- 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
-
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..
-
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.
-
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.
-
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.
-
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
- 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.
- 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?
- 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?
- 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
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%.
| Metric | CycloDiff | Direct predictor |
|---|---|---|
| Genesis-time MAE ↓ | 3.66 h | 10.17 h |
| Position error at reference genesis time ↓ | 67.34 km | 103.25 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.
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.
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.
| Settings | SSIM↑ | PSNR↑ | LPIPS↓ | CRPS↓ | FVD↓ |
|---|---|---|---|---|---|
| AlphaPre | 0.4356 | 15.70 | 0.0515 | 0.1410 | 947 |
| AlphaPre + ocean | 0.4573 | 16.21 | 0.0476 | 0.1273 | 784 |
| CycloDiff | 0.2560 | 16.56 | 0.0213 | 0.1013 | 132 |
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.
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.
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. R denotes the pearson correlation coefficient between estimates obtained from generated and ground-truth cloud images.
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.