acceptodds
Under review as a conference paper at ICLR 2027

CycloDiff:Multi-Task Conditional Diffusion Framework for Spatiotemporal Tropical Cyclone Genesis Forecasting

Abstract

Accurately predicting when and where tropical cloud clusters will develop into tropical cyclones (TCs), and what form they will evolve into, is essential for improving typhoon forecasting and enabling earlier identification of tropical cyclones genesis (TCG) and longer warning lead times. Existing TCG studies, particularly deep learning-based methods, mainly rely on classification paradigms that distinguish TC precursor clusters from non-precursors, yet struggle to model the spatiotemporal evolution of cloud systems. To address these limitations, we propose CycloDiff, a multi-task conditional diffusion framework to generate future spatiotemporal evolution of cloud systems along with the corresponding genesis time and location. To generate future cloud-image sequences that better conform to the evolution of the ocean thermal system, we incorporate subsurface temperature profiles, which provide vertical ocean thermal information that cannot be captured by sea surface temperature alone. Experiments on the Digital Typhoon dataset show that CycloDiff generates more realistic and temporally coherent cloud evolution than existing spatiotemporal prediction methods. These results demonstrate the potential of generative spatiotemporal forecasting for representing TC development and predicting its genesis time and location.

Primary area:
learning on time series and dynamical systems
Keywords:
Multi-Task Learning, Spatiotemporal Forecasting, Tropical Cyclone Genesis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_QmtiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CycloDiff combines conditional diffusion for future infrared cloud sequences with trajectory prediction and a classifier-based genesis procedure. The diffusion model conditions on historical imagery and ocean temperatures at three depths and predicts residuals relative to the latest observation. Its frozen historical cloud-ocean representation conditions an adapted MMSTN trajectory predictor. A fixed classifier identifies a persistent transition in generated frames, and the corresponding trajectory position supplies the predicted genesis location. Experiments primarily evaluate cloud forecasting and downstream diagnostic consistency, with one case illustrating the complete genesis workflow.

Critical strengths
  1. The application formulation is useful and scientifically motivated.** Connecting cloud-system evolution, subsurface ocean conditions, and explicit genesis timing/location is a worthwhile extension of image-only forecasting. Video diffusion for tropical cyclones already exists, and the relevance of subsurface conditions to genesis is also established; the contribution is their predictive integration rather than either idea in isolation. See Ren et al. and Gao et al..

  2. The ocean-conditioning result is stronger than a superficial extra-input claim.** Relative to SST alone, subsurface conditioning reduces FVD from 193 to 132, or 31.6%, and CRPS from 0.1079 to 0.1013, or 6.1%. Adding SST to the no-ocean model reduces FVD by only 2.0%. This contrast supports the practical value of depth-resolved information in the studied configuration. It does not prove a causal ocean-atmosphere mechanism, but a useful predictive contribution does not require such a causal claim.

3.The forecasting improvements are substantial on several metrics.** Against the best baseline for each metric, FVD improves by 83.6% for eight frames and 89.0% for five frames; CRPS improves by 19.3% and 21.0%, respectively. PSNR increases by 0.855 dB and 2.611 dB. The residual ablation reduces FVD from 183 to 132, a 27.9% reduction. These findings deserve credit despite the SSIM trade-off and incomplete evaluation of the final integrated task.

4.The supplementary analyses add relevant evidence and disclose their limitations.** Cloud-structure metrics, fixed-estimator consistency, and ZPR comparisons offer complementary views of the generated sequences. As an auxiliary check of Table 6, averaging the absolute discrepancy from its GT column over the six eight-frame threshold rows gives 0.0152 for CycloDiff versus 0.0768 for PredRNN. This supports closer classifier-derived agreement, although it is an audit calculation rather than a paper-defined metric or independently verified genesis accuracy. The authors appropriately acknowledge the restricted case selection and single-case genesis demonstration.

Critical weaknesses

1.Algorithmic originality is primarily in integration and conditioning.** Residual diffusion is an established forecasting approach, MMSTN is adapted from prior work, and the two branches are trained sequentially. The depth code in Appendix B represents channel order; it is not an explicit model of physical vertical interactions. Predicting residuals from the last observed frame is a reasonable adaptation, but should be distinguished from a new diffusion principle. Relative to DiffCast, the potentially distinctive contribution is ocean-conditioned cloud-to-genesis integration. That contribution is promising, but its benefit beyond cloud prediction remains insufficiently established.

  1. The cross-branch contribution is not isolated.** The trajectory model receives frozen features extracted from historical inputs, not the generated future cloud sequence. This is a valid form of transfer, but no quantitative comparison shows the incremental value of those features over MMSTN with trajectory and local ocean inputs. Consequently, the paper does not yet demonstrate that its multi-task connection is materially better than two separately useful predictors. This is a question about the central algorithmic contribution, not a demand for exhaustive experimental coverage.

  2. The interpretation of genesis predictions depends on an underspecified classifier and coordinate protocol.** The classifier's training data, independent accuracy, eta, M, and the meaning of the thresholds in Table 6 are not sufficiently specified. Likewise, storm-centered targets may be a legitimate relative-coordinate benchmark, but it is unclear how predicted location errors enter the geographically located product. One VAMCO example cannot characterize whether combining the branches improves timing and location over a simpler direct predictor. Evaluation on additional basins or a complete operational negative-case population would be valuable future work, but is not a prerequisite for establishing a narrower conditional forecasting contribution.

  3. The metric trade-off warrants focused investigation.** Eight-frame SSIM is 0.2560 versus the best baseline's 0.5027, an absolute gap of 0.2467, despite the large FVD gain. This is not inherently inconsistent: distributional realism and sample-specific correspondence are different. However, genesis timing is sample-specific, so the trade-off matters. The sampling/CRPS and FVD protocols, baseline access to ocean conditioning, and uncertainty estimates should be explicit. IDOL agreement is useful supporting evidence, but remains estimator consistency, especially after single-channel replication and fixed auxiliary inputs; the authors already acknowledge this limitation.

Questions for authors
  1. What is the incremental trajectory/genesis benefit of the frozen cloud-ocean feature over original MMSTN and MMSTN with local ocean inputs? Please provide a targeted ablation that tests this connection.
  2. How do the selected depths and ordinal depth codes affect performance? Does the gain persist with simple channel concatenation, and is it concentrated in physically interpretable ocean regimes?
  3. What are the classifier's training split, genesis-label definition, independent performance, eta, and M? What exactly do the Table 6 thresholds measure? 4.Are future cloud targets centered on observed future positions? How are generated clouds placed using predicted coordinates, and how sensitive is the result to trajectory error?
  4. On a defined set of pre-genesis initializations, how do timing/location errors compare with a direct predictor? Please also clarify frame spacing, validation split, ensemble size, sampling cost, and common baseline inputs.
4 comments
AuthorsOpenReview ↗

Response to Reviewer Qmti - Q1、Q2

Q1 What is the incremental trajectory/genesis benefit of the frozen cloud-ocean feature over original MMSTN and MMSTN with local ocean inputs? Please provide a targeted ablation that tests this connection. That contribution is promising, but its benefit beyond cloud prediction remains insufficiently established.

Response

Algorithmic originality. Our distinctive contribution is ocean-conditioned cloud-to-genesis forecasting: predicting how a precursor develops and using this development to infer when and where genesis occurs. This formulation makes the predicted development available for event interpretation and downstream diagnosis. We build on established residual diffusion and MMSTN rather than claim a new diffusion principle or a new trajectory-forecasting method. Residual prediction is well suited to this task because the observed cloud field provides a structural reference while diffusion represents uncertain subsequent development.

Benefit of cloud-ocean feature. After incorporating the cloud–ocean feature, a historical cloud–ocean representation is extracted from the cloud-condition encoder and, after global average pooling and projection, injected into the trajectory encoder and decoder. This provides the trajectory and genesis-location prediction with a compact representation of cloud-system structure and evolution trends, as well as subsurface ocean thermal information. Trajectory Error with ADE, FDE illustrate the impact of applying cloud-ocean feature.

ModelADE ↓ (km)FDE ↓ (km)
MMSTN without feature transfer56.20108.17
w/o ocean temperature55.40108.11
CycloDiff53.90107.98

The MMSTN feature-off control retains the same direct historical trajectory and local ocean inputs. The ocean-off model retains cloud-feature transfer.

As for the benefit beyond cloud prediction, we added experiments on four tasks beyond cloud prediction, including maximum sustained wind (MSW), minimum sea-level pressure (MSLP), radius of maximum wind (RMW), and radius of 34-knot winds (R34). The following results address predictive value beyond cloud-image quality. RR denotes the pearson correlation coefficient between estimates obtained from generated and ground-truth cloud images.

VariableModelMAE ↓RMSE ↓|Bias| ↓R ↑
RMWw/o ocean temperature1.1971.5360.4280.520
RMWCycloDiff0.901.150.100.639
R34w/o ocean temperature3.304.320.290.687
R34CycloDiff3.134.060.520.722
MSWw/o ocean temperature3.164.320.250.534
MSWCycloDiff2.292.920.230.681
MSLPw/o ocean temperature1.672.300.220.482
MSLPCycloDiff1.231.540.020.650

Q2 How do the selected ocean depths and ordinal depth codes affect performance?

Response

To verifying the effectiveness of depth codes, we added an ablation study w/o depth encoding on five metrics. The results show a slight benefit from ordinal depth codes over concatenating the same ocean-temperature channels. We clarify that the code identifies relative channel depth; it is a conditioning choice rather than our principal innovation, and it is not an explicit model of vertical physical interactions.

SettingsSSIM ↑PSNR ↑LPIPS ↓CRPS ↓FVD ↓
w/o depth encoding0.245516.3700.02110.1045137
CycloDiff0.256016.5590.02130.1013132
AuthorsOpenReview ↗

Response to Reviewer Qmti - Q3 Q5

Q3 What are the classifier’s training data and split, genesis-label definition, and independent performance? Please specify η, M, the ZPR definition, and the meaning of the thresholds in Table 6.

Response

Classifier training data and genesis-label definition. The event evaluator is a frozen binary CrossViT (crossvit_small_224). Its loader pairs infrared cloud images with the corresponding storm-record grade, normalizes each image, and replicates it into three channels. In this implementation, grade codes 2 and 6 map to the negative class, and other recorded grades map to the positive class. Training uses two-class cross-entropy and image augmentation. These frame-level labels provide the classifier decisions used by the persistent event-identification rule.

Classifier training. The classifier is trained with supervised image–grade pairs using AdamW, a cosine learning-rate schedule, and cross-entropy loss. Training augmentation includes horizontal and vertical flips and rotation. The saved configuration uses an initial learning rate of 0.0002 and batch size 512. Once trained, the classifier remains fixed for all forecasting methods in the event evaluation. We do not report a separate independent classifier Accuracy/F1 here; the event-level results evaluate forecasts under this common fixed classifier.

ZPR and thresholds in Table 6. For the event evaluation, η = 0.5 and M = 3 consecutive positive frames, separated by three hours. The earliest fully contained qualifying run determines the event time, assigned to its first frame. ZPR counts a window as successful only if an event is detected and its absolute timing error is within the stated tolerance. All eligible positive windows remain in the denominator, including missed events. The Table 6 thresholds 0, 3, 6, 9, 12, and 15 are tolerances in hours, rather than classifier probabilities, persistence lengths, or forecast lengths. Zero requires an exact match on the time grid. This positive-window evaluation does not measure false alarms in windows without genesis.

Q4 Are future cloud targets centered on observed future storm positions? How are generated clouds geographically located using predicted coordinates, and how sensitive are the final predictions to trajectory errors?

Response

Future cloud targets and predicted coordinates. Yes. Future cloud targets are centered on the observed storm positions at the corresponding forecast times. At inference, each generated cloud field is geographically placed by assigning its center to the predicted coordinates at that time. Historical cloud–ocean features provide environmental context for trajectory prediction. The cloud field and trajectory position are associated through the same forecast time and storm-center convention, yielding a geographically located prediction.

Regarding the sensitivity of the final predictions to trajectory errors, we thank the reviewer for the constructive comment. We will conduct this setting in future work.

Q5 How do genesis-time and location errors compare with a simpler direct predictor?

Response

Comparing with a direct predictor. The added input-matched comparison favors CycloDiff over directly predicting genesis time and location. Keeping the inputs unchanged, we adapt the model to output these event quantities directly. CycloDiff achieves a timing MAE of 3.66 hours versus 10.17 hours for this predictor, and a reported reference-genesis-time position error of 67.34 km versus 103.25 km, corresponding to reductions of approximately 64.0% and 34.8%.

MetricCycloDiffDirect predictor
Genesis-time MAE ↓3.66 h10.17 h
Position error at reference genesis time ↓67.34 km103.25 km

Q Does combining cloud forecasting and trajectory prediction improve event-level accuracy beyond the VAMCO case study?

Genesis timing and location errors. This comparison supplies dataset-level evidence beyond VAMCO and supports explicitly representing precursor development and geographic motion as useful intermediate information for event prediction. The event cohort contains 420 eligible initializations from 84 storms. Timing MAE is conditional on detected events, and the reported location quantity is evaluated at the reference genesis time. These quantities should be kept distinct from a joint event-location error evaluated at the predicted event time.

AuthorsOpenReview ↗

Response to Reviewer Qmti - Q6

Q6 How should the substantially lower SSIM despite improved FVD be interpreted, and does this trade-off affect sample-specific genesis timing? What can the IDOL results establish about downstream usefulness?

Response

SSIM and FVD. Lower FVD is relevant to our task because genesis forecasting requires representing the progressive organization of a developing cloud system, whose evolution also supports downstream diagnosis. FVD assesses distributional similarity between forecast and observed cloud developments at the sequence level, complementing the local structural correspondence measured by SSIM. Under uncertain precursor evolution, a realistic development need not reproduce the exact observed cloud arrangement, so these metrics can favor different methods. We therefore seek realistic cloud development alongside useful event predictions. The lower SSIM still indicates weaker correspondence to the particular observed future, and FVD alone cannot establish sample-specific genesis accuracy; the added timing and position results provide the direct evidence for that task.

Sample-specific genesis timing. Across 420 eligible evaluations, CycloDiff has a timing MAE of 3.66 hours and 8 missed events, compared with 5.46–5.89 hours and 14–161 misses for the evaluated baselines. This is evidence of useful event identification despite the lower SSIM. Because MAE uses each model’s detected subset, the error and missed count must be considered together; the result does not imply that sample-specific reconstruction errors are unimportant.

Downstream usefulness. IDOL provides supporting evidence that the forecast clouds retain information relevant to estimating storm intensity and size, both of which are important for the study of tropical cloud clusters and tropical cyclones. The same frozen estimator is applied to forecast and observed clouds, with infrared-channel replication and fixed auxiliary inputs. Agreement therefore establishes estimator consistency under this protocol, rather than independent accuracy.

AuthorsOpenReview ↗

Response to Reviewer Qmti - Q7

Q7 Please clarify the sampling, CRPS and FVD protocols, frame spacing, validation split, ensemble size, sampling cost, and uncertainty estimates. Do all baselines have comparable access to ocean conditioning and other inputs?

Response

Sampling frame spacing and ensemble size. We use four historical observations at a three-hour interval. The original model predicts five future cloud fields, and longer forecasts use sliding-window rollout. The archived configuration uses 200 denoising steps per sampling call and retains one realization per input. Random forecast initialization and stochastic diffusion sampling are distinct sources of variation; a single realization does not provide a forecast ensemble.

CRPS. With one retained sample, empirical CRPS reduces to absolute error: the pairwise ensemble term is zero. The reported value is averaged over the normalized cloud field and forecast times. It therefore measures single-member forecast error in this evaluation and does not establish ensemble calibration or probabilistic reliability.

FVD protocols. We compute FVD from I3D features of generated and reference sequences, replicating infrared images into three channels. The reproduced eight-frame value of approximately 132 corresponds to independent min–max normalization of prediction and reference arrays. For five- and eight-frame forecasts, every frame is repeated three times before I3D extraction to meet its temporal-length requirement. The short-horizon evaluation repeats frames for lengths three through eight, whereas nine- and ten-frame sequences use their native length; FVD is not reported for one or two frames. Thus, forecast length and temporal preprocessing change across some horizon comparisons, and these FVD values must not be interpreted as a strictly controlled horizon-only comparison. Within a given comparison, normalization and temporal processing must be identical across methods.

Baseline access to ocean conditioning. The original baselines do not all receive the same depth-resolved ocean inputs. We therefore add an AlphaPre experiment with ocean conditioning as an input control. Conditioning improves its SSIM from 0.4356 to 0.4573 and lowers FVD from 947 to 784 and CRPS from 0.1410 to 0.1273. CycloDiff retains lower FVD and CRPS, while AlphaPre with ocean conditioning has higher SSIM. This control shows that adding ocean inputs to this baseline does not close the observed gaps, but it does not make every baseline input-matched.

SettingsSSIM↑PSNR↑LPIPS↓CRPS↓FVD↓
AlphaPre0.435615.700.05150.1410947
AlphaPre + ocean0.457316.210.04760.1273784
CycloDiff0.256016.560.02130.1013132

Validation split. The cloud-training implementation does not reserve a separate validation partition: the evaluation loader is also used for periodic loss monitoring, and checkpoints are saved by iteration. This does not prevent optimization of the training objective. However, an independent validation set would provide a cleaner basis for checkpoint selection and monitoring generalization. We therefore disclose this limitation rather than claim that validation is unnecessary for diffusion models.

Sampling cost and uncertainty estimates. An archived A6000 run generated five-frame forecasts for 84 inputs in 172.6 seconds for the complete sampling run; this is not a measured single-forecast latency. The available results use one retained realization per input and do not include repeated-seed confidence intervals. They should be interpreted as point estimates. Because multiple initializations share a storm, any future uncertainty interval should account for storm-level clustering rather than treat all 420 evaluations as independent.

Reviewer_Hat9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

CycloDiff is a multi-task framework for tropical cyclone genesis forecasting. It uses residual conditional video diffusion, conditioned on historical infrared images and depth-resolved ocean temperatures, to predict future cloud sequences. Features from the frozen cloud-ocean encoder are transferred to an adapted MMSTN trajectory predictor. A pretrained classifier then detects a persistent transition in the generated frames to infer genesis time; the corresponding predicted trajectory point supplies genesis location. Experiments on held-out tropical cyclone cases show strong cloud-sequence results on FVD, LPIPS, CRPS, and PSNR, with ablations for ocean-temperature conditioning and residual prediction. The final joint genesis forecast is illustrated by a VAMCO case study.

Critical strengths
  1. The work moves beyond static genesis classification to a concrete pipeline for forecasting cloud evolution, trajectories, and a genesis event. This is a potentially useful application setting.

  2. Using subsurface temperatures addresses a limitation of SST-only inputs. Table 2 provides a relevant ablation, and the evaluation includes temporal/perceptual and meteorological diagnostic measures rather than image similarity alone.

Critical weaknesses
  1. The extensive results largely concern cloud-sequence quality. Genesis time and location are demonstrated for VAMCO, however, dataset-level timing/location errors and test-set trajectory metrics are not reported.

  2. The trajectory branch uses a frozen historical cloud-ocean feature, not generated future cloud frames, and the branches are trained sequentially. No ablation directly establishes the benefit of feature transfer for trajectory or genesis-location accuracy. It therefore remains unclear whether the proposed integration adds predictive value beyond a modular pipeline.

  3. Forecast starting points are randomly sampled, but the paper does not state whether they precede the true genesis event or whether genesis lies within the forecast horizon. Furthermore, the pretrained ZPR classifier, its threshold η, the persistence length M, and the threshold semantics in Table 6 are not described sufficiently to assess or reproduce event-level timing predictions.

Questions for authors
  1. Can the authors report genesis-time MAE, genesis-location distance error, and test-set trajectory ADE/FDE over all eligible storms, together with event-detection success and failure counts?

  2. How much does the transferred cloud-ocean feature improve trajectory and genesis-location prediction relative to the original MMSTN or a version without this feature? What precisely is the “relationship” being modeled?

  3. Are all forecast initialization times before genesis, and how are windows without a genesis event handled? Please specify the pretrained classifier, η, M, and the definition of ZPR and the thresholds in Table 6.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Hat9 - Q1、Q2

Q1 Can the authors report genesis-time MAE, genesis-location distance error, and test-set trajectory ADE/FDE over all eligible storms, together with event-detection success and failure counts?

Response

We have added the corresponding results across 420 eligible forecast initializations over all eligible 84 storms cases from 2019--2022. As for genesis-time MAE and event-detection counts, comparing with another methods, our CycloDiff perform the best. CycloDiff detects 412 events and misses 8, giving a detection rate of 98.1%. Its genesis-time MAE is 3.66 hours on detected events, and its mean position error at the reference genesis time is 67.34 km. The 420 evaluations comprise five initializations per storm, rather than 420 independent storms.

ModelTiming MAE ↓ (h)Undetected ↓
PredRNN5.8918
PhyDNet5.7498
PreDiff5.46161
EarthFormer5.7184
DiffCast5.7135
CasCast5.7014
AlphaPre5.8967
CycloDiff3.668

As for genesis-location distance error, and test-set trajectory ADE/FDE, the corresponding results are reported in the following table.

MetricCycloDiff
Genesis-time MAE on detections ↓3.66 h
Position error at reference genesis time ↓67.34 km
Trajectory ADE ↓53.90 km
Trajectory FDE ↓107.98 km

Q2 What precisely is the “relationship” being modeled? How much does the transferred cloud-ocean feature improve trajectory and genesis-location prediction relative to a version without this feature?

Response

The “relationship” being modeled. With the aim to generate future spatiotemporal evolution of cloud systems along with the corresponding genesis time and location, we design a multi-task framework. In details, genesis forecasting is not merely a binary decision about whether a tropical disturbance will develop. A practically informative forecast should also determine when and where genesis will occur while representing how the underlying cloud system evolves toward an organized cyclone. Thus, our formulation connects the evolving cloud structure with explicit genesis-time and location estimates in which cloud evolution, genesis time, and genesis location are intrinsically connected. Further, the predicted cloud evolution can also support downstream meteorological diagnosis. These complementary requirements motivate the architecture.

Impact of applying cloud-ocean feature. After incorporating the cloud–ocean feature, a historical cloud–ocean representation is extracted from the cloud-condition encoder and, after global average pooling and projection, injected into the trajectory encoder and decoder. This provides the trajectory and genesis-location prediction with a compact representation of cloud-system structure and evolution trends, as well as subsurface ocean thermal information. Trajectory Error with ADE, FDE, and the error from +3 h to +15 h, as wall as the genesis-location error illustrate the impact of applying cloud-ocean feature.

Trajectory Error

ModelADE ↓FDE ↓+3 h+6 h+9 h+12 h+15 hgenesis-location error
w/o cloud–ocean feature56.2108.1710.8528.5753.3579.96108.1772.61
CycloDiff53.9107.988.6325.4250.1677.54107.9867.34

All distances are in km.

AuthorsOpenReview ↗

Response to Reviewer Hat9 - Q3

Q3 Are all forecast initialization times before genesis, and how are windows without a genesis event handled? Please specify the pretrained classifier, η, M, and the definition of ZPR and the thresholds in Table 6.

Response

Forecast initialization times are randomly sampled and are not constrained to precede genesis, so a reference genesis event is not guaranteed to lie within every forecast horizon. For the classifier-based event evaluation, we use eligible positive windows containing a reference genesis event within the horizon. Windows without a new genesis event, including post-genesis windows, are excluded from this positive-event timing evaluation. Within the eligible cohort, a forecast without a qualifying event is counted as a missed detection, rather than discarded.

For pretrained classifier, we use the frozen binary CrossViT classifier (crossvit_small_224), with infrared images replicated into three channels, to identify genesis in predicted cloud development and compute ZPR. A frame is positive when its positive-class probability exceeds η = 0.5; ties are negative. We require M = 3 consecutive positive frames at three-hour intervals and assign genesis to the first frame of the earliest qualifying run fully contained within the forecast horizon. If no such run exists, the eligible event is counted as undetected and no genesis time or location is assigned. ZPR summarizes these classifier-based decisions as the fraction of eligible positive windows with a detected event within an absolute timing-error tolerance τ. In the expression below, N includes all eligible positive windows, including missed detections; Dᵢ = 1 indicates detection, and t̂ᵢ and tᵢ denote the predicted and reference genesis times. The Table 6 thresholds 0, 3, 6, 9, 12, and 15 are tolerances in hours, distinct from η, M, and forecast length. Zero requires an exact match on the evaluation time grid. This positive-window evaluation measures event detection and timing, rather than false alarms in windows without genesis.

ZPRM(τ)=1N∑i=1NI[Di=1∧∣t^i−ti∣≤τ]\mathrm{ZPR}_{M}(\tau)=\frac{1}{N}\sum_{i=1}^{N}\mathbb{I}\left[D_i=1\land\left|\hat{t}_i-t_i\right|\le\tau\right]

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

This paper presents a conditional diffusion model for tropical cyclone genesis and path forecasting.

Critical strengths

The proposed method addresses an important application beyond global weather modeling. The proposed method seems to be sound. But there are some issues need to be addressed. See the next section.

Critical weaknesses

Intuitively the proposed architecture and method makes sense. However the author(s) should provide more details on the rationale of such a choice.

Out of three sets of input (cloud images, ocean-temperature fields, and historical storm-center trajectory), the author(s) only performed two ablation studies on the first two sets of input. I believe the historical trajectory should be included

Questions for authors
  1. As discussed above, an ablation study on the historical trajectory would provide more insight to the effectiveness of the proposed method, since the cyclones often follow historical ensembles. The inclusion of historical trajectory may introduce a strong inductive bias, which can be revealed by an ablation study.

  2. I am not convinced the experiments have strong "signal" so that the model are not trained to learn a weak case. For example, I had expected to see a strong performance difference between 5 frames and 8 frames. Would it be possible to provide results for 1, 3, 5, and 10 frames?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Zfyk - Q1、Q2

Q1 Intuitively the proposed architecture and method makes sense. However the author(s) should provide more details on the rationale of such a choice.

Response

Rationale of the proposed architecture and method. With the aim to generate future spatiotemporal evolution of cloud systems along with the corresponding genesis time and location, we design a multi-task framework. In details, genesis forecasting is not merely a binary decision about whether a tropical disturbance will develop. A practically informative forecast should also determine when and where genesis will occur while representing how the underlying cloud system evolves toward an organized cyclone. Thus, our formulation connects the evolving cloud structure with explicit genesis-time and location estimates in which cloud evolution, genesis time, and genesis location are intrinsically connected. Further, the predicted cloud evolution can also support downstream meteorological diagnosis. These complementary requirements motivate the architecture.

Rationale of applying cloud images and ocean-temperature fields. To represent uncertain precursor development while maintaining consistency across forecast times, we use conditional diffusion to predict future cloud structures. Residual prediction takes the latest observed cloud field as a structural reference, allowing the model to focus on subsequent changes rather than reconstruct persistent structures from scratch. This choice is supported by the residual ablation, which reduces FVD from 183 to 132. To distinguish ocean columns with similar sea surface temperature (SST) but different subsurface thermal structures, we incorporate depth-resolved ocean temperatures; replacing SST-only conditioning with these inputs reduces FVD from 193 to 132.

Rationale of providing genesis time and location. To connect precursor development with geographic motion, we condition the MMSTN trajectory branch on historical positions, local ocean inputs, and encoded historical cloud–ocean information. Past positions describe displacement, while the transferred representation supplies cloud organization and thermal context. A fixed classifier with a persistence rule identifies genesis in the predicted cloud development, and the corresponding forecast trajectory position supplies its location. The generated cloud fields are also evaluated with the fixed IDOL model to assess whether they preserve information relevant to storm intensity and size. These components connect precursor development with explicit when-and-where estimates and downstream meteorological diagnosis.

Q2 An ablation study on the historical trajectory would provide.

Response

Following the reviewer’s meaningful suggestions, we added an ablation study on cloud images, ocean-temperature fields, and historical trajectory. As shown in the following table, cloud-prediction perform better when historical trajectory information is included.

SettingsSSIM↑PSNR↑LPIPS↓CRPS↓FVD↓
w/o ocean temperature0.236315.340.02480.1128197
w/o trajectory0.244116.530.02320.1067157
CycloDiff0.256016.560.02130.1013132
AuthorsOpenReview ↗

Response to Reviewer Zfyk - Q3

Q3 I had expected to see a strong performance difference between 5 frames and 8 frames. Would it be possible to provide results for 1, 3, 5, and 10 frames?

Response

Performance between 5 frames and 8 frames. We first clarify that the performance of 5 frames and 8 frames without strong difference reports the robustness of our predicting model. We forecast the observed cloud structure, preserving information relevant to the transition toward genesis. This task formulation helps explain the sustained quality over nearby forecast horizons. The short-range advantage remains compatible with increasing forecast difficulty at longer lead times, as the expanded results show.

Results for 1 3 5 and 10 frames. We report all forecast lengths from one to ten frames, including the requested settings. Two training configurations are used. For forecast lengths of 1–4 frames, we train a model that takes four observations as input and predicts a single future frame; at inference time, we apply autoregressive rollout to obtain the required forecast length. For forecast lengths of 5–10 frames, we use the original model trained to predict five future cloud fields from four observations; at inference time, forecast lengths beyond five frames are obtained via sliding-window rollout. As shown in the table, our method consistently maintains better performance across different forecast lengths.

Each cell reports CycloDiff / AlphaPre.

FramesSSIM↑PSNR↑LPIPS↓CRPS↓FVD↓
10.2878 / 0.487818.18 / 16.170.0121 / 0.03190.0795 / 0.0985N/A / N/A
20.2646 / 0.473417.45 / 15.830.0144 / 0.03820.0872 / 0.1091N/A / N/A
30.2554 / 0.459417.04 / 15.240.0161 / 0.04130.0921 / 0.1135172 / 1191
40.2478 / 0.422916.72 / 15.010.0176 / 0.04330.0965 / 0.1212239 / 1224
50.3060 / 0.538119.29 / 16.220.0185 / 0.03380.0727 / 0.0920125 / 1487
60.2684 / 0.521716.83 / 16.030.0193 / 0.03910.0980 / 0.1112129 / 1231
70.2572 / 0.465216.53 / 15.980.0198 / 0.04110.1014 / 0.1264126 / 977
80.2560 / 0.435616.56 / 15.700.0213 / 0.05150.1013 / 0.1410132 / 947
90.2398 / 0.419416.25 / 14.760.0218 / 0.05780.1057 / 0.1426199 / 1072
100.2376 / 0.406716.11 / 14.370.0225 / 0.06010.1084 / 0.1445248 / 1003

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.