acceptodds
Under review as a conference paper at ICLR 2027

Discrete Diffusion Priors for Rain-Field Reconstruction from Commercial Microwave Links

Abstract

Reconstructing rain fields from commercial microwave links (CMLs) requires inferring spatially distributed intensities from sparse, path-integrated attenuation measurements. Rainfall intermittency and spatial dependence make the prior central to this inverse problem. We introduce discrete diffusion priors for rain fields: masked and uniform diffusion models trained on fields encoded with an explicit dry class, which preserves rainfall occurrence exactly, and a codebook learned on positive intensities. We formulate CML reconstruction under three observation models (midpoint and extended virtual rain gauges, and a path-integrated likelihood) and benchmark conditioning methods: masked in-filling, CML adaptations of G2D2 and GILC, and D-MGPS, a new discrete variant of midpoint guidance. On the OpenMRG dataset, masked in-filling reaches an RMSE comparable to that of continuous diffusion priors at one to two orders of magnitude lower runtime and attains the highest wet/dry accuracy. With midpoint gauges, its intervals are much narrower than kriging intervals and cover of all locations, but only of wet locations. Likelihood-guided discrete samplers, including D-MGPS, are currently less accurate than in-filling.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Discrete Diffusion Models, Rain field reconstruction, Commercial Microwave Links, Bayesian inverse problems

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_DqodOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies rain-field reconstruction from commercial microwave links using discrete diffusion priors with an explicit dry class and quantized positive rain intensities. It compares in-filling and likelihood-guided methods, introduces D-MGPS, and evaluates reconstruction accuracy, wet/dry classification, runtime, and uncertainty quantification.

Critical strengths

The paper addresses a practically relevant inverse problem, and the explicit dry class is well motivated by the distribution of rainfall data.

In the reported experiments, masked in-filling achieves RMSE comparable to continuous diffusion baselines at substantially lower computational cost.

Reporting coverage over both all locations and wet locations exposes shortcomings obscured by aggregate metrics. The authors also transparently acknowledge the limitations of their proposed sampler.

Critical weaknesses

Lack of a unified framework overview. The encoding scheme, prior training, and conditional sampling procedures are scattered across the main text and appendices, making the overall pipeline and relationships between components difficult to follow.

Insufficiently established novelty. The method primarily combines and adapts existing techniques without clearly demonstrating that it addresses a substantive gap left unresolved by prior work. The newly proposed D-MGPS also fails to demonstrate a clear advantage in the reported experiments.

The manuscript reads more like an experimental report. Substantial space is devoted to listing methods, configurations, and results, with insufficient analysis of why the design works, what drives the performance differences, and what generalizable insights the experiments provide.

Questions for authors

Beyond combining an explicit dry class with existing discrete diffusion and conditional sampling methods, what is the substantive methodological novelty of this work?

Why does D-MGPS underperform simpler in-filling methods, and what value does the proposed D-MGPS contribute given these results?

Could the authors provide a unified framework diagram illustrating the overall pipeline?

1 comment

We appreciate the reviewer's feedback, below we address their concerns.

Q1 and W2:

To our knowledge, this is the first study to formulate CML-based rain-field reconstruction using discrete diffusion priors. The framework separates zero rainfall from strictly positive intensities, quantizes the latter, and learns a prior over the resulting categorical fields. It then conditions this prior on CML-derived information for reconstruction. This formulation preserves an explicit dry state while learning spatial rainfall structure.

Quantization and discrete diffusion are established techniques. Our contribution is their formulation and evaluation for the CML-based rain-field reconstruction problem, together with the D-MGPS sampling construction revisited below in Q2. The reported performance of VRG masked in-filling supports the practical value of the framework: it achieves reconstruction accuracy comparable to continuous diffusion baselines at substantially lower inference cost in the evaluated setting.

In the revisited version, we will strengthen the scope and contribution of the paper.

Q2:

D-MGPS explores where likelihood information is introduced along the discrete diffusion sampling trajectory. For a reverse transition from tt to t−1t-1, G2D2 uses likelihood information at time zero, whereas GILC uses information at the current noisy state. D-MGPS instead introduces likelihood information at an intermediate time 0<u<t0 < u < t. Its contribution is the resulting sampling construction and derivation.

The reported experiments do not demonstrate an empirical advantage for D-MGPS. We acknowledge that we do not yet have a conclusive explanation for its weaker performance. We are investigating several hypotheses as part of ongoing work.

W1, W3, Q3:

We agree that a unified overview would make the relationships between the components clearer. We will add a diagram showing rainfall encoding, prior training, and CML-conditioned reconstruction, including the two priors, observation models, and conditioning approaches. We will also revise the discussion to distinguish the methodological contribution from the experimental findings and to explain the limitations of D-MGPS and the current evaluation.

Reviewer_FzoQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper reconstructs rain fields from commercial microwave link (CML) measurements using discrete diffusion priors. Rain intensities are encoded as categories: one dry class plus K−1 rain levels learned with Lloyd’s algorithm. Masked and uniform diffusion priors are trained on these fields and then conditioned on link data in three ways: in-filling with virtual rain gauges, adapted versions of G2D2 and GILC, and a new sampler, D-MGPS. On OpenMRG, simple in-filling reaches an RMSE similar to continuous diffusion, runs much faster, and has the best wet/dry accuracy. The guided samplers, including D-MGPS, are less accurate.

Critical strengths
  1. The topic is interesting and practical. Commercial microwave links already exist almost everywhere as part of cellular networks. Turning their signals into rain maps provides extra rainfall information at no additional sensor cost, which can support flood warning and water management. The paper also brings discrete diffusion models to a new and meaningful application.
  2. The motivation is good. Rainfall is intermittent: many locations have exactly zero rain, and a continuous prior cannot represent exact zeros without an extra mechanism.
  3. The method is thoroughly argued. The appendix derives the discrete bridges, the expected CML likelihood, and each sampler step by step. It also states clearly which steps are exact and which are approximate, and why exact local computations do not guarantee exact posterior sampling.
  4. The results are reported transparently and are reproducible.
Critical weaknesses
  1. The main claim is not tested directly. The paper argues that the dry-class encoding helps, but there is no ablation without the dry class. The best method (MDM + VRG in-filling) differs from the continuous baselines in three ways at once: the prior, the observation model and the conditioning method. The gain therefore cannot be credited to the discrete prior. With the same path-integrated likelihood, the discrete samplers are actually worse than the continuous ones (Table 1). Please add (a) a dry-class ablation, (b) the continuous prior with the same VRG in-filling, and (c) the size and training cost of both priors.
  2. The guided samplers may be under-tuned. G2D2, GILC and D-MGPS each use a single hyperparameter setting (Table 4), with no validation set and no search. G2D2 also drops the warm start of the original method. The authors say they do not know why D-MGPS underperforms. The conclusion that likelihood guidance is worse than in-filling may therefore come from tuning. Please tune on a validation split and show sensitivity to the main hyperparameters (guidance strength, η, gradient estimator). Please also discuss related samplers such as SGDD (NeurIPS 2025).
  3. The metrics favor the proposed encoding, and the differences are not significant. Wet/dry accuracy counts any value above zero as wet. Smooth interpolators therefore score 0.22–0.38, while predicting “dry” everywhere would score about 0.80. All RMSE differences are within the confidence intervals. These intervals are not paired, even though every method is evaluated on the same test problems. For uncertainty, the 95% intervals of in-filling cover only 42% of wet locations, so their narrowness reflects overconfidence, not an advantage. Coverage for D-MGPS is missing. Please use one wet/dry threshold for all methods, report trivial baselines , use paired tests, and add CRPS or interval scores.
  4. The evaluation is narrow. There is one dataset, one link network and one noise level, and the link data are simulated with the same model used for inference. Moufad et al. (2026), whose protocol the paper follows, also test heteroscedastic noise and different link configurations; these are missing here. A test with few long links versus many short links would also check the authors’ own explanation for why guidance fails on path-integrated data. The speed comparison is not matched either: a small unguided model is compared with gradient-guided samplers on a continuous prior of unreported size. Please add these settings and report model sizes and function evaluations. Results on real CML data would be a plus.
  5. The paper is hard to follow.There is no overview figure, although the paper combines two priors, three observation models and three conditioning approaches. Table 1 does not show which rows are new, which are adapted and which are external. The baseline numbers appear identical to those in Moufad et al. (2026), yet l. 453 says all runtimes were measured on the same GPU. Please add an overview figure, mark the origin of each row in Table 1, say which baselines were re-run, and state clearly what is new compared with Moufad et al. (2026).
Questions for authors
  1. Does the benefit of the dry class remain in an ablation without it, and when the continuous prior is conditioned on the same VRG pseudo-observations by in-filling? What are the size and training cost of the continuous prior?
  2. How were the hyperparameters in Table 4 chosen? After tuning on a validation split, are the guided samplers, including D-MGPS, still
  3. With one wet/dry threshold for all methods and paired tests, is the advantage of in-filling still significant? What score does an all-dry prediction get?
  4. How do the methods rank under heteroscedastic noise, and with few long links versus many short links? What are the units of σ = 0.05?
  5. Were the baselines (MGPS, TDS, IDW, GMZ, OK) re-run, or taken from Moufad et al. (2026)? Their numbers appear identical to Tables 2–3 of that paper, while l. 453 says all runtimes were measured on the same GPU.

All five are requests for the rebuttal. If the authors can address them, especially 1–3, I will raise my score.

1 comment

We thank the reviewer for their assessment and suggestions. Below, we address their concerns.

Q1: The continuous baseline learns directly from rain fields using Gaussian corruption, Xt=X0+σtZX_t = X_0 + \sigma_t Z, where Z∼N(0,I)Z \sim N(0, I). Its formulation does not explicitly separate the probability mass at zero rainfall from positive intensities. It therefore serves as an experimental setting illustrating reconstruction without an explicit dry class. Our encoding scheme represents dry locations with a dedicated category and quantizes strictly positive intensities. We also tested alternative encoding schemes, including joint quantization of all rainfall values, without reserving a separate dry category. This produces a codebook with a representative value close to zero, rather than an explicit dry state. The overall reconstruction performance remains comparable, while wet/dry accuracy is lower, as expected when dry locations are not represented explicitly. We will add a discussion of these quantization schemes and their results to the appendix.

The continuous prior from Moufad et al. (2026) has approximately 4 million parameters and required approximately 10 hours of training on two H100 GPUs; see appendix of Moufad et al. (2026). The discrete prior has approximately 1 million parameters and requires approximately 30 minutes on one L40S GPU. We will report these model sizes and training costs explicitly.

Q2: For the inverse-problem algorithms using discrete priors, hyperparameters were selected manually on 10 samples, using RMSE and PCC as selection criteria. Our goal is to assess the baselines out of the box with a reasonable amount of tuning. With this moderate effort, diffusion-based reconstruction achieves performance comparable to or better than standard meteorological methods such as IDW and GMZ in the evaluated setting. We acknowledge that this tuning procedure does not replace careful selection on an independent validation split, which would definitely translate into performance gains. We will clarify this protocol in the revised version of the paper.

Q3: The suggestion to introduce a threshold below which predictions from the continuous baseline are treated as dry is valid. However, this threshold must be tuned and depends on the data and the relative importance of dry versus wet classifications for the intended application. We argue that incorporating an explicit dry class through a discrete prior is a central advantage of our method. Our method explicitly predicts a dry class, avoiding the need to select and tune a threshold. We agree that including a threshold-based evaluation would strengthen the comparison.

Q4: The value σ=0.05\sigma = 0.05 is the standard deviation of the Gaussian noise added to CML attenuation measurements and is expressed in dB. We appreciate the suggested experiments, including comparisons between networks with few long links and many short links, but cannot provide additional results in this rebuttal.

Q5: We adopted the experimental protocol of Moufad et al. (2026) and reused their continuous-baseline results, including runtimes. We will identify the reused results explicitly and distinguish them from results obtained in this work

Finally, as suggested by the reviewer, we will add an overview figure to illustrate the inference pipeline. We agree with their comment in Weakness 3 regarding RMSE and coverage. We will add clarifying comments and temper our claims about statistically significant improvements.

We acknowledge the valuable additions and improvements suggested by the reviewer, including the SVDD baseline and ablations on CML link lengths, among others. However, we regret that we cannot provide additional experimental results in this rebuttal.

Reviewer_EPGTOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper aims to reconstruct rainfall from commercial microwave links (CMLs) using discrete diffusion priors.

Critical strengths

1.The method is clearly explained and seems plausible.

  1. Every question I had while reading the paper was addressed either in the main paper or in the supplementary material, so I have no remaining questions about the technical details.
Critical weaknesses

1.Many meteorological satellites are now in orbit, and their data are readily accessible, for example through EUMETSAT’s GEORing project. However, the introduction does not explain why rainfall should be reconstructed from commercial microwave links.

  1. The method and its explanation are good, but the motivation is insufficiently explained. For example, wealthy countries already have sufficient radar coverage, while estimating precipitation from geostationary satellites seems much more promising for countries with less infrastructure. I would really like the authors to explain convincingly why commercial microwave links should be used.

  2. The ground truth comes from radar. In that case, why use commercial microwave links instead of simply using radar?

I realize that these comments are repetitive, but I would be inclined to give the paper a score of 4 if the authors could demonstrate that reconstructing precipitation from commercial microwave links remains important today, or that it is a better alternative to satellite-to-radar reconstruction.

Questions for authors

N/A

2 comments

We thank the reviewer for their positive assessment. Below, we address their concerns.

W1:

CMLs provide rainfall information that complements satellite observations through three features:

  • Near-ground measurements: CMLs measure rain-induced attenuation along near-ground paths, whereas geostationary satellite retrievals infer surface precipitation indirectly from cloud and atmospheric properties.
  • Existing infrastructure: CMLs use communication networks already deployed for telecommunications, reducing the need for dedicated rainfall sensors.
  • Spatial and temporal resolution: Dense link networks provide additional spatial information, particularly in urban areas. OpenMRG records CML measurements every 10 seconds, compared with 30 minutes to several hours for many meteorological satellites. Previous evaluations against meteorological satellite and radar support the value of CML rainfall estimates [R.1, R.2]. These observations motivate CML reconstruction as a complementary source of rainfall information.

W2:

Where radar coverage is limited, CMLs provide near-ground rainfall information through existing telecommunications infrastructure. Where radar is available, CMLs can supplement it with observations at different spatial and temporal scales. For example, [R.3] reports that CML estimates can match or exceed the performance of a state-of-the-art radar-gauge product in areas with dense link coverage. Therefore using the value of CMLs-based reconstruction extends to regions with established meteorological infrastructure.

W3:

Radar fields serve as a proxy for actual rain fields during prior training, allowing the model to learn spatial rainfall structure and intermittency. During reconstruction, CML observations condition the generation process; a contemporaneous radar field is not required. This distinction allows the method to use a learned rainfall prior while incorporating the temporal and spatial information provided by CMLs. For example, OpenMRG provides CML measurements every 10 seconds, compared with 5-minute radar sampling. We will revise the introduction to explain these complementary roles and the practical motivation for CML-based reconstruction.


References

[R.1] Gaona, M. F. R., Overeem, A., Brasjen, A. M., Meirink, J. F., Leijnse, H., and Uijlenhoet, R. (2017). Evaluation of rainfall products derived from satellites and microwave links for the Netherlands. IEEE Transactions on Geoscience and Remote Sensing, 55(12), 6849–6859.

[R.2] Roversi, G., Alberoni, P. P., Fornasiero, A., and Porcù, F. (2020). Commercial microwave links as a tool for operational rainfall monitoring in Northern Italy. Atmospheric Measurement Techniques, 13(11), 5779–5797.

[R.3] Van Weverberg, K., Neutens, N., De Corte, S., Passtoors, A., Calderan, S., Ghilain, N., Reinoso-Rondinel, R., Reyniers, M., Overeem, A., Van de Vyver, H., and Van Schaeybroeck, B. (2026). Country-wide rainfall estimates from a commercial microwave link network in Belgium.

Thank you for the clarification. Your explanation of how CMLs complement satellite and radar observations, and how radar data are used during training but are not required at inference time, has improved my understanding of the motivation. I am therefore raising my score from 2 to 3.

However, the practical value I identified as a condition for raising my score to 4 has not yet been sufficiently substantiated. In particular, the response provides limited discussion of access to telecom data in radar-sparse regions and the applicability of a rainfall prior trained on radar data from another region. It also remains unclear under what conditions the proposed method provides additional value over existing satellite-based rainfall estimation. These remaining uncertainties prevent me from raising my score to 4 at this stage.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.