acceptodds
Under review as a conference paper at ICLR 2027

DAWN-Cast: Dynamical Adaptive Wavelet Network for Precipitation Nowcasting

Abstract

Precipitation nowcasting, the prediction of rain over the next minutes to hours, is critical for warning of flash floods and severe convective storms. Typically, weather radar data is used for this task and modeled as a video prediction task. Within an individual storm, rainfall patterns do not evolve uniformly: coherent, large-scale structures drift smoothly, whereas small-scale convective variations change suddenly and lose correlation within just a few frames. Existing nowcasting models capture multiscale spatial structure through Fourier decomposition, wavelets, convolutions, or generative modeling, yet they do not assign each precipitation scale its own temporal pathway. We propose DAWN-Cast, a ynamical daptive avelet etwork that lets every scale learn its own temporal evolution. Its Wavelet Guided Temporal Modeling (WGTM) block splits radar features into wavelet scale families and assigns each its own Frequency Adaptive Temporal (FAT) block, with no interaction between scales before reconstruction. Each FAT block pairs a nonlinear MLP with an adaptive, history-selective Gabor mapping. A refinement stage then reconciles the independently evolved scales by combining global spectral and local spatial interactions. Across four benchmark datasets, DAWN-Cast achieves the best result on 19 of 20 reported metrics, and ablations show that per-scale temporal operators outperform a shared one. DAWN-Cast also generalizes zero-shot to unseen radar domains and adapts to a new dataset by updating only 0.71% of its parameters.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Precipitation Nowcasting, Spatiotemporal forecasting, Scientific machine learning

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_1PshOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies precipitation nowcasting from radar sequences. DAWN-Cast decomposes features into wavelet scale families, assigns each family an independent Frequency Adaptive Temporal (FAT) block combining a Gabor mapping and an MLP, and reconstructs forecasts with spectral and local spatial refinement. The authors report competitive results across several datasets and investigate the design through ablations, zero-shot transfer, and adaptation of a small parameter subset.

Critical strengths

Separating scale-dependent temporal evolution from subsequent spatial refinement has a clear task-specific motivation. The scale-persistence analysis includes a pure-advection control and states limitations of fine-scale resolution and lifetime estimation, making the rationale concrete.

SEVIR has an explicit chronological split, and zero-shot transfer using SEVIR-selected models provides useful evidence. The nearly parameter-matched shared-FAT comparison and small-parameter adaptation results are also meaningful, although their interpretation remains limited by the concerns below.

Critical weaknesses

1.Independent temporal operators are not isolated by the ablation. The main ablation changes both operator sharing and wavelet depth, while the existing parameter-matched comparison also leaves depth settings unclear. Please support the benefits of independent operators with a comparison at the same decomposition depth, or narrow the attribution.

2.The physical intensity mapping needs clarification. The paper and code do not clearly reconcile their intensity ranges, leaving high-intensity thresholds and target-domain normalization difficult to verify. The discrepancy alone does not prove an incorrect conversion; please explain the pixel-to-physical-intensity mapping used in the actual experiments and align the descriptions.

Questions for authors

Q1 : What evidence isolates the contribution of independent temporal operators?

Q2 : How were stored pixels mapped to physical intensity in the reported experiments?

Q3 : Do the performance advantages over the main competing methods remain stable across repeated training runs?

3 comments
AuthorsOpenReview ↗

Response to reviewer 1Psh.

W1/Q1. Independent temporal operators...the attribution. / What evidence...temporal operators?

We agree. The Shared FAT of the original Table 2 used J = 1, and the parameter-matched zero-shot variant (59.51M) was the same model, so both mixed depth with sharing. We now share one FAT block (Gabor and MLP weights and λ) across all families at the model's own depth (J = 2), with the same parameter count as ours to two decimals (59.54M SEVIR, 15.32M CIKM).

SEVIR: CSI-M / 219 / HSS / MSECIKM: CSI-M / 40 / HSS / MSE
Shared, J=10.3565 / 0.0892 / 0.4530 / 354.30.3239 / 0.1401 / 0.4151 / 35.81
Shared, J=20.3718 / 0.1173 / 0.4747 / 358.60.3354 / 0.1526 / 0.4301 / 35.18
Ours0.3787 / 0.1184 / 0.4847 / 340.90.3411 / 0.1549 / 0.4385 / 34.13
Zero-shotMeteoNet: CSI-M / 32 / HSS / MSEShanghai: CSI-M / 40 / HSS / MSE
Shared, J=10.3565 / 0.1892 / 0.4792 / 10.610.4163 / 0.2734 / 0.5491 / 24.41
Shared, J=20.3905 / 0.2319 / 0.5199 / 10.820.4375 / 0.2993 / 0.5719 / 26.01
Ours0.3929 / 0.2159 / 0.5206 / 9.760.4422 / 0.3061 / 0.5786 / 23.40

At equal depth, ours(separate FAT variant) wins over (shared FAT variant) every in-domain metric and, over five CIKM seeds, leads 0.3358 ± 0.0049 vs. 0.3260 ± 0.0120 (2.9%, Table 14) on cikm (refer Table 14 of the paper).

Depth settings

J is selected per dataset (SEVIR 2, MeteoNet 1, Shanghai 3, CIKM 2, App. M). Retraining MeteoNet at other depths changes CSI-M by at most 0.003:

MeteoNetCSI-MCSI-32HSSMSE
J=1 (paper)0.45290.27580.58909.31
J=20.45570.26340.59239.49
J=30.45310.27250.59099.44

Paper updated: Tables 2, 4 and 14 (J = 2 variant, seeds), App. F (J = 1 variant), App. Q (depth), Secs. 5.1 and 5.3.1.


W2/Q2. The physical intensity...align the descriptions. / How were...reported experiments?

All models use the data loaders and evaluator from the AlphaPre [2] release (DiffCast [3] protocol). App. B of the paper has been updated to provide a clearer explanation of the dataset details. Stored frames are divided by a fixed constant; outputs and ground truth are rescaled before scoring based on the protocol as shown below:

DatasetInputEval. scaleCSI/HSS thresholds
SEVIRVIL level ÷ 255× 25516, 74, 133, 160, 181, 219 = 0.15, 0.78, 3.5, 7.1, 12.1, 32.2 kg m⁻²
MeteoNetdBZ ÷ 90× 9012, 18, 24, 32 dBZ
Shanghai8-bit ÷ 255× 9020, 30, 35, 40
CIKM8-bit ÷ 255, zero-padded 101→128× 8020, 30, 35, 40

The SEVIR thresholds follow from SEVIR's VIL encoding (Veillette et al., 2020). MeteoNet is stored in dBZ, so its thresholds are physical reflectivity values. Shanghai and CIKM are stored as 8-bit radar echo images (0–255) and are not converted to dBZ: their thresholds apply on the protocol's scale, identically for every model.

Target domains

In zero-shot transfer, each dataset uses its own normalization constant from the table, with no statistics from target data, and is scored on its own scale and thresholds. SEVIR (VIL) and the targets (radar echo) are not physically matched, so Sec. 5.3.1 describes transfer of normalized dynamics, not physical calibration.

Paper updated: App. B states each dataset's stored range, physical units, normalization constant, evaluation scale, and thresholds.


Q3. Do the performance...repeated training runs?

Yes. We repeated DAWN-Cast and the strongest competitor over seeds on two datasets: EarthFormer (FACL) on CIKM and AlphaPre (FACL) on Shanghai.

Refer to Tables 14 and 30 of the paper for the tables below.

Mean ± stdSeedsCSI-MCSI-40HSS
CIKM, DAWN-Cast50.3358 ± 0.00490.1598 ± 0.00720.4334 ± 0.0059
CIKM, EarthFormer50.3183 ± 0.00720.1426 ± 0.01260.4126 ± 0.0096
Shanghai, DAWN-Cast30.4476 ± 0.00750.3042 ± 0.00640.5883 ± 0.0067
Shanghai, AlphaPre30.4370 ± 0.00470.2953 ± 0.00360.5731 ± 0.0065
Shanghai, seeds 0 / 2 / 4CSI-MCSI-40
DAWN-Cast0.4525 / 0.4390 / 0.45120.3087 / 0.2968 / 0.3071
AlphaPre0.4420 / 0.4328 / 0.43620.2988 / 0.2916 / 0.2954
  • CIKM: +5.5% CSI-M over EarthFormer.
  • Shanghai: +2.4% CSI-M over AlphaPre, the strongest baseline. DAWN-Cast wins every matched seed on CSI-M and CSI-40, and mean MSE is also lower (26.96 vs. 27.23).

Paper updated: App. F and App. O (Tables 14 and 30).


References

[1] Veillette, Mark, Siddharth Samsi, and Chris Mattioli. "Sevir: A storm event imagery dataset for deep learning applications in radar and satellite meteorology." Advances in neural information processing systems 33 (2020): 22009-22019.

[2] AlphaPre (Lin et al., 2025)

[3] DiffCast (Yu et al., 2024)

AuthorsOpenReview ↗

Response to reviewer 1Psh.

We hope our responses and the revised paper address the reviewer’s concerns. If any point remains unclear, or if further experiments would help, we would be glad to provide them as much as possible. If the reviewer finds the concerns resolved, we would be grateful if they would consider updating their score. Thankyou.

AuthorsOpenReview ↗

Response to reviewer 1Psh.

As the author–reviewer discussion closes in a few hours, we wanted to check whether our responses address all the reviewer's concerns, in particular the same-depth comparison of shared and separate FAT blocks, the intensity mapping in App. B, and the multi-seed results. If any point remains unclear, we would be glad to clarify it before the deadline. If the reviewer finds the concerns resolved, we would be grateful if they would consider updating their score.

Reviewer_h3uWOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DAWN-Cast for precipitation nowcasting. The overall framework first uses multi-level wavelet decomposition to decompose radar features into multiple spatial scale families, rather than simply separating them into low-frequency and high-frequency components. Each scale family then performs temporal forecasting through an independent Frequency Adaptive Temporal (FAT) module. The predicted features are reconstructed using the inverse wavelet transform. Finally, a refinement module combining global spectral and local convolutional operations refines the reconstructed features, and a Gabor residual is added to generate the final predictions. The proposed method is evaluated on four radar precipitation datasets, with additional ablation studies and cross-domain transfer experiments to validate its effectiveness.

Critical strengths
  1. Interesting motivation based on scale-dependent temporal evolution. The observation that coarse- and fine-scale precipitation structures exhibit different temporal persistence provides an interesting and meaningful motivation for scale-specific temporal modeling.
  2. Promising experimental results and cross-domain generalization. The method achieves strong performance across four datasets, with particularly impressive zero-shot transfer results and promising parameter-efficient adaptation.
Critical weaknesses
  1. Insufficient explanation and validation of the core methodology. The Gabor-based temporal mapping is difficult to understand, and its role in capturing scale-dependent temporal evolution is not sufficiently explained. Moreover, wavelet decomposition is performed on learned features rather than raw precipitation fields, raising questions about whether the scale-dependent temporal characteristics motivating the method are preserved in the feature space.
  2. Concerns about the experimental design and fairness of comparisons. The ablation study does not fully isolate the contribution of independent temporal modeling, as the shared-FAT and separate-FAT variants use different wavelet decomposition levels ((J=1) vs. (J=2)). Additionally, most baselines are retrained using FACL instead of their original losses, without sufficient comparisons to establish the fairness of this choice. These issues weaken the evidence supporting the claimed improvements.
Questions for authors

1.Sections 3.1.2 and 3.2.1 could benefit from clearer and more intuitive explanations of the Gabor-based temporal mapping. Although the authors motivate Gabor through its frequency-selective and history-selective properties, it remains unclear how these properties specifically help capture scale-dependent temporal evolution in precipitation nowcasting. The methodological description is rather compact, making the role and necessity of the Gabor component difficult to understand.

2.WGTM performs wavelet decomposition on convolutionally lifted features rather than raw precipitation fields. Thus, the low- and high-frequency components represent spatial frequencies of learned feature maps rather than the original precipitation fields. However, the paper's motivation is based on the different temporal evolution patterns of precipitation structures at different spatial scales. These characteristics may not necessarily be preserved after feature mapping. Could the authors clarify whether the learned features still exhibit scale-dependent temporal persistence?

3.The authors retrain most baselines using FACL rather than their original training objectives, while noting that FACL improves performance for the majority of baselines. The motivation for this choice is not entirely clear. Since training objectives can be important components of the original methods, retaining their original settings may provide a more representative comparison. Could the authors explain this choice and provide comparisons between FACL and the original losses to demonstrate the fairness of the evaluation?

4.The ablation study does not fully isolate the contribution of scale-specific temporal modeling. In Table 2, the shared-FAT variant uses a single-level wavelet decomposition ((J=1)), whereas the proposed separate-FAT model uses (J=2). Consequently, the observed improvements may arise from both finer scale decomposition and independent temporal operators. A more controlled comparison using the same decomposition level and comparable model capacity would better substantiate the central contribution. If the authors adequately address these concerns, I would consider raising my rating during the rebuttal phase. Minor Comments The visual presentation of the framework figure could be improved, as it does not fully meet the presentation standards expected at ICLR. Some fonts also appear inconsistent and could be standardized.

suggestions for future work: The zero-shot transfer results are impressive. It would be interesting to further investigate why the proposed method achieves such strong cross-domain generalization, particularly whether the learned scale-specific temporal representations exhibit consistent characteristics across different radar domains. Such an analysis could provide additional insights into the generalization ability of the proposed method.

5 comments
AuthorsOpenReview ↗

Response to reviewer h3uW.

W1a/Q1. Insufficient explanation...sufficiently explained. / Sections 3.1.2...difficult to understand.

Each Gabor neuron maps the TinT_{in}-frame history of a coefficient to one output for one future frame. It projects the history onto a learned direction (e.g., the recent trend), passes it through a sinusoid with phase slope λsf\lambda_s f, and gates it with a Gaussian centered on a learned prototype history. Secs. 3.1.2 and 3.2.1 now explain its frequency and history selectivity, and App. G works through an example.

Each family has its own temporal operator, so the past-to-future mapping can differ across scales, and the trained models show this. We measured each FAT block's share of input sensitivity (squared Jacobian) on the latest of five frames (0.2 = equal weight, full table: Table 9):

Latest-frame shareLLD3D2D1 (finest)
CIKM (J = 2)0.24–0.390.59
Shanghai0.240.400.330.28

In both datasets, every detail family relies more on the latest frame than LL, consistent with the faster decorrelation of fine scales (W1b/Q2, App. E). Replacing FAT blocks with linear per-family maps lowers this share in every family (CIKM D1 0.59 → 0.33, Table 9).

The Gabor form also matters: replacing only the Gabor unit (MLP stream, fusion and residual path kept) with an MLP (more parameters), a sinusoid, or a Gaussian lowers CIKM CSI-M from 0.3411 to 0.3189–0.3238 (App. G, table in reply to XJU6 W3/Q3a).

Paper updated: Secs. 3.1.2, 3.2.1, App. E (Table 9), App. G (worked example).


W1b/Q2. Moreover, wavelet decomposition...feature space. / WGTM performs...temporal persistence?

Yes. The lifted features keep the scale-dependent temporal persistence of the raw fields, verified on the trained models without retraining. Lifting acts on each frame separately, so it mixes no information across frames.

Lifting acts on each frame separately, mixing nothing across frames.

Lagrangian view

(App. E pipeline), SEVIR test, raw frames vs. the 64 lifted channels:

Bandρ_L at 15 min, raw → liftedLifetime (min), raw → lifted
A3 (> 48 km)0.960 → 0.929212 → 133 (extrap.)
D3 (24–48 km)0.759 → 0.70436.9 → 30.5
D2 (12–24 km)0.539 → 0.51021.2 → 19.6
D1 (6–12 km)0.323 → 0.33413.6 → 13.7

Lifted bands still decorrelate in strict order of scale up to 60 min, with lifetimes falling tenfold from A3 to D1. A pure-advection control remains above 0.95 at 15 min, indicating that the decay is physical.

Eulerian view

On the model's own decomposition (what each FAT block sees), lag-4 correlation, raw → lifted:

LLD2D1
SEVIR0.732 → 0.6340.055 → 0.0600.041 → 0.049
MeteoNet*0.794 → 0.7450.105 → 0.1190.015 → 0.025
Shanghai*0.762 → 0.6650.058 → 0.0560.030 → 0.023
CIKM0.466 → 0.3640.015 → 0.0150.002 → 0.052

*Lifting and decomposition of the SEVIR model, applied zero-shot.

Paper updated: App. E (Fig. 5 and 6, Table 8).


W2a/Q4. Concerns about...(J=2)). / The ablation study...rebuttal phase.

We agree. At the model's depth (J = 2) we now share one FAT block (Gabor and MLP weights and λ) across all families, with our parameter count (59.54M / 15.32M).

CSI-MShared, J=1Shared, J=2Ours
SEVIR0.35650.37180.3787
CIKM0.32390.33540.3411
CIKM, 5 seeds–0.3260 ± 0.01200.3358 ± 0.0049

Ours wins across all 10 metrics (see Table 2 of the paper).

Paper updated: Table 2 and 4, App. F, Fig.4, Section 5.1.


W2b/Q3. Additionally, most... improvements. / The authors...the evaluation?

We retrained each baseline on its native loss and recipe and appended the results in Table 1 of the paper. We chose FACL to counter MSE blurring, but we agree that each method's own loss is the fairer reference.

From Table 1 in the paper, we observe that FACL raises CSI-M for 4 of 8 baselines (i.e., ConvLSTM, PhyDNet, SimVP, AlphaPre) and for 20 of 32 pairs overall, so neither loss is uniformly better.

Native runs follow each paper's loss and recipe. Native loss training details and hyperparameter tuning details are provided in Appendix N (Tables 27 and 28) of the paper. Against each baseline’s better loss, DAWN-Cast still has the highest CSI-M and HSS on all four datasets (Table 1) and leads 7 of 8 zero-shot metrics (Table 4).


Minor. The ... standardized.

We have redrawn Fig. 1 with one font, consistent sizes, a per-level FAT stack, and a legend. We hope the


Future work. The zero-shot ... method.

We agree that this is a valuable direction, and we are already working on it, as evidenced by Table 23 in the paper.

Paper updated: Tables 1 and 4, Fig. 2, Appendix N (Tables 27 and 28), Section 5 and 6, abstract, and contributions changed accordingly.

AuthorsOpenReview ↗

Response to reviewer h3uW.

We thank the reviewer for the constructive review. In summary, the revised paper and the response to the review questions address each point as follows:

  1. Temporal evolution of lifted features: analysed on the trained model's lifted channels (App. E, Tables 7–8, Fig. 6).
  2. Gabor explanation: Refined sections Sec. 3.1.2 and 3.2.1, with a worked example and a Gabor-replacement ablation (App. G, Tables 16–17).
  3. Baseline losses: every baseline is now reported with both its native loss and FACL (Tables 1 and 4, App. N).
  4. Shared FAT block at J = 2: compared at equal depth and parameter count, including five seeds (Table 2, App. F).
  5. Other revisions: Fig. 1 redrawn and future work added (Sec. 6).

We hope these revisions address the reviewer's concerns and would be glad to clarify anything further. If the reviewer finds them resolved, we would be grateful if the reviewer would consider raising the rating, as kindly suggested in the review. Thankyou

AuthorsOpenReview ↗

Response to reviewer h3uW.

As the author–reviewer discussion closes in a few hours, we wanted to check whether our responses and the revised paper address the reviewer's concerns. If any point remains unclear, we would be glad to clarify it before the deadline. We thank the reviewer again for the constructive review.

Reviewer_XJU6OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DAWN-Cast performs precipitation nowcasting framework motivated by the observation that temporal persistence varies across spatial scales in precipitation fields. It decomposes lifted features into wavelet scale families, applies an independent temporal streams to each family, and refines the reconstructed predictions through spectral and local spatial processing. Across four datasets, the paper reports the best performance in 19 of the 20 evaluated dataset–metric combinations against 11 baselines. It also evaluates zero-shot cross-domain transfer and parameter-efficient adaptation that updates only 0.71% of the model’s parameters.

Critical strengths
  1. Consistent performance improvements across multiple datasets. Under the reported experimental settings, CSI-M is approximately 2.2~4.6% higher than the strongest baseline on each dataset (Table 1).

  2. Analysis supporting the design rationale. Appendix E distinguishes Eulerian from Lagrangian correlations and uses a pure-advection control and event-level bootstrapping to examine differences in persistence across scales (Table 7, 8). Although this does not directly establish the superiority of independent temporal operators, it provides empirical support for the design rationale. Assigning an independent temporal operator to each detail level also distinguishes the approach from WADEPre’s shared temporal MLP and cross-scale interactions.

  3. Promising transfer and adaptation results. The model leads on all eight reported metrics across MeteoNet and Shanghai in the zero-shot comparison (Table 4). On Shanghai, after SEVIR pretraining, it exceeds the CSI-M of the same model trained from scratch on the full target training set while using only 20% of the target training data and updating 0.71% of its parameters (Table 5). These results provide useful empirical evidence of its potential for adaptation to new domains.

Critical weaknesses
  1. The benefit of independent FAT blocks is not sufficiently isolated. The shared FAT variant in Table 2 combines scale families at J=1J=1, whereas the proposed model processes them independently at J=2J=2. Thus, decomposition depth and cross-scale processing change alongside weight sharing, which is inconsistent with the statement in Section 5.1 that only the temporal operator is varied. Matching the total parameter count in Table 4 does not eliminate these architectural confounds. Establishing the central contribution requires a comparison that varies temporal weight sharing while keeping JJ fixed.

  2. The fairness of the baseline comparisons is difficult to assess fully. Table 11 describes WADEPre as using Tout=TinT_{\mathrm{out}}=T_{\mathrm{in}}, but the modifications made to adjust the forecast horizons in these experiments are unclear. Furthermore, Table 1 uses FACL for all baselines except WADEPre and states only that it improves over the original losses for the majority of models, without providing model-specific comparisons or tuning budgets. These omissions leave an important reproducibility gap when assessing the claimed advantage over the closest prior method.

  3. The specific benefit of the Gabor functional form is not sufficiently isolated. The existing results support the effectiveness of the complete architecture, including the Gabor pathway. However, no comparison replaces Gabor with another temporal predictor while preserving the same fusion and residual pathways. Consequently, the benefit of the Gabor functional form is not sufficiently distinguished from the benefits of an additional branch and bypass connection.

  4. Evidence for robustness across random seeds and independence from Gabor initialization search remains limited. The five-seed CIKM result in Appendix L is encouraging, but corresponding statistics are not provided for the baselines or key ablations. In particular, the two-stage CIKM result in Appendix F is the best result among five seeds, whose scores range from 0.3150 to 0.3313, spanning EarthFormer’s single-run result of 0.3260. The claim that the gains do not depend on dataset-specific Gabor search is therefore stronger than the current evidence supports.

Questions for authors
  1. Keeping J=2J=2, per-family fusion, and refinement unchanged, do the in-domain and zero-shot gains persist when only the sharing of Gabor/MLP temporal weights is varied? Could the authors report the mean and standard deviation across multiple seeds for this tied-versus-untied comparison and the strongest baseline on a representative dataset?

  2. For clarification, please explain how WADEPre was adapted to the 5→205\rightarrow20 and 5→105\rightarrow10 forecasting settings, including its curriculum and loss configurations. Please provide comparisons between the original losses and FACL for the major baselines, together with their tuning budgets, and clarify how FACL was incorporated into DiffCast while retaining its diffusion objective.

  3. Could the authors provide a comparison that replaces Gabor with an alternative temporal predictor, such as an MLP with a comparable parameter count, while preserving the same fusion and residual pathways? For the two-stage CIKM experiment in Appendix F, please also report the seed-selection criterion and the mean and standard deviation across all five seeds.

8 comments
AuthorsOpenReview ↗

Response to reviewer XJU6.

W1/Q1. The benefit...JJ fixed. / Keeping JJ=2...dataset?

Thanks to the reviewer for pointing this out. The original Table 2, Shared FAT, used J = 1, thereby confounding depth and sharing. We now share one FAT block (Gabor and MLP weights) across all families at J = 2, with per-family fusion, refinement, training, and parameter count unchanged (59.54M / 15.32M parameters, as ours).

SEVIR: CSI-M / 219 / HSS / MSECIKM: CSI-M / 40 / HSS / MSE
Shared FAT, J = 20.3718 / 0.1173 / 0.4747 / 358.60.3354 / 0.1526 / 0.4301 / 35.18
Ours0.3787 / 0.1184 / 0.4847 / 340.90.3411 / 0.1549 / 0.4385 / 34.13

Ours outperforms the shared variant on every metric (Table 2). Shared FAT at J = 1 gave 0.3565 on SEVIR (App. F, Table 13), so most of the gap was depth.

CIKM, mean ± stdSeedsCSI-MCSI-40HSS
Ours50.3358 ± 0.00490.1598 ± 0.00720.4334 ± 0.0059
Shared FAT, J = 250.3260 ± 0.01200.1411 ± 0.01690.4195 ± 0.0175
EarthFormer50.3183 ± 0.00720.1426 ± 0.01260.4126 ± 0.0096

Sharing lowers mean CSI-M by 2.9%.

Yes, zero-shot gains still persist and are shown in the table below.

Zero-shotMeteoNet: CSI-M / 32 / HSS / MSEShanghai: CSI-M / 40 / HSS / MSE
Shared FAT, J = 20.3905 / 0.2319 / 0.5199 / 10.820.4375 / 0.2993 / 0.5719 / 26.01
Ours0.3929 / 0.2159 / 0.5206 / 9.760.4422 / 0.3061 / 0.5786 / 23.40

Paper updated: Table 2 (J = 2 variant), Table 4 and 14, Fig. 4, App.F.


W2/Q2. The fairness...prior method. / For clarification...diffusion objective.

WADEPre predicts 5-frame blocks, so we roll it out autoregressively for four blocks (5→20) or two (CIKM, 5→10). Its multi-scale loss is applied to every block, with its curriculum: the weight of its zero-normalized cross-correlation (ZNCC) term decays from 0.1 to 0.01 over 3k steps (App. N, Table 27, footnote c).

FACL vs. native. Table 1 is updated and now contains results from both losses(FACL and native).

Below are the results on the SEVIR dataset on the CSI-M metric:

MetricConvLSTMTrajGRUPhyDNetSimVPMAUEarthFormerEarthFarseerAlphaPre
Native0.28180.37360.33340.33710.33840.34060.31500.3482
FACL0.31660.36760.34770.35790.33400.32840.31480.3634

From Table 1 in the paper, we can observe that FACL raises CSI-M for 4 of 8 baselines (i.e., ConvLSTM, PhyDNet, SimVP, AlphaPre) and for 20 of 32 baseline–dataset pairs overall, indicating its importance, so neither loss is uniformly better. Against each baseline’s better loss, DAWN-Cast has the highest CSI-M and HSS on all four datasets and is best on 15 of 20 metrics.

DiffCast keeps its native objective (MSE loss) on every dataset, because FACL compares predicted and true images, while DiffCast’s diffusion loss is defined on the predicted noise, not on images. The footnote wrongly listed it under FACL. This is now corrected.

Tuning. For the FACL and native runs, each baseline’s learning rate and, where applicable, model-specific hyperparameters were tuned per dataset by validation CSI-M (App. N, Table 28). Native runs follow each paper’s recipe (Table 27).

Paper updated: Tables 1 and 4, Fig. 2, Appendix N (Tables 27 and 28)


W3/Q3a. The specific... connection. / Could the ...pathways?

The original ablation removed the Gabor stream and its residual path together. On CIKM, we now replace only the Gabor unit in every FAT block, keeping the MLP stream, fusion, and residual path, so only the temporal function changes. The Gabor unit was replaced by the following: an MLP(170 vs. 130 parameters per family), the sinusoid alone, or the envelope alone (RBF) (Full table in Table 17 in the appendix).

CIKMCSI-MCSI-40MSE
Gabor (ours)0.34110.154934.13
MLP0.31960.140536.25
Sinusoid only0.32380.153735.77
RBF only0.31890.148734.74

All three fall below the Gabor variant. Gabor, without its residual path (0.3279; Table 3), still beats every replacement on CSI-M. So the gain comes from the Gabor form, not from extra path.

Paper updated: Appen. G (Table 17)


W4/Q3b. Evidence ...evidence supports. / For the two-stage... seeds.

We agree. The reported CIKM run (0.3313) was the best of four seeds (0, 1, 3, 4; not five), each at its best validation checkpoint.

CIKM, mean ± stdSeedsCSI-MHSS
Two-stage40.3226 ± 0.00680.4147 ± 0.0108
EarthFormer50.3183 ± 0.00720.4126 ± 0.0096

Two-stage is 1.4% above EarthFormer (best FACL baseline) in CSI-M, within seed variation. App. H now reports every seed and no longer claims the gains are independent of the search. Baseline and ablation seeds: Tables 14, 30.

Paper updated: Tables 14, 18, 19, 30, App. H.

Thank you for your revisions. Could the authors clarify the following? Note that I am writing based on the version you have recently revised.

  1. For the FAT ablation, the validation frequency is different across experiment types. Appendix C states validations were performed every five epchs, while Table 14 reports every ten epochs for shared FAT block cases. Could the authors tell how independent FAT block is validated, and report the best case when validated with 5 epoch validation? If the validation schedules are different, one could think mean CSI-M advantage is not fully persisted.
  2. The search grids for WℓW_\ell and WhW_h in Table 26 does not contain the reported selections. For example, Wℓ=0.1W_\ell = 0.1 is selected for all cases, while the minimum grid is 0.5. Could the authors share the grids with the actual ones and report how many configurations were compared for each dataset?
AuthorsOpenReview ↗

Response to reviewer XJU6.

We thank the reviewer for the careful reading.

1. For the .... not fully persisted.

All runs in the revised paper, including separate and shared FAT blocks and all baselines, are validated every 5 epochs and tested at their best-validation checkpoint (App. C). The “every ten epochs” note for the shared FAT block was an error in our text, not a different validation schedule, and has been removed. As a check, the shared FAT block ((J=2)) on SEVIR reaches its best validation CSI-M at epoch 25, which a 10-epoch schedule could not select. The results in Table 14 are absolutely correct and therefore already the 5-epoch-validated best cases, and the comparison with the separate FAT blocks uses identical validation for both. We are happy to share the checkpoints and training logs, if the reviewer wishes, to confirm the situation.

2. The search grids... each dataset?

Table 26 in the previous version did not show the values correctly (corrected in the current version). Because of limited compute, we used a one-at-a-time search: three values of W_ℓ with W_h fixed, then three values of W_h at the selected W_ℓ.

Table 26 has been corrected in the current revision with the exact search values.

DatasetW_ℓ triedW_h triedConfigurationsSelected (W_ℓ, W_h)
Shanghai{0.1, 0.75, 1.0}{0.1, 0.75, 1.0}5(0.1, 1.0)
CIKM{0.1, 0.5, 2.0}{0.1, 0.25, 2.0}5(0.1, 0.25)
SEVIR, MeteoNetShanghai's selection1 each(0.1, 1.0)

Configurations used per dataset are given below:

#Shanghai (W_ℓ, W_h)CIKM (W_ℓ, W_h)SEVIR (W_ℓ, W_h)MeteoNet (W_ℓ, W_h)
1(0.1, 1.0) selected(0.1, 0.25) selected(0.1, 1.0) selected, from Shanghai(0.1, 1.0) selected, from Shanghai
2(0.75, 1.0)(0.5, 0.25)––
3(1.0, 1.0)(2.0, 0.25)––
4(0.1, 0.1)(0.1, 0.1)––
5(0.1, 0.75)(0.1, 2.0)––
Configurations5511

SEVIR and MeteoNet reuse the Shanghai selection. Among the three datasets with 20 output frames, Shanghai has the smallest training set (1,534 sequences, compared with 6,308 for MeteoNet and 23,808 for SEVIR), so we conducted the search there. Within these ranges, CSI-M changed little between configurations. Our search is deliberately limited (five configurations on two datasets, with the selected configurations reused on the other two) due to limited computational resources, so the reported gains do not rely on extensive tuning of the Gabor initialization, and a wider search may improve them further.

We would be happy to provide any additional details, checkpoints, or training logs to further clarify and verify these results. Thankyou

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Author Remarks

We thank Area Chair and the reviewers for their time and constructive feedback. We ran the requested experiments and revised the paper accordingly.

Summary of contributions

  1. Per-scale temporal modeling. WGTM splits lifted radar features into wavelet scale families and gives each its own FAT block (Gabor and MLP streams), with no cross-scale interaction before reconstruction. The motivating scale-dependent persistence is measured on raw and lifted fields (App. E).
  2. Accuracy. Against 11 baselines, reported with their native losses and, for the eight FACL-trained ones, also with FACL, DAWN-Cast has the highest CSI-M and HSS on all four datasets and is best on 15 of 20 metrics (Table 1).
  3. Transfer. DAWN-Cast leads all baselines on 7 of 8 zero-shot metrics (Table 4) and adapts to a new domain by updating 0.71% of its parameters (Table 5).

Strengths recognized by the reviewers

  • XJU6: "Consistent performance improvements across multiple datasets", "Analysis supporting the design rationale" (App. E), and "Promising transfer and adaptation results".
  • h3uW: "Interesting motivation based on scale-dependent temporal evolution" and "particularly impressive zero-shot transfer results".
  • 1Psh: "a clear task-specific motivation", a persistence analysis with "a pure-advection control", and zero-shot transfer that "provides useful evidence".

Below we list the concerns raised and their resolutions.

ConcernReviewerStatus
Shared vs. separate FAT confounded with depth (J=1 vs. J=2)XJU6, h3uW, 1PshResolved. Shared FAT re-run at J=2 with equal parameters (59.54M). Separate FAT wins all 10 in-domain metrics (Table 2, App. F) and zero-shot CSI-M on both targets (Tables 4, 15, Sec. 5.3.1). CIKM, 5 seeds: 0.3358 ± 0.0049 vs. 0.3260 ± 0.0120 (Table 14).
Fairness of FACL-trained baselinesXJU6, h3uWAddressed. Each FACL-trained baseline was also retrained with its native loss and recipe and is reported next to its FACL run (Tables 1, 4). FACL raises CSI-M in 20 of 32 pairs, and four baselines (ConvLSTM, PhyDNet, SimVP, AlphaPre) improve on all four datasets, so neither loss is uniformly better. Sec. 4.3, Sec. 5.3.1, abstract and conclusion updated. Recipes and tuning: App. N (Tables 27, 28). WADEPre rollout and DiffCast objective clarified.
Robustness across seedsXJU6, 1PshAddressed. CIKM, 5 seeds: +5.5% CSI-M over EarthFormer (Table 14). Shanghai, 3 seeds: +2.4% over AlphaPre, with higher CSI-M on every seed (Table 30).
Gabor form vs. an extra branchXJU6Addressed. Replacing only the Gabor unit (fusion and residual kept) by an MLP with more parameters, a sinusoid or an RBF lowers CIKM CSI-M from 0.3411 to 0.3189–0.3238 (App. G, Table 17).
Gabor mapping hard to followh3uWAddressed. Sec. 3.1.2 revised, worked example in App. G (Table 16).
Persistence after liftingh3uWAddressed. On SEVIR, the trained model's lifted channels decorrelate in the same scale order as the raw fields, with a pure-advection control (App. E, Fig. 6, Tables 7–8).
Two-stage recipe and Gabor searchXJU6Addressed, claim narrowed. All four seeds reported: 0.3226 ± 0.0068, 1.4% above EarthFormer, within seed variation. The independence claim was removed (App. H, Table 18).
Validation schedule, W grids, λ selection (follow-ups)XJU6Resolved (rating raised 2 → 3). DAWN-Cast and both FAT variants are validated every 5 epochs (App. C). Exact W grids, configuration counts and the λ procedure are in our replies. On CIKM, the selected λ pair has the highest validation CSI-M of all nine candidates.
Intensity mapping and thresholds1PshAddressed. App. B states each dataset's stored range, units, normalization and evaluation scale. One loader and evaluator serve every model.
Framework figureh3uWAddressed. Fig. 1 redrawn.

Reviewer XJU6 engaged over three follow-up rounds and raised the rating to 3 (weak acceptance). Reviewer h3uW is inclined to raise the rating to 3. Reviewer 1Psh did not respond during the discussion. Their concerns are addressed above, including the equal-depth comparison.

Code for the model, both training recipes and evaluation is in the supplementary material.

We thank the reviewers for their constructive comments and the ICLR organizers for the discussion process, which allowed us to substantially improve the paper. We believe this physically motivated design advances nowcasting and will interest the broader ICLR community working on multiscale spatiotemporal forecasting, such as weather and climate.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.