Pith. sign in

REVIEW 3 major objections 4 minor 60 references

Bayesian Deep Learning for Convective Initiation Nowcasting Uncertainty Estimation

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Combining differently initialized networks with dropout yields the most skillful, best-calibrated probabilistic forecasts among the methods tested for 0-1 h convective initiation nowcasting.

desk verdict Useful head-to-head benchmark of five Bayesian UQ methods for CI nowcasting, with a plausible method ranking, but the headline absolute skill and calibration claims are conditional on a balanced test set that does not match operational CI frequency. read the letter →

arxiv 2507.16219 v1 pith:33TY7QGI submitted 2025-07-22 physics.ao-ph cs.AI

Da Fan , David John Gagne II , Steven J. Greybush , Eugene E. Clothiaux , John S. Schreck , Chaopeng Shen This is my paper
classification physics.ao-phcs.AI
keywords convectiveinitiationnowcastingBayesiandeeplearninguncertaintyquantificationMonteCarlodropoutensemblesGOES-16Brierskillscore
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks whether five recently proposed Bayesian deep learning methods can improve 0-1 h convective initiation (CI) nowcasting over a deterministic residual neural network (ResNet) baseline, using GOES-16 satellite infrared imagery. It claims that most of them can, and that one, the initial-weights ensemble plus Monte Carlo dropout, produces the most skillful and best-calibrated probabilistic forecasts, with Brier Skill Scores 8-20% higher than the deterministic ResNet across lead times. The results also show that a Bayesian ResNet with generic priors underperforms the deterministic baseline at longer lead times because of training convergence problems, while the MOPED variant, which initializes priors from pretrained deterministic weights, converges faster and regains skill. All Bayesian methods are reported to provide well-calibrated uncertainty that separates large-error from small-error cases, which is what forecasters need for decision support. The evaluation metrics are computed on a test set deliberately balanced by undersampling non-CI events to about 58% CI and 42% non-CI.

What carries the argument

The central object is the initial-weights ensemble plus Monte Carlo dropout, an ensemble of deterministic ResNets with different random initial weights that has dropout active during inference; MC dropout randomly deactivates neurons at prediction time and is treated as an approximate posterior sampling mechanism. The method works by generating multiple solutions in hypothesis space, so the ensemble mean approximates the predictive distribution and the variance quantifies uncertainty. Two supporting mechanisms carry the secondary claims: the hypothesis-space interpretation that contrasts point hypotheses (deterministic model), neighborhood sampling (MC dropout), and posterior distributions (Bayesian and MOPED networks), and the MOPED prior initialization, which sets the mean of each weight prior to the pretrained deterministic weight and its scale to $0.1$ times the absolute value of that weight, constraining the Bayesian search near the deterministic solution.

What would settle it

Recompute the Brier Skill Score, reliability, and discard-test curves of the initial-weights ensemble plus MC dropout on the full June 2021 Great Plains sample without undersampling, or on an operational data stream with the natural CI base rate; if the method no longer beats the deterministic ResNet or its probabilities are no longer well calibrated, the central claim is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that replacing a single deterministic ResNet with a 100-member ensemble of deterministic ResNets trained from different random initial weights and run with dropout activated during inference yields the most skillful and best-calibrated CI probability forecasts at lead times up to 1 hour. The method's Brier Skill Scores are 8-20% higher than the deterministic ResNet's, it produces the closest-to-perfect reliability curves among all methods, and its uncertainty estimates rank cases by error in discard and spread-skill tests. The paper further finds that a Bayesian ResNet using variational inference with standard Gaussian priors converges slowly and performs worse than the deterministic ResNet at 30- and 60-minute lead times, whereas initializing the priors from the pretrained deterministic weights with the MOPED method stabilizes training, accelerates convergence, and improves skill. In case studies, the initial-weights ensembles show better detection of CI in clear-sky regions but weaker spatial generalization in clear-sky and anvil-cloud regions without CI, where the deterministic ResNet and MOPED ensemble give lower false-alarm probabilities.

Load-bearing premise

The load-bearing premise is that a test set deliberately balanced by undersampling non-CI events, with about 58% CI and 42% non-CI samples, can support conclusions about forecast skill and calibration that are meant to transfer to operational conditions where convective initiation is far rarer.

Editorial extensions

If this is right

  • For 0-1 h convective initiation nowcasting, the recommended uncertainty quantification method among those tested is the initial-weights ensemble plus MC dropout, with expected Brier Skill Score gains of 8-20% over a deterministic ResNet.
  • A Bayesian ResNet should not be expected to beat a deterministic baseline without informed priors; the MOPED approach of initializing weight priors from pretrained deterministic weights is what restores and improves skill.
  • Calibrated uncertainty from any of the Bayesian methods lets a user discard the highest-uncertainty forecasts and reduce mean error, supporting use as a triage or decision-support tool.
  • The poorer spatial generalization of initial-weights ensembles in clear-sky and anvil regions means deployment should include monitoring of false-positive probabilities outside training-like conditions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: because the test set was balanced by undersampling non-CI events to 58% CI and 42% non-CI, the reported reliability and BSS values likely overstate absolute performance at the natural CI base rate; the relative ranking of methods is probably more robust than the absolute calibration.
  • Editorial extension: the same ensemble-plus-dropout recipe could be tested on other satellite-based nowcasting targets, such as precipitation onset or severe-weather indicators, where calibrated probabilities and uncertainty separation are equally valuable.
  • Editorial extension: a cheaper approximation, such as a 10-20 member ensemble or checkpoint ensembles from a single training run, would test whether the calibration gains require the full 100-member configuration.
  • Editorial extension: comparing the best method against a post-hoc calibration baseline, such as temperature scaling or isotonic regression applied to the deterministic probabilities, would clarify how much of the calibration improvement comes from the Bayesian machinery versus simple probability adjustment.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The manuscript compares five Bayesian deep learning uncertainty quantification methods—Bayesian ResNet, Bayesian-MOPED ResNet, deterministic ResNet + MC dropout, initial-weights ensemble, and initial-weights ensemble + MC dropout—against a deterministic ResNet baseline for 0–1 h convective initiation (CI) nowcasting using GOES-16 infrared brightness temperatures. Forecast skill is measured by Brier Skill Score (BSS), reliability/resolution decomposition, and hourly and spatial BSS; uncertainty quality is assessed via spread–skill diagrams and discard tests. The central empirical claim is that the initial-weights ensemble + MC dropout is the most skillful and well-calibrated method, with BSSs 8–20% higher than the deterministic ResNet, and that the Bayesian-MOPED approach improves convergence and skill relative to the plain Bayesian ResNet. The paper also presents two qualitative case studies on generalization to clear-sky and anvil-cloud scenarios.

Significance. If the results hold, the paper provides a practically useful, head-to-head comparison of modern Bayesian deep learning approaches for a high-stakes nowcasting task, with a clear recommendation (initial-weights ensemble + MC dropout) among the tested methods. The evaluation uses standard verification metrics (Brier score, reliability, spread–skill), includes bootstrap confidence intervals, and makes processed training/testing data publicly available—strengths that facilitate reproducibility. However, the headline claims of skill and calibration are computed on a deliberately balanced test set (58% CI), so the absolute skill and calibration statements, and any operational recommendation, are conditional on that artificial base rate. The generalization claims rest on two hand-picked case studies without quantitative scores. These issues do not invalidate the relative ranking among methods on the balanced protocol, but they do limit the generality of the absolute claims.

major comments (3)
  1. [Section 2b and Section 4a] The balanced test set changes the reference climatology used in the BSS. Section 2b states that non-CI events are undersampled to create a balanced dataset with 58% CI and 42% non-CI events, and the BSS in Eq. (5) is computed as 1 − BS_forecast / BS_climatology, where the climatology is the event frequency on this balanced test set. Consequently, the headline claim in Section 4a that the initial-weights ensemble + MC dropout achieves BSSs '8–20% higher than the deterministic ResNet' quantifies skill relative to a climatology that always predicts the 58% frequency, not the natural CI base rate. Under an operational low base rate, reliability curves, resolution, and even the relative ranking of methods could change, particularly in the low-probability regime that is now undersampled. The paper does not report metrics on the original imbalanced distribution or a base-rate sensitivity analysis. Please provide BSS, reliability, and discard-test results on an imbalanced test set, or at minimum report the natural base rate and discuss how the balanced evaluation affects each metric. Without this, the absolute operational claims of skill and calibration are not directly supported, though the relative comparison on the balanced protocol remains valid.
  2. [Section 4b] The generalization analysis is based on two hand-picked case studies (Figs. 10 and 11) with no quantitative skill scores, confidence intervals, or number of events. Statements such as 'the two initial-weights ensembles exhibit poor generalization' (Section 4b) and the corresponding abstract sentence are stronger than the evidence supports: two selected scenes cannot establish a generalizable pattern, especially since the text acknowledges that performance varies by region and hour. I recommend quantifying these scenarios with area-averaged Brier score or BSS over the scene, and ideally over a set of multiple CI events, with uncertainty estimates, so that the qualitative observations can be evaluated statistically. This is particularly important because the abstract's caveat about poorer generalization is one of the few statements that tempers the otherwise positive recommendation.
  3. [Section 3b and Section 4a] The MOPED prior is specified with sigma0 = 0.1|w_d| and pi = 1, and the Bayesian ResNet uses pi = 1/3, with these hyperparameters tuned only at the 10-min lead time and then applied to all lead times (Section 3b). The paper does not report any sensitivity of the results to these choices. Because the MOPED prior is centered on the deterministic baseline weights, the comparison between Bayesian-MOPED and the deterministic ResNet is not a fully independent test of the method; the improvement of Bayesian-MOPED over the plain Bayesian ResNet could depend on the particular sigma0/pi values chosen. Please add a sensitivity analysis (e.g., varying sigma0 by a factor of 2–5 and pi over a small grid) for at least the 10- and 60-min lead times, and report whether the main ranking—initial-weights ensemble + MC dropout best, Bayesian ResNet worst at long lead times—is robust to these choices.
minor comments (4)
  1. [Abstract and Section 6] The abstract says the Bayesian-MOPED ensemble 'enhanced forecast skill,' but the results in Section 4a show that it performs comparably to the deterministic ResNet + MC dropout and only slightly better than the deterministic ResNet. Please phrase this as 'improved the skill of the Bayesian ResNet ensemble' to avoid implying a large improvement over the deterministic baseline.
  2. [Section 3b] There is a typo in the sentence 'Theorange dashed arrows' (should be 'The orange dashed arrows'), and the same paragraph switches between 'we chose' and 'We chose' in reference to the Flipout method; please harmonize.
  3. [Section 5] The discussion notes that the validation BSS of the deterministic ResNet is higher than that of the Bayesian-MOPED ensemble, while the test BSS is higher for Bayesian-MOPED. This deserves an explicit interpretation—it could indicate overfitting of the deterministic model or a lucky test split, yet no comment is made.
  4. [Data availability] Only processed data are provided; no training/evaluation code is linked. Given the complexity of the five Bayesian methods and the 100-member ensembles, releasing code would substantially strengthen reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the skill and calibration claims are held-out empirical comparisons, not derivations; self-citations supply the baseline and dataset without determining the target result.

full rationale

The paper does not present a derivation chain; it evaluates five Bayesian deep learning variants against a deterministic ResNet on a held-out 2021 test set. The central claim (initial-weights ensemble + MC dropout is most skillful and well-calibrated) is supported by BSS, reliability, spread-skill, and discard-test computations on samples not used for training or hyperparameter selection. MOPED's prior is initialized from the deterministic ResNet weights, but that initialization is an experimental design choice whose outcome is then tested on independent data; the comparison does not reduce to the prior by construction. Citations to Fan et al. (2024) supply the previously published baseline architecture, preprocessing, and dataset, which are external inputs rather than conclusions assumed in the target claim. The only substantive caveat is that the test set was deliberately balanced to 58% CI / 42% non-CI, so absolute BSS and calibration numbers are conditional on that artificial base rate; this is an external-validity limitation, not a circularity, and it does not invalidate the relative method ranking on the evaluation protocol used.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central comparison depends on the validity of the CI labels, the representativeness of the balanced test set, and the standard approximations (variational inference, MC dropout) used by the methods. The hyperparameters pi, sigma0, T, and ensemble size are user-chosen and not derived from first principles.

free parameters (6)
  • pi (Bayesian ResNet prior coefficient) = 1/3
    Tuned on validation at 10-min lead time via a sensitivity experiment, then applied to all lead times (Section 3b).
  • pi (MOPED prior coefficient) = 1
    Chosen to balance prior and data terms in the ELBO; no sensitivity analysis shown (Section 3b).
  • MOPED prior scale sigma0 = 0.1|w_d|
    Set to 0.1 times the magnitude of pretrained deterministic weights; ad hoc choice (Section 3b).
  • Number of Monte Carlo samples T = 100
    Chosen to balance accuracy and cost, following Filos et al. (2019) and Ortiz et al. (2023); not derived.
  • Ensemble size for initial-weights ensembles = 100
    Number of deterministic ResNets trained; no sensitivity analysis provided (Section 3b).
  • Dropout rate for MC dropout = from Fan et al. (2024)
    Reuses the dropout rate from the deterministic ResNet of the prior paper; not re-tuned.
assumptions (5)
  • standard math MC dropout at inference approximates sampling from the Bayesian posterior (Gal and Ghahramani 2016)
    Invoked in Section 3b(1) to justify interpretation of dropout-based ensembles as Bayesian approximations.
  • domain assumption MRMS composite reflectivity >= 35 dBZ with the stated spatial and temporal conditions identifies convective initiation
    Section 2a defines CI labels from MRMS; the whole evaluation depends on these labels being a valid ground truth.
  • domain assumption GOES-16 ABI infrared brightness temperatures at about 2-km resolution contain sufficient information to predict CI 0-1 h ahead
    Section 2b and Fan et al. (2024); all predictors are these brightness temperatures.
  • domain assumption The balanced test set (58% CI) remains adequate for comparing methods and assessing calibration
    Section 2b undersamples non-CI events; calibration under natural base rates is not established.
  • ad hoc to paper The MOPED prior scale sigma0 = 0.1|w_d| is a reasonable empirical Bayes choice
    Set in Section 3b(1) without a principled derivation; the paper's conclusions about MOPED depend on this choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bayesian Deep Learning for Convective Initiation Nowcasting Uncertainty Estimation." pith.science (2026). https://pith.science/paper/33TY7QGI

@misc{pith2026250716219,
  author       = {Pith},
  title        = {Pith review of: Bayesian Deep Learning for Convective Initiation Nowcasting Uncertainty Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/33TY7QGI}},
  note         = {Machine review of arXiv:2507.16219}
}
read the original abstract

This study evaluated the probability and uncertainty forecasts of five recently proposed Bayesian deep learning methods relative to a deterministic residual neural network (ResNet) baseline for 0-1 h convective initiation (CI) nowcasting using GOES-16 satellite infrared observations. Uncertainty was assessed by how well probabilistic forecasts were calibrated and how well uncertainty separated forecasts with large and small errors. Most of the Bayesian deep learning methods produced probabilistic forecasts that outperformed the deterministic ResNet, with one, the initial-weights ensemble + Monte Carlo (MC) dropout, an ensemble of deterministic ResNets with different initial weights to start training and dropout activated during inference, producing the most skillful and well-calibrated forecasts. The initial-weights ensemble + MC dropout benefited from generating multiple solutions that more thoroughly sampled the hypothesis space. The Bayesian ResNet ensemble was the only one that performed worse than the deterministic ResNet at longer lead times, likely due to the challenge of optimizing a larger number of parameters. To address this issue, the Bayesian-MOPED (MOdel Priors with Empirical Bayes using Deep neural network) ResNet ensemble was adopted, and it enhanced forecast skill by constraining the hypothesis search near the deterministic ResNet hypothesis. All Bayesian methods demonstrated well-calibrated uncertainty and effectively separated cases with large and small errors. In case studies, the initial-weights ensemble + MC dropout demonstrated better forecast skill than the Bayesian-MOPED ensemble and the deterministic ResNet on selected CI events in clear-sky regions. However, the initial-weights ensemble + MC dropout exhibited poorer generalization in clear-sky and anvil cloud regions without CI occurrence compared to the deterministic ResNet and Bayesian-MOPED ensemble.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 43 canonical work pages

  1. [1]

    Amini,A.,W.Schwarting,A.Soleimany,andD.Rus,2020: Deepevidentialregression.Advances in Neural Information Processing Systems, Vol

    Abadi, M., and Coauthors, 2016: TensorFlow: A system for large-scale machine learning.Pro- ceedings of the 12th USENIX Symposium on Operating Systems Design and Implementation, OSDI 2016, 265–283, 1605.08695. Amini,A.,W.Schwarting,A.Soleimany,andD.Rus,2020: Deepevidentialregression.Advances in Neural Information Processing Systems, Vol. 33, 14927–14937, U...

    arXiv 2016

  2. [2]

    Controlled abstention neural networks for identifying skillful predictions for regression problems

    Barnes, E. A., and R. J. Barnes, 2021: Controlled Abstention Neural Networks for Identifying Skillful Predictions for Regression Problems.Journal of Advances in Modeling Earth Systems, 13 (12), e2021MS002575, https://doi.org/10.1029/2021MS002575, 2104.08236

    work page Pith review arXiv 2021

  3. [3]

    M., 2006:Pattern recognition and machine learning

    Bishop, C. M., 2006:Pattern recognition and machine learning. Springer

    work page 2006

  4. [4]

    Blake, B. T., D. B. Parsons, K. R. Haghi, and S. G. Castleberry, 2017: The structure, evolution, anddynamicsofanocturnalconvectivesystemsimulatedusingtheWRF-ARWmodel.Monthly Weather Review,145 (8), 3179–3201, https://doi.org/10.1175/MWR-D-16-0360.1

    work page doi:10.1175/mwr-d-16-0360.1 2017

  5. [5]

    Blei, D. M., A. Kucukelbir, and J. D. McAuliffe, 2017: Variational Inference: A Review for Statisticians.J. Am. Stat. Assoc.,112 (518), 859–877, https://doi.org/10.1080/01621459.2017. 1285773

    arXiv 2017

  6. [6]

    Cornebise, K

    Blundell, C., J. Cornebise, K. Kavukcuoglu, and D. Wierstra, 2015: Weight uncertainty in neural networks.Proceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37, JMLR.org, Lille, France, 1613–1622, ICML’15, URL https: //doi.org/10.5555/3045118.3045290. Bremnes,J.B.,2020: Ensemblepostprocessingusingquant...

    work page arXiv 2015

  7. [7]

    W., 1950: Verification of Forecasts Expressed in Terms of Probability.Mon

    Brier, G. W., 1950: Verification of Forecasts Expressed in Terms of Probability.Mon. Weather Rev.,78 (1), 1–3, https://doi.org/10.1175/1520-0493(1950)078<0001:VOFEIT> 32 2.0.CO;2, URL https://journals.ametsoc.org/view/journals/mwre/78/1/1520-0493_1950_078_ 0001_vofeit_2_0_co_2.xml

    doi:10.1175/1520-0493(1950)078 1950

  8. [8]

    Dotzek, 2008:The Spatial Distribution of Severe Convective Storms and an Analysis of their Secular Changes, 35–53

    Brooks, H., and N. Dotzek, 2008:The Spatial Distribution of Severe Convective Storms and an Analysis of their Secular Changes, 35–53. Climate Extremes and Society, https://doi.org/ 10.1017/CBO9780511535840.006

    work page doi:10.1017/cbo9780511535840.006 2008

Show all 60 references
  1. [9]

    Chapman, W. E., L. D. Monache, S. Alessandrini, A. C. Subramanian, F. Martin Ralph, S. P. Xie, S. Lerch, and N. Hayatbini, 2022: Probabilistic Predictions from Deterministic Atmospheric RiverForecastswithDeepLearning.MonthlyWeatherReview,150 (1),215–234,https://doi.org/ 10.117...

    2022 doi

  2. [10]

    GitHub, https://keras.io

    Chollet, F., and Coauthors, 2015: Keras. GitHub, https://keras.io

    2015

  3. [11]

    Cintineo, J. L., M. J. Pavolonis, J. M. Sieglaff, L. Cronce, and J. Brunner, 2020: Noaa prob- severe v2.0—probhail, probwind, and probtor.Weather and Forecasting,35 (4), 1523–1543, https://doi.org/10.1175/WAF-D-19-0242.1. Colbert,M.,D.J.Stensrud,P.M.Markowski,andY.P.Richardson...

    2020 doi

  4. [12]

    Delaunay, A., and H. M. Christensen, 2022: Interpretable Deep Learning for Probabilistic MJO Prediction.Geophysical Research Letters,49 (16), e2022GL098566, https://doi.org/10.1029/ 2022GL098566

    2022

  5. [13]

    V., and Coauthors, 2017: TensorFlow Distributions.arXiv, URL https://doi.org/10

    Dillon, J. V., and Coauthors, 2017: TensorFlow Distributions.arXiv, URL https://doi.org/10. 48550/arXiv.1711.10604

    2017 doi arXiv

  6. [14]

    Dixon, P. G., A. E. Mercer, J. Choi, and J. S. Allen, 2011: Tornado risk analysis: Is dixie alley an extension of Tornado alley.Bulletin of the American Meteorological Society,92 (4), 433–441, https://doi.org/10.1175/2010BAMS3102.1

    2011 doi

  7. [15]

    Fan, D., S. J. Greybush, D. J. Gagne, and E. E. Clothiaux, 2024: Physically Explainable Deep LearningforConvectiveInitiationNowcastingUsingGOES-16SatelliteObservations.Artificial Intelligence for the Earth Systems,3, e230098, https://doi.org/10.1175/AIES-D-23-0098.1. 33

    2024 doi

  8. [16]

    Kifer, K

    Fang, K., D. Kifer, K. Lawson, and C. Shen, 2020: Evaluating the potential and challenges of an uncertainty quantification method for long short-term memory models for soil moisture predictions.Water Resources Research,56 (12), e2020WR028095, https://doi.org/10.1029/ 2020WR028095

    2020

  9. [17]

    Filos, A., and Coauthors, 2019: A systematic comparison of bayesian deep learning robustness in diabetic retinopathy tasks.4th Workshop on Bayesian Deep Learning (NeurIPS 2019), Vancou- ver, Canada, URL https://bayesiandeeplearning.org/2019/papers/12.pdf. Foster,D.,D.J.Gagne,a...

    2019 doi

  10. [18]

    Gal, Y., and Z. Ghahramani, 2016: Dropout as a bayesian approximation: representing model uncertaintyindeeplearning.Proceedingsofthe33rdInternationalConferenceonInternational Conference on Machine Learning - Volume 48, JMLR.org, 1050–1059, ICML’16, URL https: //doi.org/10.5555...

    2016

  11. [19]

    Gelman, A., J. B. Carlin, H. S. Stern, and D. B. Rubin, 2003:Bayesian Data Analysis, Second Edition(ChapmanandHall/CRCTextsinStatisticalScience).ChapmanandHall/CRC,696pp

    2003

  12. [20]

    Zhang, D

    Ghazvinian, M., Y. Zhang, D. J. Seo, M. He, and N. Fernando, 2021: A novel hybrid artificial neural network - Parametric scheme for postprocessing medium-range precipitation forecasts. Advances in Water Resources,151, 103907, https://doi.org/10.1016/j.advwatres.2021.103907

    2021

  13. [21]

    M., and E

    Gordon, E. M., and E. A. Barnes, 2022: Incorporating Uncertainty Into a Regression Neural Net- workEnablesIdentificationofDecadalState-DependentPredictabilityinCESM2.Geophysical Research Letters,49 (15), e2022GL098635, https://doi.org/10.1029/2022GL098635. Graves,A.,2011: Prac...

    2022 doi

  14. [22]

    P., and L

    Guillaumin, A. P., and L. Zanna, 2021: Stochastic-Deep Learning Parameterization of Ocean MomentumForcing.JournalofAdvancesinModelingEarthSystems,13 (9),e2021MS002534, https://doi.org/10.1029/2021MS002534

    2021 doi

  15. [23]

    Lagerquist, M

    Haynes, K., R. Lagerquist, M. McGraw, K. Musgrave, and I. Ebert-Uphoff, 2023: Creat- ing and Evaluating Uncertainty Estimates with Neural Networks for Environmental-Science Applications.Artificial Intelligence for the Earth Systems,2 (2), 1–29, https://doi.org/ 10.1175/aies-d-...

    2023 doi

  16. [24]

    Wang, and S

    Johnson, A., X. Wang, and S. Degelia, 2017: Design and implementation of a GSI-based convection-allowingensemble-baseddataassimilationandforecastsystemforthePECANfield experiment. Part II: Overview and evaluation of a real-time system.Weather and Forecasting, 32 (3), 1227–1251...

    2017 doi

  17. [25]

    Karpatne, A., and Coauthors, 2017: Theory-guided data science: A new paradigm for scientific discovery from data.IEEE Transactions on Knowledge and Data Engineering,29 (10), 2318– 2331, https://doi.org/10.1109/TKDE.2017.2720168, 1612.08544

    2017

  18. [26]

    Gal, 2017: What uncertainties do we need in bayesian deep learning for computer vision?Advances in Neural Information Processing Systems, I

    Kendall, A., and Y. Gal, 2017: What uncertainties do we need in bayesian deep learning for computer vision?Advances in Neural Information Processing Systems, I. Guyon, U. V

    2017

  19. [27]

    Bengio, H

    Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, Eds., Cur- ran Associates, Inc., Vol. 30, URL https://proceedings.neurips.cc/paper_files/paper/2017/file/ 2650d6089a6d640c5e85b2b88265dc2b-Paper.pdf

    2017

  20. [28]

    Kingma, D. P., T. Salimans, and M. Welling, 2015: Variational dropout and the local reparameterization trick.Advances in Neural Information Processing Systems, C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, Eds., Curran

    2015

  21. [29]

    28, URL https://proceedings.neurips.cc/paper_files/paper/2015/file/ bc7316929fe1545bf0b98d114ee3ecb8-Paper.pdf

    Associates, Inc., Vol. 28, URL https://proceedings.neurips.cc/paper_files/paper/2015/file/ bc7316929fe1545bf0b98d114ee3ecb8-Paper.pdf

    2015

  22. [30]

    Subedar, and O

    Krishnan, R., M. Subedar, and O. Tickoo, 2020: Specifying weight priors in bayesian deep neural networks with empirical bayes.AAAI 2020 - 34th AAAI Conference on Artificial Intelligence, 4477–4484, https://doi.org/10.1609/aaai.v34i04.5875. 35

    2020 doi

  23. [31]

    Herzog, and D

    Lakshmanan, V., B. Herzog, and D. Kingfield, 2015: A method for extracting postevent storm tracks.Journal of Applied Meteorology and Climatology,54 (2), 451–462, https://doi.org/ 10.1175/JAMC-D-14-0132.1

    2015 doi

  24. [32]

    Smith, 2010: An objective method of evaluating and devising storm- tracking algorithms.Weather and Forecasting,25 (2), 701–709, https://doi.org/10.1175/ 2009WAF2222330.1

    Lakshmanan, V., and T. Smith, 2010: An objective method of evaluating and devising storm- tracking algorithms.Weather and Forecasting,25 (2), 701–709, https://doi.org/10.1175/ 2009WAF2222330.1

    2010

  25. [33]

    Smith, K

    Lakshmanan, V., T. Smith, K. Hondl, G. J. Stumpf, and A. Witt, 2006: A real-time, three- dimensional, rapidly updating, heterogeneous radar merger technique for reflectivity, ve- locity, and derived products.Weather and Forecasting,21 (5), 802–823, https://doi.org/ 10.1175/WAF942.1

    2006 doi

  26. [34]

    Smith, G

    Lakshmanan, V., T. Smith, G. Stumpf, and K. Hondl, 2007: The warning decision support system-integrated information.Weather and Forecasting,22 (3), 596–612, https://doi.org/10. 1175/WAF1009.1

    2007

  27. [35]

    Pritzel, and C

    Lakshminarayanan, B., A. Pritzel, and C. Blundell, 2017: Simple and scalable predictive uncer- tainty estimation using deep ensembles.Advances in Neural Information Processing Systems, I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, ...

    2017

  28. [36]

    Lawson, J. R., J. S. Kain, N. Yussouf, D. C. Dowell, D. M. Wheatley, K. H. Knopfmeier, and T. A. Jones, 2018: Advancing from convection-allowing nwp to warn-on-forecast: Evidence of progress.Weather and Forecasting,33 (2), 599–607, https://doi.org/10.1175/waf-d-17-0145.1

    2018 doi

  29. [37]

    Lee, S., H. Han, J. Im, E. Jang, and M. I. Lee, 2017: Detection of deterministic and proba- bilistic convection initiation using Himawari-8 Advanced Himawari Imager data.Atmospheric Measurement Techniques,10 (5), 1859–1864, https://doi.org/10.5194/amt-10-1859-2017

    2017 doi

  30. [38]

    McGovern, A., K. L. Elmore, D. J. Gagne, S. E. Haupt, C. D. Karstens, R. Lagerquist, T. Smith, and J. K. Williams, 2017: Using artificial intelligence to improve real-time decision-making for high-impact weather.Bulletin of the American Meteorological Society,98 (10), 2073–209...

    2017 doi

  31. [39]

    Mecikalski, J. R., J. K. Williams, C. P. Jewett, D. Ahijevych, A. LeRoy, and J. R. Walker, 2015: Probabilistic 0-1-h convective initiation nowcasts that combine geostationary satellite observationsandnumericalweatherpredictionmodeldata.JournalofAppliedMeteorologyand Climatolog...

    2015 doi

  32. [40]

    Nadav-Greenberg, L., and S. L. Joslyn, 2009: Uncertainty Forecasts Improve Decision Making Among Nonexperts.Journal of Cognitive Engineering and Decision Making,3 (3), 209–227, https://doi.org/10.1518/155534309X474460

    2009 doi

  33. [41]

    Neal, R.M., 2012:Bayesianlearningforneuralnetworks, Vol.118.SpringerScience&Business Media

    2012

  34. [42]

    A., and A

    Nix, D. A., and A. S. Weigend, 1994: Estimating the mean and variance of the target probability distribution.IEEE International Conference on Neural Networks - Conference Proceedings,1, 55–60, https://doi.org/10.1109/icnn.1994.374138

    1994

  35. [43]

    Casas, M

    Ortiz, P., E. Casas, M. Orescanin, S. W. Powell, V. Petkovic, and M. Hall, 2023: Uncer- tainty Calibration of Passive Microwave Brightness Temperatures Predicted by Bayesian Deep Learning Models.Artificial Intelligence for the Earth Systems,2, e220056, https://doi.org/ 10.1175...

    2023 doi

  36. [44]

    Orescanin, V

    Ortiz, P., M. Orescanin, V. Petkovic, S. W. Powell, and B. Marsh, 2022: Decomposing Satellite- BasedClassificationUncertaintiesinLargeEarthScienceDatasets.IEEETransactionsonGeo- science and Remote Sensing,60, https://doi.org/10.1109/TGRS.2022.3152516

    2022

  37. [45]

    Wallach, H

    Ovadia, Y., and Coauthors, 2019: Can you trust your model's uncertainty? evaluating pre- dictive uncertainty under dataset shift.Advances in Neural Information Processing Systems, H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, Eds., Cur- ran...

    2019

  38. [46]

    Raftery, A. E., T. Gneiting, F. Balabdaoui, and M. Polakowski, 2005: Using Bayesian model averaging to calibrate forecast ensembles.Monthly Weather Review,133 (5), 1155–1174, https://doi.org/10.1175/MWR2906.1

    2005 doi

  39. [47]

    Gerrish, and D

    Ranganath, R., S. Gerrish, and D. Blei, 2014: Black Box Variational Inference.Proceedings of the Seventeenth International Conference on Artificial Intelligence and Statistics, S. Kaski, and J. Corander, Eds., PMLR, Reykjavik, Iceland, Proceedings of Machine Learning Research,...

    2014

  40. [48]

    Lerch, 2018: Neural networks for postprocessing ensemble weather forecasts

    Rasp, S., and S. Lerch, 2018: Neural networks for postprocessing ensemble weather forecasts. Monthly Weather Review,146 (11), 3885–3900, https://doi.org/10.1175/MWR-D-18-0187.1, 1805.09091

    2018 arXiv

  41. [49]

    Scheuerer, M., M. B. Switanek, R. P. Worsnop, and T. M. Hamill, 2020: Using artificial neural networksforgeneratingprobabilisticsubseasonalprecipitationforecastsovercalifornia.Monthly Weather Review,148 (8), 3489–3506, https://doi.org/10.1175/MWR-D-20-0096.1

    2020 doi

  42. [50]

    Schreck, J. S., and Coauthors, 2024: Evidential deep learning: Enhancing predictive uncertainty estimation for earth system science applications.Artificial Intelligence for the Earth Systems, 3 (4), 230093, https://doi.org/10.1175/AIES-D-23-0093.1, URL https://journals.ametsoc...

    2024 doi

  43. [51]

    Schulz, B., and S. Lerch, 2022: Machine Learning Methods for Postprocessing Ensemble Fore- casts of Wind Gusts: A Systematic Comparison.Monthly Weather Review,150 (1), 235–237, https://doi.org/10.1175/MWR-D-21-0150.1, 2106.09512. Sensoy,M.,L.Kaplan,andM.Kandemir,2018: Evidenti...

    2022 arXiv

  44. [52]

    Sha, Y., R. A. Sobash, and D. J. Gagne, 2024: Generative ensemble deep learning severe weather prediction from a deterministic convection-allowing model.Artificial Intelligence for the Earth Systems,3, e230094, https://doi.org/10.1175/aies-d-23-0094.1

    2024 doi

  45. [53]

    Sun, F., B. Li, M. Min, and D. Qin, 2023: Toward a Deep-Learning-Network-Based Convective WeatherInitiationAlgorithmfromtheJointObservationsofFengyun-4AGeostationarySatellite 38 andRadarfor0-1hNowcasting.IEEEJournalofSelectedTopicsinAppliedEarthObservations and Remote Sensing,...

    2023

  46. [54]

    Hardmeier, and J

    Ulmer, D., C. Hardmeier, and J. Frellsen, 2023: Prior and Posterior Networks: A Survey on Evi- dentialDeepLearningMethodsForUncertaintyEstimation.TransactionsonMachineLearning

    2023

  47. [55]

    Research, URL https://openreview.net/forum?id=xqS8k9E75c, 2110.03051

    arXiv

  48. [56]

    Vicol, J

    Wen, Y., P. Vicol, J. Ba, D. Tran, and R. Grosse, 2018: Flipout: Efficient pseudo-independent weight perturbations on mini-batches.International Conference on Learning Representations, URL https://openreview.net/forum?id=rJNpifWAb. Wilks,D.S.,2019:StatisticalMethodsintheAtmosp...

    2018 doi

  49. [57]

    G., and P

    Wilson, A. G., and P. Izmailov, 2020: Bayesian deep learning and a probabilis- tic perspective of generalization.Advances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, Eds., Curran Asso- ciates, Inc., Vol. 33, 4697–4708...

    2020

  50. [58]

    Yu, Y., X. Han, M. Yang, and J. Yang, 2020: Probabilistic Prediction of Regional Wind Power Based on Spatiotemporal Quantile Regression.IEEE Transactions on Industry Applications, 56 (6), 6117–6127, https://doi.org/10.1109/TIA.2020.2992945

    2020

  51. [59]

    Zou, and S

    Zhang, X., Y. Zou, and S. Li, 2022: Bayesian neural network with efficient priors for online quality prediction.Digital Chemical Engineering,2, 100008, https://doi.org/10.1016/j.dche. 2021.100008

    2022

  52. [60]

    Zhang, Y., D. J. Stensrud, and F. Zhang, 2019: Simultaneous assimilation of radar and all-sky satelliteinfraredradianceobservationsforconvection-allowingensembleanalysisandprediction of severe thunderstorms.Monthly Weather Review,147 (12), 4389–4409, https://doi.org/10. 1175/M...

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.