PaperlayerView exact revision on arXiv

Source-linked AI summary

MW-Nowcast: Six-hour ensemble nowcasting of extreme precipitation

Ning Wang, Zuliang Fang, Weixin Jin, Zhongjian Lv, Shuang Qin, Pengcheng Zhao, Siqi Xiang, Jiang Bian, Haoyi Xiong, Nan Guan, Bin Zhang, Liangjie Zhang, Denvy Deng, Qi Zhang, Matt Corey, Jitu Keshri, Sridhar Iyer, Hongyu Sun, Kit Thambiratnam, Jonathan Weyn, Richard E. Turner, Haiyu Dong

arXiv:2609.34836v1physics.ao-phcs.LG

TL;DR

Reliable extreme-precipitation nowcasting remains less explored across the full six-hour window, despite the need for useful lead time for warnings and response. MW-Nowcast jointly learns shared deterministic precipitation structure and probabilistic residuals. Across three regions, it sustains equivalent skill for 2.3 to more than 3 h beyond retrained NowcastNet’s 3 h CSIN reference.

  • Problem

    Reliable extreme-precipitation nowcasting remains less explored across the full 0–6 h window, although forecasts need enough lead time to characterize intense-rainfall duration.

    Source: §1 · p. 2

  • Method

    MW-Nowcast jointly learns a deterministic forecast shared across ensemble members and probabilistic residuals around that forecast.

    Source: §1 · p. 3, §2 · p. 3

  • Results

    2.3 to more than 3 h of additional equivalent skill against retrained NowcastNet’s 3 h CSIN reference across nine region–threshold combinations; MW-Nowcast had the highest CSIN at every evaluated lead time and threshold in all three regions.

    Source: §5 · p. 9

  • Takeaways & Limitations

    Forecasts far enough ahead to characterize intense-rainfall duration are relevant to warning and emergency-response planning.

    Source: §1 · p. 2

    Evaluation covers three regions with dense radar networks; extending nowcasting to regions with sparse or no operational radar coverage remains future work.

    Source: §6 · p. 12

Abstract

from arXiv · show

Extending reliable nowcasting of extreme precipitation could provide critical additional time for warnings and emergency response during high-impact events such as flash floods. Radar-based generative machine-learning models have enabled skilful hyperlocal precipitation nowcasting, but accurate prediction of intense precipitation remains confined to the first few hours. Because storm-scale structure is predictable for longer than individual cells, a natural strategy is to predict that structure while generatively modelling only the uncertain local growth, decay, reorganisation and initiation of storms. Here we present Microsoft Weather Nowcast (MW-Nowcast), a six-hour ensemble radar nowcasting model that jointly learns a deterministic predictor to capture organised precipitation structure shared across ensemble members, and a generator to produce diverse local residuals around this shared prediction. Across independent test data from the United States, Europe and China, MW-Nowcast achieves higher detection skill than leading methods for heavy and extreme precipitation throughout the 6 h horizon. For the most intense rainfall, MW-Nowcast doubles the available warning time across all three regions, delivering 6 h forecasts with skill previously limited to 3 h for the leading generative baseline. A cost-loss decision analysis shows that MW-Nowcast retains substantial value for a broad range of applications even at 4-6 h, where alternative methods offer little benefit. These additional hours can give forecasters and emergency managers the time to warn and act before extreme rainfall strikes, helping to protect lives and property.

1 Introduction

MW-Nowcast separates predictable storm structure from uncertain local evolution to extend skilful extreme-precipitation detection across the 0–6 h nowcasting window. Longer-range forecasts can support warnings and other protective actions before intense rainfall strikes.

1 Introduction

  • Nowcasting provides local forecasts up to 6 h ahead, enabling warnings and operational mitigation; extreme-rainfall forecasts need enough lead time to characterize intense-rain duration over affected locations.Potential responses include positioning emergency resources and managing drainage, roads, aviation and energy operations.Source: §1 · p. 2
  • Radar-advection methods such as STEPS provide fast, interpretable ensembles, but transported echoes limit representation of local storm growth, decay, reorganisation and initiation, particularly beyond 3 h [5, 6].Numerical weather prediction can assimilate recent radar observations, but its computational cost makes updates at radar cadence challenging.Source: §1 · p. 2
  • DGMR produced sharper probabilistic ensembles, and NowcastNet conditioned generation on deterministic advection-informed evolution, but reported horizons were 90 min and 3 h, respectively.Earlier deterministic deep-learning forecasts improved nonlinear evolution over extrapolation but often smoothed at longer lead times, weakening intense cores and providing limited uncertainty information [7–9].Source: §1 · p. 2
  • Existing generative systems forecast complete future radar sequences, coupling storm-scale structure with uncertain local cell evolution as uncertainty broadens at longer lead times.The introduction proposes making the allocation between predictable dynamics and stochastic variability learnable to help maintain coherent storm organisation and ensemble variability.Source: §1 · pp. 2–3
  • MW-Nowcast jointly learns a deterministic forecast shared across ensemble members and member-dependent probabilistic residuals, extending skilful heavy- and extreme-precipitation detection to 6 h.In two high-impact events, it kept intense rain cores over affected areas longer than baselines; cost–loss analysis showed value at 4–6 h across protective decisions.Source: §1 · p. 3

2 MW-Nowcast

MW-Nowcast combines a shared deterministic radar forecast with sampled probabilistic residuals to produce six-hour ensemble nowcasts. Joint training and flow matching model common storm evolution alongside plausible local variability.

2 MW-Nowcast

  • A shared history encoder feeds a deterministic full-horizon forecast and a flow-matching residual decoder, whose sampled residuals are added to the common forecast to form ensemble members.Independent Gaussian noise draws generate different residual sequences conditioned on the observed radar history.Source: §2 · p. 3, Figure 1 · p. 4
  • Joint optimisation trains the encoder and both decoders to allocate future radar evolution between a common predictable component and conditional residuals for local variability.The paper reports that joint optimisation yields stronger long-lead high-threshold skill than training the deterministic decoder before the residual decoder.Source: §2 · p. 4
  • Training combines a deterministic prediction loss with a probabilistic residual flow-matching loss whose target is recomputed as the deterministic prediction changes.This makes the learned residual distribution change as the common forecast is updated.Source: §2 · p. 5
  • Flow matching trains a velocity field along a probability path from Gaussian noise to target residuals, and ODE integration transforms noise draws into plausible residual sequences rather than averaged futures.The sampled residual sequence is added to the deterministic prediction to provide detailed local structure.Source: §2 · p. 5

3 Experimental Setup

MW-Nowcast is trained and evaluated on three regional radar datasets using a common six-hour task, with strong-precipitation test sequences from later observation periods. The setup compares it with three representative nowcasting baselines using threshold-based, probabilistic, structural, and decision-utility diagnostics.

Experimental Setup

  • Each US, EU, and CN dataset uses a common 6 h task over 256 km × 256 km at 2 km resolution, with 130,000–175,000 training sequences and 4,000 later-period test sequences per region.Test sequences were selected under matching criteria to focus on strong-precipitation and severe convective regimes.Source: §3 · p. 5
  • MW-Nowcast is compared with STEPS, SimVPv2, and NowcastNet, with all baselines evaluated under the same protocol and SimVPv2 and NowcastNet retrained using official implementations.STEPS represents operational advection-based ensemble nowcasting; SimVPv2 is a deterministic prediction baseline; and NowcastNet is a generative ensemble model.Source: §3 · p. 5
  • Threshold-based verification uses CSIN for rain-rate exceedances at 16, 32, and 64 mm h−1 within a 5 × 5 grid-cell neighbourhood.Ensemble methods use four members and are aggregated with Probability Matched Mean when a single field is required.Source: §3 · p. 5
  • Additional diagnostics assess probabilistic skill with CRPS and CRPSS, spatial structure with PSD, and decision utility with REV across cost–loss ratios.Source: §3 · p. 6

4 Precipitation Cases

The section examines two high-impact cases—a tornadic convective storm in the United States and mountainous heavy rainfall in China—to assess how MW-Nowcast preserves organised storm structure while representing uncertain local evolution. The cases show how that balance differs as storm-scale predictability changes.

4 Precipitation Cases

  • At 32 and 64 mm h−1, MW-Nowcast maintained higher CSIN than baselines over much of Mississippi’s final three forecast hours, with the clearest separation at 64 mm h−1.It also better preserved multiscale spatial variability and achieved higher probabilistic skill than ensemble baselines over most lead times.Source: §4 · p. 6, Figure 2 · p. 7
  • Across the six-hour forecast, MW-Nowcast preserved the rain band, heavy-rain cores and system extent, while PySTEPS faded, SimVPv2 smoothed cores, and NowcastNet later weakened cores and over-spread lighter rain.Its deterministic predictor learns predictable storm-scale structure, while sampled residuals add local detail without generating the entire field.Source: §4 · p. 6, Figure 2 · p. 7
  • In Mississippi, the deterministic branch carried shared band organisation, while ensemble residuals varied convective-core locations, intensities and morphologies across plausible local pathways.The residual branch represents possible local intensification, weakening and reorganisation while preserving storm-scale evolution supported by observed history.Source: §4 · pp. 6–7, Figure 2 · p. 7
  • In Miyun, MW-Nowcast retained higher skill over the final three hours, especially at 64 mm h−1, and sustained heavy rainfall to about 5 h as baselines lost late intense cores.This case involved progressively scattered storm structure and rapidly declining deterministic predictability.Source: §4 · pp. 7–8, Figure 3 · p. 9
  • At 64 mm h−1, deterministic CSIN fell to zero within about 3 h, while the ensemble retained skill to about 5 h, supporting continued monitoring rather than de-escalation.By 6 h, ensemble members produced intense cores in varied locations as the deterministic signal faded, although members increasingly diverged from observed cores.Source: §4 · p. 8, Figure 3 · p. 9

5 Quantitative Evaluation

Across the US, EU and CN test sets, MW-Nowcast retained extreme-precipitation detection skill through six hours and outperformed the baselines on ensemble probability, spatial variability and decision-value evaluations.

5 Quantitative Evaluation

  • MW-Nowcast achieved the highest CSIN at every evaluated lead time and threshold across all three regions, retaining measurable skill through six hours.Against the retrained NowcastNet baseline’s 3 h CSIN, it sustained equivalent skill for an additional 2.3 to more than 3 h across nine region–threshold combinations. Jointly learned residual generation contributes to sustained long-lead, high-threshold skill.Source: §5 · p. 9, Figure 4 · p. 10
  • MW-Nowcast achieved lower grid-point CRPS than both NowcastNet and PySTEPS across all three regions, while PySTEPS outperformed NowcastNet.Grid-point CRPS measures predictive-distribution agreement with observations; because it weights locations equally, conservative forecasts can remain competitive despite lower high-threshold skill.Source: Figure 4 · p. 10, §5 · p. 11
  • MW-Nowcast remained close to the observed precipitation spectrum through six hours, matching observations more closely than PySTEPS at 2 and 4 h and comparably at 6 h.In the US, SimVPv2 progressively lost medium- and short-wavelength spectral power; similar spectral behavior for MW-Nowcast was observed in the EU and CN.Source: Figure 4 · p. 10, §5 · p. 11
  • Across all three regions, MW-Nowcast provided substantially greater REV than baselines at 16 and 32 mm h−1 thresholds and at 2, 4 and 6 h lead times.In the US, MW-Nowcast retained positive value across a broad range of cost–loss ratios at 4 and 6 h, while baseline values were close to zero.Source: Figure 4 · p. 10, §5 · p. 11

6 Conclusion

MW-Nowcast combines a shared deterministic prediction with conditional probabilistic residuals to produce six-hour ensemble radar nowcasts, sustaining skill for extreme precipitation across the full horizon. The conclusion also identifies radar-plus-satellite and numerical-weather-prediction inputs as a future direction for improving forecasts and extending coverage.

6 Conclusion

  • MW-Nowcast doubles warning time for the most intense rainfall from three hours to six and retains decision value at 4–6 h, where existing methods offer little.Its skilful extreme-precipitation ensemble nowcasts span the full six-hour window across the United States, Europe and China.Source: §6 · p. 12
  • End-to-end learning of a shared deterministic prediction and conditional probabilistic residuals is essential to sustaining skill at longer lead times.The model uses this division to provide six-hour ensemble radar nowcasting.Source: §6 · p. 12
  • Future work could combine radar with geostationary satellite imagery and numerical-weather-prediction guidance to improve forecasts of convective initiation and rapidly evolving storms.Satellite and numerical inputs could also support models trained on radar observations in regions with sparse or no operational radar coverage, extending the reach and decision value of long-lead nowcasting.Source: §6 · p. 12

7 Methodology

The methodology describes MW-Nowcast’s architecture, training and evaluation, including its datasets and baselines, for predicting 36 future radar frames from nine observations. It details how the model jointly separates shared, history-supported forecast structure from probabilistic residuals representing increasingly uncertain local precipitation evolution.

7.1 Deterministic-probabilistic formulation

  • The deterministic mapping predicts the full 36-frame future in parallel, while a jointly trained residual decoder models what that prediction omits using conditional flow matching and ODE integration.The residual target is the future sequence not represented by the current deterministic prediction, not an observed physical forcing; the history representation is shared by both mappings.Source: §7.1 · p. 16, §7.3 · p. 19

7.2 Model details

  • The encoder, deterministic decoder and residual decoder share history context; independent noise enters only the residual decoder, so ensemble spread represents local conditional uncertainty around the shared prediction.The modules contain approximately 18.6M, 11.2M and 23.4M trainable parameters, respectively.Source: §7.2 · pp. 17–18

7.3 Training objective function

  • Joint training uses an intensity-weighted reconstruction loss for the deterministic decoder and conditional flow matching for the residual decoder.Training ran for 200,000 optimization steps with batch size 8 using AdamW; no data augmentation beyond random sample shuffling was used.Source: §7.3 · p. 19, §7.3.1 · p. 19, §7.3.2 · p. 20

7.4 Datasets

  • All three regional datasets use nine input frames and 36 target frames at 10-minute cadence; US and EU samples are event-record-centred, while China uses radar-native peak-anchored sampling.Each region's main evaluation uses a 4,000-sample test set; the China test set requires strong echoes in every forecast hour and centre-region maxima of at least 45 dBZ.Source: §7.4 · pp. 20–21

7.5 Baselines

  • The comparisons include PySTEPS, SimVPv2 and NowcastNet under a common forecast protocol, with PySTEPS given a larger spatial context and four ensemble members.SimVPv2 uses the same intensity-weighted reconstruction loss as MW-Nowcast's deterministic head, while NowcastNet is retrained for the nine-frame input and 36-frame output setting.Source: §7.5 · p. 22, §7.5.1 · p. 22, §7.5.2 · p. 22

7.6 Evaluation settings and metrics

  • Grid verification uses 5 × 5 neighbourhood CSIN at 16, 32 and 64 mm h−1 and pointwise POD; probabilistic CRPS is computed from ensemble distributions in precipitation space.For ensemble grid metrics, CSIN uses the PMM field unless otherwise specified; CRPS uses four members for MW-Nowcast, PySTEPS and NowcastNet.Source: §7.6.1 · pp. 23–24, §7.6.2 · p. 24

7.6.3 Relative economic value

  • Economic value is assessed at 2-, 4- and 6-hour leads for 16 and 32 mm h−1 exceedances, while PSD diagnoses spatial structure; the formulation separates more predictable storm displacement from less predictable local evolution.Relative economic value is evaluated with four-member forecasts and pooled over grid cells and samples in each chronologically held-out compact test set.Source: §7.6.3 · pp. 24–25, §7.6.4 · p. 25, §7.7 · p. 26

7.7 Motivation for the residual formulation

  • Storm displacement and broad precipitation organization inform deterministic S1:T, while probabilistic R(k) represents member-dependent local departures, including growth, decay, merging, splitting and initiation.Local evolution becomes increasingly uncertain with lead time.Source: §7.7 · p. 26
  • The allocation between deterministic structure and probabilistic residual variability is learned jointly from data, not imposed by an explicit advection equation.Equation (25) provides physical motivation only; the model neither explicitly estimates v(t) nor uses the transport equation as a training constraint.Source: §7.7 · p. 26

8 Data availability

The study combines regional severe-weather records and radar observations from US, European and Chinese providers, with access governed by provider-specific arrangements.

8 Data availability

  • US severe-weather records came from NOAA/NCEI’s Storm Events Database, and radar observations came from NOAA’s MRMS system.The records were obtained through the official NCEI bulk-data archive; MRMS access information is available from NOAA.Source: §8 · p. 26
  • European severe-weather records came from ESWD under its User’s Agreement, while OPERA radar mosaics require EUMETNET application and approval.Source: §8 · p. 26
  • The China composite-reflectivity dataset was provided by CMA, and access requires application and approval from CMA.Source: §8 · p. 26

11 Contributions

The authors collectively conceived and designed the study, developed and evaluated its methodology, and contributed to scientific framing, manuscript revision, operational context, guidance, and oversight.

Study design and data

  • H.D., N.W. and Z.F. conceived the study and designed the overall methodology; Z.F. and Z.L. curated and processed the datasets.Source: §11 · p. 27

Experiments and writing

  • Z.F. and N.W. implemented the experimental pipeline, performed experiments and analysed results; N.W. drafted the initial manuscript.Source: §11 · p. 27

Scientific framing and revision

  • W.J. led the scientific framing and substantially revised the manuscript; J.W., R.E.T., K.T., N.W., Z.F. and H.D. contributed to framing, interpretation and revision.Source: §11 · p. 27

Operations and product context

  • S.Q., P.Z. and S.X. maintained operational data pipelines, while M.C., H.S., J.K. and K.T. contributed product requirements and operational context.Source: §11 · p. 27

Guidance and review

  • N.G., B.Z., L.Z., D.D., Q.Z., H.S., K.T. and S.I. provided scientific and strategic guidance and project oversight; H.D. supervised and coordinated the study.J.B. and H.X. reviewed the manuscript and provided feedback, and all authors reviewed and approved the final manuscript.Source: §11 · p. 27

12 Competing interests

The authors report no competing interests. Extended Data Figures document varied six-hour precipitation cases and show that the PMM ensemble summary recovers the observed high-intensity tail across three regions.

12 Competing interests

Extended Data Figures

  • US cases span organized rain-band movement, convective-core reorganization, and intense precipitation that expands, contracts, or weakens over six hours.The examples include a Tennessee flash-flood band, New York and Nebraska convection, and mountainous rainfall in western Hebei, Hunan, and Guizhou.Source: Extended Data Figures · pp. 28–30
  • Additional US examples show persistent or expanding mountainous rain systems and a moving system that contracts while intense cores persist.The cases are western Hebei, Hunan Shimen, and Guizhou Guiding–Majiang.Source: Extended Data Figures · pp. 31–33
  • European cases include expanding convective and heavy-precipitation systems with persistent, reorganizing intense cores, alongside modest contraction in one case.The examples are lightning-associated convection in southern Germany, tornado-associated convection in southern France, and heavy precipitation in southwestern France.Source: Extended Data Figures · pp. 34–36
  • Across the US, Europe, and China, PMM recovers substantially more high-intensity probability mass than the deterministic branch and closely approaches the observed tail.The distributions are pooled over 4,000 held-out events per region; the shaded tail is R ≥16 mm h−1.Source: Extended Data Figures · p. 37

1 Related works

Radar nowcasting has progressed from motion extrapolation to deterministic and generative deep-learning methods, while longer-lead extreme-precipitation forecasts still require both predictable storm structure and uncertain local evolution. MW-Nowcast combines these components in a jointly trained, pixel-level flow-matching ensemble model.

1.1 Historical perspective

  • At 3–6 h, local storm growth, decay, merging, splitting and initiation limit pure extrapolation, motivating models that retain predictable radar structure while representing uncertain evolution.STEPS and PySTEPS combine scale-dependent extrapolation with stochastic perturbations for fast, interpretable ensembles; optical-flow and tracking methods also exploit recent radar motion.Source: §1.1 · p. 38
  • Deterministic deep-learning models learn nonlinear spatiotemporal evolution, but point-wise losses can average plausible futures and weaken intense echoes at longer leads; generative models sample multiple futures [9].ConvLSTM, PredRNN, EarthFormer and SimVPv2 advanced deterministic prediction, while DGMR and NowcastNet demonstrated generative forecasting and physically structured extreme-precipitation nowcasting [9].Source: §1.1 · p. 38

1.2 Diffusion-based and flow-based nowcasting

  • Recent diffusion and flow-based methods represent forecast uncertainty, while cascaded or residual formulations connect deterministic evolution with stochastic refinement [17] [18] [19] [34] [35] [36].PreDiff and LDCast use latent diffusion; StormDiT and FlowCast explore flow-based approaches, while CasCast and DiffCast use cascaded or residual designs [17] [18] [19] [34] [35] [36].Source: §1.2 · p. 38
  • MW-Nowcast is a pixel-level, end-to-end flow-matching ensemble model that jointly learns shared deterministic structure and probabilistic local residuals, unlike fixed or separately trained residual cascades.Its residual target is recomputed as the deterministic prediction changes during joint training; ablations compare pixel versus latent space, explicit decomposition versus a single generative head, joint versus two-stage training, and flow matching versus EDM diffusion.Source: §1.2 · p. 38

2 Ablation study

The ablations support MW-Nowcast’s design: an explicit deterministic forecast anchor combined with jointly trained probabilistic residual generation performs best, with full pixel-space generation and flow matching providing additional advantages.

2.1 Decoder coupling and training strategy

  • OneStage’s lower CSIN5 and TwoStage’s remaining gap show that an explicit deterministic anchor and joint optimisation support stronger intense-precipitation detection.TwoStage recovers part of OneStage’s skill at 32 and 64 mm h−1, but remains weaker than MW-Nowcast at representative lead times and across the forecast-horizon mean.Source: §2.1 · p. 39

2.2 Generation space and diffusion objective

  • The Latent variant’s lower CSIN5 supports full pixel-space generation, while the qualitative comparison illustrates differences in retaining intense cores and storm structure over time.Compression and reconstruction may smooth compact precipitation cores or attenuate local extremes, especially at high rain-rate thresholds.Source: §2.2 · pp. 39, 41
  • MW-Nowcast exceeds both EDM-ODE and EDM-SDE in every reported CSIN5 column, although the differences are modest at the highest threshold and longest lead times.The variants retain the two-branch architecture and joint training; they replace flow matching with EDM and differ in deterministic ODE versus stochastic sampling.Source: §2.2 · p. 39

2 Ablation study

  • MW-Nowcast performs best among the evaluated ablations, with gains strongest for the jointly optimised deterministic-anchor and probabilistic-residual design.Supplementary Table 1 reports CSIN5 at representative lead times and as a mean over the six-hour horizon; higher scores are better.Source: §2.2 · pp. 39–40

3 Flow-matching visualisation

MW-Nowcast’s residual decoder transforms sampled noise into structured corrections around a shared deterministic forecast, while differences among ensemble members reveal where residual uncertainty is concentrated.

3.1 Single-member residual construction

  • For T+4–6 h US hail and European thunderstorm cases, the deployment Heun ODE solver transforms Gaussian noise (τ=0) into a structured residual (τ=1) added to the deterministic prediction.The residual modifies precipitation cores and fine-scale structure around the deterministic forecast.Source: §3.1 · pp. 42–44

3.2 Ensemble divergence along the flow

  • Because ensemble members share the deterministic forecast but start from different Gaussian samples, their spread visualizes conditional uncertainty; at τ=1, standard deviation clusters near precipitation cores and edges.Four independently sampled members are shown for the same cases and lead times.Source: §3.2 · p. 45

A Ensemble members B Mean & std · 4 Additional quantitative results · 5 Additional extended precipitation cases

The supplementary material expands regional verification and adds examples of ensemble uncertainty, storm evolution, and ablation variants. It reports MW-Nowcast’s regional skill alongside diagnostic metric definitions and additional cases.

A Ensemble members B Mean & std

  • The supplementary ensemble examples show four independently seeded members and residual standard deviation localised along precipitation cores and edges in a US hail case.The European thunderstorm example follows the same layout, showing four members and the ensemble mean and standard deviation of the residual.Source: A Ensemble members B Mean & std · pp. 46–47

4.1 Supplementary verification metrics

  • FAR measures forecast threshold exceedances that are not observed; CSIN allows spatial tolerance, while FSS compares neighbourhood event fractions, with higher values better.Source: §4.1 · p. 48

4.2 Regional quantitative diagnostics

  • MW-Nowcast has substantially higher POD across regions and thresholds, generally lower FAR than PySTEPS and comparable FAR to NowcastNet; SimVPv2’s lower FAR accompanies markedly lower POD.For EU and CN, MW-Nowcast remains closer to observed spectra at T+2 h, T+4 h and T+6 h, and retains positive economic value across broader cost–loss ratios, particularly at longer lead times.Source: §4.2 · pp. 48, 49, 51

5 Additional extended precipitation cases

  • The appendix adds two regional cases each for the US, EU and CN, plus two further US ablation examples, broadening qualitative coverage across storm types and settings.Source: §5 · p. 52

4.2 Regional quantitative diagnostics

  • MW-Nowcast has the highest CSIN5 across regions and thresholds, and from T+1 h the highest FSS across every evaluated region–threshold–neighbourhood combination.The CSIN5 comparison includes T+2 h, T+4 h, T+6 h and the mean over 36 lead times; FSS neighbourhoods range from 5 × 5 to 25 × 25 grid cells.Source: §4.2 · pp. 48, 50, §5.1 · p. 53

5.1 Additional regional forecast examples

  • The six regional examples span merging, splitting or fragmenting convective cores, locally renewed Alpine echoes, and organized systems that broaden, connect or shift downstream.Source: §5.1 · pp. 54–59

5.2 Additional ablation examples

  • Two additional US strong-precipitation cases compare MW-Nowcast with Latent, OneStage, TwoStage, EDM-ODE and EDM-SDE ablation variants through T+6 h.Source: §5.2 · pp. 60–62
Loading 2609.34836v1…