acceptodds
Withdrawn from ICLR 2027

Multi-Scale Truncated Diffusion with Horizon-Targeted Reforecasting for Precipitation Nowcasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Precipitation nowcasting aims to forecast the distribution of rainfall over the upcoming 0–2 hours based on historical spatio-temporal sequences. However, current mainstream methods still struggle to accurately capture the evolutionary characteristics of precipitation, mitigate the progressive degradation of forecast accuracy with increasing lead time, and adapt to complex mountainous terrains. To address these challenges, we propose a multi-scale truncated diffusion with horizon-targeted reforecasting for precipitation nowcasting. Specifically, a multi-scale truncated diffusion module is designed to jointly model macroscopic precipitation evolution and fine-grained local structures through multi-scale spatio-temporal guidance, while shortening the reverse diffusion trajectory. Subsequently, a horizon-targeted reforecasting module is introduced to selectively reforecast the terminal frames, which are more susceptible to accumulated errors at longer lead times, thereby improving late-horizon forecast accuracy. Furthermore, to enhance precipitation forecasting capabilities in complex mountainous terrains, we construct the Southern Anhui Mountain Radar Dataset. Extensive experiments conducted on our self-constructed dataset, alongside two public datasets (SEVIR and CIKM), demonstrate that our method achieves state-of-the-art performance across all evaluation metrics, enabling robust precipitation nowcasting under complex mountainous conditions.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
multi-scale, diffusion model, precipitation nowcasting, complex mountainous terrains

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_JXctOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a precipitation nowcasting framework combining multi-scale truncated residual diffusion with horizon-targeted reforecasting. A deterministic backbone first produces coarse forecasts, which are refined through fragment-wise residual diffusion guided by optical-flow information and multi-scale spatiotemporal features extracted using a Mamba encoder. A separate latent diffusion model with a Cuboid Transformer then regenerates the final five forecast frames to improve late-horizon predictions. The authors also introduce the Southern Anhui Mountain Radar Dataset and evaluate the framework on this dataset alongside SEVIR and CIKM. The reported results outperform the compared baselines across the evaluated metrics, with particularly notable improvements at high-intensity thresholds.

Critical strengths
  1. Targeted refinement of late-horizon forecasts. The proposed reforecasting module explicitly focuses on terminal frames, providing an intuitive approach to improving predictions at longer lead times.
  2. Evaluation across multiple datasets and baselines. Experiments cover SEVIR, CIKM, and the newly constructed dataset, with comparisons against deterministic, probabilistic, and hybrid methods. The reported improvements are particularly noticeable at high-intensity thresholds.
  3. Additional data for mountainous precipitation scenarios. The Southern Anhui Mountain Radar Dataset extends the evaluation to a practically relevant setting. Component ablations and supplementary analyses also provide empirical evidence for the benefits of motion cues, contextual guidance, and terminal-frame refinement.
Critical weaknesses
  1. Insufficiently articulated challenges and motivation in the abstract. The abstract lists precipitation evolution, performance degradation with increasing lead time, and complex mountainous terrain as key challenges, but does not clearly explain the specific limitations of existing methods or how the proposed components address them. In particular, the connection between the proposed architecture and terrain-related forecasting difficulties remains unclear. Constructing and evaluating a mountainous-region dataset alone does not establish terrain-specific adaptation.

  2. Insufficient justification for the combination of modules. The framework combines a deterministic backbone, multi-scale spatiotemporal guidance, truncated residual diffusion, and horizon-targeted reforecasting. Although component ablations are provided, they do not fully establish why this particular combination is necessary or how the modules complement one another. Comparisons with simpler alternatives and configurations under comparable computational budgets would help distinguish architectural benefits from gains attributable to additional capacity or refinement steps.

  3. Incomplete disclosure of computational complexity. The parameter counts and FLOPs in Appendix B require more detailed accounting. The authors should specify the profiling tool, input resolution, sequence lengths, sampling schedules, and whether FLOPs represent a single denoising step or the complete forecasting pipeline. A component-wise breakdown should include the backbone, context encoder, residual diffusion, latent reforecasting, and VAE computation. Both total and trainable parameter counts should be reported, together with inference latency under consistent hardware and evaluation settings.

  4. Insufficient guidance for interpreting Figure 2. The intended takeaway from Figure 2 is difficult to identify from the current presentation. The caption and accompanying discussion should explain which visual differences readers should examine and how these observations support the claimed improvements. Clearer references to forecast lead times, annotations of relevant regions, and enlarged views where appropriate would make the comparison more informative.

  5. Unclear provenance and comparability of baseline results. All 130 baseline entries on SEVIR and CIKM in Table 1 exactly match those reported in Table 2 of DuoCast. The manuscript should explicitly distinguish results taken from prior publications from independently reproduced results. For reused results, the authors should establish that data splits, preprocessing, forecasting horizons, and metric implementations are consistent with their own evaluation. Backbone variants and inference settings should also be specified, particularly for DiffCast and CasCast. Reusing published results is acceptable, but their provenance and comparability must be transparent.

Questions for authors
  1. Provenance and comparability of baseline results. All 130 baseline entries on SEVIR and CIKM in Table 1 exactly match those reported in Table 2 of DuoCast. Were these results taken from that publication or independently reproduced? If reused, can the authors confirm that their own evaluation follows identical data splits, preprocessing, forecasting horizons, and metric implementations? Please also specify the backbone variants used for DiffCast and CasCast.

  2. Accounting for computational complexity. Does the reported 0.32 TFLOPs cover the complete forecasting pipeline, including all temporal fragments, repeated residual denoising, 50-step latent reforecasting, and VAE encoding and decoding? What input dimensions, batch size, and profiling tool were used? Can the authors provide a component-wise breakdown of FLOPs and parameters, reporting frozen VAE parameters separately from the stated 94.40M trainable parameters?

  3. Necessity of the combined architecture and terminal reforecasting. What evidence establishes that multi-scale guidance, truncated residual diffusion, and terminal reforecasting provide complementary benefits beyond increased model capacity or additional refinement? In particular, what supports refining exactly the final five frames, and do the available comparisons distinguish the benefit of targeting late horizons from that of simply allocating additional computation?

  4. Dataset construction and evidence for mountainous-terrain performance. How is the 8:1:1 split of the Southern Anhui dataset constructed—by individual sequence, precipitation event, or chronological period? Can overlapping sequences or observations from the same event occur across partitions? Additionally, what evidence links the reported improvements specifically to terrain-related forecasting difficulties, rather than demonstrating stronger overall performance on another regional dataset?

Reviewer_o1VzOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a nowcasting framework that combines multi-scale truncated diffusion with horizon-targeted reforecasting. The proposed method uses a deterministic backbone (SimVP) to generate coarse predictions and applies fragment-wise residual diffusion with multi-scale spatio-temporal truncate to capture both large-scale evolution and fine-grained local structures. Additionally, a horizon-targeted reforecasting module regenerates the final forecast frames to mitigate prediction degradation at longer lead times. The authors also introduce the Southern Anhui Mountain Radar Dataset to evaluate the model under complex mountainous conditions. Experiments on SEVIR, CIKM, and the proposed dataset demonstrate improvements over existing deterministic and generative forecasting methods.

Critical strengths

The horizon-targeted reforecasting module provides a practical approach to selectively refining late-horizon predictions without regenerating the entire forecast sequence. The ablation results show improvements in the final forecast frames, particularly for high-intensity precipitation.

Critical weaknesses
  1. Insufficient validation of the proposed mechanisms. The authors claim that horizon-targeted reforecasting mitigates autoregressive error accumulation. However, many diffusion-based precipitation nowcasting models employ MIMO forecasting, where performance degradation is not necessarily caused by error accumulation. Moreover, the proposed method itself uses a MIMO backbone, and reforecasting simply replaces the final five frames without correcting errors propagated through earlier fragments. Therefore, the experiments demonstrate improved terminal-frame accuracy, but do not substantiate the claimed mitigation of error accumulation. The independent contributions of truncated diffusion and multi-scale guidance are also insufficiently validated.

  2. Limited reliability of experimental evaluation. Despite achieving the best reported scores, some improvements are marginal, particularly on SEVIR. The lack of sampling variability, ensemble evaluation, and detailed baseline configurations limits the reliability and reproducibility of the comparisons. Furthermore, although DiffCast and CasCast are included as hybrid forecasting baselines, the proposed truncated diffusion strategy is not sufficiently distinguished from prior work such as DGDM (ECCV 2024). A more explicit comparison of the truncation mechanisms and their computational benefits would help clarify the methodological contribution.

  3. Insufficient validation of the proposed mountainous radar dataset. The authors introduce the Southern Anhui Mountain Radar Dataset to evaluate forecasting under complex terrain. However, limited statistical or qualitative evidence is provided to demonstrate its distinctive precipitation characteristics, such as irregular displacement, rapid intensity changes, and fragmented echoes.

  4. Insufficient methodological clarity and manuscript completeness. Important implementation details, including optical-flow estimation, Mamba configurations, and VAE pretraining, are insufficiently described. The figures and tables also suffer from limited readability and apparent errors, such as "Mamba Fueature Block" in Figure 1 and the questionable displacement direction of "Ours" in Figure 2(a). These issues raise concerns about the completeness of the submission.

Questions for authors
  1. How is temporal consistency maintained between frames 15 and 16, given that horizon-targeted reforecasting regenerates only the final five frames without explicitly conditioning on the immediately preceding fragment? Could the authors provide evidence that this does not introduce temporal discontinuities?
  2. In addition to the question above, I would appreciate the authors' responses to the concerns raised in the Critical Weaknesses.
Reviewer_dXaoOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a multi-scale truncated diffusion and horizon-targeted reforecasting framework for precipitation nowcasting.

The method consists of two main components:

  1. Multi-Scale Truncated Diffusion: First, a deterministic predictor such as SimVP is used to generate a coarse prediction. Then, a multi-scale Mamba encoder incorporating optical flow information is employed to extract spatiotemporal conditions, and preliminary predictions are generated through truncated residual diffusion performed by time segments.
  2. Horizon-Targeted Reforecasting: To address the issue of increasing errors in the later stages of prediction, historical observations and preliminary predictions are used to construct conditions, and the last five frames are regenerated via latent space diffusion to improve long-horizon predictions. The paper achieves promising results on three datasets.

Furthermore, the authors constructed the Southern Anhui Mountain Radar Dataset and conducted experiments on SEVIR, CIKM, and this new dataset. However, the paper has not sufficiently demonstrated whether these improvements stem from the proposed key mechanisms or from additional model capacity, the second-stage reforecasting, training settings, or other engineering factors.

Critical strengths

S1. The research problem is of practical significance. The localization of high-intensity precipitation, the recovery of local details, and the error growth over long horizons are indeed critical challenges in precipitation nowcasting. The paper designs modules to address these issues, with clear research motivation and practical value for meteorological applications.

S2. Multi-dataset experiments demonstrate strong predictive capabilities. The paper not only evaluates on SEVIR and CIKM but also introduces a new radar dataset. From the reported results:

  • The CSI-M on CIKM reaches 0.3511, which is higher than that of AlphaPre and DuoCast.
  • The CSI-40 on the Southern Anhui dataset improves from 0.1447 (the strongest baseline) to 0.1643.
  • Ablation studies indicate that optical flow, contextual features, and horizon-targeted reforecasting all have positive effects on the results. These results are commendable.

S3. Horizon-targeted reforecasting has a sound engineering motivation. Compared to re-forecasting all future frames, correcting only the final segment is a targeted allocation of computational resources. Appendix Table C.2 reports that the CSI-40 for the last five frames improves from 0.0730 to 0.1455. This is a phenomenon worthy of further investigation. If it can be proven through more rigorous ablation that selecting later frames is indeed superior to other correction strategies under the same computational budget, this would become a valuable contribution of the paper.

Critical weaknesses

W1. The methodological novelty is insufficient, particularly lacking independent validation of the truncated diffusion mechanism. The paper integrates multi-scale spatiotemporal features, residual diffusion, truncated sampling, and horizon-targeted reforecasting. However, it remains unclear what new diffusion modeling mechanisms are introduced compared to existing works. Appendix Table C.1 only varies tstartt_{\mathrm{start}}, which merely demonstrates the impact of the truncation time step on performance, but does not sufficiently justify that truncated diffusion has an independent advantage over standard residual diffusion.

W2. The initial distribution of the truncated diffusion lacks a sufficient theoretical explanation. In Equation (4) of Section 3.3.2, regarding the residual initialization at a specific time step, the authors argue that since the residual distribution is concentrated around zero, sampling can start from a zero-residual prior. However, the authors fail to provide the underlying reason for this. The authors need to explain why this approximation is reasonable at their chosen truncation time step.

W3. There are critical ambiguities in the training and inference processes of Horizon-Targeted Reforecasting. Equation (6) in Section 3.4 defines CrC_r, but the paper does not explicitly explain whether the ++ here denotes element-wise addition or another fusion operation. If it is element-wise addition, historical observations and future predictions would be directly mixed, which requires an explanation of the physical and statistical motivations for doing so. More importantly, the authors subsequently state that during training, ground-truth final frames are used to construct latent space training samples, whereas during inference, preliminary prediction frames are used instead. This makes it impossible to determine what the exact reforecasting conditions are during the training phase. It is recommended that the authors explicitly provide training and inference pseudocode, along with the actual training objective. Additionally, in Section 3.3.2, the residual condition of the previous segment comes from the ground-truth residual during training but from the generated residual during inference, which also introduces a teacher-forcing exposure bias. The authors should explain how these two types of training-inference discrepancies will affect error propagation.

W4. There are numerical inconsistencies in the experiments that need verification. Appendix Table C.2 reports the CSI-40 for four equal-length prediction segments on the Southern Anhui dataset: 1-5: 0.2021; 6-10: 0.1648; 11-15: 0.1286; 16-20: 0.1455. The average of these four values is 0.16025, whereas Table 1 reports 0.1643, indicating a discrepancy. The authors need to verify the original experimental results, clarify the statistical methods used in each table, and correct the inconsistencies.

W5. Lack of sufficient ablation studies for the main innovative modules. Table 2 only provides three ablations:

  • w/o Optical Flow
  • w/o Local–Global Context
  • w/o Horizon-Targeted Reforecasting While these experiments show that the three components contribute, they do not sufficiently explain why the core mechanisms proposed in the paper are effective. Specifically, regarding Multi-Scale Truncated Diffusion, there is no direct comparison between full diffusion and truncated diffusion, no independent comparison between single-scale and multi-scale, no separate ablation of local context and global context, and no proof that the Mamba structure outperforms standard spatiotemporal encoders with a similar number of parameters. Regarding Horizon-Targeted Reforecasting, there is no comparison of different reforecasting lengths KK, no comparison between correcting all frames versus only correcting the final frames, and no comparison of different reforecasting strategies under the same computational budget. For instance, if a standard diffusion refiner achieves similar results on the last five frames, then the innovation of this paper is closer to a computational resource allocation strategy rather than a novel generative modeling mechanism. The authors should clarify the actual contribution of their method through the aforementioned comparisons.

W6. The contribution of the new mountainous dataset has not been sufficiently validated. The Southern Anhui Mountain Radar Dataset is a significant contribution of the paper, but the current dataset description is insufficient to validate its uniqueness. The authors describe the geographical scope, time span, and basic preprocessing, but fail to explicitly provide the total number of samples, the number of independent precipitation events, the proportion of samples of different intensities, whether the training, validation, and test sets are split by event, time, or window, whether adjacent or overlapping radar sequences might end up in different splits, and the proportion and stratified performance of mountainous versus non-mountainous samples. In particular, if the data is randomly split using highly overlapping time windows, the training and test sets may have strong spatiotemporal correlations, thereby overestimating the generalization capability. Furthermore, the paper does not introduce explicit terrain elevation or topographical conditions, nor does it present a stratified comparison between mountainous and plain areas. Therefore, while the current experiments can support the model's performance on data from this region, they are insufficient to demonstrate its specialized adaptability to complex terrain.

Questions for authors

Q1. Is there any target information leaking into the training conditions during the reforecasting phase? Please explicitly provide the construction method of CrC_r, along with training and inference pseudocode, and clarify the meaning of the plus sign in Equation (6).

Q2. Why can a zero-residual prior approximate the true marginal distribution at intermediate time steps of the diffusion process? Please provide corresponding theoretical justification or empirical analysis of the initialization distribution.

Q3. Can you explain the numerical discrepancy in CSI-40 between Table C.2 and Tables 1/2?

Q4. Is truncated diffusion genuinely superior to standard DDIM with the same number of steps? It is necessary to control for model capacity, number of sampling steps, and other modules to isolate the benefits brought by the truncation strategy.

Q5. Why reforecast only the last five frames? Please explain the choice of K=5K=5 and the advantages of this strategy over reforecasting with different lengths and at different positions.

Q6. Could the authors provide a more detailed description of the Southern Anhui Mountain Radar Dataset to demonstrate the advantages of this dataset?

Reviewer_q3tNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a multi-scale truncated diffusion framework for precipitation nowcasting. It first generates coarse predictions using a deterministic backbone, then employs optical-flow-guided multi-scale contextual features to guide fragment-wise residual diffusion. Additionally, a horizon-targeted reforecasting module is introduced to regenerate terminal forecast frames and mitigate error accumulation at longer prediction horizons. Experiments on SEVIR, CIKM, and a newly constructed Southern Anhui Mountain Radar Dataset demonstrate consistent improvements over existing methods.

Critical strengths
  1. The integration of multi-scale spatio-temporal contextual guidance with truncated residual diffusion offers a promising way to capture both large-scale precipitation evolution and fine-grained local structures, while reducing the number of diffusion sampling steps.

  2. The proposed horizon-targeted reforecasting mechanism selectively refines terminal forecast frames, offering a targeted solution to long-horizon prediction degradation. The substantial improvement in high-intensity precipitation forecasting on the mountainous dataset provides encouraging empirical evidence for this.

Critical weaknesses
  1. Insufficient justification of the truncated diffusion initialization. During training, the diffusion process is applied to ground-truth residuals, whereas during inference, it starts from a zero-residual prior at an intermediate noise level. This may introduce a distribution mismatch between training and inference. Further theoretical justification and empirical analysis are needed to establish the validity of this initialization strategy and its impact on generation quality.

  2. Insufficient validation of the differences from existing methods and the contributions of individual components. Both this work and existing approaches such as DiffCast adopt a combination of deterministic prediction and residual diffusion. However, the specific differences in their modeling mechanisms and the respective roles of the proposed multi-scale Mamba architecture and diffusion process are not sufficiently established through comparative and ablation studies. In addition, more details regarding baseline training configurations and the fairness of comparisons would strengthen the experimental evaluation.

Questions for authors
  1. During training, ground-truth residuals are used for diffusion, whereas inference starts from a zero-residual prior. Given the potential distribution mismatch between these two settings, could the authors further explain the rationale behind this design and provide theoretical or empirical evidence regarding its impact on prediction quality?

  2. Considering the similarities between the proposed framework and existing residual diffusion approaches such as DiffCast, could the authors further clarify their differences in modeling mechanisms? Additionally, would it be possible to provide controlled comparisons across different backbone architectures (e.g., Mamba, CNN, and Transformer) and diffusion configurations to better understand the individual contributions of the proposed design choices?

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

Related papers

Discussion (0)

Sign in to comment.