Pith. sign in

REVIEW 5 major objections 4 minor 50 references

FuXi-Nowcast claims that conditioning a deep-learning nowcaster on 3D atmospheric forecasts, not just radar, beats an operational 3-km numerical model for severe convection up to 12 hours ahead.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 18:47 UTC pith:2TIHYSTG

load-bearing objection A credible new ML nowcasting system with a real but under-verified NWP comparison; worth refereeing, but the headline claim needs stronger baselines and uncertainty. the 5 major comments →

arxiv 2512.08974 v2 pith:2TIHYSTG submitted 2025-12-03 physics.ao-ph cs.LG

FuXi-Nowcast: Environment-conditioned deep learning for severe convection nowcasting

Lei Chen , Zijian Zhu , Xiaoran Zhuang , Tianyuan Qi , Yuxuan Feng , Xiaohui Zhong , Hao Li This is my paper
classification physics.ao-ph cs.LG
keywords nowcastingconvective initiationdeep learningsevere convectioncomposite reflectivitywind gustsenvironmental conditioningmulti-task forecasting
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that a deep-learning nowcasting system can beat a 3-km operational numerical weather prediction model by conditioning its forecasts on three-dimensional atmospheric fields rather than radar only. It reports higher Critical Success Index for composite reflectivity, total precipitation, and wind gusts across multiple intensity thresholds and lead times up to 12 hours, with the biggest gains for heavy rainfall. The authors argue that environmental moisture and dynamics from a machine-learning forecast model carry the signals for convective initiation—storms that appear where no radar echo existed—and for maintaining the strength of intense convection that radar-only models tend to decay. If correct, this would give forecasters a multi-hazard, long-lead nowcasting tool that works before and during storm development.

Core claim

FuXi-Nowcast is an autoregressive multi-task Swin-Transformer model that jointly predicts composite reflectivity, precipitation, 2-meter temperature, wind speed, and wind gusts at 1-km resolution over East China. Its inputs combine radar, station observations, a high-resolution land analysis, and 70 three-dimensional variables from the machine-learning forecast model FuXi-2.0. The central claim is that this coupling lets the model predict convective initiation—first echoes above 30 dBZ—and sustain strong convection longer than the operational regional numerical model, whose diagnostic reflectivity and gust estimates are less skillful. Case studies and ablations support the interpretation tha

What carries the argument

The central mechanism is the fusion of large-scale three-dimensional atmospheric state (temperature, winds, humidity, geopotential at 13 pressure levels) with high-resolution local observations inside an autoregressive decoder. A convective-signal-enhancement module applies threshold-based pooling to amplify strong echo patterns before the Swin-Transformer feature extractor. A hybrid loss—Charbonnier for regression, Balanced L1 for extreme precipitation, LPIPS and SSIM for perceptual and structural fidelity, Focal and Dice for precipitation classification—keeps the model from dulling sharp convective cores. The workhorse idea is the information flow: the 3D environment guides when and where

Load-bearing premise

The result stands or falls on the fairness of the comparison: the operational model's wind gusts and reflectivity are proxies estimated from coarse fields, and precipitation is verified against the same high-resolution analysis product that also appears in model training.

What would settle it

A verification run that scores the operational model using its direct 10-meter wind and raw radar reflectivity predictions against station gusts and observed radar, rather than proxies, would test the headline CSI advantage. If the operational model's skill rises to match or exceed the deep-learning system once its own direct output is used, the paper's main claim would lose support.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the CSI advantage is real, a data-driven nowcaster can serve as the primary 0–12-hour severe-weather guidance, exceeding a 3-km operational numerical model especially for heavy rain thresholds.
  • Forecasters gain the ability to anticipate convective initiation before radar echoes appear, which is what allows earlier severe-thunderstorm warnings.
  • One model can issue reflectivity, precipitation, gusts, and temperature simultaneously, replacing separate single-variable nowcast products.
  • Running on a fast machine-learning forecast instead of a slow data-assimilation cycle could give more frequent updates and longer effective lead times in operations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the same environmental-conditioning recipe could transfer to other regions where pre-convective moisture is the limiting factor; the architecture is generic but would likely need local re-training and re-calibration.
  • Editorial inference: a sharper test of the intensity-decay claim would be to compute the slope of CSI or peak-intensity error versus lead time for ablated and full models; the paper shows the module helps but does not isolate its effect on the decay curve.
  • Editorial inference: if the moisture-infusion hypothesis is right, forecasts should degrade notably on days with strong capping inversions when the 3D fields fail to capture the cap; this is a falsifiable prediction for independent evaluation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. FuXi-Nowcast is a 1-km, 12-h multi-task deep-learning nowcasting system for eastern China. It fuses radar composite reflectivity, station wind observations, HRCLDAS surface analyses, and 70 three-dimensional fields from FuXi-2.0 forecasts in a Swin Transformer architecture with a convective signal enhancement module and a hybrid regression/perceptual/classification loss. The model is trained on April–September 2019–2023 and evaluated over April–July 2024. The central claim is that FuXi-Nowcast surpasses the operational CMA-MESO 3-km model in Critical Success Index for composite reflectivity, total precipitation, and wind gusts across thresholds and lead times up to 12 h, with the largest gains for heavy rainfall, and that a case study shows improved convective initiation. The quantitative comparison is given in Fig. 1, with model and evaluation details in Sec. 4.

Significance. If substantiated, the result is significant: it would show that conditioning a radar-based nowcaster on three-dimensional machine-learning forecast fields can mitigate known radar-only failure modes (convective initiation, intensity decay) and extend skillful lead times well beyond 0–2 h while producing multi-hazard outputs. The paper has real strengths: explicit multi-source fusion, an autoregressive formulation, clear description of the hybrid loss design, a large training dataset, and an independent one-season test period. However, the quantitative evidence is not yet robust. The NWP baseline uses diagnostic proxy fields for reflectivity and gusts; TP verification uses the same HRCLDAS analysis that serves as the training target; there are no uncertainty estimates or significance tests; and the abstract's claim about persistence/extrapolation baselines is not supported by any result in the main text. The Discussion itself lists remaining limitations (precise CI localization, generalization across regions), consistent with the need for a stronger evaluation before the headline claim can be accepted.

major comments (5)
  1. [§4.2, Fig. 1] The NWP baseline uses proxy fields: CMA-MESO CR is the diagnosed reflectivity field and GS is estimated from WS10M, rather than the observed radar mosaic and station gust observations used to evaluate FuXi-Nowcast. Diagnosed reflectivity depends on the model's microphysics and Z–q relation and may be systematically biased in convective cells, while gusts derived from 10-m wind cannot capture sub-grid convective gust processes. If these proxies suppress CMA-MESO's hit rates at high thresholds, the CSI advantage shown in Fig. 1 is inflated. Please add a sensitivity analysis using observed radar reflectivity and station gusts as verification for CMA-MESO, or restrict the conclusions to 'vs. CMA-MESO's operational diagnostic output'.
  2. [§4.1/§4.5, Table 1] Total precipitation is verified against HRCLDAS, which is also FuXi-Nowcast's training target and an input source (Table 1). Comparing both models with the same analysis used to train one model risks giving the ML model an advantage if HRCLDAS contains systematic biases or smoothness that the model has learned. Please add verification against independent hourly gauge or gauge-adjusted QPE data for a subset of the evaluation period, and report HRCLDAS error characteristics for heavy precipitation. Without this, the headline 'largest gains for heavy rainfall' claim is not fully established.
  3. [Fig. 1, §2.1] The CSI differences are presented without uncertainty quantification or significance testing. The evaluation covers only April–July 2024, and for TP>20/40 mm the number of qualifying events is likely small. Please provide bootstrap confidence intervals, significance tests per threshold and lead time, or at least a table of event counts. This is load-bearing because the advantage at high precipitation thresholds may be within sampling variability.
  4. [Abstract, §2] The abstract states that FuXi-Nowcast outperforms 'operational numerical, persistence and extrapolation baselines,' but the main text contains no quantitative comparison with persistence or radar extrapolation. Section 1 discusses these methods qualitatively, and Fig. 1 shows only FuXi-Nowcast vs. CMA-MESO. Please either add the missing baseline results (e.g., optical-flow nowcast CSI for CR/TP thresholds) or remove the claim from the abstract.
  5. [§4.1, §4.3] During training the three-dimensional atmospheric fields come from ERA5, while at inference they come from FuXi-2.0. The paper does not quantify the distribution shift between ERA5 and FuXi-2.0 or validate FuXi-2.0's accuracy over the target region. Since environmental conditioning on 3D ML forecasts is the paper's central mechanism, this is not cosmetic. Please add an assessment of FuXi-2.0 vs. ERA5 during the evaluation period, or a sensitivity experiment in which inference inputs are drawn from ERA5, and discuss how FuXi-2.0 error affects nowcast skill.
minor comments (4)
  1. [Fig. 1 caption] The third row is labeled 'wind gust (WS)' but the variable is GS; use consistent variable names. Also, because the y-axes vary by threshold, please state this in the caption and consider using a common scale for at least the lower thresholds.
  2. [§1] Typos and grammar issues: 'developing an multi-variable nowcasting models', 'The urgent for accurate nowcasting', and 'a deep learning-based nowcasting model capable of...' should be corrected.
  3. [§2.2] The case study is dated 16 June 2025, while the evaluation period is April–July 2024. If this is a separate illustrative case, state that explicitly to avoid confusion with the statistical evaluation.
  4. [Table 1] The 'Role' column 'I & O' is ambiguous for HRCLDAS and radar/station data. Please specify which variables are used as inputs, which as outputs, and which as both.

Circularity Check

0 steps flagged

No derivation-equals-input circularity; the CSI evaluation is external and held out. The FuXi-2.0 self-citation and the HRCLDAS-as-reference design are validity caveats, not circular steps.

full rationale

The paper's central result is an empirical comparison, not a derived identity. FuXi-Nowcast is trained on 2019-2023 data with the hybrid loss in Eq. (1), and its 2024 skill is measured by CSI (Eq. 8) against observed CR/GS and the HRCLDAS reference for TP. CSI is not a term in the training loss and no closed-form relation forces the model's thresholded CSI scores; the evaluation year is outside the training interval, so the score is not an in-sample fit by construction. The FuXi-2.0 references [34,35] are self-citations, but they supply the 3-D input fields rather than the verification outcome; if FuXi-2.0 forecasts were inaccurate, the environment-conditioning attribution would weaken, but the prediction would not reduce to the input. The HRCLDAS statement ('HRCLDAS data ... serve as the reference dataset for model training and evaluation') and the CMA-MESO proxy disclosures in Sec. 4.2 ('diagnosed CR', 'GS variable is estimated from WS10M') are legitimate baseline-fairness and external-validity limitations, not circularity: the same reference and the same held-out period are used for both systems. The abstract's claim of ablations is unsupported in the main text, but that is missing evidence, not a circular reduction. No Eq. X = Eq. Y by construction, no fitted parameter renamed as a prediction, and no load-bearing uniqueness argument were found.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim rests on several domain assumptions about data products and baseline fairness rather than on new physical entities. The only free parameters are ML loss/hyperparameters and an ad hoc training filter; no new particles, forces, or conserved quantities are introduced.

free parameters (5)
  • Hybrid loss weights (lambda_cl1, lambda_bl1, lambda_lpips, lambda_SSIM, lambda_focal, lambda_dice) = 1.0, 1.0, 0.1, 0.1, 1.0, 0.1
    Hand-chosen weights in Eq. (1); they control the balance between regression, perceptual, structural, and classification losses and therefore affect the reported CSI scores.
  • Balanced-L1 TP bin thresholds = 0.1, 0.77, 1.86, 3.61, 6.43, 10.99, 18.33, 30.16, 49.24, 80, 100 mm
    Predefined bins for L_bl1 in Eq. (3); chosen to be approximately log-uniform rather than derived from data, and they set the weighting of extreme-rain samples.
  • Focal loss hyperparameters = alpha=0.25, gamma=1.5
    Set by hand in Eq. (6); they affect rain/no-rain classification and, indirectly, the spatial overlap and CSI for precipitation.
  • Training sample filter thresholds = >10 grid points with TP>1 mm or GS>10.8 m/s
    Ad hoc criterion in Section 4.1 for retaining high-impact weather samples; it shapes the training distribution and can influence apparent skill on severe events.
  • Convective enhancement module thresholds and pooling parameters = Not reported
    Section 4.3 describes 'threshold-based pooling' but gives no thresholds or pooling hyperparameters; these tuned design choices directly affect preservation of strong convection.
axioms (5)
  • domain assumption 70-variable 3D atmospheric fields from ERA5/FuXi-2.0 capture the pre-convective environment
    Section 4.1 uses these fields to supply 'large-scale three-dimensional atmospheric information', but the paper does not separately validate that they contain the necessary moisture/dynamics for convective initiation.
  • domain assumption HRCLDAS precipitation analysis is a valid ground truth for total precipitation
    Section 4.1 uses HRCLDAS TP as both training target and evaluation reference; HRCLDAS itself blends station, satellite, and NWP data, so its errors are inherited by the evaluation.
  • domain assumption IDW interpolation of station wind/gust observations to a 1-km grid represents mesoscale gust fields
    Section 4.1 inverse-distance weights gusts from more than 1300 stations to a 768x768 grid; convective gusts vary at scales finer than station spacing, so station interpolation may miss the true gust maxima.
  • domain assumption CMA-MESO diagnosed CR and WS-estimated GS are comparable to observed reflectivity and station gusts
    Section 4.2 constructs the baseline from diagnosed composite reflectivity and gust estimates from 10-m wind; if these proxies understate real NWP skill, the comparison is unfair to CMA-MESO.
  • domain assumption Autoregressive 12-step rollout remains stable and skillful
    Section 4.3 appends the model's own predictions and continues autoregressively to 12 h, but no error-accumulation diagnostics are reported.

pith-pipeline@v1.3.0-alltime-deepseek · 11687 in / 16133 out tokens · 143385 ms · 2026-08-03T18:47:29.025221+00:00 · methodology

0 comments
read the original abstract

Severe convection produces localized hazards that often require warnings before radar echoes fully reveal storm development. Convective initiation and the maintenance of intense convection remain challenging for radar-only nowcasting because pre-convective signals may be absent from recent radar observations and strong echoes often decay rapidly in forecasts. Here we present FuXi-Nowcast, an environment-conditioned deep learning system that combines high-resolution observations with three-dimensional atmospheric forecasts to predict composite reflectivity, precipitation, wind gusts, and surface variables up to 12 h ahead. In April--July 2024 evaluations over East China, FuXi-Nowcast outperforms operational numerical, persistence and extrapolation baselines for reflectivity and precipitation. Case studies, diagnostics, and ablation experiments suggest that atmospheric moisture information and explicit preservation of strong convective signals contribute to forecasts of convective initiation and maintenance. These results show that environmental conditioning can mitigate important failure modes of radar-only nowcasting for high-impact convective weather.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 2 canonical work pages

  1. [1]

    Science Advances6(19), 1838 (2020) https://doi

    Raymond, C., Matthews, T., Horton, R.M.: The emergence of heat and humidity too severe for human tolerance. Science Advances6(19), 1838 (2020) https://doi. org/10.1126/sciadv.aaw1838 15

    work page doi:10.1126/sciadv.aaw1838 2020

  2. [2]

    Meteorological applications30(4), 2124 (2023)

    Bojinski, S., Blaauboer, D., Calbet, X., De Coning, E., Debie, F., Montmerle, T., Nietosvaara, V., Norman, K., Ba˜ n´ on Peregr ´ ın, L., Schmid, F.,et al.: Towards nowcasting in europe in 2030. Meteorological applications30(4), 2124 (2023)

    2030

  3. [3]

    Proceedings of the National Academy of Sciences120(28), 1–8 (2023)

    Tripathy, K.P., Mukherjee, S., Mishra, A.K., Mann, M.E., Williams, A.P.: Cli- mate change will accelerate the high-end risk of compound drought and heatwave events. Proceedings of the National Academy of Sciences120(28), 1–8 (2023)

    2023

  4. [4]

    Fowler, H.J., Blenkinsop, S., Green, A., Davies, P.A.: Precipitation extremes in

  5. [5]

    Nature639(8055), 667–672 (2025)

    Haslinger, K., Breinl, K., Pavlin, L., Pistotnik, G., Bertola, M., Olefs, M., Greilinger, M., Sch¨ oner, W., Bl¨ oschl, G.: Increasing hourly heavy rainfall in austria reflected in flood changes. Nature639(8055), 667–672 (2025)

    2025

  6. [6]

    Bulletin 68, 2 (2019)

    Schmid, F., Wang, Y., Harou, A.: Nowcasting guidelines—a summary. Bulletin 68, 2 (2019)

    2019

  7. [7]

    Earth and Space Science7(2), 2019–000812 (2020)

    Chen, L., Cao, Y., Ma, L., Zhang, J.: A deep learning-based methodology for precipitation nowcasting with radar. Earth and Space Science7(2), 2019–000812 (2020)

    2019

  8. [8]

    Atmosphere11(3), 267 (2020)

    Franch, G., Nerini, D., Pendesini, M., Coviello, L., Jurman, G., Furlanello, C.: Precipitation nowcasting with orographic enhanced stacked generalization: Improving deep learning predictions on extreme events. Atmosphere11(3), 267 (2020)

    2020

  9. [9]

    Bulletin of the American Meteorological Society95(3), 409–426 (2014)

    Sun, J., Xue, M., Wilson, J.W., Zawadzki, I., Ballard, S.P., Onvlee-Hooimeyer, J., Joe, P., Barker, D.M., Li, P.-W., Golding, B.,et al.: Use of nwp for nowcasting convective precipitation: Recent progress and challenges. Bulletin of the American Meteorological Society95(3), 409–426 (2014)

    2014

  10. [10]

    Nature597(7878), 672–677 (2021)

    Ravuri, S., Lenc, K., Willson, M., Kangin, D., Lam, R., Mirowski, P., Fitzsi- mons, M., Athanassiadou, M., Kashem, S., Madge, S.,et al.: Skilful precipitation nowcasting using deep generative models of radar. Nature597(7878), 672–677 (2021)

    2021

  11. [11]

    Weather and forecasting27(4), 938–953 (2012)

    Huang, L.X., Isaac, G.A., Sheng, G.: Integrating nwp forecasts and observation data to improve nowcasting accuracy. Weather and forecasting27(4), 938–953 (2012)

    2012

  12. [12]

    Advances in Science and Research15, 159–172 (2018)

    Sheridan, P.: Current gust forecasting techniques, developments and challenges. Advances in Science and Research15, 159–172 (2018)

    2018

  13. [13]

    Bulletin of the 16 American Meteorological Society102(3), 635–659 (2021)

    Majumdar, S.J., Sun, J., Golding, B., Joe, P., Dudhia, J., Caumont, O., Chan- dra Gouda, K., Steinle, P., Vincendon, B., Wang, J.,et al.: Multiscale forecasting of high-impact weather: Current status and future challenges. Bulletin of the 16 American Meteorological Society102(3), 635–659 (2021)

    2021

  14. [14]

    Journal of Advances in Modeling Earth Systems14(3), 2021– 002765 (2022)

    Hess, P., Boers, N.: Deep learning for improving numerical weather prediction of heavy rainfall. Journal of Advances in Modeling Earth Systems14(3), 2021– 002765 (2022)

    2021

  15. [15]

    Weather and Forecasting18(4), 562–584 (2003) https://doi

    Roberts, R.D., Rutledge, S.: Nowcasting storm initiation and growth using goes- 8 and wsr-88d data. Weather and Forecasting18(4), 562–584 (2003) https://doi. org/10.1175/1520-0434(2003)018⟨0562:NSIAGU⟩2.0.CO;2

    doi:10.1175/1520-0434(2003)018 2003

  16. [16]

    Geophysical Research Letters47(11), 2020–087035 (2020)

    Bai, L., Chen, G., Huang, L.: Convection initiation in monsoon coastal areas (south china). Geophysical Research Letters47(11), 2020–087035 (2020)

    2020

  17. [17]

    Journal of atmospheric and oceanic technology10(6), 785–797 (1993)

    Dixon, M., Wiener, G.: Titan: Thunderstorm identification, tracking, analysis, and nowcasting—a radar-based methodology. Journal of atmospheric and oceanic technology10(6), 785–797 (1993)

    1993

  18. [18]

    In: Handbook of Mathematical Models in Computer Vision, pp

    Fleet, D., Weiss, Y.: Optical flow estimation. In: Handbook of Mathematical Models in Computer Vision, pp. 237–257. Springer, ??? (2006)

    2006

  19. [19]

    Journal of Atmospheric and Oceanic Technology26(4), 719–732 (2009)

    Han, L., Fu, S., Zhao, L., Zheng, Y., Wang, H., Lin, Y.: 3d convective storm identification, tracking, and forecasting—an enhanced titan algorithm. Journal of Atmospheric and Oceanic Technology26(4), 719–732 (2009)

    2009

  20. [20]

    arXiv preprint arXiv:1912.12132 (2019)

    Agrawal, S., Barrington, L., Bromberg, C., Burge, J., Gazen, C., Hickey, J.: Machine learning for precipitation nowcasting from radar images. arXiv preprint arXiv:1912.12132 (2019)

    Pith/arXiv arXiv 1912

  21. [21]

    arXiv preprint arXiv:2005.04988 (2020)

    Prudden, R., Adams, S., Kangin, D., Robinson, N., Ravuri, S., Mohamed, S., Arribas, A.: A review of radar-based nowcasting of precipitation and applicable machine learning techniques. arXiv preprint arXiv:2005.04988 (2020)

    Pith/arXiv arXiv 2005

  22. [22]

    Pulkkinen, S., Nerini, D., P´ erez Hortal, A.A., Velasco-Forero, C., Seed, A., Germann, U., Foresti, L.: Pysteps: An open-source python library for probabilis- tic precipitation nowcasting (v1. 0). Geoscientific Model Development12(10), 4185–4219 (2019)

    2019

  23. [23]

    Advances in neural information processing systems30(2017)

    Wang, Y., Long, M., Wang, J., Gao, Z., Yu, P.S.: Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms. Advances in neural information processing systems30(2017)

    2017

  24. [24]

    Nature communications13(1), 1–10 (2022)

    Espeholt, L., Agrawal, S., Sønderby, C., Kumar, M., Heek, J., Bromberg, C., Gazen, C., Carver, R., Andrychowicz, M., Hickey, J.,et al.: Deep learning for twelve hour precipitation forecasts. Nature communications13(1), 1–10 (2022)

    2022

  25. [25]

    arXiv preprint arXiv:2306.06079 (2023)

    Andrychowicz, M., Espeholt, L., Li, D., Merchant, S., Merose, A., Zyda, F., Agrawal, S., Kalchbrenner, N.: Deep learning for day forecasts from sparse 17 observations. arXiv preprint arXiv:2306.06079 (2023)

    Pith/arXiv arXiv 2023

  26. [26]

    Quarterly Journal of the Royal Meteorological Society147(739), 3135–3153 (2021)

    Casellas, E., Bech, J., Veciana, R., Pineda, N., Mir´ o, J.R., Mor´ e, J., Rigo, T., Sairouni, A.: Nowcasting the precipitation phase combining weather radar data, surface observations, and nwp model forecasts. Quarterly Journal of the Royal Meteorological Society147(739), 3135–3153 (2021)

    2021

  27. [27]

    arXiv preprint arXiv:2003.12140 (2020)

    Sønderby, C.K., Espeholt, L., Heek, J., Dehghani, M., Oliver, A., Salimans, T., Agrawal, S., Hickey, J., Kalchbrenner, N.: Metnet: A neural weather model for precipitation forecasting. arXiv preprint arXiv:2003.12140 (2020)

    Pith/arXiv arXiv 2003

  28. [28]

    Nature619(7970), 526–532 (2023)

    Zhang, Y., Long, M., Chen, K., Xing, L., Jin, R., Jordan, M.I., Wang, J.: Skilful nowcasting of extreme precipitation with nowcastnet. Nature619(7970), 526–532 (2023)

    2023

  29. [29]

    Computa- tional intelligence and neuroscience2020(1), 8439719 (2020)

    Ibrahim, M., Alsheikh, A., Al-Hindawi, Q., Al-Dahidi, S., ElMoaqet, H.: Short- time wind speed forecast using artificial learning-based algorithms. Computa- tional intelligence and neuroscience2020(1), 8439719 (2020)

    2020

  30. [30]

    Atmospheric Research282, 106548 (2023)

    Chkeir, S., Anesiadou, A., Mascitelli, A., Biondi, R.: Nowcasting extreme rain and extreme wind speed with machine learning techniques applied to different input datasets. Atmospheric Research282, 106548 (2023)

    2023

  31. [31]

    https://www

    ECMWF: Atmospheric Model High Resolution 15-Day Forecast. https://www. ecmwf.int/en/forecasts/datasets/set-i Accessed 2025-06-30

    2025

  32. [32]

    In: EGU General Assembly Conference Abstracts

    Shi, C., Jiang, L., Zhang, T., Xu, B., Han, S.: Status and Plans of CMA Land Data Assimilation System (CLDAS) Project. In: EGU General Assembly Conference Abstracts. EGU General Assembly Conference Abstracts, p. 5671 (2014)

    2014

  33. [33]

    Sustainability12(10) (2020) https://doi.org/10.3390/su12104311

    Han, S., Liu, B., Shi, C., Liu, Y., Qiu, M., Sun, S.: Evaluation of cldas and gldas datasets for near-surface air temperature over major land areas of china. Sustainability12(10) (2020) https://doi.org/10.3390/su12104311

    work page doi:10.3390/su12104311 2020

  34. [34]

    Preprint at https://arxiv.org/abs/2409.07188 (2024)

    Zhong, X., Chen, L., Fan, X., Qian, W., Liu, J., Li, H.: FuXi-2.0: Advancing machine learning weather forecasting model for practical applications. Preprint at https://arxiv.org/abs/2409.07188 (2024)

    Pith/arXiv arXiv 2024

  35. [35]

    npj climate and atmospheric science6(1), 190 (2023)

    Chen, L., Zhong, X., Zhang, F., Cheng, Y., Xu, Y., Qi, Y., Li, H.: Fuxi: A cas- cade machine learning forecasting system for 15-day global weather forecast. npj climate and atmospheric science6(1), 190 (2023)

    2023

  36. [36]

    Quarterly journal of the royal meteorological society146(730), 1999–2049 (2020) 18

    Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Hor´ anyi, A., Mu˜ noz-Sabater, J., Nicolas, J., Peubey, C., Radu, R., Schepers, D.,et al.: The era5 global reanal- ysis. Quarterly journal of the royal meteorological society146(730), 1999–2049 (2020) 18

    1999

  37. [37]

    Sensors18(4), 1300 (2018)

    Suomi, I., Vihma, T.: Wind gust measurement techniques—from traditional anemometry to new possibilities. Sensors18(4), 1300 (2018)

    2018

  38. [38]

    Building and Environment267, 112323 (2025)

    Chen, X., Ishihara, T.: A study of gust wind speed using a novel unsteady reynolds-averaged navier-stokes model. Building and Environment267, 112323 (2025)

    2025

  39. [39]

    International Journal of Climatology28(6), 777–788 (2008)

    You, J., Hubbard, K.G., Goddard, S.: Comparison of methods for spatially esti- mating station temperatures in a quality control system. International Journal of Climatology28(6), 777–788 (2008)

    2008

  40. [40]

    Frontiers of Earth Science16(3), 798–808 (2022)

    Zhao, W., Zhong, Y., Li, Q., Li, M., Liu, J., Tang, L.: Comparison and correction of idw based wind speed interpolation methods in urbanized shenzhen. Frontiers of Earth Science16(3), 798–808 (2022)

    2022

  41. [41]

    Sustainability12(10), 4311 (2020)

    Han, S., Liu, B., Shi, C., Liu, Y., Qiu, M., Sun, S.: Evaluation of cldas and gldas datasets for near-surface air temperature over major land areas of china. Sustainability12(10), 4311 (2020)

    2020

  42. [42]

    Earth system science data13(9), 4349–4383 (2021)

    Mu˜ noz-Sabater, J., Dutra, E., Agust ´ ı-Panareda, A., Albergel, C., Arduini, G., Balsamo, G., Boussetta, S., Choulga, M., Harrigan, S., Hersbach, H.,et al.: Era5- land: A state-of-the-art global reanalysis dataset for land applications. Earth system science data13(9), 4349–4383 (2021)

    2021

  43. [43]

    Journal of Applied Meteorological Science 33(6), 641–654 (2022)

    Liping, H., Liantang, D., Ruichun, W., Zhaorong, Z., Yuan, J., Zhifang, X., Lijuan, Z., Jin, Z., Lili, W., Fei, Y.,et al.: Key technologies of cma-meso and application to operational forecast. Journal of Applied Meteorological Science 33(6), 641–654 (2022)

    2022

  44. [44]

    In: Proceedings of 1st International Conference on Image Processing, vol

    Charbonnier, P., Blanc-Feraud, L., Aubert, G., Barlaud, M.: Two deterministic half-quadratic regularization algorithms for computed imaging. In: Proceedings of 1st International Conference on Image Processing, vol. 2, pp. 168–172 (1994). IEEE

    1994

  45. [45]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Pang, J., Chen, K., Shi, J., Feng, H., Ouyang, W., Lin, D.: Libra r-cnn: Towards balanced learning for object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 821–830 (2019)

    2019

  46. [46]

    Advances in neural information processing systems30(2017)

    Shi, X., Gao, Z., Lausen, L., Wang, H., Yeung, D.-Y., Wong, W.-k., Woo, W.- c.: Deep learning for precipitation nowcasting: A benchmark and a new model. Advances in neural information processing systems30(2017)

    2017

  47. [47]

    Advances in neural information processing systems32(2019) 19

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high- performance deep learning library. Advances in neural information processing systems32(2019) 19

    2019

  48. [48]

    arXiv preprint arXiv:1412.6980 (2014)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    Pith/arXiv arXiv 2014

  49. [49]

    arXiv preprint arXiv:1711.05101 (2017) 20

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) 20

    Pith/arXiv arXiv 2017

  50. [2023]

    Nature Reviews Earth & Environment5(4), 250–252 (2024)

    2024