acceptodds
Under review as a conference paper at ICLR 2027

PrecMamba: State-Space Forecasting and Flow-Matching Ensembles for Extreme Precipitation

Abstract

Precipitation nowcasting provides essential short-range forecasts that serve both meteorological science and everyday public safety. Deep learning models for precipitation nowcasting are commonly trained with pixel-wise regression objectives. On sparse, heavy-tailed precipitation fields, conventional objectives tend to produce blurred predictions and suppress the heavy-rain events that matter most. We introduce PrecMamba, a two-stage precipitation forecasting framework in which the deterministic stage learns more than the centre of the intensity distribution, letting the generative stage focus on the remaining uncertainty. The first stage employs a Mamba-based deterministic backbone to model the large-scale spatio-temporal evolution of precipitation, trained with a multi-quantile pinball loss that penalises the under-prediction of high-intensity events. The second stage models the residual around the deterministic forecast with conditional flow matching, generating sharp and diverse ensemble members in a few integration steps. Experiments on SEVIR, MeteoNet, Shanghai Radar and CIKM show that PrecMamba achieves the highest critical success index on all four benchmarks, and on SEVIR it is also best at the most intense threshold.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
precipitation nowcasting, weather forecasting, extreme events, Mamba, radar nowcasting, probabilistic forecasting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_yRdLOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes PrecMamba, a two-stage precipitation nowcasting model. Stage one is a deterministic backbone: a Hilbert-curve-serialised, temporally hierarchical bidirectional Mamba encoder with untied per-lead decoders. It is trained with a multi-quantile pinball loss so that it learns more than the conditional mean of the heavy-tailed intensity distribution. Stage two is a conditional flow-matching U-Net that models the scaled residual around the median forecast, conditioned on the frozen encoder features and the deterministic forecast, and produces ensemble members in about 10 Euler steps. On SEVIR, MeteoNet, Shanghai Radar and CIKM the method reports the best CSI and SSIM, gains at the two highest SEVIR thresholds, and more than an order of magnitude faster residual sampling than DiffCast's diffusion generator. Ablations cover the loss, backbone/generator pairings, scan order, ensemble size and number of Euler steps.

Critical strengths
  1. Consistent gains on four benchmarks, especially for extreme precipitation. PrecMamba has the best CSI and SSIM on all four datasets, with the largest margins at the neighbourhood scale (CSI-pool16). On SEVIR, CSI at the highest threshold is roughly twice DiffCast's, which directly addresses the paper's stated goal of forecasting high-impact events.
  2. The motivation is simple and supported by ablations. Training the deterministic stage with a pinball loss beyond the conditional mean beats both MSE and threshold-weighted MSE by a clear margin (CSI 0.3445 vs 0.2775/0.3115; CSI-pool16 0.609 vs 0.295/0.410). This gives a useful and easily transferable insight for residual-based nowcasting models.
  3. Large practical efficiency gain for ensemble forecasting. Replacing diffusion with conditional flow matching cuts residual sampling from about 1.23 s to about 0.07 s per forecast, improves accuracy on both backbones, and reaches near-saturated quality in about 10 steps. This makes multi-member ensembles feasible for operational use.
Critical weaknesses
  1. The comparison protocol is weak and statistical robustness is limited. Most baseline numbers are copied from two different papers, and DiffCast is reported with a different best backbone on each dataset. All results come from a single stochastic sample with no seeds, variance or significance testing. Some margins are small (e.g., MeteoNet CSI 0.3941 vs 0.3892), and HSS is behind DuoCast on SEVIR and MeteoNet, so the strength of the SOTA claim depends on how reliable these comparisons are.
  2. The probabilistic claims and the mechanism are not fully validated. Ensembles are a central selling point, yet there are no probabilistic metrics such as CRPS, spread–skill ratio, rank histograms or reliability diagrams to show the ensemble is calibrated. The backbone alone also underperforms SimVP, and the paper attributes the benefit of the upper-quantile terms to indirect effects through the encoder features without an ablation that isolates this. As a result, it is unclear how much of the gain comes from the quantile supervision versus the flow generator itself.
Questions for authors

See the weaknesses part.

3 comments

Dear authors,

I have carefully read the comments from other reviewers, and I remain positive about this work at this stage. Looking forward to your response to the concerns I raised.

Best

AuthorsOpenReview ↗

Reply to Reviewer yRdL

We thank the reviewer for their positive assessment and their follow-up comment, and we are glad that they found the gains consistent and the efficiency useful for ensemble forecasting. We address both weaknesses below.

W1: Comparison protocol and statistical robustness

Baseline sources. Both sources follow the same DiffCast protocol, with the same datasets, preprocessing, splits and evaluation code. For DiffCast, we report its best backbone on each dataset, which is the most favourable choice for this baseline. We also ran the released DiffCast checkpoint on the same test windows as our model (App. A.1, Tab. 3).

Single sample. We use a single sampled member to follow the DiffCast protocol, so that our numbers are comparable with the published results. To show the variability, we report the standard deviation of CSI over 10 independent samples, next to the margin over the strongest baseline in Tab. 1:

DatasetStd of CSIMargin over strongest baseline
SEVIR0.00250.0070 (DuoCast)
MeteoNet0.00230.0049 (DuoCast)
Shanghai Radar0.00330.0239 (DuoCast)
CIKM0.00200.0087 (AlphaPre)

The margin exceeds the standard deviation on all four datasets. On MeteoNet, where the margin is smallest, it is about twice the standard deviation.

Small margins and HSS. We agree that the margin on MeteoNet is small and that DuoCast keeps a small advantage in HSS on SEVIR and MeteoNet, as stated in Sec. 4.2. We will revise the text so that the SOTA claim refers to CSI only.

W2: Probabilistic evaluation and mechanism

Probabilistic metrics. We now report CRPS and the spread–skill ratio of an 8-member ensemble (CRPS in the native units of each dataset):

DatasetCRPSSpread–skill ratio
SEVIR6.1300.575
MeteoNet0.7720.538
Shanghai Radar1.4860.509
CIKM2.8460.455

The spread–skill ratio is below 1, so the ensemble is under-dispersed. A likely reason is that the generator models only the residual around a fixed deterministic forecast, so it does not represent the uncertainty of μ itself. We will report this in the paper as a limitation. We are also computing these metrics for the released DiffCast checkpoint on SEVIR, and we will post them in a follow-up comment as soon as they are available.

Quantile supervision versus generator. With the MSE-trained backbone without the generator, we can now give the full 2 × 2 comparison on SEVIR:

ObjectiveNo generatorFlow generator
MSE0.25350.2775
Pinball0.28300.3445

The pinball objective adds 0.0295 CSI to the backbone alone, but 0.0670 to the full model. The generator adds 0.0240 to the MSE backbone, but 0.0615 to the pinball backbone. So most of the benefit of the quantile supervision appears through the generator, which is consistent with our explanation that it acts mainly through the shared encoder features.

Backbone alone below SimVP. We agree that our backbone alone is a weaker deterministic forecaster than SimVP on SEVIR, since, with the same MSE objective, it reaches 0.2535 against 0.3108. Its value appears in combination with the pinball loss and the generator, since, with the same objective and generator, SimVP reaches 0.3236 on SEVIR, against 0.3445 for PrecMamba.

Reviewer_yRdLOpenReview ↗

post-rebuttal

This paper proposes PrecMamba, a two-stage precipitation nowcasting framework that combines a Mamba backbone trained with multi-quantile supervision and a conditional flow-matching residual generator to improve heavy-precipitation forecasts and enable efficient ensemble sampling.

The strengths of this paper includes: (1) Consistent gains on four benchmarks, especially for extreme precipitation; (2) Ablations supporting the motivation of using residual-based nowcasting models; (3) Large practical efficiency gain for ensemble forecasting.

However, there exists some weaknesses:

  1. The comparison with baselines are noisy, since all baseline numbers are copied from prior papers rather than re-run under a consistent experimental setup. Moreover, some small margin gains reduce the confidence of the SOTA claim. During the rebuttal, the authors clarify that the baseline sources followed the same DiffCast protocol, and agree to narrow their SOTA claim to CSI.

  2. The probabilistic claims and the mechanism are not fully validated. During the rebuttal, the authors provide additional experiment results such as CRPS and spread-skill ratios, which address my concerns.

Overall, after carefully reading the authors' rebuttal and other reviewers' comments, I think my concerns have been well addressed. Although the work still has some limitations, I think this paper meets the standard of ICLR. Nevertheless, as I am not deeply familiar with this specific area, I leave the final decision to AC.

Reviewer_6WVYOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PrecMamba, a two-stage precipitation nowcasting framework. The first stage uses a hierarchical bidirectional Mamba encoder–decoder to model the predictable large-scale evolution of precipitation, and employs a multi-quantile pinball loss to supervise multiple precipitation quantiles so that the shared representation receives explicit training signals for the heavy-precipitation tail. The second stage models the residual uncertainty beyond the deterministic forecast using conditional flow matching, conditioned on the median prediction and encoder features. The paper evaluates PrecMamba on four radar benchmarks—SEVIR, MeteoNet, Shanghai Radar, and CIKM—against deterministic, generative, and deterministic–generative methods, and achieves the highest mean CSI on all four datasets.

Critical strengths

1. The Related Work provides a relatively comprehensive overview of the development of radar-echo extrapolation methods. The paper organizes the literature from convolutional/recurrent deterministic forecasting models, to fully generative approaches, deterministic–generative residual modeling, and more recent state-space/Mamba-based methods. The relationships between these methodological lines and the challenge of preserving intense precipitation are also presented clearly, providing a useful context for the proposed method.

2. The baseline coverage is broad, and the paper makes a serious effort to maintain consistency in the evaluation protocol. The experiments include deterministic, generative, and deterministic–generative models, and follow the data preprocessing and evaluation protocol of DiffCast. For methods with publicly available checkpoints, the authors additionally perform evaluations on the same test set to reduce discrepancies caused by different experimental settings.

Critical weaknesses

1. The independent necessity of the Mamba backbone, multi-quantile supervision, and flow-matching generator is not sufficiently disentangled. The scientific problem itself is clearly stated: in existing deterministic–generative approaches, the deterministic stage tends to suppress intense precipitation, while diffusion-based residual generation incurs high sampling costs. However, the current ablations do not sufficiently reveal how much of the final gain comes from the Mamba backbone itself. In particular, the PrecMamba deterministic backbone trained with the pinball loss achieves only 0.2830 CSI, below SimVP at 0.3108, while the existing backbone–generator cross-ablation does not evaluate another strong deterministic backbone trained with the same multi-quantile objective and coupled with the same flow generator. Therefore, the results more clearly establish the effectiveness of the combination of multi-quantile training and flow matching than the independent contribution of the Mamba architecture.

2. Model complexity and training-resource reporting are insufficient. The paper does not report parameter counts, FLOPs/MACs, peak GPU memory, or training GPU-hours. This omission is particularly relevant because the deterministic decoder uses an independently parameterized BiSSM block and regression head for every forecast lead, so its parameter count grows with the output horizon; the complete framework further requires a second-stage U-Net and sequential two-stage training.

3. Several training details required for complete reproduction are still missing in the current version. The paper states that different datasets use different numbers of epochs so that each backbone receives a similar number of optimizer steps, but the actual epochs or optimizer-step counts for individual datasets are not reported. The random seed, some optimizer configurations, and the numerical values of the dataset-specific residual scale ss are also not explicitly provided. Meanwhile, the source code and checkpoints are only promised for release after acceptance.

4. The experimental validation of the paper’s central “extreme precipitation” claim remains incomplete. A primary motivation of the multi-quantile pinball loss is to reduce the systematic suppression of intense precipitation, and extreme precipitation is central to both the title and the methodological motivation. However, the high-quantile predictions are discarded at inference, with only the median entering the residual formulation. At the same time, the main CSI/HSS results are predominantly averaged across multiple intensity thresholds and forecast leads, which may obscure differences in the heavy-precipitation tail. The paper reports per-threshold CSI on SEVIR and demonstrates strong performance at its two highest thresholds, but corresponding tail-specific results are not provided for MeteoNet, Shanghai Radar, or CIKM. Metrics that more directly characterize intense-event detection or tail-distribution quality, such as POD, FAR/precision, or quantile calibration/coverage, are also absent. Therefore, the evidence for improved extreme-precipitation forecasting is currently concentrated mainly on SEVIR rather than being demonstrated consistently across all four benchmarks.

Questions for authors

1. Role of the upper-quantile outputs in the second stage. The authors already note in the Future Work section that the second stage currently receives only the median forecast μ\mu and encoder feature cc, while the other quantile predictions are discarded at inference. Have the authors tested using the full set of quantile predictions as additional conditioning for the flow generator? If not, what was the main consideration? Such a comparison would help distinguish whether the benefit of multi-quantile supervision mainly comes from regularizing the shared encoder representation, or whether the upper-quantile predictions themselves provide useful tail information for residual generation.

2. Independent contribution of the Mamba backbone. Have the authors tried training other representative deterministic backbones with the same multi-quantile pinball objective and coupling them with the same flow-matching residual generator? If not, what was the main reason? Such a comparison would help determine how much of PrecMamba’s improvement comes from the state-space backbone rather than from the objective and generator.

3. Model complexity and training resources. Could the authors provide the parameter count, FLOPs/MACs, peak GPU memory, and approximate GPU-hours for the two-stage training of PrecMamba and, where possible, the main baselines? It would also be helpful to report the actual epochs/optimizer steps, random seeds, and residual scale ss used for each dataset to improve reproducibility.

4. Extreme-precipitation evaluation. Beyond SEVIR, are per-threshold CSI results available for MeteoNet, Shanghai Radar, and CIKM, particularly at their highest intensity thresholds? Reporting POD together with FAR/precision, or calibration/coverage results for the multi-quantile outputs, would further clarify whether the gains in mean CSI correspond to more reliable detection of intense precipitation rather than primarily to improvements in threshold-averaged performance.

5 comments
AuthorsOpenReview ↗

Reply to Reviewer 6WVY

We thank the reviewer for their detailed comments, and we are glad that they appreciated the related work and the consistency of the evaluation. We address each weakness together with the corresponding question.

W1 & Q2: Contribution of the backbone

We have now trained SimVP with the same multi-quantile pinball loss and coupled it with the same flow-matching generator on SEVIR. Together with our existing ablations, this gives the full comparison of backbone, objective and generator:

BackboneObjectiveBackbone only+ flow generator
SimVPMSE0.31080.3145
SimVPPinball0.31830.3236
PrecMambaMSE0.25350.2775
PrecMambaPinball0.28300.3445

The pinball objective also improves SimVP, from 0.3145 to 0.3236 with the generator, so the multi-quantile supervision transfers to other backbones. However, the full model with our backbone remains clearly best (0.3445), showing that the combination with the Mamba backbone performs best. We hypothesise that a CNN backbone such as SimVP benefits less from the multi-quantile supervision than the Mamba backbone, since the CNN is much weaker in modelling spatio-temporal context, but also less prone to overfitting. We will investigate the behaviour of the multi-quantile supervision and the flow generator for different backbones in more detail, and add this comparison to the paper. Nevertheless, these results, together with Tab. 2, show that the proposed multi-quantile supervision and flow-matching generator also improve other backbones.

W2 & Q3: Complexity and training resources

We will add these numbers to the paper. On SEVIR (K = 384, L_out = 20), with all models trained on 4 NVIDIA GH200 GPUs:

ModelParamsFLOPs / forecastTraining (GPU-hours)
PrecMamba backbone27.66 M1.20 T104.6
U-Net generator26.59 M76.5 G per Euler step47.2
PrecMamba (full, 10 steps)54.25 M1.97 T151.8
SimVP backbone11.08 M0.41 T63.5

Peak GPU memory at inference is 1.76 GiB, and it does not grow with the ensemble size, as members are drawn one after another. Peak memory during training is 43.9 GiB for the backbone.

The reviewer is correct that the untied decoder grows linearly with L_out. Each lead adds 1.12 M parameters, so the decoder and its heads account for 81% of the backbone parameters on SEVIR, and 743 G of its 1.20 T FLOPs. The second stage takes 31% of the total training time (47.2 of 151.8 GPU-hours), because the generator is trained once on a frozen backbone.

W3 & Q3: Training details and code

We will add these details to the appendix.

DatasetTraining samplesBackbone epochs / stepsGenerator epochs / stepsResidual scale s
SEVIR59,53050 / 186,05040 / 148,8400.111
MeteoNet5,677150 / 53,25040 / 14,2000.075
Shanghai Radar1,381300 / 26,100120 / 10,4400.147
CIKM8,00090 / 45,00040 / 20,0000.167

Steps are optimiser steps with a global batch size of 16. We use more epochs on the smaller datasets so that they still receive enough updates, and we will correct the sentence in Sec. 4.1 accordingly. Both stages use Adam (β₁ = 0.9, β₂ = 0.999, ε = 10⁻⁸, weight decay 10⁻⁶ on weights only) with a cosine schedule and no warm-up. All runs use seed 0.

The paper and the details above describe all settings needed to reproduce our results, and we will release the code, checkpoints and evaluation scripts upon acceptance.

W4, Q1 & Q4: Extreme-precipitation evaluation

Per-threshold results on all four benchmarks. We now report CSI, POD, FAR and the frequency bias FBI = (TP + FP)/(TP + FN) at each threshold, and we will move them to the main paper. On SEVIR, we also keep the comparison with DuoCast and DiffCast (App. A.1). On the other three datasets, the other methods publish no per-threshold results and have no public checkpoints, so we compare PrecMamba with variants of our own model.

SEVIR (20 leads)

VariantMetric1674133160181219
Backbone alone (pinball)CSI0.73850.58440.16940.10070.07980.0253
POD0.84700.71400.18780.11050.08750.0265
FAR0.14790.23710.36710.46770.52470.6442
FBI0.99400.93590.29680.20760.18410.0745
Full model, MSE backboneCSI0.66380.55940.21410.10310.08000.0444
POD0.76880.61300.24410.11560.08560.0473
FAR0.17070.13520.36460.51130.44840.5781
FBI0.92700.70880.38410.23640.15520.1120
PrecMamba (full)CSI0.71230.65740.21930.19140.17590.1150
POD0.80020.73840.23910.21020.19250.1278
FAR0.13360.14300.27370.31860.32950.4648
FBI0.92360.86160.32920.30850.28710.2388

MeteoNet (20 leads)

VariantMetric812161824323640
Backbone aloneCSI0.64030.55350.52060.46910.34910.12260.12050.0410
POD0.76630.67710.62660.56850.40360.12960.12170.0411
FAR0.20430.24800.24530.27150.27900.30570.07390.0422
FBI0.96320.90030.83020.78040.55980.18670.13140.0429
FBI0.96320.90030.83020.78040.55980.18670.13140.0429
Full model, MSE backboneCSI0.70810.62020.56410.46980.32070.11460.09740.0255
POD0.96180.81000.71250.57340.37330.12630.11520.0279
FAR0.27140.27420.26970.27780.30520.44710.61280.7683
FBI1.32011.11600.97560.79400.53730.22840.29750.1204
PrecMamba (full)CSI0.75200.68660.57370.53350.33690.11510.12160.0325
POD0.84870.77330.63360.58560.36350.11990.12410.0327
FAR0.13160.14040.14140.14300.17830.25770.14230.1479
FBI0.97730.89960.73790.68330.44240.16150.14470.0384

Shanghai Radar (20 leads)

VariantMetric1020303540
Backbone aloneCSI0.56900.52150.44490.33870.1915
POD0.68250.63430.56060.44090.2400
FAR0.22620.25440.31690.40640.5134
FBI0.88200.85070.82070.74270.4932
Full model, MSE backboneCSI0.58980.53400.41700.37970.1894
POD0.75120.65080.49020.45560.2139
FAR0.26700.25160.26360.30490.3769
FBI1.02480.86960.66570.65550.3433
PrecMamba (full)CSI0.63770.56590.47310.35180.2171
POD0.77150.66810.57700.43830.2715
FAR0.21380.21280.27580.35940.4801
FBI0.98130.84870.79680.68420.5222

CIKM (10 leads)

VariantMetric1020303540
Backbone alone (pinball)CSI0.54810.44750.31730.18520.1139
POD0.58180.49310.38670.21140.1212
FAR0.09550.17120.36130.40060.3449
FBI0.64320.59500.60540.35270.1850
Full model, MSE backboneCSI0.59400.48020.20990.04490.0365
POD0.60120.51270.28290.04720.0379
FAR0.01990.11650.55150.52100.4990
FBI0.61340.58030.63080.09860.0756
PrecMamba (full)CSI0.58650.47980.24680.18970.1379
POD0.68580.57080.28580.22970.1560
FAR0.19800.24940.35620.47850.4570
FBI0.85510.76050.44390.44050.2873

On SEVIR, PrecMamba gives the highest CSI at every threshold except the lowest one. Compared with the MSE variant, the pinball objective raises CSI at the highest threshold on all four benchmarks, for example from 0.0444 to 0.1150 on SEVIR, with a lower FAR (0.4648 against 0.5781). Compared with the backbone alone, the generator raises CSI at the highest threshold on SEVIR, Shanghai Radar and CIKM (0.1379 against 0.1139), but not on MeteoNet (0.0325 against 0.0410), where this threshold is rarely exceeded. For PrecMamba, the FBI stays below 1 at every threshold, so its gains do not come from over-forecasting.

Quantile calibration. The table gives the empirical coverage of each quantile channel, that is, the fraction of target values below the predicted α-quantile, which should be close to α. We compute it on pixels with observed precipitation (target > 0), because the many pixels without precipitation, where both target and prediction are zero, would inflate the coverage.

Datasetα = 0.1α = 0.5α = 0.9α = 0.95α = 0.99
SEVIR0.0830.5050.8930.9460.989
CIKM0.0840.4710.8690.9260.981
MeteoNet0.0340.2800.7690.8660.966
Shanghai Radar0.0500.3370.7790.8670.963

On SEVIR and CIKM, the coverage is within 0.04 of the nominal level for all quantiles. On MeteoNet and Shanghai Radar, it is below the nominal level, so the quantiles are too low where precipitation is observed, and the backbone still under-predicts intensity there. Part of this gap is expected from the evaluation on pixels with observed precipitation only, which excludes the pixels where the model predicts rain but none is observed.

Reviewer_5pxCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The authors present a new two-stage precipitation nowcasting framework. The first stage is a deterministic model based on a state-space Mamba model architecture, and the second stage models the residuals around the deterministic forecast with conditional flow matching.

Critical strengths
  • The method is clearly described, and the paper is well structured.
  • Splitting the modeling into two stages (although already presented in Diffcast): using a Mamba model for the first stage, which is known to perform well with long temporal sequences and has a favorable linear complexity, and flow matching only the residuals is a sound idea.
  • The method is thoroughly tested on four datasets and shows improved performance over the other 2 stage method.
Critical weaknesses

As it is written now, it is not clear what are the advantages of the PrecMamba. The authors should address:

  • The paper does not cite or compare to relevant literature. Mamba models were already proposed and evaluated in STVMamba: Precipitation Nowcasting with Spatiotemporal Prediction Model, published in Scientific Reports, 2025, and showed improved performance over other deterministic models like Eartformer, SimVP, MAU. Authors should discuss how PRECMAMBA’s Hilbert-curve tokenization, quantile loss, and flow-matching residual stage compare or contrast with STVMamba’s single-stage deterministic architecture
  • The authors claim SOTA, but they do not compare to SOTA conditional flow matching methods like Flowcast (ICLR 2026), which seems to outperform their method on all metrics.
  • Authors state that using flow matching to drastically cut down sampling steps in precipitation nowcasting is their own discovery, while FlowCast had already established this.
  • The title states that the method is tailored for extreme precipitation, but it is not evaluated as such. Performance on extreme thresholds is evaluated on SEVIR only and is not compared to the performance of other methods. This should be the main result in the main paper.
  • As HSS and CSI metrics are standard metrics to assess nowcasting models over different thresholds, along with false alarm rate (FAR), which is not reported in this paper. Demonstrating that FAR remains low (or at least competitive with baselines) would confirm that PrecMamba's behaviour at high thresholds is driven by genuine spatial and intensity accuracy rather than over-forecasting.
Questions for authors
  • The used loss function is the pinball loss over a set of quantile tiles, which should improve the modeling of extreme values. Why is only the median channel passed to the residual generator?
  • Why did the authors decide to adapt RiverMamba to nowcasting when it was developed for a completely different problem (rivers on a non-standard grid), while there are existing Mamba-based video prediction models that are more similar to the nowcasting problem? Also, the BiSSMs do not seem to be the best choice for video encoding.
  • Explain why are all datasets resampled?
  • The authors train a generative model, but then they do not evaluate it as such. PrecMamba is a generative model and should be compared to other generative models as such.
  • How is the n=10 (explicit Euler steps) selected?
  • A section about the hardware used for this experiment should be included along with training and inference times. How does this compare to other methods? Please fill in the poll. I think this is the main advantage of only modelling residuals around a deterministic forecast.
6 comments
AuthorsOpenReview ↗

Reply to Reviewer 5pxC

We thank the reviewer for their feedback, and we are glad that they found the paper clear and the residual design sound. We address each point below.

W1: STVMamba

We will cite and discuss STVMamba in the related work section. STVMamba is a single-stage deterministic model which combines a spatio-temporal selective scan with depthwise separable convolutions, and it is evaluated on the Sichuan, HKO-7 and IMERG datasets. PrecMamba differs in three ways.

  1. Training objective. STVMamba predicts one deterministic forecast. Our backbone is trained with a multi-quantile pinball loss, so it also learns the upper tail of the intensity distribution. With the same architecture and generator, this objective raises CSI on SEVIR from 0.2775 (MSE) to 0.3445 (Tab. 2, left), so the gain does not come from the state-space architecture alone.
  2. Residual generation. STVMamba has no generative stage. PrecMamba adds a flow-matching generator for the residual around the deterministic forecast, which gives an ensemble of sharp forecasts in a few Euler steps.
  3. Tokenisation. We serialise each frame with a fixed Hilbert curve, which keeps most spatial neighbours close in the scan (App. A.2). The fixed order also maps the encoder features back to the grid, so they can condition the generator.

W2: FlowCast and the SOTA claim

We will also cite and discuss FlowCast. We would like to highlight that FlowCast does not follow the standard protocol proposed by DiffCast, so the numbers are not comparable:

FlowCastStandard protocol (DiffCast)
Input frames135
Output frames12 (60 min)20 (100 min)
Resolution384 × 384128 × 128
Scored forecastmean of 8 memberssingle member
Test windowsstride 12, events after 1 June 2019DiffCast test split

The protocol has a large effect on the scores. SimVP reaches a CSI-M of 0.423 in the FlowCast paper, but 0.3108 under the standard protocol (Tab. 1). For PrecMamba, changing only two of the four settings, the horizon (12 instead of 20 leads) and the scored forecast (mean of 8 members), raises the CSI-M on SEVIR from 0.3445 to 0.4124, with the same model and test windows. The input length, the resolution and the test windows still differ, so this number is not a direct comparison with FlowCast either. We will add a comparison, but the different evaluation protocol of FlowCast compared to other works makes a direct comparison during the discussion phase infeasible.

Our SOTA claim refers to the DiffCast protocol on SEVIR, MeteoNet, Shanghai Radar and CIKM, on which FlowCast does not report results. The two methods also differ in design: FlowCast generates the whole future sequence in the latent space of a VAE, while PrecMamba generates only the residual around a deterministic forecast, directly on the radar grid. We will revise the SOTA claim.

W3: Flow matching and few-step sampling

FlowCast is certainly the first work to apply conditional flow matching to precipitation nowcasting. However, our paper does not claim few-step sampling with flow matching as a finding, but we agree that the missing citation can give this impression. We will cite FlowCast and state its contribution in the introduction and related work.

Few-step sampling is a known property of flow matching with straight-line paths (Lipman et al., 2023; Liu et al., 2023), and we use it as a design choice. Our contribution is different: a deterministic backbone trained beyond the centre of the intensity distribution, with flow matching only for the residual around its forecast. Our ablation (Tab. 2, right) also shows that the advantage of flow matching over diffusion holds in this residual setting: it improves CSI on both backbones and reduces inference time from 1.23 s to 0.070 s with SimVP, and from 0.895 s to 0.079 s with our backbone.

W4: Extreme precipitation

App. A.1 (Tab. 3) already compares PrecMamba with DuoCast and DiffCast at each SEVIR threshold. At the highest threshold (λ = 219), PrecMamba reaches a CSI of 0.1150, against 0.1074 for DuoCast and 0.0582 for DiffCast. We agree that this result belongs in the main paper, and we will move it there.

We also extend the per-threshold evaluation to all four benchmarks. On MeteoNet, Shanghai Radar and CIKM, the other methods publish no per-threshold results and have no public checkpoints, so we compare PrecMamba with variants of our own model. CSI and FAR at the highest threshold of each benchmark:

VariantMetricSEVIR (219)MeteoNet (40)Shanghai (40)CIKM (40)
Backbone alone (pinball)CSI0.02530.04100.19150.1139
FAR0.64420.04220.51340.3449
Full model, MSE backboneCSI0.04440.02550.18940.0365
FAR0.57810.76830.37690.4990
PrecMamba (full)CSI0.11500.03250.21710.1379
FAR0.46480.14790.48010.4570

Compared with the MSE variant, the pinball objective raises CSI at the highest threshold on all four benchmarks, for example from 0.0444 to 0.1150 on SEVIR, with a lower FAR (0.4648 against 0.5781). Compared with the backbone alone, the generator raises it on SEVIR, Shanghai Radar and CIKM. On MeteoNet, the highest threshold (40 dBZ) is rarely exceeded, and the full model is slightly below the backbone alone (0.0325 against 0.0410). The full per-threshold tables will be in the revised paper.

We also clarify the scope of the claim. The pinball loss targets the upper tail, but its benefit is not limited to extremes: it also raises the threshold-averaged CSI on SEVIR from 0.2775 (MSE) to 0.3445 (Tab. 2, left). We will revise the abstract and introduction to present PrecMamba as a nowcasting model with a tail-aware objective, and to state the extreme-precipitation results as above.

W5: False alarm rate

We report FAR at each threshold (see W4). To test for over-forecasting directly, we also report the frequency bias FBI = (TP + FP)/(TP + FN), the ratio of forecast to observed events, where FBI > 1 means over-forecasting. At the highest threshold, the FBI of PrecMamba is 0.24 on SEVIR, 0.04 on MeteoNet, 0.52 on Shanghai Radar and 0.29 on CIKM, and the FBI of all other variants is also below 1. PrecMamba therefore under-forecasts intense events rather than over-forecasting them, so its CSI gains at high thresholds do not come from issuing more alarms.

Q1: Why only the median channel?

The residual is defined relative to a one-point forecast, r = (Y − μ)/s (Eq. 4), so the generator needs one anchor. The anchor should give a nearly centred residual, because the flow transports a symmetric Gaussian prior along a straight path. On SEVIR, the quantile channels are well calibrated: their coverage on pixels with observed precipitation is 0.083, 0.505, 0.893, 0.946 and 0.989 for the levels 0.1, 0.5, 0.9, 0.95 and 0.99. For this reason, the upper quantiles are poor anchors. On CIKM, the residual around the median is negative 47.1% of the time, but the residual around the 0.9-quantile is negative 86.9% of the time. We tested this directly on CIKM: moving the anchor from the median to the 0.9-quantile lowers CSI at every threshold, from 0.6394 to 0.6226 at λ = 10, from 0.5448 to 0.5099 at λ = 20 and from 0.3691 to 0.2860 at λ = 30.

The upper quantiles are still used. Their pinball terms shape the shared encoder features c, which the generator receives as conditioning together with μ. We also trained a variant on CIKM in which the generator is additionally conditioned on all quantile channels, with the median still as the anchor. It raises CSI-M from 0.3281 to 0.3311 and the CSI at the highest threshold from 0.0391 to 0.0451, with a slightly lower FAR (0.5007 against 0.5095). So the upper-quantile outputs carry useful tail information beyond the encoder features, and we will add this variant to the paper and evaluate it on the other benchmarks.

Q2: Why RiverMamba, and why BiSSM?

RiverMamba. Although developed for river discharge, its design addresses multi-lead forecasting of geophysical fields, which is the structure of nowcasting. We adopt three components: one spatio-temporal token sequence with a cost linear in its length, a temporally hierarchical encoder, and an untied decoder with one block per lead time. The river-specific part, the scans over an irregular network, is exactly what we replace with a Hilbert-curve serialisation of the regular radar grid (Sec. 3.2, App. A.2). With the same pinball loss and flow-matching generator, our backbone reaches a CSI of 0.3445 on SEVIR, against 0.3236 for the video prediction backbone SimVP. We will add a discussion of Mamba-based video prediction models to the related work.

BiSSM. Bidirectional scans are the standard design in vision Mamba models, for images (Vision Mamba, Zhu et al., 2024) and video (VideoMamba, Li et al., 2024). They remove the arbitrary direction that a 1D serialisation imposes on a 2D field. Our second scan reverses only the spatial order within each frame and keeps the temporal order (Eq. 14, App. B.1), so neither scan processes time backwards.

Q3: Why are the datasets resampled?

We follow the preprocessing and evaluation protocol of DiffCast (Yu et al., 2024), as stated in Sec. 4.1, which resamples all four benchmarks to 128 × 128. All baselines in Tab. 1 use the same protocol, so the resampling allows a direct comparison with their published results. We note in App. C that this is coarser than the native resolution of SEVIR. We will add an evaluation at native resolution.

Q4: Evaluation as a generative model

We agree that a generative model should also be evaluated by the quality of its predictive distribution. Tab. 1 already compares PrecMamba with generative models (MCVD, PreDiff, STRPM, CasCast, DiffCast, DuoCast), using a single member as in their published protocol. We now also report CRPS and the spread–skill ratio of an 8-member ensemble (CRPS in the native units of each dataset):

DatasetCRPSSpread–skill ratio
SEVIR6.1300.575
MeteoNet0.7720.538
Shanghai Radar1.4860.509
CIKM2.8460.455

The other methods do not report these metrics under this protocol. The spread–skill ratio is below 1, so the ensemble is under-dispersed. A likely reason is that the generator models only the residual around a fixed deterministic forecast, so it does not represent the uncertainty of μ itself. We will report this in the paper as a limitation.

Q5: Choice of n = 10

We study the number of Euler steps in App. A.4 (Fig. 4b). LPIPS decreases from 0.356 with one step to 0.209 with ten steps, and twenty steps improve it only to 0.202 at almost twice the sampling time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.