acceptodds
Under review as a conference paper at ICLR 2027

Forecast the Envelope, Reconstruct the Details: Learning Dynamic Spectral Priors for Precipitation Nowcasting

Abstract

Precipitation nowcasting supports timely flood warnings and emergency response by forecasting the location and intensity of imminent rainfall. However, rapidly evolving echo details remain difficult to predict, and deterministic models often smooth intense local structures as lead time increases. Our analysis reveals that spectral envelopes formed by locally averaging wavelet detail magnitudes exhibit higher temporal coherence than signed coefficients averaged over the same regions. These envelopes summarise the spatial distribution and strength of local variations at multiple scales. Motivated by this observation, we propose DySP (Dynamic Spectral Prior), a deterministic framework that forecasts spectral envelopes to guide the reconstruction of future radar fields. A spectral pathway models multiscale envelope dynamics and produces latent spectral priors and envelope predictions for each forecast lead time. Our prior-guided spatiotemporal block (PGSTB) conditions spatial attention queries and keys on the latent priors and gates the attention output using envelope embeddings. To recover local structures not retained by the pooled envelopes, temporal-difference skip connections enrich the decoder with high-resolution observation features and their recent changes. Experiments on Shanghai, CIKM, and SEVIR demonstrate that DySP achieves state-of-the-art forecasting skill, improving the prediction of intense precipitation and the preservation of local echo structures.

Primary area:
learning on time series and dynamical systems
Keywords:
Precipitation nowcasting, Spatiotemporal prediction, Wavelet representations, Spectral envelopes, Dynamic spectral priors

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_Nq9eOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies precipitation nowcasting and proposes DySP, a spectral prior-guided framework that exploits the temporal coherence of spectral envelopes to improve future radar frame prediction. The method combines wavelet-based spectral representation, latent spectral prior modeling, future envelope prediction, and envelope-conditioned attention to guide content reconstruction. In addition, temporal-difference skip connections are introduced to preserve evolving spatial structures across forecast lead times. Experiments on three benchmarks demonstrate competitive forecasting performance compared with existing methods. However, some issues remain regarding the justification of the spectral envelope representation, the effectiveness of future prior prediction, the actual contribution of spectral envelopes, and the evaluation of computational efficiency. Further clarification and validation are needed to strengthen the paper's contribution.

Critical strengths
  1. The paper introduces a spectral envelope representation based on wavelet detail coefficients to characterize the temporal evolution of precipitation structures, providing an alternative perspective for modeling spectral dynamics in precipitation nowcasting.
  2. The method combines spectral guidance with temporal-difference skip connections to capture spectral dynamics while preserving evolving spatial structures during multi-step forecasting.
  3. The proposed method is evaluated on three precipitation nowcasting benchmarks, and the experimental results demonstrate competitive performance compared with existing methods.
Critical weaknesses
  1. Although spectral envelopes exhibit higher temporal correlations than signed wavelet detail coefficients, this does not necessarily demonstrate their effectiveness in improving precipitation forecasting. The authors should provide a more theoretical explanation of how spectral envelopes capture precipitation dynamics relevant to future prediction and clarify its advantages over existing frequency-domain methods.
  2. The future spectral prior is generated through linear temporal projection, and the future envelope is subsequently derived from the predicted prior. However, the suitability of linear projection for modeling complex precipitation dynamics remains insufficiently justified. The authors should compare it with more expressive temporal modeling methods and analyze how prior and envelope prediction errors affect final forecasting performance.
  3. Although spectral envelopes are presented as a central motivation in introduction section, their role in the proposed architecture is mainly limited to attention gating and auxiliary supervision. The authors should provide stronger evidence that spectral envelopes contribute meaningful information beyond latent priors and conventional attention modulation.
  4. Although the proposed method achieves competitive forecasting performance, its computational cost is not sufficiently evaluated. It would be helpful if the authors could provide the comparison of computational cost, such as model parameters and FLOPs.
  5. Although the Related Work section discusses several frequency-domain methods, the experimental comparison is limited to AlphaPre and SDIR. The authors are encouraged to include additional frequency-domain forecasting methods and provide a more comprehensive comparison to clarify the advantages of the proposed spectral envelope representation over existing methods.
Questions for authors
  1. The paper shows that spectral envelopes exhibit higher temporal correlations than signed wavelet detail coefficients. Can the authors clarify whether this property translates into improved forecasting performance, beyond the benefits of the latent spectral priors and attention modulation?
  2. The future spectral priors are generated through linear temporal projection. Can the authors clarify the rationale for this design and provide evidence that it can adequately capture the complex temporal evolution of precipitation systems?
  3. The paper emphasizes spectral envelopes as a key motivation, but their role in the architecture appears mainly limited to attention gating and auxiliary supervision. Can the authors clarify how the envelope representation contributes to the final forecasting results beyond these mechanisms?
2 comments

Thank you for the comments on the spectral representation and temporal modelling.

W1 & Q1. Why envelopes help, and how they differ from other frequency-domain methods

Envelopes average Haar-detail magnitudes over local windows. A constant patch has zero detail response regardless of its mean intensity; boundaries and intensity changes produce responses at the relevant scales and orientations. Taking magnitudes avoids cancellation between opposite signs, while pooling reduces sensitivity to position within each window. Fourier amplitudes describe the whole field; envelopes preserve where detail activity occurs. DySP predicts this map to guide reconstruction, while SDIR uses spectrally filtered predictions for iterative refinement.

In Table 4, the backbone's mean CSI is 0.4283. Adding spectral guidance alone gives 0.4516, temporal differences alone give 0.4545, and both give 0.4613. In Table 5, mean CSI is 0.4516 with envelope supervision, 0.4468 without it, and 0.4457 with pooled-intensity guidance. Envelope guidance also has higher CSI at every threshold, higher SSIM, and lower MAE than pooled-intensity guidance.

W2 & Q2. Linear temporal projection

Spectral Dynamics applies nonlinear spatial and temporal attention before TP\mathcal T_P (Sec. 3.1). For each forecast lead, the projection learns to combine these historical features; it initially repeats the last state. Predicted envelopes have higher correlation with future targets than persistence: 0.560 vs. 0.478 on SEVIR, 0.569 vs. 0.532 on Shanghai, and 0.528 vs. 0.496 on CIKM (App. D.3).

Prior errors can change attention and feature updates through the Q/K offsets, gates, and shifts in PGSTBs (Eqs. 4–5). The predicted envelopes are auxiliary outputs. Both envelope and radar losses train the priors (Eq. 12). Replacing forecast guidance with last-frame guidance lowers mean CSI by 0.0165/0.0310/0.0161 on CIKM/Shanghai/SEVIR (Table 6).

W3 & Q3. Role of envelopes beyond gating and auxiliary supervision

The Wavelet Encoder takes detail magnitudes as input. Lprior\mathcal L_{\rm prior} supervises the historical and future priors, and Lwav\mathcal L_{\rm wav} compares pooled detail magnitudes of the final forecast. The priors control Q/K offsets, output gates, and additive shifts in all four PGSTBs of the second stack (Eqs. 4–5).

With envelope targets fixed, magnitude input gives mean CSI 0.4516, compared with 0.4350 for pixel input and 0.4308 for signed wavelets (Table 5). Using both conditioning routes gives 0.4516, versus 0.4418 for gate-and-shift alone and 0.4362 for Q/K alone. Removing the envelope loss leaves the magnitude input and both routes unchanged, but lowers mean CSI to 0.4468 and CSI at 40 dBZ from 0.3166 to 0.3060.

W4. Computational cost

Table 10 (App. F) reports SEVIR 5→20 costs at 256×256256\times256, batch 1, on one RTX 4090 D. FLOPs include sampling and VAE encode/decode. Measurements use FP32; DySP's training-memory measurement includes gradient checkpointing.

ModelParams (M)FLOPs (G)Infer. (s)Train mem. (GB)
ConvLSTM7.51610.0470.49
PredRNN14.53240.1050.90
PhyDNet3.16410.0722.07
Earthformer1.52850.0274.94
DiffCast (250 steps)49.4243,73614.00318.26
AlphaPre89.012,4430.34619.35
FACL14.41340.0211.70
SDIR34.896,6811.2249.13
FlowCast (10 steps)160.442,0651.36314.89
DySP25.32,3380.1006.46

DySP has fewer parameters, uses fewer FLOPs and less training memory, and runs faster than DiffCast, AlphaPre, SDIR, and FlowCast. The spectral pathway and its injection layers add 5.0M parameters and 113 GFLOPs (4.8% of total FLOPs). FACL has lower cost than DySP on all four measures.

W5. Additional frequency-domain baselines

FACL (NeurIPS 2024) uses the official SimVP implementation and FACL loss, with the same training settings as DySP (Sec. 4.1; App. B). Tables 1–3 report scores from the final scheduled checkpoint. Best scores within each dataset are bold.

DatasetModelHSS-AVG↑CSI-AVG↑SSIM↑MAE↓
ShanghaiAlphaPre0.42760.31450.75341445.3
SDIR0.58820.44970.85481129.1
FACL0.53500.39840.67571506.1
DySP0.59960.46130.85001226.2
CIKMAlphaPre0.36330.30920.4775661.4
SDIR0.47240.40430.5574600.4
FACL0.41710.35750.4727671.3
DySP0.47880.40730.5577595.5
SEVIRSDIR0.44010.34990.75441897.9
FACL0.45640.35660.69642168.5
FlowCast0.46450.35850.71972131.1
DySP0.50320.39310.75501924.9

Among the baselines, FlowCast has the highest mean HSS/CSI on SEVIR, while SDIR leads on Shanghai and CIKM. DySP's SEVIR scores are 8.3%/9.7% higher than FlowCast's. FACL adds a baseline trained with Fourier losses.

Dear Reviewer Nq9e, Thank you again for your time and detailed feedback. Do our additional experiments and clarifications address your concerns? We are more than happy to clarify any remaining questions.

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

The paper presents a novel deterministic method for precipitation nowcasting based on spectral envelopes formed by a multilevel Haar decomposition of radar frames and a spatial encoder. They introduce the prior-guided spatiotemporal block to join spectral and spatiotemporal information. To recover local structures lost by pooling, they use temporal-difference skips. The results on three datasets show an improved performance over several deterministic baselines and other spectral methods.

Critical strengths
  • As deterministic methods are known to average out the most intense precipitation over lead times, this work shows a great improvement in this field. The provided figures depicting the forecasts show that the method produces sharp forecasts even at longer lead times.
  • The prior-guided spatiotemporal block seems to be an original way to join spectral and spatial information.
  • The method is thoroughly tested on three datasets. The testing is specifically focused on high-intensity precipitation.
  • The authors present a meticulous ablation study that shows the contribution of each introduced component.
Critical weaknesses
  • Sometimes it is difficult to follow the description of the method; some important details are in the Appendix, and Figure 2, which depicts the entire method (very hard to follow).
  • The loss function should be better explained, what are the individual terms in the loss function how are they evaluated? Only the Lprior part is described in the main text. The full explanation is in the Appendix, but at least the mathematical formulations should be included in the text. Only the Lprior part is described.
  • The authors compare their method to baseline methods on three datasets: Shanghai, CIKM, and SEVIR datasets. When comparing nowcasting methods, the SEVIR dataset is usually the main dataset, but the authors do not include the full SEVIR results. Most of the SEVIR results are in the Appendix. I would advise the authors to move the SEVIR results to the paper.
  • DySP is only compared to deterministic methods (apart from DiffCast), which is fair because DySP is a deterministic method. The only generative method Diffcast is compared to is Diffcast some other methods are mentioned but not compared to, even if they usually outperform deterministic methods (e.g., FlowCast ICLR 2026). Even DiffCast is not end-to-end generative model, as it does not generate precipitation fields directly from noise. It models stochastic residuals around a deterministic base prediction. The comparison protocol should be clearly stated.
  • HSS and CSI are standard metrics to assess nowcasting models over different thresholds, along with false alarm rate (FAR), but the authors do not report these metrics. Demonstrating that FAR remains low (or at least competitive with baselines) would confirm that DySP’s high CSI and HSS at high thresholds are driven by genuine spatial and intensity accuracy rather than systematic over-forecasting.
  • One of the claims in the paper is that the exact, pixel-by-pixel arrangement of high-frequency details fluctuates unpredictably, whereas spatial distribution and strength with the envelope exhibit high temporal coherence. To prove this point, the authors should add the evaluation of the Fractions Skill Score (FSS) to the evaluation, which would directly prove that DySP captures detail activity correctly within local neighborhoods across multiple spatial scales. Add the evaluation over different thresholds.
Questions for authors
  • Why the authors used the Haar decomposition should be addressed; what do the different bands represent.
  • The caption of Figure 1 is not clear. What is c, LH1?
  • The method is composed of several components, and their connections are hard to understand. There is too much information squeezed into Figure 2, and it does not help in understanding the whole procedure. I advise the authors to remove a), b), and c) from the figure and show them somewhere else, or remove them and spread the remaining part. As they are, they do not help with understanding.
  • Why authors decided to resize different datasets should be explained.
4 comments

Thank you for the suggestions on presentation and evaluation.

W1 & Q3. Readability and Figure 2

We redrew Figure 2 as three labelled rows (spectral prior, content prediction, skip path), with one block per stage. We removed the former panel (b) and embedding details, and enlarged PGSTB and PGSA. The caption refers to Secs. 3.1–3.3, and Eq. 2 is in the main text. App. A gives J=2J=2, D=256D=256, the 32×3232\times32 token grid, and other implementation details. The four PGSTBs belong to the joint historical/future stack.

W2. Loss function

Sec. 3.4 now gives formulas for all loss terms (Eqs. 8–12). Lbal\mathcal L_{\rm bal} is intensity-weighted L1L_1, with weights 1,2,4,8,121,2,4,8,12 across the intervals defined by 20/30/35/40 dBZ on Shanghai and CIKM, and 1,2,4,8,12,24,481,2,4,8,12,24,48 across the SEVIR VIL intervals (App. B). Lwav\mathcal L_{\rm wav} matches the coarsest approximation and pooled detail magnitudes. Lflow\mathcal L_{\rm flow} regularises spatial first and temporal second differences of skip displacements. Eq. 8 defines Lpriorf\mathcal L_{\rm prior}^{f}; Lprior\mathcal L_{\rm prior} averages the historical and future terms. The full objective is L=Lbal+Lwav+0.1 Lprior+0.01 Lflow\mathcal L=\mathcal L_{\rm bal}+\mathcal L_{\rm wav}+0.1\,\mathcal L_{\rm prior}+0.01\,\mathcal L_{\rm flow} (Eq. 12).

W3. SEVIR results in the main paper

We moved the full SEVIR results to Table 3 in the main paper. DySP has the highest HSS and CSI at all six VIL thresholds.

W4. Generative baselines and comparison protocol

We score each deterministic method on one forecast, and DiffCast and FlowCast on one generated sample. All methods use the same splits, resolution, thresholds, and 100,000-iteration budget (Sec. 4.1), with DySP's AdamW settings (App. B). Tables 1–3 give the full results. CSI-high refers to CSI at 40 dBZ on Shanghai/CIKM or VIL 181 on SEVIR. DySP has higher HSS-AVG, CSI-AVG, CSI-high, and SSIM, and lower MAE, than FlowCast on all three datasets. Its CSI is higher at every threshold on Shanghai and SEVIR. On SEVIR, DySP uses about 1/18 of FlowCast's FLOPs and 1/14 of its inference time (App. F, Table 10).

W5. FAR, POD, and frequency bias

Table 8 (App. E) reports POD, FAR, and BIAS=(TP+FP)/(TP+FN)\mathrm{BIAS}=(\mathrm{TP}+\mathrm{FP})/(\mathrm{TP}+\mathrm{FN}) at all six VIL thresholds (20-lead means; Earthformer FAR at VIL 219 uses 15 leads). DySP has higher POD and bias closer to one than SDIR at every threshold. At VIL 160/181, POD is 0.380/0.340 vs. 0.252/0.204, bias 0.75/0.73 vs. 0.41/0.34, and FAR 0.518/0.563 vs. 0.426/0.445. DySP detects more events than SDIR, with a higher false-alarm ratio. Its bias remains below one, indicating fewer forecast exceedances than observed. At both thresholds, DySP has higher POD and lower FAR than FlowCast and FACL.

W6. Fractions Skill Score

We added FSS (Eq. 16) at VIL 74/133/181 for s=5,9,17s=5,9,17, averaged over 20 leads (App. E, Table 9). The table compares FACL, SDIR, and DySP; best scores are bold and second-best scores are underlined:

SEVIR FSS74, s=5s=5s=9s=9s=17s=17133, s=5s=5s=9s=9s=17s=17181, s=5s=5s=9s=9s=17s=17
FACL0.8297‾\underline{0.8297}0.86940.91110.5987‾\underline{0.5987}0.6689‾\underline{0.6689}0.7511‾\underline{0.7511}0.4167‾\underline{0.4167}0.4890‾\underline{0.4890}0.5743‾\underline{0.5743}
SDIR0.82540.85760.89520.57230.62740.69540.32630.37440.4295
DySP0.83050.8630‾\underline{0.8630}0.9003‾\underline{0.9003}0.62510.68280.75220.51130.58450.6721

DySP has the highest FSS in seven of nine settings, including every neighbourhood size at VIL 133 and 181. FACL leads at VIL 74 for s=9,17s=9,17. At VIL 181, DySP exceeds FACL by 0.0946/0.0955/0.0978 and SDIR by 0.1850–0.2426.

Q1. Why Haar, and what the bands represent

We chose Haar for its simple, local contrasts and alignment with the token grid (Sec. 3.1; App. A). LH/HL measure directional block contrasts and HH measures mixed contrast. Their pooled magnitudes describe local detail activity. Level-jj atoms have non-overlapping 2j×2j2^j\times 2^j support, so each token aggregates its own patch. On SEVIR (1.5 km/pixel after resizing), levels 1/2 correspond to about 3/6 km.

Q2. Figure 1 caption

The caption now defines c=W1,LH(x)c=\mathcal W_{1,\mathrm{LH}}(x) as the signed level-1 Haar LH coefficient (LH1). It also states that both detail representations use 4×44\times4 pooling and share a colour scale across rows within each column.

Q4. Why the datasets are resized

We resize Shanghai and SEVIR to 256×256256\times256 to reduce the memory needed for 20-frame prediction across the compared models. CIKM frames are padded from 101×101101\times101 to 128×128128\times128 without interpolation, preserving native pixel spacing, and cropped back for scoring (Sec. 4.1, App. C).

Dear Reviewer evGz, Thank you again for your time and detailed feedback. Do our additional experiments and clarifications address your concerns? We are more than happy to clarify any remaining questions.

Reviewer_evGzOpenReview ↗

Additional evaluations

You addressed all my concerns. Just one small detial: for generative models you pick one genrated sample. How do you select the sample, you just take the first one generated?

Thank you for the question. For each test input, each generative baseline produced exactly one forecast sequence. We used that sequence directly for evaluation, so there was no sample-selection step.

Reviewer_4nUpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DySP, a deterministic precipitation nowcasting framework that forecasts multiscale spectral envelopes to guide the reconstruction of future radar fields. The envelopes are obtained by locally averaging the magnitudes of Haar wavelet detail coefficients, preserving the spatial distribution of detail activity while reducing sensitivity to coefficient signs and fine-scale arrangement. A spectral pathway predicts latent priors and future envelopes. These condition spatial attention through query/key offsets and envelope-dependent output gating. Temporal-difference skip connections provide complementary observation features and recent changes for reconstruction. The reported results show improved mean HSS and CSI over the compared methods, with the largest gains on SEVIR.

Critical strengths
  1. The envelopes have a clear interpretation as spatially localized, multiscale detail activity. The motivation compares envelopes with signed coefficients under matched pooling, which controls for the effect of spatial averaging. Appendix E.2 goes beyond this temporal-coherence analysis by showing that predicted envelopes outperform last-observation persistence on all three datasets. Together with the input-representation comparisons in Table 5, this provides substantive support for the proposed representation and makes the contribution more informative than an architectural modification alone.

  2. The reported mean HSS and CSI improve across all three benchmarks. On SEVIR, mean CSI increases from 0.3561 to 0.3931 relative to SDIR, and CSI at VIL 181 increases from 0.1708 to 0.2487. These are meaningful reported improvements for high-intensity echo prediction. Table 4 also shows complementary contributions from spectral guidance and temporal differences within a common backbone, while retaining skip fusion, warping, and gating across configurations. This strengthens the empirical case for the overall design, although the comparison-protocol concerns below need to be resolved.

Critical weaknesses
  1. The comparability of the main results is not fully established. Appendix C states that CIKM and Shanghai values are converted to reflectivity by multiplying normalized values by 90, whereas the cited SDIR paper uses different dataset-specific conversions. In addition, the paper does not clearly report baseline training settings or DySP’s learning rate and training duration, making it difficult to determine whether all methods were compared under a consistent training and evaluation protocol.

  2. The contribution of explicit envelope forecasting is not fully isolated. Table 4 validates the spectral pathway as a whole, and Table 5 studies its input representation and conditioning mechanisms, but Appendix E.2 only shows that predicted envelopes are more correlated with future targets than persistence. A matched control without envelope supervision, or using the last observed envelope for future gating, would more directly verify whether forecasting dynamic envelopes itself is responsible for the final prediction gains.

Questions for authors
  1. Given that future priors are initialized through a shared linear temporal projection (Eq. 10) and decoder skips propagate observed features through warping and gating (Eqs. 5–6), how well does DySP handle newly forming or rapidly intensifying echoes, rather than primarily advecting existing structures?

  2. Table 8 shows that pooled pixels exhibit substantially higher temporal correlations than spectral envelopes. What makes envelope forecasting a more effective intermediate representation? A comparison with similarly pooled intensity maps under matched supervision and conditioning would help establish its distinctive advantage.

  3. Could the authors evaluate the model without envelope supervision or with persistence-based envelope guidance? This would help isolate the contribution of dynamic envelope forecasting from that of the additional spectral pathway.

2 comments

Thank you for your comments. We have clarified the evaluation protocol and added the requested ablations.

W1(a). Reflectivity conversion

We use the conversion in the public evaluation code of DiffCast, AlphaPre, and SDIR: divide pixel values by 255, then multiply by 90 before thresholding (Tables 1–2; App. C). The normalisation, conversion, thresholds, and evaluation metrics are identical for all compared methods.

W1(b). Training settings

We train all methods for 100,000 iterations on four RTX 4090 D GPUs, with per-GPU batch sizes of 8 for CIKM and 4 for Shanghai/SEVIR. We use AdamW with a peak learning rate of 3×10−43\times10^{-4}, 5% warmup, and cosine decay (Sec. 4.1; App. B).

W2 & Q3. Isolating dynamic envelope forecasting

We added two controls to Table 5: no envelope loss and pooled-intensity guidance. Both use the Shanghai ablation setting, with temporal differences disabled and Q/K conditioning and gate-and-shift modulation enabled:

ConfigurationHSS-AVG ↑CSI-AVG ↑SSIM ↑MAE ↓
Magnitude, both (envelope targets)0.58840.45160.85181191.6
No envelope loss0.58270.44680.85271164.3
Pooled intensity0.58460.44570.84141346.1

Removing Lprior\mathcal L_{\rm prior} leaves the spectral pathway, magnitude input, and conditioning routes unchanged. Mean HSS falls from 0.5884 to 0.5827 and mean CSI from 0.4516 to 0.4468. At 40 dBZ, CSI falls from 0.3166 to 0.3060. This control has higher SSIM and lower MAE.

In Table 6 (App. D.1), last-frame envelope guidance uses the final input frame's envelope throughout the forecast. DySP's mean CSI is 0.4073 vs. 0.3908 for this control on CIKM, 0.4613 vs. 0.4303 on Shanghai, and 0.3931 vs. 0.3770 on SEVIR. Predicted envelopes also have higher correlation with future targets than persistence (0.560 vs. 0.478 on SEVIR; App. D.3).

Q1. Newly forming and rapidly intensifying echoes

The decoder predicts future fields from learned tokens and warped observation features. Nonlinear attention precedes TZ\mathcal T_Z, and four PGSTBs then update the future tokens across space and time (Eqs. 2–3). The prior supplies a gate in (0,2)(0,2) and an additive shift that can strengthen or weaken local features (Eq. 5). Temporal differences provide recent changes in the observed frames (Eq. 6).

On SEVIR, DySP's POD at VIL 160/181 is 0.380/0.340, compared with SDIR's 0.252/0.204. CSI is also higher at both thresholds, as is FSS at VIL 181 (Table 3; App. E, Tables 8–9).

We also evaluated all 16,212 SEVIR test sequences by region. For each lead, we exclude a neighbourhood around last-frame threshold exceedances, using a 15 m/s motion allowance plus a four-pixel buffer. In the remaining regions near last-frame VIL-16 exceedances, CSI at VIL 133/181 is 0.0280/0.0259 for DySP and 0.0167/0.0084 for SDIR.

Q2. Why envelopes rather than pooled intensity

Pooled pixels have the highest temporal correlation in Table 7. They track average intensity: a constant patch and a compact core can have the same pooled mean but different local contrast. Envelopes measure this contrast through locally averaged wavelet-detail magnitudes at each scale and orientation. With identical pooling, they have higher temporal correlation than signed details (0.570 vs. 0.318 at 6 min). The input comparisons in Table 5 retain envelope targets; the pooled-intensity control changes both the branch input and auxiliary target.

The pooled-intensity control uses the same latent width, conditioning routes, and 100,000-iteration budget. Its mean HSS/CSI is 0.5846/0.4457, compared with 0.5884/0.4516 for magnitude input with envelope supervision. Envelope guidance has higher CSI at all four thresholds, including 0.3166 vs. 0.3066 at 40 dBZ, higher SSIM (0.8518 vs. 0.8414), and lower MAE (1191.6 vs. 1346.1).

Dear Reviewer 4nUp, Thank you again for your time and detailed feedback. Do our additional experiments and clarifications address your concerns? We are more than happy to clarify any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.