- Rating
- 2
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper forecasts 8-band geostationary satellite imagery (Himawari-8/9 and GOES-16) three hours ahead at 30-min steps. CloudFlow has two stages. An Evolution Network first warps the last frame with a predicted velocity field and adds a learned source term, in a semi-Lagrangian style. A flow-matching model then starts from this coarse forecast and refines it with a few deterministic Euler steps. The paper reports the best MSE and FVD on both datasets, as well as results on a cross-region test set and several robustness tests.
Critical strengths
- Forecasting multispectral satellite imagery is useful where radar coverage is poor, and the method is tested on two satellites and an additional region.
- Splitting the evolution into motion and source terms makes the model easy to inspect, and the learned speeds look plausible (Fig. 12).
- The refined forecasts are much sharper than those of the baselines (lowest FVD), at far lower cost than MCVD and PreDiff.
- The appendix includes per-band results and tests on seasons, input noise and missing frames.
Critical weaknesses
-
The task is satellite image forecasting, not precipitation nowcasting, yet the paper is motivated by extreme weather and rapid convection. At 2–4 km and 30-min steps, with daytime data only, it is unclear how much convective-scale information is left, and summer is the worst season (Fig. 22). The evaluation relies on image metrics only. MSE/MAE are averaged over bands with different units, there are no event-based scores, and no variance over runs is reported.
-
Baseline training is not described at all (adaptation to 8 bands, epochs, tuning), and some numbers suggest under-trained baselines: MCVD reaches MSE 496.6 and 1274, and PreDiff 227.3 on GOES-16. Even the simplest stage-1 variant (SSD only in Table 2, MSE 53.73, SSIM 0.682) beats every baseline in Table 1. CloudFlow also has 193.7M parameters, while most baselines have under 35M.
-
Both stages largely follow existing work. An evolution network that predicts a motion field and an intensity residual with semi-Lagrangian advection, followed by a generative refiner conditioned on its output, is very close to NowcastNet, which is not cited. Refining a deterministic nowcast with a flow has also been explored before (e.g., PercpCast), and image restoration methods such as InDI and I2SB also start the flow from a degraded estimate. PA-FM itself is built from standard parts.
-
The physics part is weakly supported. Nothing is actually constrained: S is free and the velocity field is not regularized. The physically motivated DFI ordering gives higher MSE than the reversed ordering (44.55 vs. 43.76, Table 12), and α barely changes the results (Table 13), since S can absorb it. No ablation replaces the Evolution Network with a generic deterministic prior, so it is unclear whether the physics matters for the final model.
-
The refinement stage trades accuracy for sharpness. In Table 11, MSE rises from 45.89 at N=1 to 53.87 at the default N=10 and keeps increasing with N. Table 2 also suggests that the Evolution Network alone reaches 44.55. This contradicts the claim that CloudFlow breaks the perception–accuracy trade-off. In addition, the output is deterministic and gives no uncertainty estimate, and the model fails when the latest frame is missing .
-
The efficiency analysis only covers parameters and GFLOPs (Table 17). CloudFlow needs 3–7.5× the GFLOPs of SimVP, SimVPv2 and Earthformer for small gains (e.g., 0.7% lower MSE than SimVP). The 812× saving over PreDiff assumes 1000 sampling steps for PreDiff. Latency, memory and training cost are not reported. Table 10 lists CFG, which the text never mentions and which would double the refinement cost if used at inference.
Questions for authors
- How were the baselines adapted to and trained on the 8-band data?
- Can you report the stage-1 prediction in Table 1 under the same protocol, and state the evaluation setting of each ablation?
- What happens if PA-FM starts from a generic deterministic prior (e.g., SimVP) instead of the Evolution Network?
- How does CloudFlow compare with NowcastNet and with deterministic-plus-generative refiners such as DiffCast, CasCast or PercpCast?
- Is CFG used at inference? Please also report latency, memory and training time, and explain how N was chosen for each dataset.
- Can you add event-based scores (e.g., CSI on cold cloud-top thresholds) and results at night? If these points are addressed, I am willing to raise my score.
6 comments
Rebuttal Response 2/6
3. Prior work, modern comparators, and model capacity (W2, W3, Q4)
Following your suggestions, we added NowcastNet, DiffCast, CasCast, and PercpCast comparisons; our CloudFlow has lower pixel errors than these complete refiners (Table 1). The revision credits their designs and clarifies the differences.
Differences from NowcastNet's evolution network (Section 3.1 and Appendices B and D.3):
- Motion–intensity coupling. NowcastNet’s Evolution Network predicts displacement and intensity residuals through parallel decoders with shared encoder features. Our Evolution Network also uses DFI to supply the intensity decoder with speed, deformation, vorticity, and divergence features derived from the predicted displacement field.
- Displacement field decomposition. NowcastNet’s motion decoder directly predicts a dense displacement field. Ours combines global translation and rotation from a rigid-motion head with a dense residual from a local decoder.
- Gradient input for cloud-boundary. NowcastNet’s encoder takes historical frames as input. Ours also takes their spatial gradients, giving both decoders explicit cloud-boundary cues.
- Attention gate skip connections. NowcastNet’s decoders concatenate encoder features directly with upsampled decoder features. Our local displacement decoder uses attention gates to select encoder features before fusion.
- Evolution update. NowcastNet’s Evolution Network uses nearest-neighbor warping during recursion and bilinear warping for supervision. Ours uses bilinear warping during recursion and refines each updated frame with a 5×5 depthwise convolution followed by a 1×1 convolution.
At the same base width of 32 base channels and training budget of 100 epochs, our Evolution Network outperforms NowcastNet$^{\mathrm{S1}}$ on MSE, MAE, PSNR, FVD, and cold-cloud CSI (Table 19). These gains persist after matching width, although our model uses more parameters and GFLOPs and has slightly lower SSIM.
| Model | MSE | MAE | PSNR | SSIM | FVD | Params (M) | GFLOPs | |
|---|---|---|---|---|---|---|---|---|
| Evolution Network | 47.18 | 3.373 | 24.14 | 0.683 | 111.4 | 0.490 | 11.97 | 60.74 |
| NowcastNet$^{\mathrm{S1}}$ | 51.98 | 3.530 | 23.79 | 0.690 | 124.7 | 0.458 | 8.89 | 51.62 |
Relative to existing refinement methods (Introduction and Section 3.2 / 5.1):
- DiffCast/CasCast. These cascades refine deterministic forecasts with diffusion; our PA-FM integrates conditional flow from the evolution forecast itself.
- PercpCast. Both start from deterministic forecasts. Our PA-FM jointly processes leads with clean-endpoint parameterization; PercpCast refines leads separately with LPIPS supervision. Figure 3 compares their step-dependent behavior.
- InDI/I2SB. Following your literature suggestions, the Section 3.2 Physics now cites InDI [1] and I2SB [2] as informative-input image-restoration precedents.
Selected Himawari-8/9 results from Table 1 follow.
| Model | MSE | MAE | PSNR | SSIM | FVD | Rank | |
|---|---|---|---|---|---|---|---|
| SimVP | 54.25 | 4.171 | 23.71 | 0.637 | 264.4 | 0.388 | 7 |
| NowcastNet$^{\mathrm{S1}}$ | 51.98 | 3.530 | 23.79 | 0.690 | 124.7 | 0.458 | 3 |
| PercpCast$_{N=2}^{\ddagger}$ | 74.12 | 4.855 | 22.91 | 0.635 | 52.15 | 0.481 | 6 |
| DiffCast$^{\dagger}$ | 70.05 | 4.309 | 22.51 | 0.608 | 29.11 | 0.424 | 8 |
| CasCast$^{\dagger}$ | 70.86 | 4.680 | 22.66 | 0.615 | 90.92 | 0.389 | 13 |
| NowcastNet | 67.10 | 4.821 | 22.83 | 0.626 | 692.3 | 0.436 | 16 |
| CorrDiff | 64.74 | 4.443 | 22.89 | 0.621 | 24.90 | 0.479 | 4 |
| CloudFlow$^{\mathrm{S1}}$ | 44.55 | 3.296 | 24.42 | 0.701 | 130.4 | 0.504 | 1 |
| CloudFlow$_{N=10}$ | 53.87 | 4.129 | 23.79 | 0.672 | 67.56 | 0.491 | 2 |
: first stage; : listed SimVP backbone; : official Phy-DNet. Ranks follow full Table 1. Newly added stochastic models use the mean of separately evaluated scores over three sampling seeds, not ensemble-mean predictions.
The added shared-backbone comparison shows our PA-FM best preserving MSE/PSNR/SSIM/CSI; DiffCast leads MAE/FVD (Table 2). Refiners use the same Evolution Network and 200 epochs, with architectures and compute differing. DiffCast and PercpCast attain lower FVD.
Rebuttal Response 3/6
4. Physics support, DFI, α, and a generic prior (W4, Q3)
To clarify, our Evolution Network is inspired by physical advection, without claiming physical consistency (Section 3.1; Appendix H). Its displacement field and source term are learned operator intermediates optimized for prediction.
- DFI ordering. The proposed order improves MAE/SSIM; reversing it lowers MSE (Table 14). This supports an empirical compromise, not a unique physical ordering.
- Source scale. Sensitivity to α initialization is modest (Table 15). A small initial value favors early transport, but the unconstrained source branch can absorb scale changes, precluding guaranteed sparsity or physical bounds.
- Motion diagnostics. To address your physics concern, our added comparison with operational Himawari-9 atmospheric motion vectors (AMVs) supports the plausibility of the learned displacement fields. Displacement-derived velocities show broad directional agreement and lower RMSVD than zero motion in all four selected 10.4 μm cases (Appendix E.6; Table 17; Figures 19–23).
| Date | RMSVD (m/s) | Zero-motion RMSVD (m/s) | Median angular error (°) | Speed bias (m/s) |
|---|---|---|---|---|
| 2023-01-15 | 10.858 | 23.444 | 8.290 | -1.776 |
| 2023-06-15 | 13.608 | 29.237 | 5.063 | -4.172 |
| 2023-07-14 | 13.281 | 17.754 | 11.405 | -6.798 |
| 2023-10-15 | 10.065 | 12.548 | 16.846 | -7.089 |
Speed magnitudes remain underestimated.
- Source interventions. Added interventions show lower 10.4 μm MAE when retaining the source than removing/reversing it in each selected case, supporting post-transport correction (Table 18; Appendix E.7). Weights and motion are fixed.
Thank you for suggesting SimVP as a prior; in the added test, our Evolution Network + PA-FM performs better on all six metrics under the same refiner epoch count and inference steps. Our PA-FM lowers SimVP's FVD, with little CSI change and poorer pixel and structural scores. This supports the chosen prior's value in the complete system, while first-stage architecture, capacity, and forecast quality remain part of the comparison.
Table 20 (Appendix E.8) uses 200 refiner epochs and for both priors; our Evolution Network retains its default base width 64.
| Model | MSE | MAE | PSNR | SSIM | FVD | |
|---|---|---|---|---|---|---|
| SimVP | 54.25 | 4.171 | 23.71 | 0.637 | 264.4 | 0.388 |
| SimVP + PA-FM | 73.96 | 4.889 | 22.73 | 0.618 | 112.2 | 0.387 |
| Evolution Network | 44.55 | 3.296 | 24.42 | 0.701 | 130.4 | 0.504 |
| Evolution Network + PA-FM | 56.41 | 4.437 | 23.60 | 0.664 | 83.49 | 0.490 |
5. Refinement trade-off and deterministic output (W5)
Our CloudFlow model offers a favorable empirical perception–accuracy operating point among the evaluated models (Figure 1). Against SimVP, it maintains similar MSE with substantially lower FVD; against CorrDiff, it has lower MSE but higher FVD.
Refinement sacrifices pixel accuracy for lower FVD, as the first-stage rows in Table 1 show. Figure 1's connecting lines are comparisons, not a proven bound; refinement does not eliminate the accuracy–perception trade-off.
Our CloudFlow model targets deterministic prediction, with fixed initialization and integration yielding reproducible outputs. Flow matching describes the training objective, not a claim of probabilistic forecasting. The lack of uncertainty estimates limits risk-sensitive applications but is distinct from performance on the stated deterministic task. Point-mass CRPS equals MAE and does not measure uncertainty calibration (Appendix C).
Rebuttal Response 1/6
Dear Reviewer KZFz,
We sincerely thank you for your careful review and for recognizing the forecasting task, inspectable evolution components, and sharper predictions. To further address your concerns, we provide the following clarifications and additional experimental evidence. All changes described below have been incorporated into the revised manuscript. All section, table, and figure references below refer to this revised version.
1. Baseline adaptation, training, and convergence (W2, Q1)
Thank you for suggesting that we provide implementation details. Appendix D.3 now documents baseline adaptation and training. All baselines share evaluation data and forecast targets: six eight-band frames from six historical frames at 256 × 256. Adaptations differ as follows.
Internal resolution, context, capacity, and budgets differ. The ten original deterministic baselines use 100 epochs, AdamW, effective batch size 64, and warm-up/cosine decay; Appendix D.3 supplies model-specific settings.
To address your concern of insufficient training of PreDiff and MCVD, we added loss curves (Figures 9–10), which flatten substantially on both datasets without establishing optimal tuning. On GOES-16, MCVD is trained for 1,000,000 steps; PreDiff’s VAE/diffusion trained 250/480 epochs. These figures show that the losses have converged.
The source-sink-only comparison alone is insufficient to assess baseline convergence. NowcastNet$^{\mathrm{S1}}$ has lower MSE than this variant, while added competitive refiners extend the comparison beyond MCVD and PreDiff.
2. First-stage results and ablation settings (Q2)
Thank you for suggesting that we report first-stage results. Table 1 now includes our CloudFlow$^{\mathrm{S1}}$ under the full-model protocol. Section 3 gives Himawari results; ablations follow.
Our added common-width test favors our Evolution Network over NowcastNet$^{\mathrm{S1}}$ in MSE, MAE, PSNR, FVD, and CSI (Appendix E.8, Table 19). Both use 32 base channels and 100 epochs.
Variant V leads pixel/structural metrics; II has slightly lower FVD. II/V and IV/V change both inputs, so neither isolates initialization nor establishes a conditioning-weight trend.