Pith. sign in

REVIEW 3 major objections 4 minor 2 references

Improving Post-Processing for Quantitative Precipitation Forecasting Using Deep Learning: Learning Precipitation Physics from High-Resolution Observations

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a generative post-processor trained on radar precipitation beats conventional NWP and GraphCast for heavy and intense rainfall over the Korean Peninsula.

desk verdict A credible and potentially useful Patch-cGAN QPF post-processor over Korea, but the verification target is the same radar product used as training labels, so the reported heavy-rain gains need an independent check before I'd trust the margin. read the letter →

arxiv 2506.03842 v1 pith:7OLO2FGH submitted 2025-06-04 physics.ao-ph

ChangJae Lee , Heecheol Yang , Byeonggwon Kim This is my paper · ORCID
classification physics.ao-ph
keywords quantitativeprecipitationforecastingdeeplearningpost-processinggenerativeadversarialnetworksradar-estimatedperfectprognosticmethodheavyrainfallfrequencybiasGraphCast
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

The paper aims to establish that a deep-learning post-processor, DL-QPF, can turn operational NWP meteorological fields into precipitation forecasts that are closer to observed rainfall than the host model's own precipitation output, and also closer than regional convection-permitting models and the AI model GraphCast. It trains a Patch-cGAN with a U-Net generator on three warm seasons of 500-meter radar-estimated precipitation, using IFS-HRES forecast variables as inputs under the Perfect Prognostic Method. Verification over the summer of 2024 at thresholds from 1 to 50 mm/6h shows DL-QPF keeping frequency bias near one and achieving the highest success ratios, with the strongest advantage at heavy and intense rainfall. The paper presents this as evidence that direct learning from high-resolution observations is a viable route to de-biased, realistic QPF.

What carries the argument

The central mechanism is a Patch-cGAN, a conditional generative adversarial network whose discriminator judges local patches instead of whole fields, paired with a U-Net generator and a CNN-based discriminator. The generator maps 80 meteorological input variables to 2-km precipitation fields, the discriminator distinguishes observed radar precipitation from generated precipitation over 270 km by 270 km mesoscale patches, and the total loss combines L1 and L2 regression losses with a GAN cross-entropy loss. This adversarial setup is what pushes the output histogram toward the observed rainfall distribution, and the CLEANER radar-estimated precipitation product supplies the high-resolution labels that make the learned relationship observational rather than reanalysis-driven.

What would settle it

Verify DL-QPF on the same summer 2024 cases using independent rain-gauge observations, or a different high-resolution quantitative precipitation estimate, as the truth instead of the radar product used for training; if its frequency bias and CSI at 25 and 50 mm/6h no longer beat KIM-Regional and GraphCast, the reported superiority is partly an artifact of training and verifying on the same benchmark.

Watch

Extended reading notes

Core claim

The central claim is that a conditional generative adversarial network can learn the mapping from forecast meteorological variables to observed precipitation well enough to outperform both conventional NWP models and a state-of-the-art AI weather model for the most impactful rainfall categories. DL-QPF maintains a frequency bias near the ideal value of 1 across light, moderate, heavy, and intense thresholds, and on average achieves a higher success ratio than the baselines; at the intense threshold of 50 mm/6h it performs best among all models, despite a slight under-forecasting tendency on day 1 forecasts. The paper attributes this extreme-rain skill to the adversarial loss recovering the true precipitation distribution, counteracting the smoothing and underestimation that L1/L2-based training induces in models such as GraphCast. Case studies show DL-QPF reproducing observed rainfall patterns better than IFS-HRES for a synoptic low, a stationary front with a meso-low, and a tropical depression, while the paper also reports a clear weakness in underestimating small-scale convective showers.

Load-bearing premise

The same radar rainfall product is used both to teach the model and to grade it, so if that product has systematic errors, the model may be learning and reproducing those errors rather than forecasting true rainfall.

Editorial extensions

If this is right

  • If DL-QPF is correct, operational QPF post-processing can remove most precipitation frequency bias across all intensity categories without changing the NWP dynamical core.
  • Heavy and intense rainfall skill, measured by CSI and success ratio, can be improved beyond what regional convection-permitting models and GraphCast achieve using only global-model input fields.
  • An adversarial loss component is sufficient to recover the extreme precipitation tail that pure L1/L2-trained AI weather models systematically under-forecast.
  • The same IFS-HRES inputs can yield 2-km rainfall fields that are structurally closer to observed patterns than the host model's own precipitation output.
  • The model's documented weakness on convective showers implies that resolved meteorological relationships alone are insufficient for convective initiation, so adding NWP precipitation guidance is a natural next step that the paper itself suggests.

Reading between the lines

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

  • Because the paper trains under the Perfect Prognostic Method, pairing each observation with its coincident forecast cycle, a natural extension is to retrain on actual forecast lead-time errors; perfect-prog training likely overstates skill at the longer lead times used in verification.
  • Adding the NWP precipitation field as an auxiliary input could fix convective-shower underprediction but would risk re-introducing exactly the parameterization bias the model is designed to remove, making the trade-off a testable design question.
  • The same architecture could be retrained on gauge-merged or differently quality-controlled precipitation and evaluated outside the Korean warm season to test whether the radar-label advantage generalizes beyond this dataset.
  • An ablation that turns off the discriminator would isolate how much of the extreme-rain improvement comes from the adversarial loss versus the U-Net's L1/L2 regression, and would make the causal role of the GAN component explicit.
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 proposes DL-QPF, a Patch-cGAN post-processing model that maps IFS-HRES meteorological fields to 2-km precipitation fields over the Korean Peninsula. Training labels are 3-hour accumulated CLEANER radar-estimated precipitation for May-October 2021-2023. Verification for summer 2024 uses performance diagrams at 1, 10, 25, and 50 mm/6h thresholds against IFS-HRES, KIM, KIM-Regional, KIM-LENS, and GraphCast, plus four qualitative case studies. The paper's central claim is that DL-QPF achieves near-unity frequency bias and the highest success ratio across thresholds, and outperforms all baselines for heavy and intense rain.

Significance. If the central claim holds, the paper is a useful contribution: it shows that a relatively simple GAN post-processor trained on high-resolution radar labels can de-bias NWP precipitation and outperform strong global, regional, and AI baselines in a real-world warm-season setting. The paper is commendable for releasing code and data, for using a temporally disjoint verification period (2024 vs 2021-2023 training), and for explicitly discussing failure modes such as convective showers. The main caveat is that the verification target is the same radar product used for training labels, and no independent gauge-based verification is provided at the thresholds where superiority is claimed; additionally, the reported metrics lack uncertainty estimates. These issues are fixable and do not invalidate the approach, but they currently weaken the headline claim.

major comments (3)
  1. [Section 2.2 and Section 3; Eq. (1)] The training labels (CLEANER radar-estimated precipitation, Section 2.2) and the verification ground truth (Section 3) are the same product. Because the adversarial loss in Eq. (1) explicitly trains the generator to reproduce the label distribution, the near-unity frequency bias reported in Sections 3.1-3.5 is partly a convergence diagnostic rather than an independent measure of forecast quality. The quantitative claim that DL-QPF outperforms NWP at 25 and 50 mm/6h thresholds therefore needs support from verification against rain gauges or a different observational product, ideally stratified by threshold; the all-intensity scatter in Figure 2 is only qualitative and does not establish performance at extreme thresholds.
  2. [Sections 3.1-3.4, Figures 5-8] No confidence intervals, bootstrap resamples, or significance tests are reported for POD, SR, CSI, or frequency bias. The verification set is a single summer (JJA 2024), and intense rain (>=50 mm/6h) is rare, so the differences in Figure 8 may be within sampling uncertainty. The authors should either add uncertainty estimates or temper the wording 'best performance among all models' until the differences are shown to be statistically distinguishable.
  3. [Section 2.1 and Section 3] The model is trained on IFS-HRES lead times of 6-15 h (Section 2.1), but verification is performed for lead times of 6-72 h (Section 3). Under the perfect-prognosis assumption the diagnostic mapping can in principle be applied at any lead time, but forecast fields at 72 h have larger errors and different spatial characteristics than those at 6-15 h, so the inputs are out-of-distribution during verification. The paper should either train on the full lead-time range used in verification, or analyze performance as a function of lead time and demonstrate that the PPM relationship is stable.
minor comments (4)
  1. [Abstract] The sentence 'learning directly from high-resolution radar estimates precipitation' is missing a function word; it should read 'radar-estimated precipitation.'
  2. [Section 2.4, Eq. (1)] The expectation in Eq. (1) is written without specifying the distributions over which expectations are taken; please define x, y, and the sampling procedure. Also, the learning rate '10e-5' is likely a typo for 1e-5 or 1e-4; please clarify.
  3. [Section 2.2 and Figure 1/Figure 2] Figure 2 is cited in Section 2.2 before Figure 1 is introduced; renumber or reorder the figures so they appear in the order cited.
  4. [Section 2.3] Bilinear interpolation is used to regrid all fields to 2 km, including precipitation; for precipitation fields, conservative or nearest-neighbor remapping is generally preferred to avoid smoothing maxima, which may matter at the heavy-rain thresholds emphasized in the results. Please justify this choice or test its sensitivity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DL-QPF is trained on 2021–2023 CLEANER radar labels and verified on a 2024 temporal holdout against independent baselines; using the same radar product for labels and verification is a limitation, not a derivation loop.

full rationale

I find no circularity in the paper's derivation chain. DL-QPF is a supervised mapping from IFS-HRES meteorological fields to CLEANER radar-estimated precipitation, trained on May–October 2021–2023 and verified on June–August 2024, a temporal holdout. The verification target is the same CLEANER radar product used as training labels (compare Section 2.2 with Section 3), which is a genuine limitation: the model may learn radar-specific artifacts and then score well against the same flawed benchmark. However, this is evaluation self-reference, not a reduction of the prediction to its inputs. The model is not fitted to the 2024 verification data, and the comparison against IFS-HRES, KIM variants, and GraphCast is computed on a common benchmark, so the central claim of superior skill has independent content. The near-unity frequency bias is partly encouraged by the adversarial loss in Section 2.4, whose objective is to match the label distribution, but the paper also reports POD, SR, and CSI, and admits a convective-shower weakness in Section 4.3, showing the match is not vacuous. The only self-citation is Lee et al. (2024), used for the Spatial Aligned Mean of the KIM-LENS baseline; it is a methodological choice for one comparison model and is not load-bearing for the central claim. The absence of independent gauge-based verification is a correctness/limitation concern, not a circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim is an empirical ML result, so there are no fitted physical constants. The free parameters are the hand-chosen hyperparameters of the GAN and the regridding resolution. The axioms are all domain assumptions about the trustedness of radar data, the perfect-prog shortcut, and the sufficiency of the chosen input variables. No new physical entities are proposed.

free parameters (5)
  • lambda (loss weight) = 100
    Chosen by hand to balance GAN and pixel losses; no sensitivity analysis is reported.
  • alpha (L1/L2 weighting) = 0.8
    Chosen by hand in the combined loss function; no ablation is shown.
  • learning rate = 10e-5 (ambiguous notation)
    Fixed optimizer setting; the paper states it without justification or search, and the notation 10e-5 is ambiguous.
  • number of epochs = 50
    Training length chosen without early stopping or convergence analysis.
  • output grid resolution = 2 km x 2 km
    All data are regridded to 2 km; the choice affects the spatial structure of the target and the comparison.
assumptions (4)
  • domain assumption CLEANER radar-estimated precipitation is an accurate ground-truth representation of surface rainfall.
    Used as both training labels and verification truth (Sections 2.2 and 3); any systematic radar bias is learned and scored as skill.
  • domain assumption The Perfect Prognostic Method applies: concurrent NWP atmospheric fields can be paired with observed precipitation as if the forecasts were perfect.
    Stated in Section 2: 'This study assumes a perfect NWP forecast, following the Perfect Prognostic Method.' This ignores forecast errors in the input fields at inference time.
  • domain assumption The 80 selected IFS-HRES variables contain sufficient information to diagnose precipitation.
    Section 2.1 and Table 1 list the variables; no feature importance or sensitivity analysis is provided to test this.
  • domain assumption Bilinear interpolation to a 2 km grid preserves precipitation-relevant information.
    Section 2.3 describes regridding from 500 m radar data and 0.1 degree IFS data; this smooths fine-scale structure and may affect heavy-rain peaks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Post-Processing for Quantitative Precipitation Forecasting Using Deep Learning: Learning Precipitation Physics from High-Resolution Observations." pith.science (2026). https://pith.science/paper/7OLO2FGH

@misc{pith2026250603842,
  author       = {Pith},
  title        = {Pith review of: Improving Post-Processing for Quantitative Precipitation Forecasting Using Deep Learning: Learning Precipitation Physics from High-Resolution Observations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7OLO2FGH}},
  note         = {Machine review of arXiv:2506.03842}
}
read the original abstract

Accurate quantitative precipitation forecasting (QPF) remains one of the main challenges in numerical weather prediction (NWP), primarily due to the difficulty of representing the full complexity of atmospheric microphysics through parameterization schemes. This study introduces a deep learning-based post-processing model, DL-QPF, which diagnoses precipitation fields from meteorological forecasts by learning directly from high-resolution radar estimates precipitation. The DL-QPF model is constructed using a Patch-conditional Generative Adversarial Network (Patch-cGAN) architecture combined with a U-Net generator and a discriminator. The generator learns meteorological features relevant to precipitation, while the adversarial loss from the discriminator encourages the generation of realistic rainfall patterns and distributions. Training is performed on three years of warm-season data over the Korean Peninsula, with input variables derived from ECMWF's Integrated Forecasting System High-Resolution forecast (IFS-HRES). Model verification is conducted against multiple reference models, including global (IFS-HRES, KIM), regional (KIM-Regional, KIM-LENS), and AI-based (GraphCast) forecasts. Verification across multiple rainfall thresholds shows that DL-QPF achieves a frequency bias near one and superior success ratios. Particularly for heavy and intense rainfall events, DL-QPF outperforms both conventional NWP and an AI model, demonstrating improved skill in capturing high-intensity precipitation. This study highlights the potential of observational data-driven deep learning approaches in post-processing QPF. By directly learning from observations, DL-QPF reduces systematic biases and enhances the realism of forecasted rainfall distributions. These results demonstrate the model's potential to enhance QPF realism.

Figures

Figures reproduced from arXiv: 2506.03842 by the authors.

Figure 1
Figure 1. Locations of S-band radar sites (blue dots) and their coverage areas (skyblue outline) (left), and locations (gray dots) of rain gauges (right) operated by the Korea Meteorological Administration (KMA) [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Scatter plot comparing radar-estimated precipitation with observed rainfall from rain gauges (left), and comparison between ERA5-predicted precipitation fields and observed rainfall (right). Six-hour accumulated rainfall data from June to August 2024 is used. The comparison is based on the nearest-point method [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Architecture of the generator. Each blue box represents a feature map, with the [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Architecture of the discriminator. Each blue box represents a feature map, with the [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Performance diagram for each model at the 1 mm/6 h rainfall threshold, from day 1 [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 9
Figure 9. Figure 9: Observed radar-estimated precipitation (left), IFS-HRES forecast (middle), and DL￾QPF prediction (right) for 3-h accumulated precipitation associated with a synoptic low case, from the 42-h forecast valid at 18 UTC on 9 July 2024. Mean sea level pressure composites are…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 1 linked inside Pith

  1. [1]

    Copyright in this Work may be transferred without further notice

    22 This Work has been submitted to Artificial Intelligence for the Earth Systems. Copyright in this Work may be transferred without further notice. Farr, T. G., P. A. Rosen, E. Caro, R. Crippen, R. Duren, S. Hensley, M. Kobrick, M. Paller, E. Rodriguez, L. Roth, D. Seal, S. Shaffer, J. Shimada, J. Umland, M. Werner, M. Oskin, D. Burbank, and D. Alsdorf, 2...

    arXiv doi:10.48550/arxiv.2006.11239 2007

  2. [2015]

    J., 2007: Parameterization Schemes: Keys to Understanding Numerical Weather Prediction Models

    vol 9351, https://doi.org/10.1007/978-3-319-24574-4_28 Stensrud, D. J., 2007: Parameterization Schemes: Keys to Understanding Numerical Weather Prediction Models. Cambridge University Press, https://doi.org/10.1017/CBO9780511812590 Wilks, D. S., 1995: Statistical Methods in the Atmospheric Sciences: An Introduction. Academic Press. Zhou, K. H., J. S. Sun,...

    doi:10.1007/978-3-319-24574-4_28 2007

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.