precipitation nowcasting is the task of learning to predict future precipitation fields at high spatial and temporal resolution, typically over the next few hours, from heterogeneous observations. Existing approaches struggle to leverage diverse observation sources, to remain robust when sources are unavailable, and to efficiently generate ensemble forecasts. RainOps addresses these challenges with an efficient framework that separates deterministic forecasting from stochastic refinement. An availability-conditioned backbone produces a deterministic forecast from the available observations, while lightweight stochastic residual refinement generates diverse ensemble members without rerunning the backbone. A post-hoc calibration stage further refines precipitation intensities. Thorough empirical evaluation on benchmark data demonstrates that RainOps outperforms state-of-the-art baselines, in particular for challenging rare high-intensity precipitation. Without any retraining, RainOps successfully maintains high performance when auxiliary observations are missing, while exploiting additional data sources where available.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Precipitation nowcasting, probabilistic forecasting, multi-source data fusion, spatio-temporal prediction, generative model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_zkiHOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The authors target the task of precipitation nowcasting, a well-known problem in meteorology which has been addressed in multiple ways over the years by the deep-learning community. While the problem is well-known, the authors address a number of issues with the existing solutions:

  • proposal of forecast-sample the correction-calibrate pipeline
  • availability conditioning by distinguishing between no event and no record of event

The authors highlight higher performance on benchmarks than SOTA.

Critical strengths

Clarity: The paper is clearly written, outlining the architecture and the experiments in a reproducible way

Quality: A number of experiments support the claims of superior performance of the method; it is especially important to highlight that the use of availability masks for handling missing data is shown, through the ablation studies, to be beneficial for the performance

Soundness: The approach is sound, and is grounded in existing body of literature on precipitation nowcasting

Originality: the way the work treats the missing data addresses a research gap and is a valuable addition to the existing methodology.

Critical weaknesses

Concerns about originality: there is an issue that this work seems to frame the approach as the novel view of the precipitation nowcasting as forecast-refine-calibrate pipeline. However, it seems to apply to multiple works, such as DiffCast, that decompose the forecasting into deterministic forecast, stochastic refinement, with subsequent calibration at the deployment time. Therefore, the key point would be to emphasise why this is a substantial step beyond works such as DiffCast.

Significance: The authors state that "Future work could relax the restriction of ensemble corrections to variations around the backbone forecast and expand the input area to provide additional spatial context, especially at longer horizons." However, these limitations should be more thoroughly described, opening the door for future research. Not sure it is clear from this text why the relaxation of this restriction and expansion of input area is technically challenging (which I assume is implied).

Questions for authors

I think the whole set of questions would be about clarifications of concerns above. If the authors can give a persuasive answer why (1) the forecast-refine-calibrate pipeline is substantially different, and (2) address concerns on the significance, that would address current concerns.

3 comments

We thank Reviewer zkiH for their feedback and for recognizing the clarity of the paper, the soundness of the approach, and the value of our treatment of missing data. Below, we address the two questions raised.

(1) Why RainOps is substantially different from DiffCast. DiffCast and RainOps both separate a deterministic forecast from a stochastic correction, but they differ in four ways.

  • Cost: DiffCast generates each sample from noise with a 250-step diffusion chain for every 5-frame segment (1,000 network calls for one 20-frame forecast), whereas RainOps starts from the backbone forecast and needs only two steps of a small U-Net per member, producing a 10-member, 3-hour ensemble in 146 ms.
  • Objective: RainOps trains the ensemble spread directly with a proper score (fair CRPS). Without the ensemble, CRPS worsens by 24.6%.
  • Calibration: RainOps calibrates intensities with a quantile map per lead time instead of clamping them.
  • Sources: a single RainOps model handles every combination of available sources, while DiffCast assumes a fixed input set.

To isolate the refiner, we also built a DiffCast-style variant of RainOps (new Appendix G.1). It keeps all our other contributions, namely the same frozen multi-source, availability-aware backbone, the same inputs and conditioning, the quantile calibration, and the 10-member ensemble, and replaces only our flow refiner with a residual diffusion refiner of the same U-Net. Its members are clearly blurrier (LPIPS 0.30–0.35 vs. 0.245), and it needs 12–50× more steps to clearly improve CRPS.

(2) Significance and limitations. We added an extended Limitations paragraph (Section 5) that explains why these extensions are not straightforward. Ensemble members are corrections around the backbone forecast. This is exactly what makes two-step sampling possible, but it also means that members cannot move a misplaced storm (Appendix G.2). Sampling such displacements requires starting from noise with many more steps, which brings back DiffCast-like cost. At long lead times, rain entering from outside the domain cannot be predicted from the inputs, and expanding the input area is limited by SEVIR's fixed 384 km patches and increases training cost. Perturbing the backbone's latent state and training on larger domains are promising directions for future work.

In terms of my comments, the following substantial concern remains: it seems to me that the argument goes that a number of architectural decisions help outperform existing methods. An arguably better argument you may have would be to say what are the fundamental limitations of the current approaches which make it a principally different approach from the current works that separate a deterministic forecast from a stochastic correction.

Now the abstract says: "Existing approaches struggle to leverage diverse observation sources, to remain robust when sources are unavailable, and to efficiently generate ensemble forecasts. RainOps addresses these challenges with an efficient framework that separates deterministic forecasting from stochastic refinement." So, the argument goes that the framework is efficient, as opposed to proposing a step change. I would like the authors to say exactly why is it a step change, addressing the gap in the literature, and not four separate architectural improvements cited in the rebuttal.

Neither does related work contrast this work against other works that separate deterministic forecasts and stochastic corrections...

The limitations paragraph answers my question, thank you! I have also checked the responses to the other reviewers, I believe they mostly address the gap in the experimental design.

Thank you for the follow-up, and for reading the other responses as well. We're glad the limitations paragraph answered your question. We agree that the key question is what current methods that separate a deterministic forecast from a stochastic correction fundamentally cannot do. We see two limitations.

  1. The ensemble is never trained. In DiffCast, CasCast and PercpCast, the stochastic stage is a generative model trained one sample at a time, to make each sample look realistic. An ensemble only appears at inference, when the model is sampled repeatedly. Nothing in training asks the members, taken together, to cover the observed outcome. Good spread therefore depends on sampling the generative model accurately, which takes many steps per member (1,000 network calls for one DiffCast forecast). PercpCast avoids this cost by giving up the ensemble altogether. RainOps instead trains the ensemble itself. Members are generated together during training and scored jointly with a proper scoring rule (fair CRPS), which directly rewards members that cover the observed outcome without drifting away from it. Because the spread is learned rather than recovered through sampling, two refinement steps are enough. In a controlled comparison (Appendix G.1), we kept everything else fixed and replaced only this training: the per-sample refiner gives blurrier members and needs 12–50× more steps to clearly improve CRPS.

  2. Each model is tied to one set of inputs. Current methods assume that every input source is present at forecast time. If a satellite or lightning source is missing, the options are to train another model for that configuration, repeating the expensive generative training, or to feed in data the model was never trained to handle. With k sources, there are 2^k possible configurations. In RainOps, which sources are available is an input to the deterministic backbone, and the stochastic refiner does not depend on it at all. A single trained model therefore serves every configuration. In Appendix F, one RainOps model evaluated in all eight availability states improves with every added source and performs comparably to specialists trained for each state, without any retraining.

Together, these two differences give one model that produces an ensemble from whatever observations are available. A 10-member ensemble takes 146 ms, 13.5% longer than a single deterministic PercpCast forecast.

Reviewer_oRjpOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes RainOps, a multi-source precipitation nowcasting model built around a "forecast, refine, calibrate" decomposition. (i) An availability-conditioned backbone, adapted from RainPro-8 and trained with a lead-time-weighted MSE, predicts all 36 future frames in a single pass from radar, auxiliary sources with per-source availability masks, and static inputs. (ii) A per-frame rectified-flow refiner produces ensemble members with two Euler steps. It is conditioned on the detached backbone forecast, the lead time, the last radar frame and a 32-dimensional member-specific noise vector injected through FiLM, and is trained with fair CRPS, a velocity-matching loss and LPIPS. (iii) Lead-time-specific quantile maps calibrate the marginal intensity of the members and of the ensemble mean. On the 3-hour SEVIR protocol (12→36 frames at 128×128), RainOps reports the best scores on all metrics in Table 1 except LPIPS. A single model trained with source dropout comes close to a dedicated specialist when the auxiliary sources are removed (Table 2), and a 10-member ensemble takes 146 ms.

Critical strengths
  1. The problem is practically relevant. Operational nowcasting has to cope with delayed or missing observations, and explicitly distinguishing missing inputs from observed zeros is a sensible design choice.
  2. Ensemble generation is cheap: the backbone runs once, and each additional member only needs two refiner passes per lead time .
  3. The paper is clearly written and openly reports some of its own limitations.
  4. In the one reduced-availability setting tested, a single model trained over the availability lattice comes within 1.1% CSI-m of a dedicated specialist.
Critical weaknesses
  1. Unclear source of the gains: RainOps uses more input sources than most baselines, including DEM topography and latitude/longitude (L354–356), and appears to be the only calibrated model. With radar-only input, it improves over PercpCast by only 4.6% CSI-m and 1.7% CSI4-m. Based on Table 3, without calibration its CSI4-m and CSI-219would fall below HARECast. Removing the refiner even increases CSI-m and CSI-219 by 4.3% and 11.0% .

  2. No comparison with a diffusion refiner: The diffusion baselines are quoted from other papers and differ from RainOps in backbone, inputs and calibration. A controlled experiment is needed that replaces only the refiner with a diffusion model, keeping everything else and the ensemble size fixed. The efficiency advantage is argued only through step counts , and CRPS is never compared with another ensemble method.

  3. Dependence on the backbone is not analysed: The refiner only perturbs the backbone forecast, and calibration cannot move precipitation . Backbone errors therefore propagate to all members, which is consistent with the under-dispersion in App. E. No sensitivity analysis is provided.

  4. Limited missing-data evaluation: Only one of the eight availability states is evaluated, and the baselines are never evaluated with missing inputs. Time-varying or partial missingness is not considered.

  5. Evaluation and figures: Fig. 2 and App. G show no baselines. Common metrics are missing: SSIM, POD/FAR, FSS, and probabilistic scores against other ensembles. LPIPS and CRPS also serve as training losses . All results come from a single run, and 11 of the 12 baseline results are quoted from other papers.

  6. Missing recent baselines: The paper does not compare with SOTA methods (e.g., MMCast), although it cites the first four. In addition, somebaselines in Table 1 (SimVP, PastNet, EarthFarseer, Rain-F, STJointNet) are not in the reference list.

Questions for authors
  1. Could you add radar-only RainOps to Table 1 with all metrics (including HSS-m and CSI16-m), and compare the multi-source models under identical inputs?
  2. What are the results when the same quantile mapping is applied to the baselines (at least the retrained PercpCast), or when RainOps is reported without calibration?
  3. Which input sources does each multi-source baseline use? Could you run the controlled diffusion-refiner comparison described in Weakness 2?
  4. How does the probabilistic skill (CRPS, spread-skill, rank histograms, reliability at the 133/219 thresholds) compare with ensemble baselines? Why does removing the refiner increase high-threshold CSI, and which metric demonstrates the refiner's value for early warning?
  5. How sensitive is the final ensemble to the quality of the backbone forecast? Could you report all eight availability states, evaluate the baselines with missing inputs, test time-varying and partial missingness, and run the mask ablation under missing inputs?
  6. Could you add RainPro-8 as a baseline, report SSIM, POD/FAR and FSS, and give the variance over several runs? Could you add baselines to Fig. 2 and state how the examples were selected?
3 comments

We thank Reviewer oRjp for their feedback and for recognizing the practical relevance of the problem, the explicit distinction between missing inputs and observed zeros, the efficient ensemble generation, and the clarity of the paper. We ran all requested experiments and answer each question below. All changes are in blue in the revised manuscript.

Q1 - Radar-only RainOps and identical inputs (W1). Radar-only RainOps is now in Table 1 with all metrics. It outperforms all radar-only baselines on every metric except CSI₁₆-m, including HSS-m (0.3474 vs. 0.3324 for PercpCast). Under identical inputs, RainOps trained with exactly PiMMNet's inputs (VIL, IR069, IR107) improves CSI-m by 16.7% over PiMMNet (0.2778 vs. 0.2380), with better LPIPS (Appendix D). PiMMNet and MMCast also use future satellite frames as training targets, which RainOps does not.

Q2 - Calibration of baselines and of RainOps (W1). The same quantile map adds only 1.1% CSI-m to PercpCast, and RainOps still leads calibrated PercpCast by 0.028 CSI-m. Without calibration, with the checkpoint selected on raw CSI, RainOps still outperforms PercpCast by 5.2% CSI-m (Appendix D).

Q3 - Baseline inputs and controlled diffusion refiner (W2). PiMMNet, Rain-F, and STJointNet use VIL, IR069, and IR107, and HARECast uses VIL and IR107. For the controlled comparison, we kept the backbone (frozen), inputs, calibration, and ensemble size fixed, and replaced only the refiner with a residual diffusion refiner of the same U-Net (Appendix G.1). The diffusion members are clearly blurrier (LPIPS 0.30–0.35 vs. 0.245) and less accurate at the pooled scale (CSI₄-m 0.326–0.329 vs. 0.333). Diffusion needs 25–100 steps to clearly improve CRPS, which takes 2.4–8.7 h on the test set versus 38 min for RainOps. We use our own diffusion refiner because DiffCast's public code does not include training, so DiffCast cannot be retrained on our protocol.

Q4 - Probabilistic skill and the refiner (W1, W2). Appendix H now reports spread-skill, rank histograms, and reliability diagrams at VIL 133 and 219. Among ensembles, the diffusion refiner clearly lowers CRPS only with 25–100 steps (Q3). Removing the refiner increases pixel CSI because CSI favors smooth forecasts (the double penalty). The refiner's value for early warning shows in CRPS, which worsens by 22.1% without it, and in the heavy-precipitation probabilities of the ensemble. Their observed frequency increases steadily with the forecast probability, while a deterministic forecast can only give 0 or 1.

Q5 - Backbone sensitivity and missing data (W3, W4). Appendix G.2 adds a sensitivity table. Displacing the backbone forecast by 6, 12, or 24 km degrades all metrics, lowering CSI-m to 0.249, 0.222, or 0.184, as the refiner does not correct displacement. Blurring it barely changes pixel CSI and CRPS but worsens LPIPS, so the refiner only partly recovers lost detail. Appendix E reports all eight availability states with all metrics. Every additional source improves CSI-m, CSI-133, and HSS-m. With all satellite and lightning data missing, RainOps still reaches 0.2722 CSI-m, above PiMMNet and HARECast with all their inputs. As availability is set per source and sample, partial and time-varying outages mark the affected sources as missing, and RainOps falls back to the corresponding state in Appendix E. Per-pixel and per-frame masks are a direct extension. The multi-source baselines cannot be evaluated with missing inputs on this protocol, as PiMMNet and MMCast have no public code and HARECast only releases a 5→20 checkpoint.

Q6 - RainPro-8, metrics, variance, and figures (W5). RainPro-8, retrained on our pipeline, reaches 0.2635 CSI-m versus 0.2829 for RainOps. Tuning its thresholds for CSI, as in the original work, raises CSI only by forecasting heavy rain over much larger areas. When forecast frequencies are matched to the observations, RainOps outperforms both RainPro-8 variants, and it is also better in bias, SSIM, LPIPS, and CRPS (Appendix D). RainOps improves on PercpCast in SSIM (0.621 vs. 0.593), POD, FAR, and FSS, none of which are training losses (Appendix D). Three training seeds give CSI-m 0.2812 ± 0.0015, far below the margins over the baselines. Fig. 2 and Appendix J now include PercpCast, RainPro-8, and the CSI-m of each sample. All examples are randomly drawn test sequences, which the captions now state.

Missing baselines and references (W6). MMCast uses a different split and metrics and has no public code, and the same DiffCast model scores 0.2418 CSI-m in its evaluation versus 0.2259 in ours. We now state this in Section 4 and added the missing references.

The controlled diffusion comparison and the matched-input experiments address my main concerns, and I am raising my score. One question remains about the explicit missingness mask. Under the same training and test missingness patterns, how does RainOps compare with a version trained using zero-filled inputs alone? This ablation would quantify the benefit of distinguishing missing values from observed zeros and can be performed entirely within your own pipeline.

Thank you for raising your score and for the suggestion. We will add the zero-fill ablation to Table 3. It matches our learned-token variant with the token fixed at zero, and that variant already scores below the masks in all eight availability states.

Reviewer_5UtbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents RainOps, a multi-source precipitation nowcasting framework designed for missing observations and efficient ensemble forecasting. It combines an availability-aware deterministic backbone, stochastic residual refinement using rectified flow, and post-hoc intensity calibration. Experiments on SEVIR show improvements in precipitation forecasting accuracy, particularly for high-intensity events.

Critical strengths

The paper considers practical problems in precipitation nowcasting, particularly missing observation sources and the computational cost of ensemble generation.

The experiments include recent baselines, source availability analysis, ablations, and inference cost comparisons. The reported performance is competitive.

Critical weaknesses

RainOps mainly integrates existing techniques, including the RainPro-8 backbone, modality dropout, rectified-flow refinement, and quantile calibration. The contribution beyond their combination is not sufficiently clear.

Removing the refiner improves CSI-m and high-intensity CSI, suggesting that this component does not consistently improve forecasting accuracy.

The experiments mainly simulate missing auxiliary observations, while the robustness to realistic sensor failures, partial outages, and more complex availability patterns is not sufficiently demonstrated.

The generated ensembles remain under-dispersed, as shown in Appendix E. This raises concerns about uncertainty calibration despite the improvements in CRPS and other forecasting metrics.

Questions for authors

Clarify the methodological novelty beyond existing approaches.

Explain why removing the refiner improves several CSI metrics.

Discuss the limitations of the missing-source evaluation and ensemble calibration.

3 comments

We thank Reviewer 5Utb for their feedback and for recognizing the practical relevance of missing sources and ensemble cost, as well as our recent baselines, availability analysis, and ablations. Below, we answer the three questions raised.

Q1 - Methodological novelty (W1). Each design choice in RainOps is validated against its alternative. Availability masks are not modality dropout: dropout alone cannot distinguish a missing source from observed zeros. Replacing the masks with a learned missing-value token degrades all seven metrics in Table 3. The refiner produces each ensemble member with only two steps of a small U-Net, compared with 100–1,000 steps in DiffCast, PiMMNet, and MMCast, and a diffusion refiner on the same backbone gives blurrier members and needs 12–50× more steps (Appendix G.1). Quantile mapping itself is a standard tool. What RainOps adds is applying it per lead time after stochastic refinement, with separate maps for the members and the ensemble mean, since averaging compresses intensities. It adds 7.4% CSI-m to the same RainOps checkpoint (0.2633 to 0.2829), but RainOps still outperforms PercpCast by 5.2% CSI-m without it (Appendix D). Relative to RainPro-8, RainOps changes the training target and adds availability conditioning and the stochastic refiner. Retrained on our pipeline with its default thresholds, RainPro-8 scores 0.2635 CSI-m versus 0.2829 for RainOps (Appendix D). Finally, the gains do not come from additional inputs either: with exactly PiMMNet's inputs, RainOps improves CSI-m by 16.7% over PiMMNet.

Q2 - Why removing the refiner improves CSI (W2). Pixel CSI favors smooth forecasts because of the double penalty. The refiner instead targets sharp, realistic members and probabilistic skill: without it, LPIPS worsens by 49.8% and CRPS by 22.1%, and heavy-precipitation probabilities reduce to 0 or 1. We revised the text to present the refiner accordingly.

Q3 - Limitations of the missing-source evaluation and ensemble calibration (W3, W4). Appendix E now evaluates all eight availability states on all metrics. With all satellite and lightning data missing, RainOps keeps CSI-m at 0.272, well above PiMMNet and HARECast with all their inputs. Availability is one flag per source and sample, so a partial outage in space or time marks the affected source as missing, and RainOps falls back to the corresponding state. Per-pixel or per-frame masks are a direct extension. For ensemble calibration, Appendix H adds reliability diagrams. The ensemble is under-dispersed because members are corrections around the backbone forecast, which is also what makes two-step sampling possible. Its probabilities are still informative, as the observed frequency of heavy precipitation increases steadily with the forecast probability. The new Limitations paragraph (Section 5) discusses both points and directions to address them.

Thanks for explaining the drop in pixel-level CSI. The CRPS results make the benefit of the refiner clearer to me. My remaining concern is how the model handles a partial outage, since it still drops the whole source when some observations are usable. I’m keeping my score.

Thank you for the follow-up. Dropping the whole source is the conservative case, and Appendix E shows it costs at most 0.6% CSI-m for any single source, without retraining. Even with all satellite and lightning data missing, RainOps outperforms every baseline in Table 1. Keeping the usable part of a source can only narrow this small gap, and it needs no architectural change: the availability mask is already a per-frame H×WH\times W plane (Section 3.1), which we would sample per pixel or per frame during training instead of per source.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of responses and revisions

We thank the reviewers for their detailed feedback. RainOps addresses operational nowcasting from intermittently available sources, combining a single availability-conditioned model, a two-step refiner that generates a 10-member ensemble in 146 ms, and intensity calibration. The revised manuscript marks all changes in blue. We used the additional page to clarify details that support the performance of RainOps. The method and the reported RainOps results are unchanged, and the new results are in Appendices D, E, G, and H.

We here summarise our main responses:

  • Novelty: Each design choice is validated against its alternative: availability masks against a learned missing-value token, and our two-step flow refiner against a diffusion refiner.
  • Source of the gains: The gains do not come from calibration or inputs alone. With exactly PiMMNet's inputs, RainOps improves CSI-m by 16.7% over PiMMNet, and our calibration adds only 1.1% CSI-m to PercpCast. RainOps also outperforms the added RainPro-8 baseline.
  • Missing data: RainOps improves with every additional source, and even radar-only RainOps outperforms all baselines in HSS-m, CSI-m, CSI-133, and CSI-219. Under partial outages, it falls back to the state without the affected source.
  • Refiner: Removing it raises pixel CSI only because CSI favors smooth forecasts, while LPIPS worsens by 49.8% and CRPS by 22.1%. A diffusion refiner on the same backbone gives blurrier members and needs 12–50× more steps to improve CRPS.
  • Evaluation: We added SSIM, POD/FAR, FSS, reliability diagrams, three training seeds, and an extended Limitations paragraph.

In conclusion, RainOps provides accurate, efficient, and robust probabilistic nowcasting when observation sources can be missing, making it a strong fit for the applications to physical sciences track.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.