pith. sign in

arxiv: 2606.18436 · v2 · pith:TYAMSUW4new · submitted 2026-06-16 · 📊 stat.ML · cs.LG

Pointwise is Pointless? A Multimodal Ablation Study for Precipitation Nowcasting with Graph Neural Networks

Oph\'elia Miralles , M\'at\'e Mile , Christoffer Artturi , Thomas Nipen , Ivar Seierstad This is my paper

Pith reviewed 2026-06-26 22:17 UTC · model grok-4.3

classification 📊 stat.ML cs.LG
keywords precipitation nowcastinggraph neural networksmultimodal ablationpoint observationsradar forecastsCRPS lossNetatmo datasatellite channels
0
0 comments X

The pith

Each data source improves a different part of precipitation nowcasting rather than point observations proving uninformative

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

The paper runs an ablation on a multimodal graph neural network that forecasts rain rate every five minutes up to two hours ahead over the Nordic domain. It tests six configurations that add or remove radar history, MEPS numerical weather prediction, Netatmo surface stations, MSG satellite channels, stochastic noise, and CRPS ensemble losses. Each added source strengthens a distinct diagnostic: MEPS stabilizes extrapolation, stations sharpen local and onset predictions, satellite cuts some biases, and CRPS yields the steadiest radar-grid scores. The central finding is that local station skill and spatially coherent radar-field skill are separate targets, so sparse point data supplies useful local constraints once the loss and observation encoding are matched to that target.

Core claim

The ablation of radar-only, NWP-informed, station-informed, satellite-informed, noise-augmented, and CRPS-based configurations shows that each source improves a different part of the forecast problem. MEPS stabilises radar-only extrapolation, Netatmo observations improve local station and onset diagnostics, satellite predictors reduce some station-level biases but may activate rain too early when used deterministically, and CRPS-based configurations provide the most consistent radar-grid gains. The combined satellite and CRPS setup gives the best overall oracle/DAS score. These results indicate that local observational skill and spatially coherent radar-field skill are distinct targets.

What carries the argument

Multimodal graph neural network nowcasting system that ingests combinations of radar history, MEPS fields, Netatmo point observations, and MSG satellite channels, trained under deterministic or CRPS ensemble losses and evaluated on radar-grid, station, onset, oracle, displacement, and amplitude diagnostics.

If this is right

  • MEPS data stabilises radar-only extrapolation
  • Netatmo observations improve local station and onset diagnostics
  • Satellite predictors reduce some station-level biases but may trigger rain too early
  • CRPS-based training yields the most consistent radar-grid gains
  • The satellite-plus-CRPS combination produces the best overall oracle and DAS scores

Where Pith is reading between the lines

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

  • Models could be trained with explicit multi-objective losses that separately reward station-level accuracy and radar-field coherence instead of a single scalar loss
  • The way point observations are attached to graph nodes likely determines how much local skill they can transfer to the dense field, suggesting targeted experiments on different graph-construction rules
  • Repeating the ablation over other radar domains would test whether the separation between local and spatial targets is domain-specific or general
  • Operational nowcasting pipelines might route point data only to the loss terms or output heads that target the diagnostics where it adds value

Load-bearing premise

The ablation assumes that swapping input sources and loss functions produces isolated effects without interference from how the graph neural network encodes the spatial support of each observation type or from unstated hyperparameter differences across the six runs.

What would settle it

A controlled re-run of all six configurations in which hyperparameter search is performed separately and identically for each and the graph construction for point observations is held fixed, checking whether the reported source-specific gains survive.

Figures

Figures reproduced from arXiv: 2606.18436 by Christoffer Artturi, Ivar Seierstad, M\'at\'e Mile, Oph\'elia Miralles, Thomas Nipen.

Figure 1
Figure 1. Figure 1: Radar sampling height over the Nordic domain. The radar products include several quality￾control variables that describe missing coverage, beam blocking, clutter, convective classification, and physically suspicious precipitation rates ( [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Example of the Nordic radar quality-control bitmask on 2 October 2020. Several binary radar-quality fields, including no-data areas, beam blocking, clutter, con￾vective classification, extreme rates, and in￾valid high values, are compressed into a sin￾gle integer flag field. The encoded flags are used both to mask unreliable radar targets and to retain information on the trustwor￾thiness of each pixel. The… view at source ↗
Figure 3
Figure 3. Figure 3: Schematic of the multimodal graph neural network nowcaster. Radar, station observa [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Pixelwise and pointwise verification as a function of lead time. The first row is evaluated [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Example radar-field evolution for a representative precipitation case at T+5 min and [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Rain-onset diagnostics against Netatmo observations. Deterministic satellite-guided fore [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Spatial distribution of rain-onset bias at Netatmo stations. Negative values indicate [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Spatial distribution of rain-onset false-alarm rate at Netatmo stations. False alarms [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Spatial distribution of rain-onset miss rate at Netatmo stations. Misses occur when rain [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
read the original abstract

Sparse point observations are increasingly available for precipitation nowcasting, but it is unclear how much they improve dense radar-field forecasts. We partially address this question with a multimodal graph neural network nowcasting system over the Nordic radar domain. The model predicts rain rate every five minutes up to two hours ahead and is trained with different combinations of radar history, MEPS numerical weather prediction, Netatmo surface observations, MSG satellite channels, stochastic noise, and CRPS-based ensemble losses. The study is designed as an ablation of operationally relevant information sources and training objectives. We compare radar-only, NWP-informed, station-informed, satellite-informed, noise-augmented, and CRPS-based configurations using complementary diagnostics on the radar grid, at station locations, for rain onset, and through oracle, displacement, and amplitude scores. The results show that each source improves a different part of the forecast problem. MEPS stabilises radar-only extrapolation, Netatmo observations improve local station and onset diagnostics, and satellite predictors reduce some station-level biases but may activate rain too early when used deterministically. CRPS-based configurations provide the most consistent radar-grid gains, while the combined satellite and CRPS setup gives the best overall oracle/DAS score. These results do not support the conclusion that point observations are uninformative for nowcasting, but they show that local observational skill and spatially coherent radar-field skill are distinct targets. The practical implication is that sparse observations can provide useful local constraints, but their benefit for radar-like fields depends on the training loss, uncertainty representation, and how observation support is encoded in the model.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper presents a multimodal ablation study of a graph neural network for precipitation nowcasting over the Nordic radar domain. Models are trained on combinations of radar history, MEPS NWP fields, Netatmo point observations, MSG satellite channels, stochastic noise, and CRPS-based ensemble losses. Using complementary diagnostics on the radar grid, at station locations, for rain onset, and via oracle/displacement/amplitude scores, the work claims that each source improves distinct forecast aspects (MEPS stabilizes extrapolation, Netatmo aids local/onset scores, satellite reduces some biases but can activate rain early, CRPS yields consistent grid gains) and that the combined satellite+CRPS configuration performs best overall. It concludes that point observations provide useful local constraints even if local observational skill and spatially coherent radar-field skill remain distinct targets.

Significance. If the ablation isolates modality effects, the results would usefully demonstrate that sparse point observations can supply local constraints in nowcasting without automatically translating to radar-field improvements, while highlighting the role of loss function and uncertainty representation. The complementary diagnostic suite (grid, station, onset, oracle/DAS) is a positive feature that avoids over-reliance on any single metric.

major comments (1)
  1. [Study design (abstract and methods description)] The ablation design compares six configurations but does not demonstrate that GNN observation encoding (node/edge construction for point observations, support radius, feature normalization, message-passing depth) and hyperparameter tuning were held identical across radar-only, station-informed, and combined runs. Observed differences in station-level CRPS or onset scores could therefore arise from encoding or tuning variations rather than the added data sources themselves. This directly undermines attribution of improvements to specific modalities and is load-bearing for the central claim that each source improves a different part of the forecast problem.
minor comments (2)
  1. [Abstract] The abstract reports results only qualitatively (e.g., 'improves local station and onset diagnostics', 'best overall oracle/DAS score') without numerical values, confidence intervals, or effect sizes, which limits assessment of practical significance.
  2. A summary table listing the exact input combinations, loss functions, and any per-configuration hyperparameter or encoding choices would clarify the ablation controls.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the importance of controlled ablation design. We address the concern on study design below and will revise the manuscript accordingly to strengthen the attribution of results to specific modalities.

read point-by-point responses

  1. Referee: [Study design (abstract and methods description)] The ablation design compares six configurations but does not demonstrate that GNN observation encoding (node/edge construction for point observations, support radius, feature normalization, message-passing depth) and hyperparameter tuning were held identical across radar-only, station-informed, and combined runs. Observed differences in station-level CRPS or onset scores could therefore arise from encoding or tuning variations rather than the added data sources themselves. This directly undermines attribution of improvements to specific modalities and is load-bearing for the central claim that each source improves a different part of the forecast problem.

    Authors: We agree that the manuscript should explicitly demonstrate identical encoding and hyperparameter settings to support modality attribution. The methods describe a single GNN architecture (identical message-passing depth, support radius, feature normalization, and node/edge rules) applied across input combinations, with hyperparameters tuned once on the radar-only baseline and frozen thereafter; point observations are incorporated solely by adding nodes in the relevant runs without altering other components. However, this control was not stated with sufficient detail. We will revise the methods section to add an explicit paragraph confirming that all six configurations used the exact same encoding pipeline, support radius, normalization, and 3-layer message passing, with no per-configuration retuning. This change directly addresses the load-bearing concern and will be reflected in the revised abstract and results discussion. revision: yes

Circularity Check

0 steps flagged

Empirical ablation study with no derivation chain

full rationale

The paper is an empirical ablation study comparing GNN configurations with different input sources and losses. No equations, derivations, or predictions are presented that could reduce to inputs by construction. All claims rest on experimental comparisons across radar-grid, station, and onset diagnostics. No self-citations are load-bearing for any uniqueness theorem or ansatz. This matches the default expectation for non-circular empirical work.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claims rest on the modeling assumption that a single GNN architecture can fairly compare multimodal inputs and that the chosen diagnostics separate local versus field skill. No explicit free parameters, axioms, or invented entities are stated in the abstract.

pith-pipeline@v0.9.1-grok · 5846 in / 1183 out tokens · 34398 ms · 2026-06-26T22:17:11.241307+00:00 · methodology

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

300 extracted references · 114 canonical work pages · 1 internal anchor

  1. [1]

    Observation-Guided Interpolation Using Graph Neural Networks for High-Resolution Operational Nowcasting in Switzerland , journal =

    Miralles, Oph. Observation-Guided Interpolation Using Graph Neural Networks for High-Resolution Operational Nowcasting in Switzerland , journal =. 2026 , volume =

    2026

  2. [2]

    Latent diffusion models for generative precipitation nowcasting with accurate uncertainty quantification , url =

    Leinonen, Jussi and Hamann, Ulrich and Nerini, Daniele and Germann, Urs and Franch, Gabriele , month = apr, year =. Latent diffusion models for generative precipitation nowcasting with accurate uncertainty quantification , url =

  3. [3]

    and Tomasi, E

    Franch, G. and Tomasi, E. and Wanjari, R. and Poli, V. and Cardinali, C. and Alberoni, P. P. and Cristoforetti, M. , TITLE =. Geoscientific Model Development , VOLUME =. 2025 , NUMBER =

    2025

  4. [4]

    Advances in Neural Information Processing Systems , year =

    Shi, Xingjian and Gao, Zhihan and Lausen, Leonard and Wang, Hao and Yeung, Dit-Yan and Wong, Wai-kin and Woo, Wang-chun , title =. Advances in Neural Information Processing Systems , year =

  5. [5]

    Geoscientific Model Development , year =

    Ayzel, Georgy and Scheffer, Tobias and Heistermann, Maik , title =. Geoscientific Model Development , year =

  6. [6]

    2013 , note =

    Sánchez Pérez, Javier and Meinhardt-Llopis, Enric and Facciolo, Gabriele , journal =. 2013 , note =

    2013

  7. [7]

    , title =

    Keil, Christian and Craig, George C. , title =. Weather and Forecasting , year =

  8. [8]

    , title =

    Gneiting, Tilmann and Raftery, Adrian E. , title =. Journal of the American Statistical Association , year =

  9. [9]

    Ferro, Christopher A. T. , title =. Quarterly Journal of the Royal Meteorological Society , year =

  10. [10]

    Anemoi: A Framework for Machine Learning Weather Forecasting , year =

  11. [11]

    and Hering, A

    Kopp, J. and Hering, A. and Germann, U. and Martius, O. , TITLE =. Atmospheric Measurement Techniques , VOLUME =. 2024 , NUMBER =

    2024

  12. [12]

    doi: 10.1002/qj.4777

    Vahid Yousefnia, Kianusch and Bölle, Tobias and Zöbisch, Isabella and Gerz, Thomas , title =. Quarterly Journal of the Royal Meteorological Society , volume =. doi:https://doi.org/10.1002/qj.4777 , url =. https://rmets.onlinelibrary.wiley.com/doi/pdf/10.1002/qj.4777 , year =

    work page doi:10.1002/qj.4777

  13. [13]

    Flora and Patrick Skinner and Corey K

    Montgomery L. Flora and Patrick Skinner and Corey K. Potvin and Brian Matilla and Anthony Reinhart. Assessing the Impact of Biased Target Variables on Machine Learning Models of Severe Hail. Weather and Forecasting. 2025. doi:10.1175/WAF-D-24-0051.1

    work page doi:10.1175/waf-d-24-0051.1 2025

  14. [14]

    Adams-Selin and Conrad L

    Rebecca D. Adams-Selin and Conrad L. Ziegler. Forecasting Hail Using a One-Dimensional Hail Growth Model within WRF. Monthly Weather Review. 2016. doi:10.1175/MWR-D-16-0027.1

    work page doi:10.1175/mwr-d-16-0027.1 2016

  15. [15]

    2024 , month = nov, note =

    Ruoyi Cui and Iris Thurnherr and Patricio Velasquez and others , title =. 2024 , month = nov, note =

    2024

  16. [16]

    and Germann, Urs , title =

    Leinonen, Jussi and Hamann, Ulrich and Sideris, Ioannis V. and Germann, Urs , title =. Geophysical Research Letters , volume =. doi:https://doi.org/10.1029/2022GL101626 , url =. https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1029/2022GL101626 , note =

    work page doi:10.1029/2022gl101626

  17. [17]

    Seamless Lightning Nowcasting with Recurrent-Convolutional Deep Learning

    Jussi Leinonen and Ulrich Hamann and Urs Germann. Seamless Lightning Nowcasting with Recurrent-Convolutional Deep Learning. Artificial Intelligence for the Earth Systems. 2022. doi:10.1175/AIES-D-22-0043.1

    work page doi:10.1175/aies-d-22-0043.1 2022

  18. [18]

    and Leinonen, J

    Rombeek, N. and Leinonen, J. and Hamann, U. , TITLE =. Natural Hazards and Earth System Sciences , VOLUME =. 2024 , NUMBER =

    2024

  19. [19]

    Lightning Fatalities in Europe (2001–20)

    Thilo Kühne and Bogdan Antonescu and Pieter Groenemeijer and Tomáš Púčik. Lightning Fatalities in Europe (2001–20). Weather, Climate, and Society. 2025. doi:10.1175/WCAS-D-24-0038.1

    work page doi:10.1175/wcas-d-24-0038.1 2001

  20. [20]

    upfer, K. and M\

    Mohr, S. and Ehret, U. and Kunz, M. and Ludwig, P. and Caldas-Alvarez, A. and Daniell, J. E. and Ehmele, F. and Feldmann, H. and Franca, M. J. and Gattke, C. and Hundhausen, M. and Knippertz, P. and K\"upfer, K. and M\"uhr, B. and Pinto, J. G. and Quinting, J. and Sch\"afer, A. M. and Scheibel, M. and Seidel, F. and Wisotzky, C. , TITLE =. Natural Hazards...

    2023

  21. [21]

    Convolutional

    Xingjian Shi and Zhourong Chen and Hao Wang and Dit. Convolutional. CoRR , volume =. 2015 , url =. 1506.04214 , timestamp =

    Pith/arXiv arXiv 2015

  22. [22]

    Nature Communications , year =

    Lasse Espeholt and Shreya Agrawal and Casper Sønderby and Manoj Kumar and Jonathan Heek and Carla Bromberg and Cenk Gazen and Rob Carver and Marcin Andrychowicz and Jason Hickey and Aaron Bell and Nal Kalchbrenner , title =. Nature Communications , year =. doi:10.1038/s41467-022-32483-x , url =

    work page doi:10.1038/s41467-022-32483-x

  23. [23]

    Atmosphere , VOLUME =

    Franch, Gabriele and Nerini, Daniele and Pendesini, Marta and Coviello, Luca and Jurman, Giuseppe and Furlanello, Cesare , TITLE =. Atmosphere , VOLUME =. 2020 , NUMBER =

    2020

  24. [24]

    Ortland and Michael J

    Stephanie M. Ortland and Michael J. Pavolonis and John L. Cintineo. The Development and Initial Capabilities of ThunderCast, a Deep Learning Model for Thunderstorm Nowcasting in the United States. Artificial Intelligence for the Earth Systems. 2023. doi:10.1175/AIES-D-23-0044.1

    work page doi:10.1175/aies-d-23-0044.1 2023

  25. [25]

    2024 , organization =

    ICON Tutorial 2024: Working with the ICON Model , author =. 2024 , organization =

    2024

  26. [26]

    2022 , howpublished =

    Leinonen, Jussi , title =. 2022 , howpublished =

    2022

  27. [27]

    Science , volume =

    Remi Lam and Alvaro Sanchez-Gonzalez and Matthew Willson and Peter Wirnsberger and Meire Fortunato and Ferran Alet and Suman Ravuri and Timo Ewalds and Zach Eaton-Rosen and Weihua Hu and Alexander Merose and Stephan Hoyer and George Holland and Oriol Vinyals and Jacklynn Stott and Alexander Pritzel and Shakir Mohamed and Peter Battaglia , title =. Science...

    2023

  28. [28]

    Mesoscale Meteorology in Midlatitudes , volume =

    Markowski, Paul and Richardson, Yvette , address =. Mesoscale Meteorology in Midlatitudes , volume =

  29. [29]

    Radford and Imme Ebert-Uphoff and Jebb Q

    Jacob T. Radford and Imme Ebert-Uphoff and Jebb Q. Stewart and Kate D. Musgrave and Robert DeMaria and Natalie Tourville and Kyle Hilburn. Accelerating Community-Wide Evaluation of AI Models for Global Weather Prediction by Facilitating Access to Model Output. Bulletin of the American Meteorological Society. 2025. doi:10.1175/BAMS-D-24-0057.1

    work page doi:10.1175/bams-d-24-0057.1 2025

  30. [30]

    Open Data Documentation , year =

  31. [31]

    2024 , url =

    Orage dévastateur du 24 juillet 2023 à La Chaux-deFonds , institution =. 2024 , url =

    2023

  32. [32]

    2025 , url =

    Ackermann, Florian , title =. 2025 , url =

    2025

  33. [33]

    Geophysical Research Letters , volume =

    Feldmann, Monika and Beucler, Tom and Gomez, Milton and Martius, Olivia , title =. Geophysical Research Letters , volume =. doi:https://doi.org/10.1029/2024GL110960 , url =. https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1029/2024GL110960 , note =

    work page doi:10.1029/2024gl110960

  34. [34]

    2025 , eprint=

    Observation-guided Interpolation Using Graph Neural Networks for High-Resolution Nowcasting in Switzerland , author=. 2025 , eprint=

    2025

  35. [35]

    Hering, A. M. and Germann, U. and Boscacci, M. and S. Operational nowcasting of thunderstorms in the Alps during MAP D-PHASE , booktitle =

  36. [36]

    Nowcasting of Thunderstorm Hazards with Deep Learning: Performance Report of the First Convective Season in Operations , booktitle =

    Ulrich Hamann and Luca Nisi and Irina Mahlstein and Matteo Buzzi and Michele Cattaneo and N. Nowcasting of Thunderstorm Hazards with Deep Learning: Performance Report of the First Convective Season in Operations , booktitle =. 2025 , address =

    2025

  37. [37]

    2022 , month =

    Hagelereignis überschattet das Geschäftsjahr 2021 , howpublished =. 2022 , month =

    2021

  38. [38]

    Shi, Xingjian and Gao, Zhihan and Lausen, Leonard and Wang, Hao and Yeung, Dit-Yan and Wong, Wai-kin and WOO, Wang-chun , editor =. Deep. Advances in

  39. [39]

    Remote Sensing , VOLUME =

    Albu, Alexandra-Ioana and Czibula, Gabriela and Mihai, Andrei and Czibula, Istvan Gergely and Burcea, Sorin and Mezghani, Abdelkader , TITLE =. Remote Sensing , VOLUME =. 2022 , NUMBER =

    2022

  40. [40]

    Meteorological Applications , volume =

    Sivle, Anders Doksæter and Agersten, Solfrid and Schmid, Franziska and Simon, André , title =. Meteorological Applications , volume =. 2022 , keywords =. doi:https://doi.org/10.1002/met.2053 , url =

    work page doi:10.1002/met.2053 2022

  41. [41]

    Nipen and Ivar A

    Thomas N. Nipen and Ivar A. Seierstad and Christian Lussana and Jørn Kristiansen and Øystein Hov , journal =. CITIZEN OBSERVATIONS FOR WEATHER FORECASTS , urldate =

  42. [42]

    Verification of Solid Precipitation Forecasts from Numerical Weather Prediction Models in Norway

    Morten Køltzow and Barbara Casati and Thomas Haiden and Teresa Valkonen. Verification of Solid Precipitation Forecasts from Numerical Weather Prediction Models in Norway. Weather and Forecasting. 2020. doi:10.1175/WAF-D-20-0060.1

    work page doi:10.1175/waf-d-20-0060.1 2020

  43. [43]

    arXiv preprint arXiv:2511.23043 , year=

    High-Resolution Probabilistic Data-Driven Weather Modeling with a Stretched-Grid , author=. arXiv preprint arXiv:2511.23043 , year=

    arXiv

  44. [44]

    2026 , howpublished =

    2026

  45. [45]

    2019 , note =

    Weather and Forecasting , author =. 2019 , note =. doi:10.1175/WAF-D-19-0030.1 , abstract =

    work page doi:10.1175/waf-d-19-0030.1 2019

  46. [46]

    Learning Diffusion Priors from Observations by Expectation Maximization , url =

    Rozet, Fran. Learning Diffusion Priors from Observations by Expectation Maximization , url =. Advances in Neural Information Processing Systems , doi =

  47. [47]

    2017 , keywords =

    Weather and Forecasting , author =. 2017 , keywords =. doi:10.1175/WAF-D-16-0099.1 , abstract =

    work page doi:10.1175/waf-d-16-0099.1 2017

  48. [48]

    Nipen and Ivar A

    Thomas N. Nipen and Ivar A. Seierstad and Cristian Lussana and Jørn Kristiansen and Øystein Hov. Adopting Citizen Observations in Operational Weather Prediction. Bulletin of the American Meteorological Society. 2020. doi:10.1175/BAMS-D-18-0237.1

    work page doi:10.1175/bams-d-18-0237.1 2020

  49. [49]

    Convolutional

    Shi, Xingjian and Chen, Zhourong and Wang, Hao and Yeung, Dit-Yan and Wong, Wai-kin and WOO, Wang-chun , editor =. Convolutional. Advances in

  50. [50]

    Making and evaluating point forecasts.Journal of the American Statistical Association, 106(494):746–762, 2011

    Tilmann Gneiting , title =. Journal of the American Statistical Association , volume =. 2011 , publisher =. doi:10.1198/jasa.2011.r10138 , URL =

    work page doi:10.1198/jasa.2011.r10138 2011

  51. [51]

    Hering and Marc G

    Amanda S. Hering and Marc G. Genton , title =. Technometrics , volume =. 2011 , publisher =. doi:10.1198/TECH.2011.10136 , URL =

    work page doi:10.1198/tech.2011.10136 2011

  52. [52]

    Roberts and Humphrey W

    Nigel M. Roberts and Humphrey W. Lean. Scale-Selective Verification of Rainfall Accumulations from High-Resolution Forecasts of Convective Events. Monthly Weather Review. 2008. doi:10.1175/2007MWR2123.1

    work page doi:10.1175/2007mwr2123.1 2008

  53. [53]

    Sideris, I. V. and Foresti, L. and Nerini, D. and Germann, U. , title =. Quarterly Journal of the Royal Meteorological Society , volume =. doi:https://doi.org/10.1002/qj.3766 , url =. https://rmets.onlinelibrary.wiley.com/doi/pdf/10.1002/qj.3766 , year =

    work page doi:10.1002/qj.3766

  54. [54]

    P. K. Andersen and R. D. Gill , title =. The Annals of Statistics , number =. 1982 , doi =

    1982

  55. [55]

    Reinert, Daniel and Rieger, Daniel and Prill, Florian , year =

  56. [56]

    and Sideris, I

    Gabella, Marco and Panziera, L. and Sideris, I. and Boscacci, M. and Wolfensberger, D. and Clementi, L. and Germann, U. , editor =. Twelve years of operational real-time hourly precipitation estimation in the. Rainfall. 2019 , note =. doi:10.3929/ethz-b-000347536 , language =

    work page doi:10.3929/ethz-b-000347536 2019

  57. [57]

    Annals of Applied Statistics , publisher =

    Improved inference on risk measures for univariate extremes , author =. Annals of Applied Statistics , publisher =. doi:10.1214/21-aoas1555 , url =

    work page doi:10.1214/21-aoas1555

  58. [58]

    arXiv preprint arXiv:2205.01138 , year =

    Transformers in time-series analysis: A tutorial , author =. arXiv preprint arXiv:2205.01138 , year =. 2205.01138 , archiveprefix =

    arXiv

  59. [59]

    Nature , publisher =

    Liability for climate change , author =. Nature , publisher =

  60. [60]

    doi:10.5194/gmd-12-3085-2019 , url =

    Aznar-Siguan, Gabriela and Bresch, David N , year = 2019, journal =. doi:10.5194/gmd-12-3085-2019 , url =

    work page doi:10.5194/gmd-12-3085-2019 2019

  61. [61]

    Layer normalization , author =

  62. [62]

    Journal of the Royal Statistical Society: Series C (Applied Statistics) , volume = 69, number = 4, pages =

    Inference for extreme values under threshold-based stopping rules , author =. Journal of the Royal Statistical Society: Series C (Applied Statistics) , volume = 69, number = 4, pages =. doi:https://doi.org/10.1111/rssc.12420 , url =

    work page doi:10.1111/rssc.12420

  63. [63]

    Annual Review of Statistics and Its Application , author =

    Is there a cap on longevity?. Annual Review of Statistics and Its Application , author =. 2022 , note =. doi:10.1146/annurev-statistics-040120-025426 , number =

    work page doi:10.1146/annurev-statistics-040120-025426 2022

  64. [64]

    Tellus A: Dynamic Meteorology and Oceanography , volume=

    Local ensemble Kalman filtering in the presence of model bias , author=. Tellus A: Dynamic Meteorology and Oceanography , volume=. 2006 , publisher=

    2006

  65. [65]

    and Alexe, M

    Lang, S. and Alexe, M. and Chantry, M. and Dramsch, J. and Pinault, F. and Raoult, B. and Clare, M.C. and Lessig, C. and Maier-Gerber, M. and Magnusson, L. and Bouallègue, Z.B. , journal=. 2024 , archivePrefix=

    2024

  66. [66]

    2022 , journal=

    GraphCast: Learning skillful medium-range global weather forecasting , author=. 2022 , journal=

    2022

  67. [67]

    arXiv preprint arXiv:2309.17370 , year=

    Graph-based Neural Weather Prediction for Limited Area Modeling , author=. arXiv preprint arXiv:2309.17370 , year=

    arXiv

  68. [68]

    doi:10.48550/ARXIV.2003.12140 , shorttitle =

    2020 , author =. doi:10.48550/ARXIV.2003.12140 , shorttitle =

    work page doi:10.48550/arxiv.2003.12140 2020

  69. [69]

    Nature , author =

    Skilful nowcasting of extreme precipitation with. Nature , author =. 2023 , langid =. doi:10.1038/s41586-023-06184-4 , pages =

    work page doi:10.1038/s41586-023-06184-4 2023

  70. [70]

    Machine Learning for Precipitation Nowcasting from Radar Images , url =

    Agrawal, Shreya and Barrington, Luke and Bromberg, Carla and Burge, John and Gazen, Cenk and Hickey, Jason , urldate =. Machine Learning for Precipitation Nowcasting from Radar Images , url =. doi:10.48550/ARXIV.1912.12132 , journal =

    work page doi:10.48550/arxiv.1912.12132 1912

  71. [71]

    Nature , author =

    Skilful precipitation nowcasting using deep generative models of radar , volume =. Nature , author =. 2021 , langid =. doi:10.1038/s41586-021-03854-z , pages =

    work page doi:10.1038/s41586-021-03854-z 2021

  72. [72]

    Weather and Forecasting , author =

    The Integrated Nowcasting through Comprehensive Analysis (. Weather and Forecasting , author =. doi:10.1175/2010WAF2222451.1 , pages =

    work page doi:10.1175/2010waf2222451.1

  73. [73]

    Deep Learning for Day Forecasts from Sparse Observations , url =

    Andrychowicz, Marcin and Espeholt, Lasse and Li, Di and Merchant, Samier and Merose, Alexander and Zyda, Fred and Agrawal, Shreya and Kalchbrenner, Nal , year =. Deep Learning for Day Forecasts from Sparse Observations , url =

  74. [74]

    Theoretical and Applied Climatology , publisher =

    Downscaling precipitation extremes , author =. Theoretical and Applied Climatology , publisher =

  75. [75]

    A Conceptual Introduction to Hamiltonian Monte Carlo

    A conceptual introduction to. arXiv e-prints , pages =. doi:10.48550/arXiv.1701.02434 , archiveprefix =. 1701.02434 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1701.02434

  76. [76]

    Resource and Energy Economics , volume = 32, number = 3, pages =

    Climate change and hailstorm damage: Empirical evidence and implications for agriculture and insurance , author =. Resource and Energy Economics , volume = 32, number = 3, pages =. doi:https://doi.org/10.1016/j.reseneeco.2009.10.004 , issn =

    work page doi:10.1016/j.reseneeco.2009.10.004 2009

  77. [77]

    Weather, Climate, and Society , publisher =

    Evaluating hail damage using property insurance claims data , author =. Weather, Climate, and Society , publisher =

  78. [78]

    Journal of Climate , volume = 25, number = 12, pages =

    Downscaling extremes—An intercomparison of multiple statistical methods for present climate , author =. Journal of Climate , volume = 25, number = 12, pages =

  79. [79]

    Journal of Climate , volume = 26, number = 10, pages =

    Downscaling extremes: An intercomparison of multiple methods for future climate , author =. Journal of Climate , volume = 26, number = 10, pages =

  80. [80]

    Climate Research , author =

    Tornado hazards in the. Climate Research , author =. 2003 , pages =

    2003

Showing first 80 references.