acceptodds
Under review as a conference paper at ICLR 2027

SDRNet: Coupled Rainfall-State Evolution for Multi-Step Tropical Cyclone Precipitation Forecasting

Abstract

Tropical cyclone rainfall fields undergo spatial deformation and local intensity changes beyond the motion of the storm center. Multi-step forecasting must therefore respond to nonuniform evolution while retaining spatial skill at later lead times. Direct regression of rainfall fields leaves the evolution relation implicit in the output mapping, whereas explicit net-residual updates do not separately coordinate the weights assigned to existing rainfall and additions within each step. We propose the Storm-relative Deformation–Survival–Renewal Network (SDRNet), which treats the current rainfall field as an explicit state and constructs a coupled survival–renewal recurrence derived from closed-form integration along characteristics. The model redistributes existing rain areas through backward warping, uses shared decay to coordinate exponential retention and nonnegative effective renewal, and feeds predicted rainfall fields back to refresh structural conditions and the hidden state for subsequent evolution. In four-step forecasting over 24 h on the Corrected and Raw rainfall products of LagTCPMoS, using given retrospective best tracks, SDRNet improves the aggregate fractions skill score (FSS) by approximately 6% over the best external comparison method on both products and achieves the highest critical success index at 20 mm/6 h (CSI20). At +24 h, it retains 93.5% and 93.7% of its first-step FSS, respectively. Comparisons using the same backbone and analyses of rainfall evolution further support the roles of explicit state updates and dynamic feedback in retaining spatial skill across forecast steps.

Primary area:
learning on time series and dynamical systems
Keywords:
tropical cyclone precipitation forecasting, spatiotemporal forecasting, multi-step forecasting, storm-relative rainfall evolution, dynamical systems

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

The paper studies multi-step tropical cyclone precipitation forecasting on storm-centered grids conditioned on a given storm track, using historical rainfall fields to predict the next 24 hours. SDRNet treats rainfall as an explicit state and combines backward deformation, survival and renewal, and prediction feedback. Experiments on the Corrected and Raw products of LagTCPMoS use retrospective best tracks and report improvements in spatial matching and heavy-rainfall identification. Controls using the same backbone examine the update parameterization and dynamic conditioning.

Critical strengths

The design has a clear connection to the forecasting problem. Deformation, intensity adjustment, and feedback address rain-area motion and deformation, local intensification or weakening, and the conditions for subsequent predictions. The state remains nonnegative, and the quantities are limited to effective changes in rainfall-product space, avoiding overly strong physical interpretations.

The predictive results and design controls are useful. Chronological testing with intact storms reports improvements in spatial matching and heavy-rainfall identification. With the same backbone and objective, Full outperforms Warp+Residual and Uncoupled, while paired storm-level analyses support the predictive value of the proposed parameterization.

Critical weaknesses

1.The external baseline comparison needs clearer support. NowcastNet-Evolution is a reimplemented baseline, but the available materials are insufficient to assess whether its adaptation and training are appropriate. Please clarify the implementation and training evidence underlying this comparison to support the advantage reported in the main results.

2.The interpretation of skill retention needs verification. The first and final lead times use partly different storm sets, so their score ratio does not directly represent skill retention on the same storms. Please verify retention on fixed storm sets; this concern does not negate comparisons between methods at the same lead time.

Questions for authors

Q1 : What implementation and training evidence supports the NowcastNet-Evolution comparison?

Q2 : Does skill retention hold on a fixed storm set?

Q3 : How are the "forecast-like" track perturbations generated, and how do they relate to actual track forecast errors?

1 comment

1. NowcastNet-Evolution implementation and training evidence (Weakness 1; Question 1)

Thank you for requesting clearer implementation and training evidence for the NowcastNet-Evolution comparison. The baseline is a task-adapted PyTorch reimplementation of NowcastNet's deterministic Evolution Branch, excluding the conditional generator and discriminator. A shared Evolution Encoder, Motion Decoder, and Intensity Decoder predict motion fields and intensity residuals for recursive evolution. Inputs and outputs are four 6 h fields on a 20×2020\times20 storm-centered grid with 50 km spacing. The accumulation loss constrains the bilinearly warped and final evolved fields; motion regularization uses Sobel spatial gradients.

The added implementation paragraph in Appendix C.6 specifies independent training on Corrected and Raw using common storm splits and seeds 9101/9102/9103: Adam, batch size 16, learning rate 10−310^{-3}, reduced to 10−410^{-4} at iteration 200,000, a 300,000-iteration budget, and motion-regularization coefficient 0.01. Checkpoints are selected solely by validation aggregate FSS across four leads; test data are not used for model selection.

Common data, available information, and evaluation and selection rules align the forecasting task. External methods retain their respective training objectives, whereas the same-backbone controls share soft-FSS. External comparisons therefore assess complete adapted methods, while internal controls isolate the update parameterization. Claims about NowcastNet are restricted to this adapted Evolution Branch.

2. Skill retention on a fixed storm set (Weakness 2; Question 2)

Thank you for pointing out that lead-dependent storm sets complicate the interpretation of skill retention. The conclusion holds on fixed storm sets. We agree that the original first-to-last ratio alone does not establish retention on the same storms. We intersect target-eligible storm IDs across all four leads and three thresholds within each product, yielding 138 Corrected storms and 132 Raw storms. All methods, seeds, and threshold–neighborhood components use the same product-specific set, with the original equal weighting of storms and components.

In new Figure S1 (Appendix D.6.1), SDRNet's +24 h FSS is 0.7169/0.7223, compared with 0.6642/0.6686 for NowcastNet-Evolution. Retention relative to each model's +6 h seed-mean score is 93.38%/94.13%, versus 89.82%/90.43%. Their FSS gaps widen from 0.0282/0.0280 at +6 h to 0.0527/0.0537 at +24 h. Thus, higher final-step absolute skill and slower relative degradation both hold on fixed storm sets.

3. Synthetic perturbations and actual forecast-track errors (Question 3)

Thank you for asking how the forecast-like perturbations are generated and how they relate to actual track forecast errors. Original Table D3 changes track conditions through synthetic offsets to Best Track centers: stepwise Gaussian position perturbations with approximately 25 km RMS displacement and forecast-like perturbations with approximately 45 km median position error at +24 h. The latter is a sensitivity setting at a specified error magnitude; it is not presented as a calibrated stochastic model of operational errors. The two magnitudes also use different statistics.

The historical official forecast-track replay in new Table D6 (Appendix D.6.5) separately examines practical relevance. Tracks come from the NHC ATCF archive and are matched to rainfall initialization by actual release time. We prioritize OFCI, the adjusted previous-cycle OFCL forecast. Missing required times are interpolated between valid positions within the selected forecast cycle to obtain 6 h intervals. Historical rainfall inputs and checkpoints remain fixed; all methods share replay samples and the corresponding grids.

SDRNet achieves FSS of 0.714/0.707, exceeding NowcastNet-Evolution by 0.050/0.049. This supports a relative advantage under these historical forecast-track conditions without calibrating the original synthetic perturbations. Best Track and forecast-track conditions also define different evaluation grids, so cross-condition score changes cannot be attributed solely to one source of error.

Table D6. Aggregate FSS under Best Track and historical official forecast-track conditions.

MethodCorrected (Best Track)Corrected (Forecast Track)Raw (Best Track)Raw (Forecast Track)
SDRNet0.7340.7140.7270.707
NowcastNet-Evolution0.6920.6640.6860.658
Earthformer0.6810.6540.6750.648
LUPIN0.6740.6470.6670.640
Reviewer_5XNMOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SDRNet, a multi-step tropical cyclone precipitation forecasting framework that explicitly models rainfall evolution through storm-relative deformation, survival, and renewal. Rather than directly predicting future rainfall fields, SDRNet combines learned spatial warping with coupled decay and renewal terms to represent the persistence and local changes of rainfall structures. The model also incorporates dynamic feedback from predicted rainfall fields to update subsequent forecasts. Experiments on LagTCPMoS datasets demonstrate improvements in spatial forecasting skill, particularly at longer lead times.

Critical strengths
  1. The proposed approach offers an interesting way to model rainfall evolution without explicitly requiring atmospheric moisture-source information. By leveraging historical rainfall fields and storm-track information, SDRNet learns spatial deformation, decay, and renewal, providing a physically motivated structure for precipitation forecasting.

  2. Explicitly distinguishing the retention of existing rainfall from newly generated rainfall is a promising direction for multi-step forecasting. T

Critical weaknesses
  1. Limited contribution of survival–renewal coupling. Although the proposed formulation is promising, its performance gains over Warp+Residual and Uncoupled variants are relatively modest (approximately 0.010 and 0.008 in FSS, respectively). The results do not sufficiently demonstrate the distinct advantages of the proposed coupling over existing rainfall evolution formulations.

  2. Limited evaluation of rainfall evolution and forecasting accuracy. The evaluation primarily relies on FSS, CSI20, and AW-MAE, which provide limited insight into the accuracy of rainfall deformation, weakening, and regeneration. Additional metrics, such as POD, FAR, and intensity-dependent errors, would help substantiate the claimed improvements, particularly for heavy rainfall at longer lead times.

  3. Insufficient qualitative comparisons. Despite explicitly modeling spatial deformation and rainfall evolution, the paper provides limited qualitative analysis, with forecast visualizations placed entirely in the appendix. Moreover, the case studies compare SDRNet with SimVP or persistence rather than stronger evolution-based baselines such as NowcastNet-Evolution and LUPIN, making the claimed structural improvements difficult to assess.

Questions for authors
  1. Given the relatively small performance differences between the full model and the Uncoupled or Warp+Residual variants, what specific advantages does the survival–renewal coupling provide beyond an alternative parameterization of rainfall intensity changes?

  2. Could the authors visualize and analyze the learned displacement, decay, and renewal fields to demonstrate how these components contribute to rainfall evolution, particularly in cases where rainfall coverage contracts while peak intensity increases?

  3. Could the authors provide side-by-side qualitative comparisons with the strongest evolution-based baselines, particularly NowcastNet-Evolution and LUPIN, across all four forecast steps? This would help clarify whether the improvements in FSS correspond to more realistic spatial deformation and intensity evolution.

3 comments

1. What survival–renewal coupling contributes (Weakness 1; Question 1)

Thank you for asking us to clarify the distinct contribution of survival–renewal coupling beyond an alternative parameterization of intensity changes. The contribution is a state-update parameterization grounded in within-step time integration. Under the approximation that effective coefficients remain constant within a step, existing rainfall experiences attenuation over the entire step, while renewal generated within the step experiences attenuation over its remaining duration. Both therefore share Λ\Lambda but receive different weights:

\widehat P_{k+1}&=M_{k+1}\odot\left[e^{-\Lambda_k}\odot\widetilde P_k+\phi_1(-\Lambda_k)\odot Q_k\right],\\ \phi_1(-\Lambda)&=\frac{1-e^{-\Lambda}}{\Lambda},\qquad \phi_1(0)=1. \end{aligned}$$ Appendix A.4 shows that, when renewal can take arbitrary nonnegative values, the Full and Uncoupled update formulas have equivalent single-step output sets. The contribution thus concerns how these terms are organized and their predictive value, without claiming a larger single-step output space. The same-backbone, same-objective controls directly assess this value. Full improves FSS over Warp+Residual by approximately **0.010/0.010**, with paired 95% confidence intervals of **[0.007, 0.013]/[0.008, 0.014]**, and over Uncoupled by **0.008/0.009**, with intervals of **[0.004, 0.012]/[0.005, 0.013]**. CSI20 and AW-MAE also improve over both variants (Table D1). These are modest but consistent incremental benefits. The intervals reflect test-storm sampling uncertainty conditional on the three training runs. ### 2. Heavy-rainfall detection, intensity-dependent errors, and evolution (Weakness 2) Thank you for recommending additional detection and intensity-dependent metrics, particularly for heavy rainfall at longer lead times. Original Table D2 reports +24 h POD20 and FAR20. SDRNet has higher POD20 than NowcastNet-Evolution (**0.564/0.557 versus 0.522/0.519**) and lower FAR20 (**0.681/0.692 versus 0.748/0.758**), demonstrating improved detection alongside a lower false alarm ratio. **New Table D5 (Appendix D.6.2)** extends the comparison to +18/+24 h and target-defined Heavy ($[5,20)$ mm/6 h) and Extreme ($\geq20$ mm/6 h) intensity bins. At +24 h, SDRNet's Extreme AW-MAE is **14.770/14.073 mm/6 h**, below NowcastNet-Evolution's 15.719/15.079. However, Corrected Extreme AW-MAE remains higher than Earthformer's, Heavy-bin errors are not uniformly lowest, and Extreme bias remains negative. The supported conclusion is improved detection and lower errors in selected extreme-intensity comparisons, with continuing underestimation. Evolution is examined through complementary evidence: original Figure D2 evaluates joint area–peak changes; **new Figures D7–D9** show the model's computational decomposition; and **new Figure D10** compares four-step outputs with strong evolution-based baselines. These address detection, amplitude, and structure separately. Neither POD/FAR nor internal-field visualizations directly validate atmospheric processes. **Table D5. Late-lead heavy-rainfall detection and target-intensity-conditioned errors.** **(a) Detection at 20 mm/6 h** | Product | Method | +18 h POD20 ↑ | +18 h FAR20 ↓ | +24 h POD20 ↑ | +24 h FAR20 ↓ | |---|---|---:|---:|---:|---:| | Corrected | SDRNet | **0.590** | **0.650** | **0.564** | **0.681** | | Corrected | NowcastNet-Evolution | 0.555 | 0.700 | 0.522 | 0.748 | | Corrected | Earthformer | 0.545 | 0.720 | 0.513 | 0.764 | | Raw | SDRNet | **0.582** | **0.665** | **0.557** | **0.692** | | Raw | NowcastNet-Evolution | 0.540 | 0.700 | 0.519 | 0.758 | | Raw | Earthformer | 0.525 | 0.735 | 0.501 | 0.780 | **(b) Errors conditioned on target rainfall intensity (mm/6 h)** | Product | Method | Lead | Heavy AW-MAE ↓ | Extreme AW-MAE ↓ | Extreme bias | |---|---|---|---:|---:|---:| | Corrected | SDRNet | +18 h | 5.958 ± 0.166 | 14.056 ± 0.118 | **−10.977** | | Corrected | NowcastNet-Evolution | +18 h | **5.372 ± 0.130** | 15.060 ± 0.180 | −12.641 | | Corrected | Earthformer | +18 h | 5.948 ± 0.140 | **13.707 ± 0.160** | −13.176 | | Corrected | SDRNet | +24 h | 6.169 ± 0.166 | 14.770 ± 0.079 | **−11.851** | | Corrected | NowcastNet-Evolution | +24 h | **5.457 ± 0.140** | 15.719 ± 0.200 | −13.572 | | Corrected | Earthformer | +24 h | 6.153 ± 0.150 | **14.471 ± 0.180** | −13.995 | | Raw | SDRNet | +18 h | 5.294 ± 0.086 | **13.014 ± 0.063** | **−9.567** | | Raw | NowcastNet-Evolution | +18 h | **4.838 ± 0.120** | 14.122 ± 0.160 | −12.611 | | Raw | Earthformer | +18 h | 5.143 ± 0.120 | 13.386 ± 0.160 | −12.860 | | Raw | SDRNet | +24 h | 5.530 ± 0.132 | **14.073 ± 0.080** | **−10.843** | | Raw | NowcastNet-Evolution | +24 h | **4.963 ± 0.130** | 15.079 ± 0.180 | −13.776 | | Raw | Earthformer | +24 h | 5.354 ± 0.130 | 14.398 ± 0.170 | −13.924 |

3. Learned fields during area contraction and peak intensification (Question 2)

Thank you for suggesting that we visualize the learned displacement, decay, and renewal fields when rainfall coverage contracts while peak intensity increases. New Figures D7–D9 (Appendix D.6.3) show rainfall outputs, learned fields, and contribution maps for Raw storm AL122021, initialized at 12 UTC on 3 September 2021, using seed 9101. At +24 h, target heavy-rainfall area decreases by 21.05% and peak intensity increases by 24.95% relative to the last input. SDRNet predicts a 21.05% area decrease and 28.66% peak increase, reproducing these opposing directions of change.

At the final predicted peak location, the retained contribution is 39.69 mm/6 h and effective renewal is 5.14 mm/6 h, summing to 44.83 mm/6 h; decay loss is 0.022 mm/6 h. Renewal supplies the principal positive increment relative to the warped field at this location, while retained rainfall accounts for most of the final peak. Decay loss is already incorporated in the retained term and is not subtracted again.

This case shows how spatial rearrangement, retention, and local renewal form the output, including the small late-step decay at the peak location. Displacements are backward-sampling offsets, and the components are effective quantities in rainfall-product space, not a unique physical decomposition. The controlled ablations, rather than this case alone, establish the coupling's predictive benefit over Uncoupled.

4. Four-step qualitative comparisons with strong evolution baselines (Weakness 3; Question 3)

Thank you for requesting side-by-side comparisons with NowcastNet-Evolution and LUPIN across all four forecast steps. New Figure D10 (Appendix D.6.4) presents side-by-side targets and predictions from NowcastNet-Evolution, LUPIN, and SDRNet at +6, +12, +18, and +24 h for Raw storm AL072016, initialized at 12 UTC on 28 August 2016. Methods share historical inputs, track conditions, grids, and valid domains.

The two baselines progressively weaken the initial heavy-rainfall core. SDRNet retains a more concentrated core and more closely follows the target's location changes and intensification. It has higher case-level FSS at all four leads; at +24 h the scores are 0.872, 0.612, and 0.590 for SDRNet, NowcastNet-Evolution, and LUPIN, respectively. SDRNet still underestimates late-stage heavy-rainfall coverage and peak intensity, so improved spatial matching does not imply fully accurate intensity evolution.

This figure extends the original comparisons with SimVP and persistence to the requested evolution-based baselines. The visible differences in core location, coverage, and intensity explain this case's scores; overall performance remains supported by the full-test-set results.

Thank you for your rebuttal. While I appreciate the additional experiments, I believe the main manuscript still needs further refinement. Therefore, I have decided to maintain my original rating for this round.

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

This paper presents a deep learning based method, SDRnet, for multi-step tropical cyclone precipitation forecasting. Comparing with existing methods, SDRnet treats the current rainfall field as an explicit state. It constructs successive rainfall fields through learned spatial displacement and backward warping, coupled survival and local rainfall renewal through shared decay, and dynamic conditioning based on the evolving rainfall structure and cyclone track. It also introduces a storm-balanced multiscale learning objective to encourage spatial skill across lead times, rainfall thresholds, and spatial scales. Experiments are conducted on both the Corrected and Raw rainfall products of LagTCPMos to evaluate the performance of SDRnet, in comparison with 12 baseline methods.

Critical strengths
  1. The methodological design is well justified. The explicit state-update formulation, survival-renewal coupling, and dynamic conditioning form a coherent approach to multi-step rainfall forecasting. Although each individual component of the proposed method may not be new, the authors have done a very good job of integrating them into a coherent framework and presenting the overall methodology clearly.

  2. The experiments are well designed and very comprehensive, with extensive baseline comparisons and targeted ablation studies that provide empirical support for the main methodological choices.

Critical weaknesses
  1. The use of given retrospective best tracks may limit the practical applicability of the proposed method, as future storm tracks would typically be uncertain in an operational forecasting setting.

  2. The performance gain from each individual component of the proposed method appears to be relatively small from the ablation studies, although their integration results in a more noticeable overall improvement.

Questions for authors
  1. Could you provide some justification or discussion on the applicability of the assumption of access to given retrospective best tracks?
  2. Have you conducted any experiments on the sensitivity of SDRnet's performance to the quality of the tracks?
  3. The ablation study results (Figure 4 and Table D1) suggest that the performance gain from each individual component is relatively modest, whereas their combination leads to a more noticeable overall improvement. Could you provide some insight into why this occurs? For example, is this due to complementary or interacting effects among the proposed components?
1 comment

Thank you for recognizing the coherence of the method and the experimental design. We address the track assumption, sensitivity, and component contributions below.

1. Best-track assumption and practical applicability (Weakness 1; Question 1)

Our task is multi-step rainfall prediction conditioned on a given cyclone track. Best tracks supply both conditioning information and the locations of storm-centered grids, allowing methods to be compared under common track conditions. This setting does not assume that future best tracks are available operationally.

The historical NHC official forecast-track replay in new Table D6 (Appendix D.6.5) evaluates the effect of replacing future best tracks with forecast guidance available at initialization. Historical rainfall inputs and model checkpoints remain fixed; prediction and target grids follow the respective tracks. On the common replay samples, SDRNet achieves aggregate FSS of 0.714/0.707, compared with 0.664/0.658 for NowcastNet-Evolution, 0.654/0.648 for Earthformer, and 0.647/0.640 for LUPIN.

These results support SDRNet's relative advantage under historical forecast-track conditions. Because historical rainfall inputs and training checkpoints remain unchanged, and track conditions also define evaluation grids, the replay evaluates conditional rainfall prediction rather than a complete real-time operational system.

Table D6. Aggregate FSS under Best Track and historical official forecast-track conditions.

MethodCorrected (Best Track)Corrected (Forecast Track)Raw (Best Track)Raw (Forecast Track)
SDRNet0.7340.7140.7270.707
NowcastNet-Evolution0.6920.6640.6860.658
Earthformer0.6810.6540.6750.648
LUPIN0.6740.6470.6670.640

2. Sensitivity to track quality (Question 2)

Original Table D3 and Appendix D.4.4 report two track-perturbation tests. SDRNet's FSS changes from 0.734/0.727 to 0.725/0.718 under Gaussian center-position perturbations and 0.713/0.706 under forecast-like synthetic perturbations. The respective decreases, 0.009 and 0.021, are smaller than Earthformer's 0.016 and 0.031.

Track errors therefore reduce skill, while SDRNet retains higher FSS and degrades less in these settings. The approximately 25 km RMS Gaussian displacement and 45 km median position error at +24 h for forecast-like perturbations use different statistics. These sensitivity tests do not cover the full operational error distribution; new Table D6 provides a complementary test with historical forecast tracks.

3. Modest individual gains and complementarity (Weakness 2; Question 3)

The variants in Figure 4 and Table D1 are independently retrained. Their differences measure incremental effects within specific configurations and cannot be added as independent contributions. Full improves FSS over Uncoupled by 0.008/0.009, with paired storm-level 95% confidence intervals of [0.004, 0.012]/[0.005, 0.013]. Removing renewal reduces FSS by 0.016/0.018, whereas No survival reduces it by 0.003/0.003. Specifically, No survival sets Λ=0\Lambda=0, making both weights one; it retains the existing-rainfall term while removing attenuation and its weighting effect.

Functionally, deformation adjusts spatial placement and shape, survival and renewal adjust local intensity, and dynamic feedback updates subsequent conditions from the evolving rainfall state. Their roles are connected through the recurrence.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.