acceptodds
Under review as a conference paper at ICLR 2027

Probabilistic 3D Radar Nowcasting via Sparse Volumetric Transport

Abstract

Probabilistic 3D radar nowcasting forecasts how storms may move, develop vertically, and change in intensity. At high resolution, current methods reconstruct radar volumes from compressed representations such as 3D Gaussians and autoencoder latents, which can degrade high-intensity echoes. Diffusion and flow-matching models also repeat iterative sampling for every ensemble member, making high-resolution ensembles costly for time-critical nowcasting. We introduce Probabilistic Sparse Volumetric Transport (PSVT), which transports observed echoes along a small set of \(K\) 3D transport paths and estimates their growth and decay. Each path carries the full-resolution observed reflectivity through the native voxel space, so high-intensity echoes come directly from the observation rather than from reconstruction. The \(K\) path-wise predictions describe alternative evolutions of the same storm. Their deviations serve as \(K\) shared factors that define the spatiotemporal dependence of echo variations. PSVT models the local uncertainty of reflectivity at each voxel with a Gaussian distribution. Each member takes its variations from one linear combination of the \(K\) factors. This couples echo variations across space and time while avoiding iterative sampling and repeated network evaluations. On NEXRAD-3Y, PSVT improves single-forecast CSI at 40 dBZ by 114.0% over FlowCast and generates 16-member ensembles about \(30\times\) faster.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Precipitation Nowcasting, 3D Nowcasting

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_5kwDOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses the challenge of probabilistic nowcasting of three-dimensional (3D) radar data. Compared to two-dimensional (2D) data, generation is ensembles is often expensive, which presents a particular challenge given that nowcasting is time-critical. The existing works use different techniques to reduce the amount of computational resources required to produce the ensemble of forecasts. To this end, the state-of-the-art works use low-rank approximations, sampling voxels independently in a way that sacrifices between-voxel correlations. The authors pursue an alternative strategy, which they demonstrate not only to be computationally efficient but also, on some metrics, outperforming the state-of-the-art FlowCast. The authors implement the ideas of copula learning, separating the dependency structure from marginal distributions, using an optimal-transport technique for observed echoes. This corresponds to the known practice in motion-based video prediction, where the motion is modelled separately from the appearance.

Critical strengths

Soundness: the problem of making 3D nowcasting more efficient is an important conceptual problem in radar nowcasting. The authors propose a principled solution to this problem.

Clarity: the paper is clearly written (in general, but see the question below)

Quality: the evaluation includes a number of benchmarks, such as SEVIR, NEXRAD-3Y, TRRD. They reproduce baselines faithfully, including CasCast and performing adaptations of FlowCast. The adaptations look reproducible and fair, and the same can be said about the proposed method. They also go to an effort of reproducing the work on 2D SEVIR data.

Significance: it is significant to see the method that works faster and comparable in terms of scores most of the time, with significant improvements on some scores on some datasets.

Critical weaknesses

Clarity: see the question below. Also, while most of the work is written well, some caption need some improvements. This includes: "SEVIR comparison with deterministic and probabilistic models. " I know that SEVIR is a dataset, but the captions needs to be clearer.

Questions for authors

Clarity: while the idea of the copula modelling appears important to the proposed solution, it is only mentioned briefly in the Introduction and at the beginning of Section 3.5. I suggest there should be some reference in the appendix aimed for the clarity of existing solutions in copula models, introducing some form of review of current works in the area, and justifying the implementation in this paper (why is it exactly that you build on Salinas et al's implementation?).

6 comments

Q (1/3). Why build on Salinas et al. (2019)?

We appreciate your recognition of PSVT's principled design and reproducible evaluation.

Reviewer. ... I suggest there should be some reference in the appendix aimed for the clarity of existing solutions in copula models, introducing some form of review of current works in the area, and justifying the implementation in this paper (why is it exactly that you build on Salinas et al's implementation?).

Answer. We adopt the low-rank Gaussian covariance structure of Salinas et al. (2019) because it allows each ensemble member to reuse a small set of shared factors ak(q)a_k(q). Combining KK factors over nn forecast coordinates costs O(nK)O(nK) operations per member (Eq. 6). PSVT computes the factors once and generates 16 members in 3.3 s. FlowCast generates 16 members in 98.6 s (Table 1).

Low-rank covariance structure. Salinas et al. (2019) represent latent Gaussian dependence with covariance Σ=D+VV⊤\Sigma=D+VV^\top for nn variables. The diagonal entries of the nonnegative diagonal matrix DD give the variances of the independent Gaussian variation. The matrix V∈Rn×rV\in\mathbb R^{n\times r} contains rr shared factors with r≪nr\ll n. Sampling combines these factors with the independent variation at a cost of O(nr)O(nr) operations (App. A.4.1).

A forecast on NEXRAD-3Y contains 1.9×1081.9\times10^{8} values across 20 lead times, 36 height levels and a 512×512512\times512 grid. PSVT computes K=3K=3 shared factors ak(q)a_k(q) and the Gaussian reference mean μ(q)\mu(q) and standard deviation σ(q)\sigma(q) once per radar history. Each member reuses these network outputs and is censored to [0,70][0,70] dBZ (Eq. 6).

PSVT factors, marginals and member construction. We subtract the weighted mean of the path-wise predictions from each prediction Pτ(k)P^{(k)}_\tau. We then weight and normalize these deviations to obtain the shared factors ak(q)a_k(q) (Sec. 3.4). The local predictive distribution is a Gaussian censored to [0,70][0,70] dBZ (Eq. 3). PSVT draws M/2M/2 Gaussian coefficient vectors c\mathbf c and adds their negatives. We then whiten the paired set to obtain the member coefficient vectors c(m)\mathbf c^{(m)}. With equal member weights 1/M1/M, these vectors have zero empirical mean and identity empirical covariance (Eq. 7). The closed-form CRPS term in Eq. (9) fits the local predictive distributions without sampling ensemble members (Sec. 3.6 and App. A.4.3).

Transport in PSVT is advection of radar reflectivity. PSVT samples the latest observed volume XTinX_{T_{\mathrm{in}}} along the KK predicted 3D displacement fields Δτ(k)\Delta_\tau^{(k)} (Eq. 1).

Reference.

  • Salinas et al. (2019). High-Dimensional Multivariate Forecasting with Low-Rank Gaussian Copula Processes. NeurIPS.

Q (2/3). Review of existing copula models

Reviewer. ... some form of review of current works in the area ...

Answer. Copula models separate marginal predictive distributions from dependence between forecast variables. Weather-ensemble post-processing reconstructs this dependence from ensemble ranks or historical data. Learned copula models estimate dependence with neural networks or Gaussian processes (App. A.4.1 and A.4.2).

Copulas as models of dependence. By Sklar's theorem, a joint distribution can be written in terms of its marginal distributions and a copula (Sklar, 1959; Nelsen, 2006). The copula describes dependence between the variables. A Gaussian copula uses the correlation matrix of a latent Gaussian vector to specify this dependence (App. A.4.1).

  • Post-processing of weather ensembles. These methods first calibrate the marginal predictive distributions of a numerical weather prediction ensemble and then restore dependence across variables. Ensemble copula coupling (Schefzik et al., 2013) uses the rank order of the raw ensemble. The Schaake shuffle (Clark et al., 2004) uses the rank order of past observations. The Gaussian copula approach (Möller et al., 2013) estimates a Gaussian copula from past data.

  • Learned copulas for multivariate time series. The low-rank Gaussian copula process (Salinas et al., 2019) predicts the diagonal and low-rank covariance components with a recurrent network. Each series is transformed to a Gaussian scale by applying its empirical distribution function followed by the inverse standard-normal CDF. TACTiS and TACTiS-2 (Drouin et al., 2022; Ashok et al., 2024) learn a nonparametric copula with an attention-based network. Vine copulas construct multivariate dependence from pair copulas. Lopez-Paz et al. (2013) learn the parameters of the pair copulas with Gaussian processes.

  • Generative nowcasting models. CasCast (Gong et al., 2024) refines a deterministic forecast with latent diffusion. FlowCast (Perrone Ribeiro & Faganeli Pucer, 2026) generates forecasts with conditional flow matching in a latent space. Both models represent dependence across lead times and grid cells through their generative networks. In our comparisons, CasCast forecasts VIL on SEVIR and FlowCast forecasts reflectivity on the 3D radar datasets. Each ensemble member starts from a Gaussian noise sample and requires multiple network evaluations.

References.

  • Ashok et al. (2024). TACTiS-2: Better, Faster, Simpler Attentional Copulas for Multivariate Time Series. ICLR.
  • Clark et al. (2004). The Schaake Shuffle: A Method for Reconstructing Space–Time Variability in Forecasted Precipitation and Temperature Fields. Journal of Hydrometeorology.
  • Drouin et al. (2022). TACTiS: Transformer-Attentional Copulas for Time Series. ICML.
  • Gong et al. (2024). CasCast: Skillful High-Resolution Precipitation Nowcasting via Cascaded Modelling. ICML.
  • Lopez-Paz et al. (2013). Gaussian Process Vine Copulas for Multivariate Dependence. ICML.
  • Möller et al. (2013). Multivariate Probabilistic Forecasting Using Ensemble Bayesian Model Averaging and Copulas. Quarterly Journal of the Royal Meteorological Society.
  • Nelsen (2006). An Introduction to Copulas. Springer.
  • Perrone Ribeiro & Faganeli Pucer (2026). FlowCast: Advancing Precipitation Nowcasting with Conditional Flow Matching. ICLR.
  • Salinas et al. (2019). High-Dimensional Multivariate Forecasting with Low-Rank Gaussian Copula Processes. NeurIPS.
  • Schefzik et al. (2013). Uncertainty Quantification in Complex Simulation Models Using Ensemble Copula Coupling. Statistical Science.
  • Sklar (1959). Fonctions de répartition à n dimensions et leurs marges. Publications de l'Institut de Statistique de l'Université de Paris.

Q (3/3). How is the copula formulation used in PSVT?

Reviewer. ... justifying the implementation in this paper (why is it exactly that you build on Salinas et al's implementation?).

Answer. PSVT computes the shared volumetric factors ak(q)a_k(q) and the local Gaussian reference mean μ(q)\mu(q) and standard deviation σ(q)\sigma(q) once per radar history. Each member then reuses these network outputs at a cost of O(nK)O(nK) for nn forecast coordinates and KK factors. The path-wise predictions Pτ(k)P^{(k)}_\tau provide these factors without additional trainable parameters. The local predictive distributions are fitted with the closed-form CRPS term in Eq. (9). Evaluating this CRPS term requires no ensemble sampling.

PSVT as a copula model. At coordinates where the shared factor vector a(q)\mathbf a(q) is nonzero the weighted deviation vector d(q)\mathbf d(q) is normalized as a(q)=d(q)/∣d(q)∣2\mathbf a(q)=\mathbf d(q)/\\|\mathbf d(q)\\|_2 (Eq. 5 and App. A.3). The uncensored Gaussian reference variable L(q)L(q) has mean μ(q)\mu(q) and standard deviation σ(q)\sigma(q):

L(q)=μ(q)+σ(q)a(q)⊤c.L(q)=\mu(q)+\sigma(q)\mathbf a(q)^\top\mathbf c.

In the Gaussian reference, one coefficient vector c∼N(0,IK)\mathbf c\sim\mathcal N(\mathbf0,\mathbf I_K) is shared across spatial coordinates and lead times. For coordinates qq and q′q' with nonzero shared factor vectors, the covariance of L(q)L(q) and L(q′)L(q') is

Cov⁡[L(q),L(q′)]=σ(q)σ(q′)a(q)⊤a(q′).\operatorname{Cov}[L(q),L(q')]=\sigma(q)\sigma(q')\mathbf a(q)^\top\mathbf a(q').

This shared covariance component has rank at most KK and follows the low-rank structure of Salinas et al. (2019). Censoring this Gaussian construction to [0,70][0,70] dBZ gives the local predictive distribution of Eq. (3) at these coordinates.

PSVT uses 16 paired and whitened member coefficient vectors c(m)\mathbf c^{(m)} on NEXRAD-3Y. At coordinates where a(q)≠0\mathbf a(q)\ne\mathbf0, the uncensored finite ensemble exactly matches the mean and covariance of the Gaussian reference. These empirical moments use equal member weights 1/161/16. The censored members approximate the local predictive distributions of Eq. (3) (Eq. 7 and App. A.3).

  • Linear cost of member generation. Combining KK shared factors over nn forecast coordinates costs O(nK)O(nK) operations per member. PSVT uses K=3K=3 on NEXRAD-3Y and reuses the factors computed by the network. The number of ensemble members MM can be chosen separately from the number of paths KK. The M/2M/2 Gaussian coefficient vectors can span the KK coefficient dimensions for even M≥2KM\ge2K. Pairing and whitening these vectors gives the moment conditions of Eq. (7).

  • Factors from predicted echo evolution. PSVT derives the shared factors ak(q)a_k(q) from differences among the path-wise predictions Pτ(k)P^{(k)}_\tau for the current radar history. These predictions describe echo motion and changes in echo intensity. We replace the path-derived factors with random normalized factors to evaluate their contribution to dependence modelling. The local parameters μ(q)\mu(q) and σ(q)\sigma(q) remain unchanged (Table 3 and Fig. 6). Voxel CRPS changes by 0.02% in this ablation. Regional reflectivity CRPS increases by 21.5%. The CRPS scores for the temporal change (Δ20\Delta_{20}) and vertical difference (Vert. Δ3\Delta_3) in regional mean reflectivity also increase. FSS at 40 dBZ decreases by 25.0%.

  • Marginal fitting without ensemble sampling. The closed-form CRPS term in Eq. (9) fits the Gaussian reference mean μ(q)\mu(q) and standard deviation σ(q)\sigma(q) without sampling ensemble members (Sec. 3.6). At coordinates where a(q)≠0\mathbf a(q)\ne\mathbf0 the factor vector has unit norm. The paired and whitened coefficients then give the uncensored members a standard deviation of σ(q)\sigma(q) (Eq. 7). The empirical variance uses divisor MM. As an ablation, we replace the unit-length normalization of the weighted deviation vector d(q)\mathbf d(q) with one global scale. Member spread then also depends on the magnitude of the path deviations. Global normalization increases CRPS by 33.9% (Table 3). The separation of marginal distributions and dependence is also used in weather-ensemble post-processing (App. A.4.2).

Reference.

  • Salinas et al. (2019). High-Dimensional Multivariate Forecasting with Low-Rank Gaussian Copula Processes. NeurIPS.
Reviewer_JvJLOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

PSVT separates echo transport from intensity change for probabilistic 3D radar nowcasting. From encoded radar history, it predicts K displacement fields and path weights at each lead. The fields sample the latest full-resolution observation; a shared intensity predictor produces the path-wise forecasts and a separate deterministic forecast. Deviations among the path-wise forecasts define shared space-time factors, which are combined with local censored-Gaussian parameters to produce an ensemble without repeated network evaluations. The reported 40 dBZ CSI gains and 16-member inference-time reduction against the adapted FlowCast baseline are substantial.

Critical strengths
  1. Sampling reflectivity from the latest native-resolution observation avoids reconstructing the echo source from a compressed output. The high-threshold CSI results and the coarse-source ablation support the practical value of this design, although interpolation and motion errors can still affect individual strong echoes.

  2. Additional ensemble members reuse the same network output. The 16-member inference-time gain is substantial, and the path-factor ablation shows improvements in regional and temporal statistics over random factors. These results support the computational motivation and indicate that the factors carry useful dependence information.

Critical weaknesses
  1. The central forecast products and the meaning of a “path” remain difficult to determine from the main text. The relationships among the path-wise forecasts, the deterministic forecast, and the outputs evaluated in Table 1 are not clearly explained. Section 3.2 defines a path by sharing an output-channel index across leads, while Equation (1) samples each lead directly from the latest observation; no explicit cross-lead consistency condition is stated. The factor ablation supports useful dependence but does not establish that these fields are temporally continuous storm trajectories. This ambiguity affects both the interpretation of the proposed mechanism and the main comparison: the paper does not explicitly identify which output is PSVT-Single or which point forecast is used for the PSVT-Ensemble metrics in Table 1.

  2. The local probability model has an important conditional-information limit that is not evaluated. Equation (4) predicts the Gaussian parameters from a short vector f(q) of forecast-derived local quantities, plus lead and height, without directly using the history feature hτ. The history influences these quantities indirectly, but at the same lead and height, two histories producing the same local summary receive the same local marginal even if their risks of growth, decay, or initiation differ. The paper does not justify why this reduced summary is sufficient to model history-conditioned uncertainty, nor does it compare against parameter heads that directly receive history features. This limitation is particularly concerning for newly developing strong echoes. When all transported sources at a voxel are 0 dBZ, the bounded intensity update limits the deterministic forecast to at most 12 dBZ. Ensemble members can exceed this value through the local-distribution mapping, but the reported aggregate results do not establish their skill or spatial coherence for newly developing strong cores. Separately, wherever path spread is zero, the fallback uses independent variations across space-time coordinates. Its ability to represent coherent new development is also not evaluated.

Questions for authors
  1. Please clearly explain the data flow among the path weights, weighted transported volume, path-wise forecasts, deterministic forecast, local Gaussian parameters, and ensemble members. Which exact output is PSVT-Single, and which output supplies the point metrics for PSVT-Ensemble in Table 1? What, beyond a shared output-channel index, enforces or demonstrates temporal coherence of a transport path?

  2. Why do the Gaussian parameter heads in Equation (4) use only forecast-derived local summaries, lead, and height, without directly receiving history features? What evidence shows that these summaries retain the historical information needed to distinguish different risks of growth, decay, or initiation? Please compare local calibration and ensemble skill with parameter heads that also receive hτ or richer history-conditioned features.

  3. For strong echoes that initiate where the transported sources are weak or absent, what are the deterministic and ensemble hit rates and spatial coherence? Please separate these cases from the aggregate 40 dBZ results and evaluate the role of independent variations where path spread is zero.

13 comments

Q1 (and W1). Data flow and evaluation of PSVT-Single and PSVT-Ensemble

We thank the reviewer for the detailed questions on forecast construction and radar history.

Reviewer. [Q1-a] Please clearly explain the data flow among the path weights, weighted transported volume, path-wise forecasts, deterministic forecast, local Gaussian parameters, and ensemble members.

Reviewer. [Q1-b] Which exact output is PSVT-Single, and which output supplies the point metrics for PSVT-Ensemble in Table 1?

Answer. PSVT-Single is the deterministic forecast Y^det\widehat Y^{\mathrm{det}} computed from the weighted transported volume Gˉτ\bar G_\tau. PSVT-Ensemble consists of 16 members Y^(m)\widehat Y^{(m)}. CRPS uses all members. SSIM, LPIPS, HSS and CSI use the voxelwise ensemble mean for both PSVT-Ensemble and FlowCast-Ensemble (Sec. 3.1). The path-wise predictions Pτ(k)P^{(k)}_\tau supply the path spread and path support in the feature vector f(q)\mathbf f(q). Their deviations also define the shared volumetric factors ak(q)a_k(q) used to form the ensemble members.

1. Transport (Sec. 3.2). The displacement head DθΔ\mathcal D_{\theta_\Delta} maps the history feature hτh_\tau to K=3K=3 displacement fields Δτ(k)\Delta^{(k)}_\tau at each forecast lead. Each field produces a transported volume Gτ(k)G^{(k)}_\tau by backward sampling from the latest observed volume XTinX_{T_{\mathrm{in}}} at full resolution (Eq. 1). The path-weight head also receives hτh_\tau and predicts the weights wτ(k)w^{(k)}_\tau. A softmax across the paths normalizes these weights at each voxel. The weighted transported volume is Gˉτ=∑kwτ(k)Gτ(k)\bar G_\tau=\sum_k w^{(k)}_\tau G^{(k)}_\tau.

2. Reflectivity evolution (Sec. 3.2). The reflectivity predictor Rψ\mathcal R_\psi takes hτh_\tau and a transported volume and predicts a reflectivity change bounded by 12 dBZ (App. A.1). For path kk the predictor receives Gτ(k)G^{(k)}_\tau and gives

Pτ(k)=Gτ(k)+Rψ(hτ,Gτ(k)).P^{(k)}_\tau=G^{(k)}_\tau+\mathcal R_\psi(h_\tau,G^{(k)}_\tau).

The deterministic forecast uses the same predictor with Gˉτ\bar G_\tau as input. We add Rψ(hτ,Gˉτ)\mathcal R_\psi(h_\tau,\bar G_\tau) to Gˉτ\bar G_\tau and clip the result to [0,70][0,70] dBZ to obtain Y^τdet\widehat Y^{\mathrm{det}}_\tau.

3. Local Gaussian distribution (Sec. 3.3). At each space-time coordinate qq the pointwise location-correction head mϕm_\phi predicts a correction bounded by 6 dBZ. This correction is added to the deterministic forecast Y^det(q)\widehat Y^{\mathrm{det}}(q) to obtain the Gaussian mean μ(q)\mu(q). The pointwise scale head sηs_\eta predicts the Gaussian standard deviation σ(q)\sigma(q). These parameters describe the Gaussian reference before censoring. Both heads receive the feature vector f(q)\mathbf f(q) together with the lead and height embeddings (Eq. 4 and Table 4).

4. Ensemble members (Secs. 3.4 and 3.5). We subtract the weighted mean of the KK path-wise predictions from each prediction Pτ(k)P^{(k)}_\tau. We then weight and normalize these deviations to obtain the shared volumetric factors ak(q)a_k(q). Each member uses KK coefficients ck(m)c^{(m)}_k shared across all voxels and leads. At coordinates where a(q)≠0\mathbf a(q)\ne\mathbf0 the member is

Y^(m)(q)=C[0,70][μ(q)+σ(q)∑kck(m)ak(q)].\widehat Y^{(m)}(q)=\mathcal C_{[0,70]}\left[\mu(q)+\sigma(q)\sum_k c^{(m)}_k a_k(q)\right].

The operator C[0,70]\mathcal C_{[0,70]} clips reflectivity to [0,70][0,70] dBZ (Eq. 6). PSVT uses independent Gaussian variations at coordinates where the path spread is treated as zero (App. A.3). Each draw is paired with its negative across members (Sec. 3.5).

The deterministic forecast applies the reflectivity predictor after averaging the transported volumes. It therefore generally differs from the weighted mean of the path-wise predictions.

Q1-c (and W1). Temporal coherence of transport paths across forecast leads

Reviewer. [Q1-c] What, beyond a shared output-channel index, enforces or demonstrates temporal coherence of a transport path? [W1] ... Equation (1) samples each lead directly from the latest observation; no explicit cross-lead consistency condition is stated. The factor ablation ... does not establish that these fields are temporally continuous storm trajectories.

Answer. PSVT uses the same displacement head DθΔ\mathcal D_{\theta_\Delta} at every forecast lead to predict the displacement fields Δτ(k)\Delta^{(k)}_\tau from the lead-conditioned history feature hτh_\tau. We assess temporal coherence by comparing changes within a path across leads with differences between paths at the same lead. Changes within a path are smaller than differences between paths.

We also reorder the 16 complete member forecasts independently at each lead. The ensemble at each lead is unchanged. The permutation changes which forecasts belong to the same member across leads. The permutation raises the CRPS of the 20-minute change in regional mean reflectivity by 2.3% (App. F.4).

Definition of a transport path (Sec. 3.2). Transport path kk is the family of displacement fields Δτ(k)τ=1Tout\\{\Delta^{(k)}_\tau\\}_{\tau=1}^{T_{\mathrm{out}}}. The vector Δτ(k)(p)\Delta^{(k)}_\tau(p) gives the displacement from a source in the latest observed volume XTinX_{T_{\mathrm{in}}} to the target voxel pp at lead τ\tau. We measure the change of this field between adjacent leads at a fixed target voxel by Δτ+1(k)(p)−Δτ(k)(p)\Delta^{(k)}_{\tau+1}(p)-\Delta^{(k)}_\tau(p). A path describes an alternative evolution of the whole volume. It does not follow an individual echo through time. Each lead samples XTinX_{T_{\mathrm{in}}} at p−Δτ(k)(p)p-\Delta^{(k)}_\tau(p) through trilinear interpolation (Eq. 1).

Displacement prediction across leads (Sec. 3.1, App. A.1). PSVT computes the history feature hh once per radar history. An MLP predicts a channel-wise scale γτ\boldsymbol\gamma_\tau and offset βτ\boldsymbol\beta_\tau from the lead time. The lead-conditioned history feature is hτ=γτ⊙h+βτh_\tau=\boldsymbol\gamma_\tau\odot h+\boldsymbol\beta_\tau. PSVT uses no explicit consistency loss across leads. Each member keeps the same coefficients ck(m)c^{(m)}_k at all leads (Sec. 3.5). Shared temporal randomness is also used in continuous ensemble forecasting (Andrae et al., 2025).

Path and member coherence diagnostics (Table 20 in App. F.4). We compare leads that are 5 and 20 minutes apart on voxels where the weighted transported volume Gˉτ\bar G_\tau reaches 20 dBZ at either lead. The reported values are averages over all 29 test events from 2022. Events from 2020 and 2021 form the training and validation sets (App. B.1).

  • Displacement change across forecast leads. The displacement field of a path changes by 0.89 horizontal grid cells over 5 minutes. Two paths at the same lead differ by 10.97 grid cells. The change over 5 minutes is therefore 8% of the difference between paths. The change over 20 minutes is 3.52 grid cells and corresponds to 29% of the difference between paths.
  • Dominant path. The dominant path at a voxel is the path with the largest weight wτ(k)(p)w^{(k)}_\tau(p). This path remains the same at both leads for 96.3% of the voxels at a 5-minute interval and 84.1% at a 20-minute interval.
  • Member correlation across forecast leads. At a fixed voxel, we compare the standardized member variations u(m)(q)u^{(m)}(q) at two leads. We compute their correlation across the 16 members before censoring. When path spread is nonzero at both leads, this correlation is a(τ,p)⊤a(τ−ℓ,p)\mathbf a(\tau,p)^{\top}\mathbf a(\tau-\ell,p) for leads ℓ\ell steps apart (Eq. 8). This lag correlation is 0.85 for 5 minutes and 0.51 for 20 minutes.

Member permutation across leads. We independently permute the 16 complete three-dimensional member forecasts at each lead. A single permutation applies to all voxels at that lead. The set of 16 volumes, ensemble mean and all single-lead scores remain unchanged. The Δ20\Delta_{20} CRPS increases from 0.3153 to 0.3226. The paired difference is +0.0073+0.0073 with a 95% CI of [0.0050,0.0096][0.0050,0.0096]. The score increases in 25 of 29 test events.

Reference.

  • Andrae et al. (2025). Continuous ensemble weather forecasting with diffusion models. ICLR.

Q2-a (and W2, first part). Why use forecast summaries f(q)\mathbf f(q) in the Gaussian heads?

Reviewer. [Q2-a] Why do the Gaussian parameter heads in Equation (4) use only forecast-derived local summaries, lead, and height, without directly receiving history features?

Answer. The local-distribution heads mϕm_\phi and sηs_\eta use the feature vector f(q)\mathbf f(q) because its seven entries provide voxel-level predictors of reflectivity and uncertainty from quantities that PSVT already computes (Table 4 and Sec. 3.3). The history feature hτh_\tau is shared across height levels on a downsampled grid.

Different resolutions of hτh_\tau and the local distribution. The heads predict the parameters of the distribution in Eq. (3) at each voxel and lead. The history feature hτh_\tau has 80 channels on the Hf×WfH_f\times W_f grid shared by all 36 height levels (Sec. 3.1). The history feature hτh_\tau conditions both the displacement head DθΔ\mathcal D_{\theta_\Delta} and reflectivity predictor Rψ\mathcal R_\psi. Transport samples the latest observed volume XTinX_{T_{\mathrm{in}}} at full radar resolution. From these quantities, PSVT obtains the forecast reflectivity, unit reflectivity, reflectivity change, gradient magnitude, path-weight entropy, path spread and path support in f(q)\mathbf f(q). The path spread ∣d(q)∣2\\|\mathbf d(q)\\|_2 is the path-weighted standard deviation of the KK path-wise predictions Pτ(k)P_\tau^{(k)}. The path support indicates whether their deviations are nonzero (Sec. 3.4). Supplying hτh_\tau directly requires interpolating its 80 channels to the voxel grid.

Connection to EMOS. Ensemble model output statistics (EMOS) predicts a Gaussian mean as a linear function of ensemble member forecasts and a variance as a linear function of ensemble variance (Gneiting et al., 2005). EMOS estimates the coefficients by minimizing CRPS. PSVT likewise includes the deterministic forecast Y^det(q)\widehat Y^{\mathrm{det}}(q) and the path-weighted standard deviation ∣d(q)∣2\\|\mathbf d(q)\\|_2 among its predictors. The closed-form CRPS term in Eq. (9) fits the Gaussian mean μ(q)\mu(q) and scale σ(q)\sigma(q) without sampling ensemble members (Sec. 3.6). During the original model training this CRPS term propagates through f(q)\mathbf f(q) to the backbone. It also updates the displacement head DθΔ\mathcal D_{\theta_\Delta} and reflectivity predictor Rψ\mathcal R_\psi. The forecast-derived quantities are therefore trained as predictors of the local distribution.

Reference.

  • Gneiting et al. (2005). Calibrated probabilistic forecasting using ensemble model output statistics and minimum CRPS estimation. Monthly Weather Review.
Reviewer_4eS1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Probabilistic Sparse Volumetric Transport (PSVT), an efficient framework for probabilistic 3D radar nowcasting. A lightweight network predicts a small number of volumetric displacement fields and reflectivity updates, which are applied directly to the latest full-resolution observation. Deviations among the resulting path-wise predictions define shared spatiotemporal factors. These factors are combined with local censored-Gaussian distributions to generate ensemble members without repeated network evaluations. The method is evaluated on NEXRAD-3Y and TRRD, with an additional 2D comparison on SEVIR. On NEXRAD-3Y, PSVT reports a 114.0% relative improvement in single-forecast CSI at 40 dBZ over FlowCast and approximately 30× faster generation of 16-member ensembles. Its ensemble CRPS is comparable to FlowCast, while high-intensity CSI is substantially higher. Ablations examine transport resolution, intensity evolution, factor normalization, moment matching, and path-derived dependence.

Critical strengths
  1. A compelling efficiency–skill trade-off for volumetric nowcasting. PSVT improves high-intensity CSI with a lightweight model. On NEXRAD-3Y, it generates 16 members about 30× faster than FlowCast with comparable ensemble CRPS. TRRD and SEVIR experiments further support its practical value.
  2. Informative component and dependence analyses. The ablations support the roles of echo transport, intensity evolution, and ensemble construction. Random factors preserve voxel CRPS but worsen spatial and temporal scores, highlighting the value of structured dependence. The authors also acknowledge non-Gaussian coefficients and low-rank limitations.
Critical weaknesses
  1. Probabilistic validation remains incomplete. Voxelwise CRPS and point-forecast scores do not fully assess joint storm evolution. The dependence analysis compares path-derived factors with random factors, but not FlowCast. Moment matching also does not ensure calibrated probabilities after censoring. Regional and temporal comparisons with FlowCast, alongside high-intensity reliability or coverage diagnostics, would strengthen the probabilistic evaluation.
  2. The attribution of high-intensity gains needs stronger controls. Section 4.1 attributes PSVT’s high-threshold advantage to transporting observed echoes rather than reconstructing compressed representations. However, PSVT uses high-intensity-weighted classification and soft CSI losses, while FlowCast retains its original objectives. The single-forecast comparison also contrasts PSVT’s deterministic output with one prespecified FlowCast sample. These differences complicate isolating the benefit of avoiding reconstruction. Matched forecast products and training-objective controls would strengthen this interpretation.
Questions for authors
  1. How does PSVT compare with FlowCast on the regional, temporal-change, vertical-difference, and occupancy scores used in Table 3? Can the authors additionally report reliability or coverage diagnostics, particularly for high-intensity events and longer lead times?
  2. How sensitive is the reported high-intensity advantage to the training objective and forecast product? In particular, could the authors report average per-member CSI for both probabilistic methods and investigate whether comparable high-intensity supervision changes the gap?
  3. The transport bounds apply to total displacement from the latest observation, and the reflectivity update is limited to ±12 dBZ. How does skill vary for rapid storm motion, strong intensification, and newly developing convection? The limitation is acknowledged in Appendix E, but a stratified evaluation would clarify the method’s operating range.
10 comments

Q1-a (and W1). FlowCast comparison on regional, temporal, vertical and occupancy scores

We appreciate the suggestions on ensemble diagnostics and training-objective comparisons.

Reviewer. [Q1] How does PSVT compare with FlowCast on the regional, temporal-change, vertical-difference, and occupancy scores used in Table 3? [W1] The dependence analysis compares path-derived factors with random factors, but not FlowCast.

Answer. We evaluate the regional, temporal-change, vertical-difference and occupancy scores of Table 3 for both methods with 16 members each (App. B.3). PSVT has higher FSS than FlowCast at all three thresholds and improves FSS at 40 dBZ by 91.8%. FlowCast has lower values for all six CRPS scores (Table R1). Within PSVT, the path-derived shared volumetric factors ak(q)a_k(q) give lower values for all six CRPS scores than random normalized factors (App. D.3).

Table R1. Regional, temporal-change, vertical-difference and occupancy CRPS and FSS of Table 3 for PSVT and FlowCast-Ensemble on NEXRAD-3Y with M=16M=16 members. The PSVT entries repeat Table 3. Occupancy is defined at 40 dBZ, and its CRPS values are multiplied by 10310^3. The last column gives the mean paired difference over the 29 test events and its 95% interval.

ScorePSVT, path-derived factorsPSVT, random normalized factorsFlowCast-EnsemblePSVT − FlowCast [95% CI]
Regional CRPS ↓0.33710.40950.3283+0.009 [0.004, 0.015]
Temporal Δ20\Delta_{20} CRPS ↓0.31530.36140.3044+0.011 [0.007, 0.015]
Vertical Δ3\Delta_3 CRPS ↓0.22880.26130.2095+0.019 [0.014, 0.025]
Occupancy CRPS, region ↓1.1111.4250.919+0.19 [0.13, 0.26]
Occupancy CRPS, Δ20\Delta_{20} ↓1.1621.3341.082+0.08 [0.06, 0.10]
Occupancy CRPS, vertical Δ3\Delta_3 ↓0.7290.9090.579+0.15 [0.11, 0.20]
FSS at 20 dBZ ↑0.83080.79920.8109+0.020 [0.014, 0.026]
FSS at 30 dBZ ↑0.73530.67400.6638+0.071 [0.057, 0.086]
FSS at 40 dBZ ↑0.49610.37220.2587+0.237 [0.210, 0.263]

We evaluate all 29 test events from 2022. Training and validation use events from 2020 and 2021 (App. B.1). Regional CRPS evaluates mean reflectivity over 16×1616\times16 horizontal cells at each height and lead. Occupancy CRPS uses each member's fraction of voxels reaching 40 dBZ in the same region. The Δ20\Delta_{20} and Vert. Δ3\Delta_3 scores use 20-minute changes and differences three height levels apart, respectively.

For FSS, we first take the column maximum of each member and then average the 16 resulting fields. We use a 9×99\times9 neighborhood (App. B.3). The 95% CIs of the paired FSS differences remain above zero at all three thresholds.

Q1-b (and W1). Calibration after censoring for strong echoes and longer leads

Reviewer. [Q1] Can the authors additionally report reliability or coverage diagnostics, particularly for high-intensity events and longer lead times? [W1] Moment matching also does not ensure calibrated probabilities after censoring.

Answer. We assess calibration using the final censored members. We compute coverage, rank statistics and spread-skill ratios separately by reflectivity subset and lead group. PSVT has higher coverage and lower outer-bin mass than FlowCast in every reported lead group. Coverage on voxels observed above 40 dBZ is 252.9% higher over all lead times. PSVT also has a spread-skill ratio closer to 1 on all voxels and on echo voxels (Table R2 and App. D.1).

Moment constraints and calibration after censoring. At coordinates with a nonzero shared factor vector a(q)\mathbf a(q), moment matching fixes the uncensored empirical mean and variance to μ(q)\mu(q) and σ(q)2\sigma(q)^2 (Eq. 7). We evaluate probability calibration on the final censored members separately from moment matching. Clipping to [0,70][0,70] dBZ preserves whether a member reaches any threshold in (0,70](0,70] dBZ. The fraction of members reaching that threshold is therefore unchanged.

Table R2. Calibration diagnostics of the 16-member ensembles on NEXRAD-3Y, over all lead times and by lead group (minutes). Outer-bin mass is the share of observations that receive the lowest or the highest rank, with ties ranked at random. We compute coverage within each test event. For each lead group, we average these values over test events and the leads in that group. Rank histograms and spread-skill ratios are pooled over voxels.

Voxels / diagnosticModelAll+5–30+35–60+65–100
All voxels / Spread-skill ratioPSVT0.820.860.820.81
FlowCast0.430.490.430.41
Echo voxels / Spread-skill ratioPSVT0.660.680.670.65
FlowCast0.310.390.310.28
Echo voxels / CoveragePSVT0.770.810.780.74
FlowCast0.470.600.460.39
Echo voxels / Outer-bin massPSVT0.2200.1870.2110.249
FlowCast0.5200.3940.5320.597
Observed ≥ 40 dBZ / CoveragePSVT0.600.780.630.46
FlowCast0.170.380.130.05
Observed ≥ 40 dBZ / Outer-bin massPSVT0.3840.2170.3570.523
FlowCast0.8190.6050.8560.944

A voxel is covered if its observation lies between the smallest and largest of the 16 member values. The strong-echo coverage advantage persists through +65 to +100 minutes. Over all lead times, outer-bin mass (Hamill, 2001) is 53.1% lower for PSVT on voxels observed above 40 dBZ.

For each method, we select echo voxels where the observed reflectivity or that method's ensemble-mean reflectivity is at least 20 dBZ. PSVT's coverage is 63.8% higher on this subset over all lead times. Its spread-skill ratio is closer to 1 in every lead group (App. D.1).

Reference.

  • Hamill (2001). Interpretation of Rank Histograms for Verifying Ensemble Forecasts. Monthly Weather Review.

Q2-a (and W2). Sensitivity of strong-echo skill to the forecast product

Reviewer. [Q2] How sensitive is the reported high-intensity advantage to the ... forecast product? In particular, could the authors report average per-member CSI for both probabilistic methods ...? [W2] The single-forecast comparison also contrasts PSVT's deterministic output with one prespecified FlowCast sample.

Answer. To compare the same forecast product, we compute CSI separately for each of the 16 members of both methods and average the 16 values. PSVT has higher average per-member CSI at 40 dBZ than FlowCast (0.200 against 0.113). The paired difference is +0.087+0.087 (95% CI [0.081,0.093][0.081,0.093]). The advantage holds in all 29 test events (Table 5 in App. C.1).

PSVT has higher average per-member CSI at 30 and 40 dBZ. At 40 dBZ FlowCast's first member has a CSI of 0.114 and its average per-member CSI is 0.113.

Forecast products in Table 1. PSVT directly outputs the deterministic forecast Y^det\widehat Y^{\mathrm{det}} as PSVT-Single (Sec. 3.1). The released FlowCast has no separate deterministic forecast branch (Perrone Ribeiro & Faganeli Pucer, 2026). FlowCast-Single uses the first pre-specified member without selection by forecast skill. All members use the same sampling procedure with independent Gaussian noise (App. B.4). For both ensemble products, HSS and CSI use the voxelwise mean of the 16 members.

Reference.

  • Perrone Ribeiro & Faganeli Pucer (2026). FlowCast: Advancing Precipitation Nowcasting with Conditional Flow Matching. ICLR.
Reviewer_iaobOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PSVT for probabilistic 3D radar nowcasting. A lightweight network (about 0.5M parameters) encodes the radar history on a downsampled grid and predicts K 3D displacement fields together with path weights. These fields warp the latest full-resolution observed volume, and a reflectivity predictor adds bounded growth/decay, giving K path-wise predictions and a deterministic forecast.

Pointwise MLPs predict a local censored-Gaussian distribution (µ, σ) at each voxel and lead, trained with a closed-form CRPS loss. Ensemble members are formed in a Gaussian-copula-inspired way: normalized deviations of the K path-wise predictions serve as shared factors, and each member combines these factors with local distribution parameters using coefficients fixed across leads. This produces ensembles without additional network evaluations.

On NEXRAD-3Y, PSVT improves high-intensity echo prediction over FlowCast (+114% CSI at 40 dBZ for the single forecast), reaches comparable ensemble CRPS, and generates 16-member ensembles about 30× faster. The paper also reports results on TRRD and 2D SEVIR, and ablates the main design components.

Critical strengths

An effective and computationally efficient framework for high-resolution 3D nowcasting. The separation of motion estimation from reflectivity representation is a well-motivated design choice. By transporting observed reflectivity directly at native resolution rather than reconstructing it from compressed latent representations, PSVT addresses an important limitation of existing generative approaches. The experimental results are convincing, particularly for high-intensity echoes: on NEXRAD-3Y, CSI at 40 dBZ increases from 0.114 for FlowCast-Single to 0.244 for PSVT-Single, and the coarse-echo-source ablation further supports the benefit of native-resolution transport. The model also uses substantially fewer parameters and generates ensembles about 30× faster. These improvements are particularly valuable for severe-weather nowcasting, where accurate high-intensity echo prediction and computational efficiency are both important.

Critical weaknesses

More evaluation to support the main claims?

PSVT relies heavily on learned extrapolation of the latest observation with bounded growth and decay, yet it is not compared with classical extrapolation methods such as optical-flow-based nowcasting or STEPS (pySTEPS), which are natural baselines for this design. Some existing 3D baselines operate at lower resolution (128×128), while FlowCast is adapted to 3D by the authors. The evaluation is also limited to 29 test events on NEXRAD-3Y and 60 test sequences from distinct days on TRRD, with no variation across PSVT training runs reported.

On the probabilistic side, the evaluation relies mainly on CRPS, where PSVT performs comparably to FlowCast on NEXRAD-3Y (0.446 vs. 0.448) and slightly worse on TRRD (0.239 vs. 0.234). Calibration diagnostics such as rank histograms or spread–skill analysis are also missing.

These limitations matter because the paper's main claims concern both accurate high-intensity echo prediction and efficient probabilistic forecasting. Without comparisons against established extrapolation approaches and more thorough uncertainty evaluation, it is difficult to determine how much PSVT improves over simpler alternatives and whether its generated ensembles adequately represent forecast uncertainty.

Questions for authors
  1. Could the authors compare PSVT with classical extrapolation methods, such as optical-flow-based nowcasting or STEPS (pySTEPS), to better establish its advantage over traditional approaches?

  2. Could the authors report additional ensemble calibration diagnostics, such as rank histograms or spread–skill ratios, to assess whether the generated ensembles adequately represent forecast uncertainty?

Suggestion for future work: Evaluating PSVT on cases dominated by convective initiation would help clarify the limitations of transporting existing radar echoes when predicting newly developing storms.

7 comments

Q1-a (1/2) (and W). Comparison with classical optical-flow extrapolation

Thank you for recognizing PSVT's efficiency and its gains in high-intensity echo prediction.

Reviewer. [W] PSVT relies heavily on learned extrapolation of the latest observation with bounded growth and decay, yet it is not compared with classical extrapolation methods such as optical-flow-based nowcasting or STEPS (pySTEPS), which are natural baselines for this design. [Q1] Could the authors compare PSVT with classical extrapolation methods ... to better establish its advantage over traditional approaches?

Answer. We compare the deterministic forecast Y^det\widehat Y^{\mathrm{det}} (PSVT-Single) with classical optical-flow extrapolation on NEXRAD-3Y. Both methods use the evaluation protocol of Table 1 (App. B.3). PSVT-Single has higher HSS and CSI at all three thresholds. Its CSI at 40 dBZ is 47.0% higher (Table R1 and App. C.2).

Classical extrapolation baselines. We compare PSVT with optical-flow extrapolation and Eulerian persistence implemented in pySTEPS (Pulkkinen et al., 2019) on NEXRAD-3Y (App. B.1). Both baselines use observations from PSVT's five-volume input window and forecast the same 20 lead times on the native grid.

Table R1. Classical extrapolation and single forecasts on NEXRAD-3Y under the protocol of Table 1. For a single forecast, CRPS is the mean absolute error. The last column gives the paired difference over the 29 test events with its 95% confidence interval.

MetricEulerian persistenceOptical-flow extrapolationFlowCast-SinglePSVT-SinglePSVT − optical flow [95% CI]
CRPS0.9730.6840.6160.663−0.021 [−0.037, −0.004]
SSIM0.8380.8740.9230.881—
HSS at 20 dBZ0.4560.6070.6230.642+0.034 [0.025, 0.042]
HSS at 30 dBZ0.3170.4720.4610.559+0.087 [0.076, 0.097]
HSS at 40 dBZ0.1570.2810.2030.389+0.108 [0.095, 0.121]
CSI at 20 dBZ0.3140.4510.4660.486+0.035 [0.026, 0.044]
CSI at 30 dBZ0.1980.3170.3080.397+0.080 [0.069, 0.092]
CSI at 40 dBZ0.0870.1660.1140.244+0.079 [0.068, 0.089]
CSI at 40 dBZ, +5–30 min0.2160.3760.2560.408+0.032 [0.023, 0.040]
CSI at 40 dBZ, +35–60 min0.0570.1410.0810.232+0.092 [0.079, 0.104]
CSI at 40 dBZ, +65–100 min0.0300.0630.0300.128+0.065 [0.050, 0.081]

The 95% CIs of the HSS and CSI differences at all three thresholds remain above zero (Table R1). PSVT-Single has the higher CSI in all 29 test events at 30 and 40 dBZ and in 26 test events at 20 dBZ.

PSVT-Single's relative CSI improvement over optical-flow extrapolation at 40 dBZ increases across the lead groups. It is 8.5% for the +5–+30 min group and 103.2% for the +65–+100 min group (Table R1). At 30 dBZ PSVT-Single also has the higher CSI in every lead group.

At 40 dBZ, PSVT-Single has higher POD than optical-flow extrapolation (41% versus 25%) and lower FAR (62% versus 67%).

Reference.

  • Pulkkinen et al. (2019). Pysteps: An Open-Source Python Library for Probabilistic Precipitation Nowcasting (v1.0). Geoscientific Model Development.

Q1-a (2/2). What does PSVT add to classical extrapolation?

Answer. We estimate a separate optical-flow motion field at each height level to check whether PSVT's CSI advantage persists when motion can vary with height. The deterministic forecast Y^det\widehat Y^{\mathrm{det}} (PSVT-Single) has higher CSI at 40 dBZ than level-wise optical-flow extrapolation on both datasets. The relative improvement is 47.9% on NEXRAD-3Y and 51.9% on TRRD (Apps. C.2 and C.4).

Level-wise motion fields. The optical-flow baseline in Table R1 uses one motion field estimated from the column-maximum reflectivity to advect all 36 height levels. The level-wise variant reaches a CSI at 40 dBZ of 0.165. The paired difference between PSVT-Single and this variant is +0.080+0.080 (95% CI [0.070, 0.089]). PSVT-Single has the higher CSI at 40 dBZ in each of the 29 test events.

TRRD comparison with level-wise optical flow. We compare the original PSVT model with level-wise pySTEPS optical-flow extrapolation on all 60 test sequences from 60 different dates. Both methods use all five observations and forecast the same 20 leads at all 36 height levels on the native 512×512512\times512 grid (Table R2 and App. C.4).

Table R2. Single-forecast scores on TRRD after horizontal Pool4. Thresholds are in dBZ. Counts are summed over all 20 leads within each date, and the resulting scores are averaged over the 60 dates. Ratios with zero denominators are set to zero before averaging.

MethodCSI at 20 dBZ ↑CSI at 30 dBZ ↑CSI at 40 dBZ ↑POD at 40 dBZ ↑FAR at 40 dBZ ↓
Level-wise optical flow0.1710.1060.0470.1130.592
PSVT-Single0.2060.1410.0710.1950.612

Reflectivity update of PSVT. PSVT and optical-flow extrapolation transport echoes from the latest observed volume at native resolution. PSVT additionally predicts echo growth and decay with the reflectivity predictor Rψ\mathcal R_\psi (Sec. 3.2). On NEXRAD-3Y, removing this update reduces the CSI at 40 dBZ of the PSVT ensemble mean by 26.4% ("No intensity update", Table 3). The reduction in CSI at 40 dBZ is the largest among the component ablations of Table 3 (Sec. 4.2).

Optical-flow extrapolation against the baselines of Table 1. On NEXRAD-3Y, optical-flow extrapolation has higher HSS and CSI at 30 and 40 dBZ than ConvGRU3D, PhyDNet3D, SimVP3D, STC-GS and FlowCast (Table 1). At 40 dBZ the CSI of optical-flow extrapolation is 45.6% higher than that of FlowCast-Single. PSVT-Single and PSVT-Ensemble are the only entries of Table 1 that outperform optical-flow extrapolation in HSS and CSI at every threshold.

Q1-b (and W). Comparison with STEPS

Reviewer. [Q1] Could the authors compare PSVT with classical extrapolation methods, such as optical-flow-based nowcasting or STEPS (pySTEPS), to better establish its advantage over traditional approaches?

Answer. We compare PSVT with 16-member STEPS forecasts on NEXRAD-3Y in the volumetric and column-maximum settings. The column-maximum comparison retains the two-dimensional setting of STEPS (Bowler et al., 2006). In the volumetric evaluation, PSVT has 217.6% higher ensemble-mean CSI at 40 dBZ than STEPS. In the column-maximum evaluation, PSVT has 47.5% higher average per-member CSI at 40 dBZ (Table R3 and App. C.3).

Volumetric and column-maximum STEPS evaluations. Both evaluations use all 530 sequences of the NEXRAD-3Y test set (App. C.3).

  • (A) Volumetric evaluation. We run STEPS separately on each of the 36 height levels and evaluate the resulting 3D ensemble with the protocol of Table 1.
  • (B) Column-maximum evaluation. STEPS forecasts the 2D column-maximum reflectivity. We reduce each PSVT member to its column maximum. We compute CSI separately for every member of both ensembles and average the 16 CSI values within each ensemble.

Table R3. STEPS (16 members) and PSVT-Ensemble on NEXRAD-3Y. HSS, CSI and FSS are given at 20 / 30 / 40 dBZ. Differences are paired over the 29 test events. The 95% CI of each difference does not contain zero.

EvaluationScoreSTEPSPSVT-EnsemblePSVT − STEPS
(A) VolumetricCRPS ↓0.4580.446−0.012 [−0.021, −0.003]
SSIM ↑0.8620.872+0.010 [0.007, 0.013]
HSS, ensemble mean ↑0.495 / 0.317 / 0.1260.618 / 0.508 / 0.352+0.122 / +0.191 / +0.226
CSI, ensemble mean ↑0.344 / 0.195 / 0.0680.460 / 0.349 / 0.216+0.116 / +0.154 / +0.148
FSS ↑0.753 / 0.601 / 0.3050.831 / 0.735 / 0.496+0.078 / +0.135 / +0.191
(B) Column maximumCSI, average over members ↑0.464 / 0.349 / 0.1770.556 / 0.441 / 0.261+0.091 / +0.092 / +0.084

In the volumetric evaluation, PSVT-Ensemble has the higher HSS, CSI and FSS at all three thresholds in each of the 29 test events.

The regional CRPS of PSVT-Ensemble is 8.2% lower than that of STEPS (paired 95% CI of the difference [-0.039, -0.020]). The vertical-difference CRPS is 4.2% lower (paired 95% CI [-0.013, -0.007]). The paired 95% CI for the temporal-change CRPS difference between PSVT-Ensemble and STEPS is [-0.002, 0.007] (Table 3 and App. C.3).

In the column-maximum evaluation (B), PSVT has the higher average per-member CSI at all three thresholds. PSVT also has the higher average per-member CSI in each of the 29 test events.

Reference.

  • Bowler et al. (2006). STEPS: A Probabilistic Precipitation Forecasting Scheme Which Merges an Extrapolation Nowcast with Downscaled NWP. Quarterly Journal of the Royal Meteorological Society.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: changes to the paper

We sincerely thank all reviewers for their constructive feedback.

Summary. Compared with FlowCast, a strong recent baseline (ICLR 2026), PSVT makes high-resolution 3D ensemble nowcasting less costly while improving high-intensity echo prediction. The main model, PSVT-Small, has 0.524M parameters versus 163.339M for FlowCast (Table 1). On NEXRAD-3Y, a 16-member forecast on a 512×512512\times512 grid with 36 height levels takes 3.3 s for PSVT-Small versus 98.6 s for FlowCast, about 30 times faster (Table 1). At 40 dBZ, the single-forecast CSI is 0.244 versus 0.114 for FlowCast, and the ensemble-mean CSI is 0.216 versus 0.071 (Table 1). Scaling the backbone to 5.20M parameters (PSVT-Medium) reduces CRPS by 4.5% and improves ensemble-mean CSI at 40 dBZ by 6.9%. PSVT-Large (10.40M) reaches CRPS of 0.424 and ensemble-mean CSI at 40 dBZ of 0.233 (App. F.1).

The revised paper includes the additional experiments and clarifications discussed in our responses. Their locations are listed below.

Main text

  • Table 1: Optical-flow extrapolation.
  • Secs. 3.1 and 3.2, Table 1 caption: Clarification of PSVT-Single and PSVT-Ensemble and the definition of a transport path.
  • Table 2 caption: Clarification of the SEVIR task and evaluation details.

Appendix

  • App. A.4: A review of copula models and the relation of PSVT to Salinas et al. (2019).
  • App. C.1: Comparisons of matched forecast products, including average per-member CSI.
  • Apps. C.2 to C.4: Optical-flow extrapolation on both datasets and comparison with a 16-member STEPS ensemble.
  • Apps. B.5 and C.5: Paired intervals over the 29 test events and three independent PSVT training runs.
  • Apps. D.1 and D.2: Rank histograms, coverage and spread-skill ratios.
  • App. D.3: Regional, temporal-change, vertical-difference and occupancy scores against FlowCast for three PSVT backbone capacities.
  • Apps. E.1 to E.3: Strong-echo evaluations by echo source, growth of strong echoes and observed motion.
  • App. F.2: Results for models trained with MSE and MSE+CRPS, evaluated under the validation matched-FAR setting.
  • App. F.3: Comparisons of local-distribution heads with and without the history feature hτh_\tau.
  • Apps. F.4 and F.5: Diagnostics of the transport paths across lead times and member-permutation controls.
  • App. F.6: Comparison with a direct predictor using the same history backbone and decoder architectures as PSVT.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.