acceptodds
Desk rejected at ICLR 2027

Transport-Guided Factorized Evolution for Precipitation Nowcasting

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Precipitation nowcasting predicts the spatial distribution and intensity of precipitation over the next few hours from recent radar observations. Motion-aware methods combine advective transport of existing echoes with intensity updates to model localized growth and decay. However, in recursive rollout, each forecast already incorporates predicted intensity updates and is then fed back as input to the next transport step. This forecast-field feedback carries earlier evolution errors into subsequent transport steps and can compound spatial distortion. We present ransport-Guided actorized lution, an end-to-end framework that breaks this forecast-field feedback by using learned displacement fields to warp the latest observed radar field directly to each forecast horizon. Conditioned on the transported prediction, TFEvo models residual evolution by factorizing the field into spatial coverage and intensity components. This factorization accounts for the zero-inflated sparsity of radar data by representing zero-valued background through the coverage component. These updates are learned with an asymmetric training objective applied to the evolved prediction to penalize the underestimation of extreme rainfall. Across three benchmarks, TFEvo outperforms competitive baselines on most forecasting metrics. Notably, at resolution, it improves CSI-M by up to % and reduces LPIPS by up to % relative to a competitive high-resolution baseline, while using less than % of the baseline's parameters. Code is available at https://anonymous.4open.science/r/TFEvo-CECD.

Primary area:
learning on time series and dynamical systems
Keywords:
Precipitation Nowcasting, Spatiotemporal Forecasting, Extreme Precipitation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Tr7mOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an end-to-end transport-guided factorized evolution framework for precipitation nowcasting. Its transport module predicts cumulative displacement fields and directly warps the latest radar observation to each forecast horizon, avoiding the feedback of predicted precipitation fields into subsequent transport steps. The evolution module factorizes the transported field into coverage and intensity components and learns residual updates to model precipitation growth and decay. Joint supervision and an asymmetric loss further address forecasting accuracy and the underestimation of extreme precipitation. Experiments on three radar benchmarks demonstrate competitive forecasting performance with a compact model.

Critical strengths
  1. It transports the latest observation directly to each forecast horizon and learn the remaining changes to the transported field, avoiding the feedback of predicted precipitation fields into subsequent transport steps.
  2. To model this residual, a zeroinflation-aware Factorized Evolution (ZFEvo) module then factorizes each transported prediction into spatial coverage and intensity components and learns their updates conditioned on the transported field, spatiotemporal context, and the learned motion representation.
  3. An asymmetric training objective alongside joint supervision on both the transported and evolved predictions further counter the underestimation driven by rare extreme events.
Critical weaknesses
  1. The proposed framework first transports the latest observation to preserve alignable echo structures, then predicts the remaining changes to refine local details. Such a coarse-to-fine strategy is commonly used in many fields.
  2. The Introduction provides limited analysis of existing approaches and does not adequately discuss the broader range of relevant methods. The authors should discuss related approaches that adopt similar coarse-to-fine strategy and clearly explain the methodological novelty of their framework relative to them.
  3. The contribution of breaking forecast-field feedback lacks a controlled comparison between recursively transporting previous predictions and directly transporting the latest observation under identical network and training settings. Thus, the gains cannot be clearly attributed to blocking evolution-error feedback rather than avoiding repeated interpolation or other design differences. Moreover, displacement fields remain recursively updated, so motion errors may still accumulate.
  4. The physical meaning and necessity of the coverage–intensity factorization remain unclear. The decomposition mainly affects zero-valued and weak echoes, while coverage can only increase and decay is handled through intensity reduction. This asymmetric design needs stronger justification and validation.
Questions for authors
  1. Several variables lack clear definitions. What do x,τ,rx, \tau, r represent in Eq. (2)?
  2. How is the decomposition in Eq. (3) derived from the preceding advection–evolution equation? Please clarify the connection and any assumptions involved.
  3. The comparison methods are not explicitly cited. Please provide the corresponding reference for each baseline.
  4. The framework is largely implemented using standard convolutional blocks, gated state updates, displacement prediction, and residual correction heads. The additional designs mainly involve fine-grained operations such as channel-wise decay, displacement centering, and bounded factor updates. The architectural novelty may be limited.
3 comments

We appreciate the reviewers for acknowledging the contributions of this work and all the constructive suggestions. We hope the following clarifications can address your common concerns.

[W1, W2 & Q4] Modeling contribution and relation to prior work

We appreciate the opportunity to clarify our methodological contribution. TFEvo combines spatial transport and zero-inflation-aware evolution. At each forecast lead, it directly samples the latest observation, then predicts the remaining changes through coverage and intensity updates.

To clarify the distinction from coarse-to-fine strategies in related work, we compare the coarse-to-fine decompositions in CasCast and DuoCast with the transport–evolution decomposition in TFEvo. CasCast [1] cascades deterministic prediction and probabilistic generation to model mesoscale precipitation distributions and small-scale patterns. DuoCast [2] uses low- and high-frequency diffusion models, with the low-frequency forecast conditioning fine-scale refinement. TFEvo uses a transported observation as the reference for evolution, aligning observed echo structures through transport and modeling coverage and intensity changes relative to this reference. This formulation separates spatial movement from echo evolution.

Experiments across three datasets demonstrate competitive performance. High-resolution comparisons show forecasting gains over exPreCast [3] with fewer parameters, while component ablations support the complementary roles of transport and evolution. The additional comparisons below examine forecast-field feedback and factorized evolution. We will expand the Introduction and Related Work to clarify these distinctions alongside the existing discussion of motion-aware methods such as MoCast [4].

[1] Gong et al. CasCast: Skillful High-resolution Precipitation Nowcasting via Cascaded Modelling. ICML 2024.

[2] Wen et al. DuoCast: Duo-Probabilistic Diffusion for Precipitation Nowcasting. AAAI 2026.

[3] Song et al. Extreme Weather Nowcasting via Local Precipitation Pattern Prediction. ICLR 2026.

[4] Wu et al. MoCast: Learning Turbulent Motions Under Physical Guidance for Precipitation Nowcasting. AAAI 2026.

[W3] Comparison with recursive transport

We compare three transport formulations on Shanghai using the same training objective and nearly identical computational costs. The recursive variants retain the main network components, with conditioning adapted to concatenate the preceding field with history features for stepwise motion estimation. The results are shown in the table below.

MethodCSI-M ↑CSI-40 ↑HSS-M ↑SSIM ↑LPIPS ↓
Recursive transport–evolution0.41590.25970.55550.76140.3103
Recursive transport + parallel evolution0.42530.28100.56690.78320.2602
TFEvo0.43400.29260.57490.79640.1843

Recursive transport–evolution feeds the evolved forecast into the next motion-estimation and transport step. Recursive transport with parallel evolution feeds back only the transported prior and applies evolution independently at each lead. TFEvo directly samples the latest observation at each lead.

Both recursive variants retain repeated sampling, yet excluding evolved forecasts from the feedback loop improves all reported metrics. This supports the benefit of blocking evolved-forecast feedback, and the improvement cannot be attributed solely to avoiding repeated interpolation. TFEvo further improves these metrics. TFEvo’s motion-state updates and displacement composition retain temporal dependence across leads, so motion-estimation errors can still accumulate. Observation anchoring samples the latest observation directly at each lead and prevents evolved forecasts from entering subsequent transport.

[W4] Coverage–intensity factorization and asymmetric updates

The factorization is motivated by the zero-inflated structure of radar fields, where a large zero-valued background coexists with varying echo amplitudes. In Eq. (8), coverage represents soft spatial support, taking zero at zero-prior locations and one where the transported echo reaches the threshold κ\kappa. It does not directly represent physical area coverage or a calibrated precipitation probability.

This factorization couples intensity corrections to spatial support. At zero-prior locations, intensity updates alone cannot produce nonzero outputs while coverage remains zero, so coverage updates are needed to activate these locations. Where the transported echo reaches κ\kappa, coverage remains one, and intensity updates handle strengthening and weakening. The asymmetric rule assigns activation and support expansion to coverage, and weakening to intensity, reducing the overlap whereby either factor could decrease to reduce the output.

The table below reports results on Shanghai for unfactorized evolution, bidirectional coverage updates that allow both increases and decreases, and the default rule.

Evolution formulationCSI-M ↑SSIM ↑LPIPS ↓
Unfactorized evolution0.43500.76940.2594
Factorized, bidirectional coverage0.43570.79630.2215
TFEvo0.43400.79640.1843

Both factorized formulations achieve higher SSIM and lower LPIPS at comparable CSI-M. The default rule yields lower LPIPS than bidirectional coverage, with similar CSI-M and SSIM.

[Q1] Variable definitions

In Eq. (2), x\boldsymbol{x} denotes a two-dimensional spatial location, τ\tau denotes continuous time, and r(x,τ)r(\boldsymbol{x},\tau) is the source–sink rate describing local growth and decay along motion trajectories. The subsequent index tt denotes a discrete forecast lead. We will make these definitions explicit in the manuscript.

[Q2] Connection between Eqs. (2) and (3)

Eq. (2) describes intensity changes along motion trajectories, with rr representing the rate of local growth and decay. In this continuous description, integrating over time along a trajectory gives the transported initial value plus the accumulated source–sink contribution. This motivates the transport–evolution decomposition in Eq. (3).

In the discrete model, learned displacements and grid sampling construct the transported prior, and the evolution module predicts the remaining changes relative to that prior. Eq. (3) adopts this structure as a modeling approximation; the model does not solve Eq. (2) on discrete radar fields. We will clarify this connection and the nature of the approximation in the manuscript.

[Q3] Baseline references

Thank you for pointing out this omission. We will add the corresponding citations when introducing each baseline and in the comparison tables.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer Tr7m,

We appreciate your insightful comments. We have revised our paper and provided a point-by-point response, and we hope it addresses your concerns.

We are very keen to clarify any remaining points and look forward to your feedback.

Best regards,

Authors

Reviewer_1QCxOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper estimates motion displacement and directly shifts the latest observation for each forecast lead time, avoiding iterative prediction. The shifted field is then decomposed into coverage and intensity components, followed by residual correction. A weighted loss penalizes the underestimation of intense echoes to improve extreme-event detection. The model contains approximately 0.287 million parameters, which is quite efficient. The experiments on 3 benchmarks, SEVIR, MeteoNet, and Shanghai radar verify the effectiveness of the proposed method. Overall, I think this is a quite useful paper for radar image prediction.

Critical strengths
  1. This paper is well organized and easy to understand.
  2. The idea is straightforward, and the performance of the proposed model is also impressive.
Critical weaknesses

1. The benefit of interrupting the predictive feedback loop is not isolated. A controlled ablation should compare observation-based shifting with recursive prediction under identical architectures, losses, and computational budgets. Existing ablations do not establish the contribution of this core mechanism.

2. The bounded correction limits extreme intensity changes. Equations 8–9 constrain the output to approximately 0.51 when the shifted prior is zero and at least 0.4 when it is 0.9. The authors should quantify how often targets exceed these bounds and examine sensitivity to γ and the coverage rules.

3. Parameter efficiency does not imply computational efficiency. Compared with exPreCast, TFEvo requires 2.48× more computation and 2.05× higher latency at 128² resolution, and 3.12× more computation at 384². High-resolution latency and memory usage should also be reported.

Questions for authors

Please supply the following questions and answer the weaknesses above.

  1. Controlled comparison of recursive and observation-based prediction under identical loss functions.
  2. Stratified analysis of convection initiation, dissipation, and cases exceeding residual correction bounds.
  3. Joint evaluation of detection gains, false alarms, and magnitude errors.
4 comments

We appreciate the reviewers for acknowledging the contributions of this work and all the constructive suggestions. We hope the following clarifications can address your common concerns.

[W1 & Q1] Comparison with recursive prediction

We compare three transport formulations with closely matched architectures, the same training objective, and nearly identical computational costs. Removing evolved forecasts from the recursive feedback loop improves forecasting performance, with further gains from observation-anchored transport.

Recursive transport–evolution feeds the evolved forecast into the next motion-estimation and transport step. Recursive transport with parallel evolution feeds back only the transported prior and applies evolution independently at each lead. TFEvo samples the latest observation once per lead without forecast-field feedback. In the recursive variants, the preceding field is concatenated with history features to provide recursive conditioning.

MethodCSI-M ↑CSI-40 ↑HSS-M ↑SSIM ↑LPIPS ↓
Recursive transport–evolution0.41590.25970.55550.76140.3103
Recursive transport + parallel evolution0.42530.28100.56690.78320.2602
TFEvo0.43400.29260.57490.79640.1843

[W2 & Q2] Correction bounds and coverage updates

We assess the correction constraints through three analyses on Shanghai. Most targets remain within the attainable range, while relaxing the intensity bound or allowing coverage to decrease does not consistently improve the evaluated metrics.

(1) Target exceedance rates. We measure how often targets fall outside the attainable range under the default γ=0.5\gamma=0.5. Rapid appearance and disappearance are defined at fixed spatial locations by

x_{t-1}<\tau\leq x_t,\qquad x_t-x_{t-1}\geq\delta_+,$$ $$\text{Disappearance:}\quad x_t<\tau\leq x_{t-1},\qquad x_{t-1}-x_t\geq\delta_-.$$ Here, $x_{t-1}$ and $x_t$ denote echo intensities at the same spatial location, $\tau=0.01$, and $\delta_+=0.2412$ and $\delta_-=0.2397$ are the training-set 95th percentiles of increase and decrease magnitudes, respectively. These events include changes caused by echo motion. | Target subset | Below the lower bound (%) | Above the upper bound (%) | Within bounds (%) | | ------------------- | ------------------------: | ------------------------: | ----------------: | | All targets | 0.032 | 0.093 | 99.875 | | Rapid appearance | 0 | 0.429 | 99.571 | | Rapid disappearance | 1.017 | 0 | 98.983 | Percentages are calculated within each subset. More than 98.9% of targets in both rapid-event subsets remain within bounds. Exceeding the attainable range limits amplitude reconstruction but does not necessarily imply a missed detection. **(2) Sensitivity to $\gamma$.** We vary $\gamma$ to examine whether a wider correction range improves rapid-event detection. | $\gamma$ | Appearance CSI ↑ | Appearance HSS ↑ | Disappearance CSI ↑ | Disappearance HSS ↑ | | ----------------- | ---------------: | ---------------: | ------------------: | ------------------: | | 0.1 | 0.0078 | 0.0149 | 0.0082 | 0.0158 | | 0.3 | **0.0135** | **0.0254** | **0.0113** | **0.0213** | | **0.5 (default)** | **0.0135** | **0.0254** | 0.0110 | 0.0208 | | 0.7 | 0.0133 | 0.0251 | 0.0107 | 0.0201 | | 1.0 | 0.0130 | 0.0245 | 0.0103 | 0.0193 | The smaller range at $\gamma=0.1$ yields lower detection scores, whereas increasing $\gamma$ beyond the default provides no further improvement. **(3) Coverage-update rules.** We compare the default coverage rule, which assigns weakening to intensity updates, with bidirectional coverage updates that also allow coverage to decrease. | Evolution formulation | CSI-M ↑ | SSIM ↑ | LPIPS ↓ | | ---------------------------------- | ---------: | ---------: | ---------: | | Unfactorized evolution | 0.4350 | 0.7694 | 0.2594 | | Factorized, bidirectional coverage | **0.4357** | 0.7963 | 0.2215 | | TFEvo | 0.4340 | **0.7964** | **0.1843** |

[W3] Parameter efficiency and computational cost

We acknowledge the reviewer’s concern about computational efficiency. TFEvo requires less weight storage, but full-resolution feature processing across forecast leads still incurs computation. We report latency, peak GPU memory, and weight storage below, and will include this analysis in the revised manuscript to clarify the distinction between parameter efficiency and computational efficiency.

To reduce computational cost, we explore two variants. TFEvo-3D simplifies the 3D history encoder, while TFEvo-2D uses frame-wise 2D encoding with temporal fusion. Both use half-resolution evolution and retain the transport–evolution formulation. Both variants require fewer GFLOPs than exPreCast while achieving better scores on most forecasting metrics. We will include these results in the appendix and discuss compact feature processing as a direction for future computational optimization.

Resource measurements. Values separated by “/” correspond to 128×128128\times128 and 384×384384\times384.

ModelParams (M) ↓GFLOPs ↓Latency (ms) ↓Peak GPU memory (MiB) ↓Weights (MiB) ↓CSI-M ↑SSIM ↑LPIPS ↓
exPreCast (ICLR 26)32.00426.54 / 189.977.79 / 18.24169.31 / 399.61122.090.3767 / 0.40590.6421 / 0.76640.3771 / 0.3016
AlphaPre (CVPR 25)89.0263110.75 / 27995.7286.63 / 815.471339.96 / 9199.47339.610.4248 / -0.7951 / -0.3236 / -
DuoCast (AAAI 26)179.304 / 179.3199625.14 / 87918.881192.44 / 4519.981116.45 / 4362.77683.99 / 684.050.4252 / -0.7788 / -0.2223 / -
TFEvo (Ours)0.28765.90 / 593.0814.23 / 62.66101.64 / 840.051.090.4340 / 0.42220.7964 / 0.84820.1843 / 0.2035

Variant comparison on Shanghai, 128×128128\times128.

ModelParams (M) ↓GFLOPs ↓CSI-M ↑CSI-40 ↑POD-M ↑SSIM ↑LPIPS ↓
exPreCast32.00426.540.37670.22110.59500.64210.3771
TFEvo0.28765.900.43400.29260.67070.79640.1843
TFEvo-3D0.23124.820.43020.29500.58150.79950.1723
TFEvo-2D0.21221.710.42630.29520.57860.79670.1744

[Q3] Joint evaluation of detection and intensity accuracy

We jointly report CSI, POD, FAR, HSS, and MSE on MeteoNet, including the symmetric setting q=0.5q=0.5. CSI considers both missed events and false alarms, while HSS additionally accounts for chance agreement. These measures complement POD and FAR in assessing detection performance.

ModelCSI-M ↑POD-M ↑FAR-M ↓HSS-M ↑CSI-32 ↑POD-32 ↑FAR-32 ↓HSS-32 ↑MSE ↓
AlphaPre0.38890.47890.35910.52240.20210.25690.52270.311412.80
DuoCast0.35600.53710.53340.47900.20160.38420.73080.306818.09
exPreCast0.34120.49400.50760.47440.16680.30650.73420.273317.95
TFEvo, q=0.5q=0.50.39050.50420.37420.52140.17100.19750.45180.266614.31
TFEvo, q=10/11q=10/110.40940.65810.49980.55120.24710.47830.66550.383317.51

TFEvo with q=10/11q=10/11 achieves the highest average and high-threshold CSI and HSS among the compared methods. AlphaPre has the lowest FAR-M and MSE, while TFEvo with q=0.5q=0.5 has the lowest FAR-32. Changing qq from 10/1110/11 to 0.50.5 reduces FAR and MSE but also lowers CSI and POD.

Thanks for the effort in providing the new experimental results; my main concerns have been addressed. The authors also honestly present the trade-off between the POD and FAR of the proposed method. Based on this, I maintain my previous rating (3).

We appreciate your acknowledgment that the main concerns have been addressed. We have also shared additional analysis of the existing predictions, showing improved rain/no-rain detection and providing further support for the role of coverage–intensity factorization.

Reviewer_GQ6uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TFEvo, a deterministic framework for precipitation nowcasting that separates spatial transport from residual evolution. The Anchor-based Shared Transport module predicts and composes global and local displacement fields, then warps the latest observed radar field directly to each forecast horizon. This removes feedback of previously evolved forecast fields into subsequent transport steps. The Zero-inflation-aware Factorized Evolution module decomposes the transported field into coverage and intensity factors and predicts bounded corrections to them. Training combines intermediate transport supervision, an asymmetric body–tail loss on the final forecasts, and regularization. Experiments report improved detection scores with approximately 0.29M parameters, together with component ablations, lead-time analyses, and additional evaluation at 384 × 384 resolution.

Critical strengths
  1. The reported detection results are consistently strong across three benchmarks. TFEvo achieves the highest CSI-M among the methods listed in Table 1, and the locally evaluated comparisons also show substantial improvements at high thresholds. Achieving these results with approximately 0.29M parameters is a meaningful practical contribution.

  2. Warping the latest observation through composed displacement fields is a clear architectural way to exclude evolved forecast fields from the transport pathway. Table 4 supports the complementary value of transport and residual evolution. It also shows that factorized evolution improves structural and perceptual metrics over unfactorized evolution while maintaining similar detection performance. These results support the practical value of the proposed formulation.

Critical weaknesses
  1. Limited expressiveness of coverage–intensity factorization. In Eqs. (8)–(9), the coverage factor becomes fixed at one when (Bt≥κB_t \geq \kappa), making coverage updates ineffective in existing rainy regions. Moreover, the bounded intensity correction limits precipitation initiation and dissipation. For instance, a dry transported pixel cannot exceed 0.51, while a pixel above 0.5 cannot fully dissipate. Table 4 also shows negligible CSI-M improvements over unfactorized evolution, raising questions about the actual benefits of this factorization.

  2. Trade-off between precipitation detection and intensity accuracy. Despite improved CSI and POD, TFEvo produces higher MSE than AlphaPre across all three datasets. The strongly asymmetric loss in Eq. (11) may favor detection performance at the expense of squared-error accuracy. It remains unclear whether the improvements primarily reflect better precipitation modeling or the bias introduced by this objective.

Questions for authors
  1. Given the constraints imposed by Eqs. (8)–(9), how does ZFEvo handle rapid precipitation initiation and dissipation? Can the authors quantify the impact of these constraints on such events?

  2. Table 4 shows nearly identical CSI-M with and without factorization. What evidence supports the claimed benefits of zero-inflation-aware modeling, particularly for precipitation occurrence and intensity changes?

  3. How much of the high-threshold CSI improvement comes from the asymmetric loss? Could the authors compare CSI, POD, false alarm ratio, and MSE under a symmetric alternative ((q=0.5))?

  4. Can the authors directly compare observation-anchored transport with recursive forecast-field transport under an otherwise identical architecture? The current motion-state ablation in Figure 3 does not isolate the effect of eliminating forecast-field feedback.

3 comments

We appreciate the reviewers for acknowledging the contributions of this work and all the constructive suggestions. We hope the following clarifications can address your common concerns.

[W1 & Q1] Bounded evolution and rapid echo changes

We deliberately bound evolution corrections to retain the transported field as a spatial prior while learning the remaining changes. The hyperparameter γ\gamma controls the allowable intensity adjustment and can be increased to relax this constraint. On the normalized [0,1][0,1] scale, γ=0.5\gamma=0.5 allows intensity-factor adjustments of up to 0.5 in either direction before clipping.

To further evaluate how this constraint affects rapid evolution, we vary γ\gamma and assess rapid echo changes at fixed spatial locations on Shanghai. We define rapid appearance and disappearance events as threshold crossings accompanied by sufficiently large intensity increases and decreases, respectively:

\text{Appearance:}\quad &x_{t-1}<\tau\leq x_t,\qquad x_t-x_{t-1}\geq\delta_+, \end{align}$$ $$\begin{align} \text{Disappearance:}\quad &x_t<\tau\leq x_{t-1},\qquad x_{t-1}-x_t\geq\delta_-. \end{align}$$ Here, $x_{t-1}$ and $x_t$ denote echo intensities at the same spatial location, $\tau=0.01$, and $\delta_+=0.2412$ and $\delta_-=0.2397$ are the training-set 95th percentiles of increase and decrease magnitudes, respectively. | Model | Appearance CSI ↑ | Appearance HSS ↑ | Disappearance CSI ↑ | Disappearance HSS ↑ | | -------------------- | ---------------: | ---------------: | ------------------: | ------------------: | | TFEvo ($\gamma=0.5$) | **0.0135** | **0.0254** | 0.011 | 0.0208 | | $\gamma=0.1$ | 0.0078 | 0.0149 | 0.0082 | 0.0158 | | $\gamma=0.3$ | 0.0135 | 0.0254 | **0.0113** | **0.0213** | | $\gamma=0.7$ | 0.0133 | 0.0251 | 0.0107 | 0.0201 | | $\gamma=1$ | 0.013 | 0.0245 | 0.0103 | 0.0193 | | AlphaPre | 0.0001 | 0.0001 | 0.0001 | 0.0002 | | ExPreCast | 0 | 0 | 0 | 0 | In these Shanghai experiments, increasing $\gamma$ from $0.5$ to $1$ does not improve event detection, whereas the smaller correction range at $\gamma=0.1$ yields lower scores. Despite its bounded corrections, TFEvo achieves higher CSI and HSS for both rapid appearance and disappearance than AlphaPre and exPreCast. **[W1 & Q2] Benefits of coverage–intensity factorization** Coverage–intensity factorization explicitly couples intensity corrections to spatial support. Unfactorized evolution directly corrects the transported field, jointly handling changes in spatial support and intensity. Our formulation represents the same prior as ($B_t=S_t\odot A_t$) and learns updates to both factors. Defining the effective updates as ($\Delta S_t=\widehat S_t-S_t$) and ($\Delta A_t=\widehat A_t-A_t$), we obtain $$\widehat X_t-B_t =\widehat S_t\odot\Delta A_t+A_t\odot\Delta S_t.$$ Thus, intensity corrections are modulated by the updated coverage, while coverage updates allow nonzero predictions at zero-prior locations that intensity-only updates cannot activate. Where ($B_t\geq\kappa$), coverage remains one, and intensity updates handle strengthening and weakening. This formulation explicitly couples intensity adjustments to spatial support. Our ablations show improved SSIM and LPIPS, which measure structural similarity and perceptual quality, respectively, across all three datasets while maintaining comparable CSI-M.

[W2 & Q3] Effect of the asymmetric objective

Our body–tail loss is designed to counter the underestimation of intense echoes associated with conservative, MSE-based predictions. We retain MSE for transport supervision and moderate-intensity targets, while assigning a larger penalty to underestimation in the tail. The parameter qq controls this asymmetry.

The table below compares the submitted setting, q=10/11q=10/11, with the requested symmetric alternative, q=0.5q=0.5, and the baseline methods on MeteoNet. CSI combines detection and false alarms through CSI=(1/POD+1/(1−FAR)−1)−1\mathrm{CSI}=(1/\mathrm{POD}+1/(1-\mathrm{FAR})-1)^{-1}, so FAR alone cannot characterize overall detection performance. HSS additionally accounts for chance agreement.

With q=0.5q=0.5, TFEvo retains a higher CSI-M than the compared baselines and a similar HSS-M to AlphaPre. The asymmetric setting achieves the highest CSI-M, HSS-M, CSI-32, and HSS-32, demonstrating stronger detection performance under these joint criteria. Relative to the symmetric setting, it also yields higher FAR and MSE, indicating that the detection gains are accompanied by increased false alarm ratios and mse.

ModelCSI-M ↑POD-M ↑FAR-M ↓MSE ↓HSS-M ↑CSI-32 ↑POD-32 ↑FAR-32 ↓HSS-32 ↑
AlphaPre0.38890.47890.359112.80.52240.20210.25690.52270.3114
DuoCast0.3560.53710.533418.090.4790.20160.38420.73080.3068
exPreCast0.34120.4940.507617.950.47440.16680.30650.73420.2733
TFEvo, q=0.5q=0.50.39050.50420.374214.310.52140.1710.19750.45180.2666
TFEvo, q=10/11q=10/110.40940.65810.499817.510.55120.24710.47830.66550.3833

[Q4] Comparison with recursive transport

We compare TFEvo with two recursive variants: recursive transport–evolution, which feeds the evolved forecast into the next motion-estimation and transport step, and recursive transport with parallel evolution, which feeds back only the transported prior and applies evolution independently at each lead. TFEvo instead samples the latest observation once per lead, without feeding predicted fields back into transport. The recursive variants retain the main network components and training objective, with conditioning adapted to incorporate the preceding field by concatenating it with history features.

MethodCSI-M ↑CSI-40 ↑HSS-M ↑SSIM ↑LPIPS ↓
Recursive transport–evolution0.41590.25970.55550.76140.3103
Recursive transport + parallel evolution0.42530.28100.56690.78320.2602
TFEvo0.43400.29260.57490.79640.1843

Recursive transport with parallel evolution improves all reported metrics over recursive transport–evolution, supporting the benefit of excluding evolved forecasts from the transport feedback loop. TFEvo achieves further improvements across these metrics.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer GQ6u,

We appreciate your insightful comments. We have revised our paper and provided a point-by-point response, and we hope it addresses your concerns.

We are very keen to clarify any remaining points and look forward to your feedback.

Best regards,

Authors

Comments on the paper

3 comments

We thank the reviewers for their careful reading and detailed feedback. We have revised the paper in response to their comments and provided detailed responses to each reviewer separately. Changes are marked in blue\color{blue}{\text{blue}} in the revised paper.

AuthorsOpenReview ↗

Response to Common Questions—The Role of Coverage–Intensity Factorization

Further analysis of the existing MeteoNet predictions shows that TFEvo makes fewer false rain predictions while achieving higher CSI and HSS. This provides additional evidence for the role of coverage–intensity factorization in handling the large zero-valued regions of radar fields.

We examine how well the models predict whether it rains. This aspect is not directly assessed by the conventional benchmark thresholds. TFEvo improves CSI and HSS over the compared baselines and unfactorized evolution, with a lower false alarm ratio. By coupling intensity corrections to spatial coverage, ZFEvo allows intensity changes while keeping locations with inactive coverage at zero. Similar performance under q=0.5q=0.5 shows that the observed benefit persists under a symmetric tail objective. These findings complement the SSIM and LPIPS improvements in our factorization ablations.

ModelCSI-0 ↑POD-0 ↑FAR-0 ↓HSS-0 ↑CSI-1 ↑POD-1 ↑FAR-1 ↓HSS-1 ↑
AlphaPre (CVPR 2025)0.12400.99990.87600.00060.46360.95140.52580.5533
exPreCast (ICLR 2026)0.18860.97940.81030.12730.45440.90760.52510.5492
Unfactorized0.45500.94930.53430.53940.52450.91030.44850.6238
TFEvo, q=0.5q=0.50.53510.80480.39130.64300.55890.77950.34500.6691
TFEvo0.53610.81130.39350.64380.56000.78950.35020.6698

For -0, a pixel is classified as rain if its value is >0 and no rain if it is =0. For -1, rain is defined as ≥1 and no rain as <1. All thresholds use the original image scale, with rain as the positive class.

AuthorsOpenReview ↗

Summary by Authors

Dear Area Chair and Reviewers,

We sincerely appreciate the time and effort you have dedicated to reviewing our paper.

As highlighted by the reviewers, TFEvo offers a simple and effective formulation for separating echo movement from local evolution (1QCx, GQ6u, Tr7m). Its strengths include observation-anchored transport that excludes evolved forecasts from subsequent transport (GQ6u, Tr7m), the structural and perceptual benefits of zero-inflation-aware factorization (GQ6u), and strong detection performance with few parameters (GQ6u, 1QCx, Tr7m).

During the rebuttal, we strengthened the evidence for these contributions. Recursive comparisons show improved forecast skill from excluding evolved forecasts from transport, with further gains from observation anchoring. For factorized evolution, additional low-threshold analysis reveals higher CSI and HSS with lower false alarm ratios, complementing its structural and perceptual benefits at comparable benchmark CSI-M. Rapid-change analyses show better detection of echo appearance and disappearance than the compared baselines, with most targets remaining within the correction bounds. Symmetric-loss comparisons further show that TFEvo retains competitive detection performance without strong tail asymmetry. In their follow-up, 1QCx confirmed that their main concerns had been addressed.

We once again thank the Area Chair and all reviewers for your hard work and valuable comments.

Best regards,
The authors

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

Related papers

Discussion (0)

Sign in to comment.