acceptodds
Under review as a conference paper at ICLR 2027

QFSNet: A Quaternion Fourier Spectral Network for Deterministic Precipitation Nowcasting

Abstract

Precipitation nowcasting requires capturing complex, multi-scale spatiotemporal dynamics under strict latency constraints. However, existing deep learning paradigms face two fundamental bottlenecks: standard spectral representations fail to explicitly disentangle scale and orientation, leading to trajectory underfitting, and a critical mismatch exists between continuous pixel-wise optimization and discrete structural boundaries, causing models to blur sharp intensity transitions to minimize aggregate errors. To address these challenges, we propose QFSNet, a highly efficient deterministic nowcasting architecture anchored by a Quaternion Fourier Neural Operator. To resolve the directional misalignment, we introduce a lossless wavelet-quaternion representation that maps multi-scale details into a quaternion field, leveraging the non-commutative Hamilton product and a two-sided Quaternion Fourier Transform (QFT) to jointly model directional kinematics, and decouple the forecasting process into three specialized branches for fine-scale dynamics, macroscopic trends, and local phase transport. Furthermore, standard continuous optimization inherently causes structural blurring and the loss of high-contrast edges in predictions. To resolve this, we introduce a mathematically rigorous normal–tangent decomposition mechanism. This strictly decouples the boundary-reshaping dynamics required to restore sharp structural discontinuities from the intra-region intensity calibration , explicitly preventing the over-smoothing of abrupt precipitation transitions without introducing artificial artifacts. Extensive experiments across four benchmark datasets demonstrate that QFSNet achieves state-of-the-art performance, particularly in extreme precipitation detection, while requiring the lowest parameter count and computational cost, and maintaining the fastest inference speed.

Primary area:
learning on time series and dynamical systems
Keywords:
Precipitation Nowcasting, Lightweight Architecture, Quaternion Fourier Transform, Spectral Neural Operators, Normal-Tangent Decomposition

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_aYQ8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents a precipitation nowcasting framework based on Quanternion Fourier Neural Operator, to address the challenges in precipitation nowcasting such as: standard spectral representations fail to explicitly disentangle scale and orientation, the mismatch between the pixel-wise optimisation and discrete boundaries problem exist in other precipitation nowcasting models. Experiments on several benchmarks demonstrate superior performance of QFSNet to other baselines.

Critical strengths
  • The problem setting is timely. The proposed method seems to have strong performance compare to the baselines and the paper compares against a diverse range of baselines.
  • Experiments are extensive, covering several different precipitation benchmarks for evaluation of the proposed framework under different conditions
  • The results show promise of the methodology (efficient and strong performance) and is generally aligned with the motivation.
Critical weaknesses
  • In previous literature ([1]), a similar differentiable CSI-based loss has previously been proposed for precipitation nowcasting (Ko et al., 2022). There is few difference in the incorporation of false-alarm weighting but the formulation follows a similar objective. With this being said, first the authors should cite relevant literatures in section 4.5. Second, the loss seems to contribute the performance substantially. And the improvements of adding this loss are larger than those observed in the later ablations on proposed architecture. So here my question is: to what extent do the reported improvements come from the proposed framework instead of the choice of using this loss? I would need some clarifications on this. I would suggest the authors adding a new ablation, adding this loss to baselines such as SImVP to see how much it will affect its performance.
  • The similar decomposition of large-scale and fine-scale features for precipitation nowcasting model has been explored [2]. WADEPre is a new work but it is very similar to the division between the Large-scale LL and Fine-scale QFT branches in QFSNet. I would encourage the authors to add it as another baseline and depending on the results clarify what additional benefits QFSNet’s decomposition provide over WADEPre.
  • Other weaknesses see questions.

[1] Ko, J., Lee, K., Hwang, H., Oh, S.G., Son, S.W. and Shin, K., 2022. Effective training strategies for deep-learning-based precipitation nowcasting and estimation. Computers & Geosciences, 161, p.105072.

[2] Liu, B., Zhang, H., Yuan, H., Wang, D., Li, Y., Chen, F. and Wu, H., 2026. WADEPre: A Wavelet-based Decomposition Model for Extreme Precipitation Nowcasting with Multi-Scale Learning. arXiv preprint arXiv:2602.02096.

Questions for authors
  • Since CRPS reduces to MAE (up to normalisation) for deterministic point forecasts, why are we adding CRPS as one of the metrics here? Also why the ranking of CRPS seems to be inconsistent with that of MAE if they are equivalent for deterministic forecast? The author should clarify how CRPS is computed for deterministic and generative baselines.
  • Could authors clarify why “w/o LL branch” has such bad result and “Fine only” is much better? Because to me the only difference is the phase branches, does this show some interactions between two branches?
  • The "w/o fine spectral branch" variant in table 2 and the "LL only (w/o fine & phase)" variant in table 9, are they equivalent to each other?
  • Could authors provide baseline details for all experiments?
5 comments

We thank the reviewer for recognizing the practical motivation, empirical coverage, and efficiency of QFSNet. We address the relationship to prior CSI-based training and wavelet decomposition, provide the additional baseline and branch controls, and clarify the metric and implementation details below.

1. Prior CSI-based objectives and attribution to the architecture — W1

We thank the reviewer for highlighting Ko et al. (2022) [1], which we will cite in Sec. 4.5. Their nowcasting objective constructs differentiable hit, miss, and false-alarm counts from predicted class probabilities and optimizes the average soft CSI for Rain (the union of Light and Heavy) and Heavy, with equal coefficients for misses and false alarms.

QFSNet supervises continuous radar-intensity forecasts. Its event term evaluates threshold exceedances across the selected intensity thresholds and forecast lead times, while MAE directly supervises the magnitudes of the same predictions. In Eq. (15), αFA\alpha_{\mathrm{FA}} controls the false-alarm contribution relative to misses within the overlap score, whereas λe\lambda_e balances this event term against intensity reconstruction in L=LMAE+λeLevent\mathcal{L}=\mathcal{L}_{\mathrm{MAE}}+\lambda_e\mathcal{L}_{\mathrm{event}}. These two coefficients separately control the relative weighting of event-detection errors and the balance between event overlap and continuous-intensity accuracy. We will clarify these differences in Sec. 4.5.

To examine the contribution of the event term across predictors, we additionally trained SimVP and AlphaPre with this term on Shanghai. Table R1 includes both their original and event-augmented results.

Table R1. Original and event-augmented training results on Shanghai.

ModelTraining settingCSI-M ↑CSI-35 ↑CSI-40 ↑POD ↑SSIM ↑MAE ↓
AlphaPreOriginal training0.41760.38530.26140.50080.64502.1153
AlphaPreWith event term0.38730.35020.25950.50920.64512.5314
SimVPOriginal training0.37720.34430.21100.44490.75721.8788
SimVPWith event term0.38970.35450.24170.47750.70652.1647
QFSNetMAE only, λe=0\lambda_e=00.31300.23520.14740.35650.80171.6878
QFSNetMAE + event, λe=0.1\lambda_e=0.10.42520.38820.28690.60200.80401.7280

The event term improves SimVP’s CSI-M by 0.0125 and POD by 0.0326. AlphaPre shows a small POD increase of 0.0084 but lower CSI scores. QFSNet gains 0.1122 in CSI-M and 0.2455 in POD. The corresponding MAE increases are 0.2859 for SimVP, 0.4161 for AlphaPre, and 0.0402 for QFSNet. Thus, the event objective makes a substantial contribution to QFSNet, while its benefit varies across predictors. Under the tested settings, QFSNet achieves larger detection gains with a smaller reconstruction penalty.

The architectural contribution is also measurable with the objective fixed. Under MAE-plus-event training, a core predictor retaining the three branches, additive fusion, and spatial readout, but excluding PC and all calibration/Normal–Tangent refinement, achieves CSI-M/POD/MAE of 0.4023/0.5341/1.7724. The complete model achieves 0.4252/0.6020/1.7280 under the same objective. These controls establish additional gains from the consistency and calibration/refinement stages. We will present the results as contributions from both the architecture and its event-aware training.

We have added the comparison with Ko et al. (2022) to Sec. 2.2 and supplemented the original loss-weight analysis with event-loss transfer experiments on SimVP and AlphaPre and clean-core versus complete-model controls under both MAE-only and MAE-plus-event training in Appendix E.3. These additions clarify the contributions of the architecture and the training objective under the evaluated settings. Please refer to the resubmitted manuscript for details.

2. Relationship to WADEPre and additional comparison — W2

We will cite WADEPre [2] and discuss its relationship to QFSNet. Both methods use wavelet decomposition to organize precipitation structure across scales. WADEPre predicts approximation and detail coefficients using separate networks and combines the reconstructed fields through a spatial Refiner. QFSNet’s Fine branch processes the complete wavelet–quaternion field, including LL and the directional detail components, through two-sided QFT and Hamilton operators. Its dedicated LL pathway complements this joint predictor, while local phase transport and global/local PhaseConsistency provide additional transport and fusion operations. The distinction therefore lies in how the decomposed information is jointly predicted and fused.

We additionally trained WADEPre on Shanghai using its official implementation and native training objective. Table R2 reports this reproduction together with the QFSNet results.

Table R2. Additional Shanghai comparison with WADEPre.

ModelCSI-M ↑CSI-35 ↑CSI-40 ↑POD ↑SSIM ↑MAE ↓Parameters ↓
WADEPre0.39370.35410.23740.50420.72382.061636,526,226
QFSNet0.42520.38820.28690.60200.80401.7280185,433

In these runs, QFSNet has higher CSI-M by 0.0315 and CSI-40 by 0.0495, with MAE lower by 0.3336. It also uses 185,433 parameters, approximately 0.51% of WADEPre.

Within QFSNet, Table 9 provides a controlled backbone comparison: replacing the spectral backbone with a wavelet-domain convolutional backbone, while retaining the other modules and training objective, changes CSI-M from 0.4252 to 0.4146 and POD from 0.6020 to 0.5836. This supports the contribution of the spectral backbone beyond wavelet decomposition itself. The comparison evaluates the backbone as a whole; its individual quaternion and Fourier components are not separately isolated by this control.

We have expanded the related-work discussion in Sec. 2.2 to explain the differences from WADEPre, added WADEPre to the baseline list in Sec. 5.1, and extended Table 1 with its Shanghai, Meteo, and CIKM results and parameter count of 36.53M, with unavailable entries marked as “–”. Please refer to the resubmitted manuscript for details.

3. CRPS reporting and its equivalence to MAE — Q1

The inconsistent CRPS values in the original tables were transcription errors. For the point forecasts evaluated here, the correct CRPS values equal normalized MAE and therefore yield the same ranking within each dataset. We agree that reporting CRPS is redundant in this setting and have removed it from the revised manuscript’s result tables and the corresponding metric descriptions and comparisons.

We have revised the original evaluation to retain MAE for absolute reconstruction error and removed the redundant CRPS metric from Sec. 5.1, Appendix D.2, the result tables, and the associated comparisons. Please refer to the resubmitted manuscript for details.

4. Fine-only, Fine + Phase, and the conditional contribution of LL — Q2

The Fine branch processes all wavelet components, including LL, and can therefore learn a complete forecast. Its strong standalone performance does not require a dedicated LL pathway. The difference between Fine-only and Fine + Phase involves their interaction with the shared PhaseConsistency (PC) module, as shown by the additional ablation in Table R3.

Table R3. Branch configurations with and without PhaseConsistency on Shanghai.

Branch configurationWith PC: CSI-M ↑POD ↑MAE ↓Without PC: CSI-M ↑POD ↑MAE ↓
Fine only0.41780.58931.78730.40990.56861.7588
Fine + LL0.41730.58911.78660.41010.56221.7489
Fine + Phase0.25070.32172.33550.40860.56771.7664
Fine + LL + Phase0.42520.60201.72800.40940.56271.7598

Without PC, Fine-only and Fine + Phase achieve CSI-M values of 0.4099 and 0.4086, compared with 0.4178 and 0.2507 when PC is retained. The large degradation therefore occurs when Phase and PC are jointly enabled without LL.

Phase contributes signed transport increments relative to the last observation, while PC processes the entire fused prediction using historical information. Its shared consistency correction acts on the combined Fine and transport contributions. The LL pathway supplies separately parameterized coarse-intensity adjustments before this processing.

The results establish a conditional contribution of LL when Phase and PC are jointly enabled. With both present, adding LL raises CSI-M from 0.2507 to 0.4252; without PC, the corresponding change is only from 0.4086 to 0.4094. Likewise, with LL and PC retained, adding Phase improves CSI-M from 0.4173 to 0.4252. These controls support their coordinated use within the complete predictor and clarify why branch contributions depend on the fusion configuration. We will incorporate this analysis into the revised ablation discussion.

We have revised the ablation discussion in Sec. 5.3 from describing the LL branch as “irreplaceable” to explaining its configuration-dependent contribution, and added the with/without-PC branch controls and their analysis in Appendix E.1. The revised analysis identifies the pronounced Fine + Phase degradation as occurring when Phase and PC are jointly enabled without LL. Please refer to the resubmitted manuscript for details.

5. “w/o Fine” and “LL only” are distinct ablations — Q3

The “w/o fine spectral branch” variant retains both LL and Phase, whereas “LL only” removes both Fine and Phase. The other shared modules and the training/evaluation settings are retained. Their identical reported values resulted from a transcription error in the “w/o Fine” row of Table 2.

Table R4. Corrected comparison of the two ablations on Shanghai.

ConfigurationCSI-M ↑CSI-35 ↑CSI-40 ↑POD ↑SSIM ↑CRPS ↓MAE ↓
w/o Fine: LL + Phase0.38270.34080.24550.54610.78710.02151.9357
LL only0.38320.34280.24780.55530.78670.02181.9607

The corrected results distinguish the two configurations: retaining Phase alongside LL improves MAE and slightly improves SSIM, with slightly lower event-detection scores. We will correct Table 2 and align the associated discussion with these results.

We have corrected the “w/o fine spectral branch” row in Table 2: CSI-M/CSI-35/CSI-40/POD/SSIM/MAE have been changed from 0.3832/0.3428/0.2478/0.5553/0.7867/1.9607 to 0.3827/0.3408/0.2455/0.5461/0.7871/1.9357. We have also added an explicit distinction in Appendix E.1 that this variant retains LL and Phase, whereas LL-only removes both Fine and Phase. Please refer to the resubmitted manuscript for details.

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

This paper proposes QFSNet, a deterministic precipitation nowcasting architecture designed to preserve multiscale radar details while mitigating directional phase conflicts. The method maps wavelet-decomposed low-frequency and directional high-frequency subbands into a quaternion field, and uses a two-sided QFT and Hamilton products to model direction-aware spectral evolution. It then employs three branches—Fine-scale QFT, Large-scale LL, and Local Phase Transport—to model fine-scale directional information, large-scale evolution, and local nonlinear motion, respectively, followed by PhaseConsistency for fusion. Finally, a Normal–Tangent module separates boundary reshaping from intra-region intensity calibration. The paper evaluates QFSNet on four radar nowcasting datasets against multiple classes of baselines and provides component, loss, and model-capacity ablations.

Critical strengths

1. The analysis of limitations in existing paradigms is well targeted. The paper distinguishes different bottlenecks in spatial/optical-flow methods, generative methods, and standard spectral representations, including high-frequency boundary blurring, computational overhead and potentially inconsistent textures, and directional phase coupling under multi-directional motion. The proposed modules are correspondingly designed around these distinct issues rather than relying on a single mechanism.

2. The mathematical tools are mapped to specific modeling roles in a relatively coherent way. The wavelet–quaternion representation organizes low-frequency structure and directional details, the two-sided QFT and Hamilton products model direction-aware spectral evolution, local phase correlation addresses spatially heterogeneous motion, and the Normal–Tangent decomposition targets boundary reshaping and intra-region intensity calibration.

Critical weaknesses

1. The actual complementarity of the three branches and their interactions are insufficiently explained. In Appendix E, the Fine-only variant already reaches a CSI-M of 0.4178, while the full three-branch model reaches 0.4252, yielding only a marginal gain of 0.0074. More importantly, Fine + Phase (w/o LL) drops sharply to 0.2507, despite Fine-only being substantially stronger. This indicates strong non-additive interactions among the branches. The paper mainly interprets the branches as complementary physical components, but does not analyze the source of this negative interaction or isolate whether PhaseConsistency is necessary to resolve it.

2. There is substantial attribution ambiguity between the extreme-event gains and the event-specific training objective. When λe=0\lambda_e=0, CSI-M drops from 0.4252 to 0.3130 and POD from 0.6020 to 0.3565, showing that Levent\mathcal{L}_{\text{event}} itself has a very large effect on threshold-based metrics. However, representative strong baselines are not retrained under the same event-aware objective, making it difficult to separate the contribution of the QFSNet architecture from that of the training loss. In addition, the paper does not report a clean core-backbone comparison with both PhaseConsistency and the Normal–Tangent/calibration refinement removed, which would help disentangle the contributions of the backbone, fusion/refinement, and loss design.

Questions for authors

1. Full-threshold performance and false alarms. QFSNet achieves high POD across all datasets, but at some intermediate thresholds its CSI does not consistently exceed the strongest baselines, while the advantage becomes more pronounced at the highest thresholds. Could the authors report CSI at all thresholds together with FAR or precision, to clarify whether the higher event recall is accompanied by an increased false-alarm rate?

2. What causes the strong negative interaction between branches? The Fine-only variant achieves a CSI-M of 0.4178, whereas Fine + Phase without the LL branch drops to 0.2507. What specifically causes this large degradation? Is PhaseConsistency intended to mitigate such branch conflicts? If so, could the authors provide an ablation without PhaseConsistency or an analysis of branch interactions?

3. Fair comparison under the event-aware objective. Could the authors retrain at least several representative strong deterministic baselines using the same Levent\mathcal{L}_{\text{event}} and report CSI/POD, so that the contribution of the event-aware objective can be separated from that of the architecture itself? A result for the core architecture without PhaseConsistency and the Normal–Tangent/calibration refinement would also help clarify the independent contributions of these components.

5 comments

We thank the reviewer for recognizing QFSNet's motivation and the modeling roles of its components. We provide the requested threshold-specific evaluation, branch–PhaseConsistency controls, baseline loss-transfer experiments, and clean-core comparisons below.

1. Full-threshold detection performance and false alarms — Q1

We compiled the full-threshold results for AlphaPre, DiffCast, and QFSNet from the evaluation records. Table R1 reports CSI, POD, and the false-alarm ratio, FAR=F/(H+F)\mathrm{FAR}=F/(H+F), where HH and FF denote hits and false alarms. The recorded evaluation pools contingency counts within each forecast lead time and then averages the resulting scores over lead times. Precision equals 1−FAR1-\mathrm{FAR} for the reported combinations.

Table R1. Full-threshold evaluation. Each cell reports CSI ↑ / POD ↑ / FAR ↓.

DatasetThresholdAlphaPre: CSI / POD / FARDiffCast: CSI / POD / FARQFSNet: CSI / POD / FAR
Shanghai200.5547 / 0.6416 / 0.21460.4879 / 0.6123 / 0.31820.5571 / 0.7205 / 0.3029
Shanghai300.4691 / 0.5614 / 0.28230.3989 / 0.5293 / 0.41180.4688 / 0.6451 / 0.3859
Shanghai350.3853 / 0.4786 / 0.36420.3241 / 0.4540 / 0.50370.3882 / 0.5680 / 0.4711
Shanghai400.2614 / 0.3218 / 0.45690.2346 / 0.3481 / 0.62170.2869 / 0.4744 / 0.6034
SEVIR160.6861 / 0.8832 / 0.25010.6321 / 0.7676 / 0.23040.6519 / 0.7986 / 0.2281
SEVIR740.6091 / 0.7053 / 0.19290.5244 / 0.6658 / 0.30610.5587 / 0.6931 / 0.2726
SEVIR1330.3291 / 0.3858 / 0.33700.2752 / 0.3905 / 0.55400.3060 / 0.4348 / 0.5276
SEVIR1600.1929 / 0.2158 / 0.36080.1789 / 0.2613 / 0.68250.1982 / 0.2851 / 0.6592
SEVIR1810.1402 / 0.1562 / 0.40030.1443 / 0.2124 / 0.73360.1595 / 0.2472 / 0.7362
SEVIR2190.0557 / 0.0611 / 0.56550.0730 / 0.1167 / 0.86260.0879 / 0.1748 / 0.8720
Meteo120.5160 / 0.6328 / 0.28060.4440 / 0.5982 / 0.38640.4971 / 0.6642 / 0.3500
Meteo180.4493 / 0.5330 / 0.28200.3935 / 0.5322 / 0.42220.4506 / 0.6197 / 0.3937
Meteo240.3630 / 0.4280 / 0.32490.3203 / 0.4408 / 0.48990.3797 / 0.5585 / 0.4764
Meteo320.2001 / 0.2435 / 0.49480.1829 / 0.2786 / 0.68210.2233 / 0.4148 / 0.6901
CIKM200.5960 / 0.7211 / 0.23660.6312 / 0.7696 / 0.22890.6639 / 0.7979 / 0.2077
CIKM300.2606 / 0.3298 / 0.50760.3230 / 0.4568 / 0.50130.3723 / 0.5269 / 0.4626
CIKM350.1920 / 0.2315 / 0.54750.2231 / 0.3197 / 0.62000.2728 / 0.4034 / 0.5833
CIKM400.1194 / 0.1450 / 0.64310.1782 / 0.2517 / 0.67650.2165 / 0.3703 / 0.6917

QFSNet achieves higher CSI and POD than DiffCast at all 18 thresholds, with lower FAR at 14. Relative to AlphaPre, it improves CSI at 13 of 18 thresholds. At the highest threshold of every dataset, QFSNet achieves the highest CSI and POD among the three models.

The false-alarm trade-off depends on the comparator. At CIKM-40, relative to DiffCast, POD increases from 0.2517 to 0.3703 and CSI from 0.1782 to 0.2165, while FAR increases from 0.6765 to 0.6917. Relative to AlphaPre, QFSNet has higher FAR at the highest threshold of all four datasets. At Shanghai-40, for example, POD increases from 0.3218 to 0.4744 and FAR from 0.4569 to 0.6034, while CSI improves from 0.2614 to 0.2869.

This behavior is consistent with the intended emphasis on heavy-event detection. The event objective includes both misses and false alarms, while MAE maintains supervision of intensity reconstruction. The higher high-threshold CSI shows improved event overlap with both error types included; the accompanying POD and FAR specify how that improvement is achieved. AlphaPre retains higher CSI at Shanghai-30, SEVIR-16/74/133, and Meteo-12.

We have expanded the original threshold-based evaluation with full-threshold CSI/POD/FAR comparisons for AlphaPre, DiffCast, and QFSNet in Appendix D.3, and added the FAR definition and lead-time/threshold aggregation protocol in Appendix D.2. We have also corrected Table 1: DiffCast’s Shanghai CSI-35 from 0.3989 to 0.3241, its SEVIR CSI-181/CSI-219 from 0.1301/0.0582 to 0.1443/0.0730, and QFSNet’s SEVIR CSI-219 from 0.0876 to 0.0879, with the corresponding improvement over AlphaPre revised from 57.3% to 57.8% in the text. Please refer to the resubmitted manuscript for details.

2. Branch complementarity and PhaseConsistency — W1/Q2

The Fine branch processes all wavelet components, including LL, and can therefore learn a complete forecast. Its strong standalone performance does not require a dedicated LL pathway. Relative to Fine-only, the complete model improves CSI-M by 0.0074 and POD by 0.0127, while reducing MAE by 0.0593. To clarify these gains and the severe degradation of Fine + Phase, we evaluated all four branch configurations with and without global/local PhaseConsistency (PC), retaining the other shared modules.

Table R2. Branch configurations with and without PhaseConsistency on Shanghai.

Branch configurationWith PC: CSI-M ↑POD ↑MAE ↓Without PC: CSI-M ↑POD ↑MAE ↓
Fine only0.41780.58931.78730.40990.56861.7588
Fine + LL0.41730.58911.78660.41010.56221.7489
Fine + Phase0.25070.32172.33550.40860.56771.7664
Fine + LL + Phase0.42520.60201.72800.40940.56271.7598

The additional ablation shows that the severe degradation depends on retaining PC in the Fine + Phase configuration. Without PC, Fine-only and Fine + Phase achieve CSI-M values of 0.4099 and 0.4086, compared with 0.4178 and 0.2507 when PC is retained. Thus, direct Fine–Phase fusion does not reproduce the large performance drop; the drop occurs when Phase and PC are jointly enabled without LL.

Fine predicts a future field, while Phase contributes signed transport increments relative to the last observation. PC processes their summed prediction using historical information, so its shared consistency correction acts on the combined Fine and transport contributions. The LL pathway supplies separately parameterized coarse-intensity adjustments before this processing.

The ablation quantifies the conditional contribution of LL when Phase and PC are jointly enabled. With both present, adding LL raises CSI-M from 0.2507 to 0.4252; without PC, the corresponding change is only from 0.4086 to 0.4094. Likewise, with LL and PC retained, adding Phase improves CSI-M from 0.4173 to 0.4252. These comparisons establish a positive Phase contribution within the complete configuration and a strongly configuration-dependent contribution of LL.

PC is intended to improve phase coherence after branch fusion through global and local projections. Its benefit is demonstrated in the complete predictor: enabling PC increases CSI-M from 0.4094 to 0.4252 and POD from 0.5627 to 0.6020, while reducing MAE from 1.7598 to 1.7280. Together, these results support the coordinated use of the three branches and PC within the complete predictor. We will include these controls and describe the measured configuration-dependent benefits in the revised discussion.

We have revised the original characterization of the LL branch as 'irreplaceable' to a configuration-dependent contribution in Section 5.3, and expanded Appendix E.1 with all four branch configurations evaluated with and without global/local phase-consistency. The added controls show that Fine + Phase achieves CSI-M = 0.4086 without PC, compared with 0.2507 with PC, and clarify the coordinated benefit of LL, Phase, and PC in the complete predictor. Please refer to the resubmitted manuscript for details.

3. Event-aware training of representative baselines — W2/Q3

The event objective makes a substantial contribution to detection performance. We additionally trained AlphaPre and SimVP with our event term on Shanghai. Table R3 reports both their original and event-augmented results, together with QFSNet trained with and without the event objective.

Table R3. Original and event-augmented training results on Shanghai.

ModelTraining settingCSI-M ↑CSI-35 ↑CSI-40 ↑POD ↑SSIM ↑MAE ↓
AlphaPreOriginal training0.41760.38530.26140.50080.64502.1153
AlphaPreWith event term0.38730.35020.25950.50920.64512.5314
SimVPOriginal training0.37720.34430.21100.44490.75721.8788
SimVPWith event term0.38970.35450.24170.47750.70652.1647
QFSNetMAE only, λe=0\lambda_e=00.31300.23520.14740.35650.80171.6878
QFSNetMAE + event, λe=0.1\lambda_e=0.10.42520.38820.28690.60200.80401.7280

For AlphaPre, all three CSI scores decrease despite a small POD increase of 0.0084. SimVP gains 0.0125 in CSI-M and 0.0326 in POD, while QFSNet gains 0.1122 and 0.2455, respectively. The corresponding MAE increases are 0.4161, 0.2859, and 0.0402. Under the tested settings, QFSNet therefore obtains larger detection gains with a smaller reconstruction penalty, and its complete configuration outperforms both the original and event-augmented versions of these two baselines on the reported metrics.

The event term optimizes a soft, asymmetrically weighted overlap objective, which need not improve hard-threshold CSI uniformly. AlphaPre's slight recall increase accompanies lower event overlap and greater intensity error; SimVP improves detection with a larger reconstruction trade-off. These results demonstrate that the benefit of event supervision depends on the predictor and training setting. The clean-core controls below further distinguish this supervision effect from the contributions of consistency and calibration/refinement.

We have extended the original QFSNet-only event-loss analysis by adding comparisons between the original and event-augmented training of AlphaPre and SimVP, alongside the corresponding QFSNet results, in Appendix E.3. The revised discussion reports the detection changes and MAE trade-offs for each predictor and clarifies that the effect of event supervision depends on the predictor and training setting. Please refer to the resubmitted manuscript for details.

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

This paper proposes QFSNet, a compact deterministic precipitation nowcasting framework. It packs wavelet subbands into quaternion fields and combines a fine-scale spectral branch, a large-scale LL branch, and a local phase-transport branch. A subsequent normal–tangent refinement module aims to separate boundary adjustment from within-interval intensity calibration. Training combines MAE with an asymmetric event-overlap loss. Experiments on four radar datasets report improvements in high-threshold precipitation detection with a small parameter count and low computational cost.

Critical strengths
  1. The promising motivation. The paper addresses a practically important problem: efficient precipitation nowcasting with an emphasis on heavy-precipitation detection. The related-work section is clearly organized, grouping existing approaches into spatial- and spectral-domain methods and providing an accessible overview of the methodological landscape.
  2. Broad empirical evaluation and architectural details. The paper presents a broad empirical evaluation across four datasets, complemented by component ablations and analyses of model capacity, loss weighting, and random-seed sensitivity. The appendix also provides detailed architectural specifications, tensor shapes, and a breakdown of trainable parameters, which help readers inspect the proposed implementation.
Critical weaknesses
  1. The methodological novelty is not clearly established. The existing QWNet (2026) already adopts the approach of combining wavelet components with quaternions, and the standard Haar decomposition together with its lossless reconstruction property are established tools as well. Likewise, Clifford Neural Layers for PDE Modeling (2023) employs algebraically structured convolution and Fourier layers for physical-field modeling. The results support the effectiveness of the proposed combination, but its methodological advance over these related approaches remains insufficiently established.
  2. The claimed gradient orthogonality of the normal–tangent module is not established. Section 4.4 and Appendix A.1 show that the tangent map has an identity Jacobian at initialization, which preserves the gradient propagated through that stage. However, this does not imply strict gradient orthogonality between the normal and tangent stages. The relevant vectors, the space in which they are compared, and the inner product defining orthogonality are not specified. A precise definition and a separate justification are therefore needed to support this claim.
  3. Strict boundary localization is not established. The localization argument in Appendix A.1 shows that the threshold weighting peaks when the predicted intensity equals the threshold, but does not establish that pixels far from the threshold are unaffected by the shape flux. Moreover, Equation (11) includes an unmasked learned term and global mean subtraction, so the localization of the weighting function does not guarantee localization of the complete correction. The claim of strict boundary localization therefore requires additional justification or qualification.
  4. Inconsistencies in the reported metrics. Some reported CRPS and MAE values are not consistent with the stated deterministic evaluation under a common normalization. The authors should clarify the metric computation and verify the affected entries before the reconstruction-quality comparisons can be fully assessed.
Questions for authors
  1. Figure 1(a) reports 0.43M parameters for QFSNet, whereas Table 1 reports 0.18M and Appendix Table 7 lists 185,433 parameters. Could the authors clarify this discrepancy and confirm whether the accuracy, parameter count, FLOPs, and latency reported in Figure 1 and Table 1 correspond to the same model configuration?

  2. Could the authors clarify the branch ablation results in Tables 2 and 9? First, removing the fine branch in Table 2 and retaining only the LL branch in Table 9 yield identical values across all seven metrics, although the former apparently retains the phase branch. Second, removing the LL branch in Table 2 reduces CSI-M substantially from 0.4252 to 0.2507, whereas the fine-only configuration in Table 9 achieves 0.4178. Are the remaining modules and training settings identical across these experiments? Please explain these results and their implications for the claimed complementarity of the branches and the necessity of the LL branch.

  3. Appendix D.2 states that CRPS reduces to normalized MAE for deterministic predictions. In Table 9, the full model reports MAE = 1.7280 and CRPS = 0.0192, consistent with a normalization factor of 90. However, the wavelet-domain convolutional backbone reports MAE = 1.7318 and CRPS = 0.0198; under the same normalization, its CRPS would be approximately 0.01924. Could the authors explain this discrepancy and clarify whether both configurations use the same normalization and aggregation procedures?

4 comments
AuthorsOpenReview ↗

Response to Reviewer bmXu

We thank the reviewer for recognizing the practical motivation, broad evaluation, and architectural detail, and for identifying specific issues that require correction and clarification. We acknowledge the reporting errors and agree that the manuscript overstates the orthogonality and localization properties. Below, we provide the corrected results, clarify the scope of the mathematical claims, and explain the contribution supported by the architecture and existing experiments.

1. Methodological advances over QWNet and Clifford Neural Layers — W1

QFSNet develops forecasting operators that connect compact spectral evolution, explicit local phase transport, and interval-preserving intensity calibration within a wavelet–quaternion representation. Its methodological distinctions from the two related works are as follows.

Relative to QWNet, which uses quaternion wavelet features for feature enhancement and cross-modal interaction and fusion, QFSNet constructs explicit temporal prediction mechanisms in this representation. The spectral encoder–translator–decoder predicts the future wavelet–quaternion field, while a separate transport branch estimates local displacements and converts them into phase updates. The resulting signed transport increments are combined through confidence-weighted overlap-add and fused with the learned forecasts. This provides a concrete mechanism for combining learned field evolution with spatially heterogeneous motion.

Relative to Clifford Neural Layers, the distinction includes both the spectral parameterization and the transport operation. Clifford Fourier layers apply multivector weights to retained Fourier modes. QFSNet parameterizes its global mixer using radially shared, grouped Hamilton weights (Eq. 3), tying parameters across frequency locations while retaining scale-dependent responses. An exact local bi-sided Hamilton convolution, implemented through pair-complex factorization, supplements this global mixer at the translator bottleneck (Eq. 4). This construction controls spectral parameter growth while providing a separate local processing operation.

Furthermore, QFSNet uses the two-sided QFT operationally for motion prediction: the estimated vertical and horizontal patch displacements determine left and right phase factors associated with distinct quaternion units (Eq. 8). This supplies an explicit displacement-conditioned transport path alongside learned spectral evolution.

The output refinement provides an additional distinction. A dedicated LL pathway supplies coarse-intensity predictions, and the final tangent calibration adjusts intensities within the threshold intervals selected after the normal update. This imposes an explicit output constraint that connects continuous intensity calibration with precipitation-event thresholds.

The existing controls support these design choices within QFSNet. Replacing the spectral backbone with wavelet-domain convolution reduces CSI-M from 0.4252 to 0.4146; removing the exact local convolution reduces it to 0.4011; and the real/complex Fourier variant obtains 0.4036. These results support the contribution of the proposed spectral backbone and its local operator, with the complete forecasting model using 185,433 parameters. We will cite both related works and revise the contribution statement to make these operator-level and forecasting-specific distinctions explicit.

In Section 2.2, we have replaced the general claim that quaternion packing resolves directional conflicts with an explicit comparison to QWNet and Clifford Neural Layers, identifying radially shared grouped Hamilton mixing, exact local bi-sided convolution, and displacement-conditioned two-sided phase transport as the forecasting-specific distinctions. Please refer to the resubmitted manuscript for details.

2. Identity initialization and gradient orthogonality — W2

Thank you for requesting a precise justification of the claimed gradient relationship. The structural property central to our refinement is that intensity calibration preserves the threshold-defined event regions of the intermediate prediction. Specifically, Eq. (13) maps each input within a threshold interval back into the same interval. Consequently, calibration preserves the pixelwise event assignments at the prescribed thresholds while allowing within-interval intensity adjustment. This property follows directly from the calibration map and does not require orthogonality between the gradients of the two stages.

The identity-Jacobian derivation establishes gradient preservation at initialization, but does not establish mutual gradient orthogonality. We will state the gradient guarantee precisely as gradient-preserving initialization and ground the structural guarantee in interval-preserving calibration, with their assumptions and scope made explicit. We will align Section 4.4, Appendix A.1, the ablation interpretation, and the contribution statement with these established properties. This revision requires no change to the specified forward computations or training objective.

In Appendix A.1, we have replaced the claim of strict gradient orthogonality with gradient preservation at initialization and clarified the assumptions underlying interval preservation and the identity-Jacobian derivation; in the ablation discussion, we have also changed “this orthogonality” to “this separation.” Please refer to the resubmitted manuscript for details.

3. Boundary weighting and the scope of conservation — W3

Localization of the shape correction and preservation of threshold-defined regions — W3

Thank you for requesting a precise characterization of the complete correction. We provide a sufficient condition under which pixels away from the thresholds retain their event membership, explicitly accounting for both the unmasked learned term and global mean subtraction.

Suppressing the forecast-step index, write Eq. (11) as

+\frac{1}{K}\sum_{k=1}^{K}s_k w_k(x), \qquad \mu=\langle U\rangle_{xy}, \qquad J(x)=U(x)-\mu.$$ Define the threshold margin $d(x)=\min_k|P^0(x)-\lambda_k|$. For pixels with $d(x)\ge\delta>0$, $$w_k(x)\le \varepsilon_\delta := \operatorname{sech}^{2}(\kappa\delta/2) \le4e^{-\kappa\delta}.$$ Let $S$ bound $K^{-1}\sum_k|s_k|$ over the frame, and define $$C_\delta= \sup_{d(x)\ge\delta}|J^{\mathrm{learn}}(x)-\mu|, \qquad r_\delta= \sup_{d(x)\ge\delta}|bM(x)|.$$ The complete normal update therefore satisfies $$|P^n(x)-P^0(x)| \le |\alpha_{\mathrm{shape}}| (C_\delta+S\varepsilon_\delta)+r_\delta.$$ Consequently, whenever $$|\alpha_{\mathrm{shape}}| (C_\delta+S\varepsilon_\delta)+r_\delta<\delta,$$ all specified threshold indicators are preserved outside the threshold band: $$\mathbf{1}P^n(x)\ge\lambda_k= \mathbf{1}P^0(x)\ge\lambda_k, \qquad d(x)\ge\delta,\quad\forall k.$$ The proof follows because the complete intensity change is smaller than the distance to every threshold. For the thresholds defining the tangent-calibration intervals, the subsequent interval-preserving mapping retains these indicators. Global centering also has an exact spatial property: $$J(x)-J(z)=U(x)-U(z).$$ Thus, its uniform offset enforces zero-mean redistribution without itself attenuating spatial differences. The learned correction and reaction term remain explicitly included in the event-preservation condition above. Strictly vanishing correction outside the band would additionally require $U(x)=\mu$ there, which Eq. (11) does not enforce. The result established here is a conditional guarantee for the complete update: event-mask changes are confined to the threshold neighborhood when the stated bound holds, while continuous intensity adjustment remains possible elsewhere. This directly specifies when nonzero off-threshold corrections preserve the predicted event regions. In Appendix A.1, we have replaced strict boundary localization with threshold-centered, decaying weighting and a sufficient condition for preserving threshold-defined event membership under the complete normal update. We have also made the assumptions for conservative shape redistribution explicit and limited that guarantee to the normal update before subsequent calibration or clipping. Please refer to the resubmitted manuscript for details.

4. Consistency between deterministic CRPS and MAE — W4 and Q3

We agree with the reviewer’s calculation. The inconsistent CRPS entries are transcription errors. For a deterministic point prediction, CRPS equals absolute error when both are evaluated in the same units. Consequently, using the same predictions, targets, masks, and aggregation weights,

CRPSnorm=MAEevals\mathrm{CRPS}_{\mathrm{norm}} = \frac{\mathrm{MAE}_{\mathrm{eval}}}{s}

where ss is the dataset-specific evaluation scale.

For the wavelet-domain convolutional backbone in the original Table 9, the correct normalized CRPS is

1.7318/90=0.0192422…,1.7318/90=0.0192422\ldots,

which rounds to 0.0192, rather than the reported 0.0198. The full model gives 1.7280/90=0.01921.7280/90=0.0192. Thus, the two configurations tie at the displayed four-decimal precision, while their unrounded CRPS values follow the MAE ordering.

Since CRPS is a rescaled MAE under this point-forecast evaluation protocol, it provides no independent evaluation information. We have therefore removed CRPS from the revised manuscript’s evaluation metrics, result tables, and associated comparisons, retaining MAE to report absolute reconstruction error.

The backbone comparison supports improved event detection with comparable reconstruction quality: the proposed spectral backbone increases CSI-M from 0.4146 to 0.4252 and CSI-40 from 0.2816 to 0.2869, while MAE decreases slightly from 1.7318 to 1.7280. We will make this interpretation explicit in the revised discussion.

We have changed the evaluation from reporting both CRPS and MAE to retaining MAE for absolute reconstruction error, removing the CRPS definition, table columns, and associated comparisons, including the erroneous 0.0198 entry for the wavelet-domain convolutional backbone. Please refer to the resubmitted manuscript for details.

5. Parameter count and the reported configuration — Q1

The correct model is the compact QFSNet configuration reported as approximately 0.18M parameters in Table 1. Appendix Table 7 gives its exact count of 185,433 trainable parameters, corresponding to 0.185433M. The 0.43M annotation in Figure 1(a) is erroneous. We will standardize the presentation to 185,433 parameters / 0.185M to remove ambiguity in numerical precision.

Tables 1 and 12 report the default configuration with Shanghai CSI-M of 0.4252, 8.04 GFLOPs, and 20.21 ms latency. We will regenerate Figure 1 from the corresponding tabulated configuration, including its FLOPs coordinate and parameter annotation, and make the inference dimensions and profiling settings explicit. We apologize for the inconsistent reporting.

In the capacity-ablation table, we have changed the displayed parameter count of the default QFSNet configuration from 0.18M to 0.18M to match Table 1, while retaining the exact total of 185,433 parameters in the parameter-allocation table. Please refer to the resubmitted manuscript for details.

6. Branch ablations and the role of the LL branch — Q2

The original branch ablations retain the same shared-module architectures, including PhaseConsistency and the calibration/refinement modules, and use identical training and evaluation settings. Each configuration is trained with its specified branches.

The “w/o fine spectral branch” row in Table 2 contains a transcription error. The corrected results, together with the LL-only control from Table 9, are:

ConfigurationCSI-M ↑CSI-35 ↑CSI-40 ↑POD ↑SSIM ↑CRPS ↓MAE ↓
w/o Fine: LL + Phase, corrected0.38270.34080.24550.54610.78710.02151.9357
LL only0.38320.34280.24780.55530.78670.02181.9607

These configurations therefore do not produce identical results. Adding Phase to LL-only improves MAE and slightly improves SSIM, while reducing the event-detection scores. We will correct the erroneous row and describe this trade-off.

Regarding the difference between Fine-only and Fine + Phase, the Fine branch processes all wavelet components, including LL, and can therefore learn a complete forecast. Its strong standalone performance does not require a dedicated LL pathway. The severe degradation involves its interaction with Phase and the shared PhaseConsistency module, as clarified by our additional ablation: without PhaseConsistency, Fine-only and Fine + Phase achieve CSI-M values of 0.4099 and 0.4086, compared with 0.4178 and 0.2507 when PhaseConsistency is retained.

Phase contributes signed transport increments relative to the last observation, while PhaseConsistency processes the entire fused prediction using historical information. Adding transport therefore changes the representation subjected to the shared consistency correction, which can also modify structures already predicted by Fine. The LL pathway supplies separately parameterized coarse-intensity adjustments before this processing, providing an additional means to coordinate the predicted field with transport and consistency corrections.

The results support a conditional contribution of LL when Phase and PhaseConsistency are jointly enabled. With both present, adding LL raises CSI-M from 0.2507 to 0.4252; without PhaseConsistency, the corresponding change is only from 0.4086 to 0.4094. Likewise, with LL and PhaseConsistency retained, adding Phase improves CSI-M from 0.4173 to 0.4252.

We will clarify that the branches provide configuration-dependent benefits within the complete predictor. These results support their coordinated use, while the strong Fine-only result and the corrected LL-only comparison do not support universal necessity of the dedicated LL branch or uniformly positive gains from every branch addition.

We have corrected the “w/o fine spectral branch” row in Table 2, including CSI-M from 0.3832 to 0.3827 and MAE from 1.9607 to 1.9357; added branch controls with and without phase-consistency in Appendix E.1; and changed the LL branch’s claimed “irreplaceable” role to a “configuration-dependent” contribution in the ablation discussion. Please refer to the resubmitted manuscript for details.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.