acceptodds
Withdrawn from ICLR 2027

Neural Stochastic Processes for Satellite Precipitation Refinement

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurate precipitation estimation is critical for flood forecasting, water resource management, and disaster preparedness. Satellite products provide global hourly coverage but contain systematic biases; ground-based gauges are accurate at point locations but too sparse for direct gridded correction. We propose Neural Stochastic Process (NSP), pairing a set-conditioned spatial latent encoder with a Neural SDE transition regularizer to improve spatial refinement with single-hour inference. An Euler-discretized variational model yields an analytic conditional transition KL, used in a regularized training objective without simulating SDE trajectories. We also introduce QPEBench, a benchmark of 43,756 hourly samples over the Contiguous United States (2021–2025), with four aligned data sources and six deterministic metrics. Compared with 17 baselines, NSP achieves the best performance across all six deterministic metrics and surpasses JAXA's operational gauge-calibrated product. NSP also transfers to two regions excluded from training without parameter updates, using sparse local gauges. Controlled experiments show benefits from temporal coupling during training and learned latent transitions during temporary gauge outages. Our project page is available at https://nsp-rlc9n.kinsta.page/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural processes, neural SDEs, satellite precipitation, sparse observations, spatio-temporal modeling, benchmark

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_YSRjOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose Neural Stochastic Processes (NSP), which extend a spatial Neural Process with stochastic latent transitions to correct satellite precipitation using sparse ground stations and topography. They also introduce QPEBench, a benchmark focused on the CONUS region with satellite, gauge, elevation, and radar data. NSP outperforms 17 baselines and enables latent extrapolation during gauge outages.

Critical strengths
  1. Promising stochastic temporal modeling. The latent SDE regularizer is an interesting idea, particularly for predicting during gauge outages. This could complement or replace temporal conditioning in existing spatio-temporal Neural Processes.
  2. Valuable benchmark contribution. QPEBench integrates relevant data sources, temporal splits, and evaluation procedures for precipitation estimation. The extensive supplementary experiments and planned code and dataset release are positive contributions.
Critical weaknesses
  1. Evaluation protocol and benchmark robustness. The main radar errors are restricted to radar-rainy pixels, ignoring false-positive rainfall, while gauge metrics include context stations rather than focusing on held-out locations. The benchmark also lacks the Critical Success Index (one of the main precipitation forecasting metrics), CRPS (for probabilistic assessment), and spatial generalization analysis (since these are not complete and do not compare all methods). The Gaussian likelihood is also poorly suited to zero-inflated, heavy-tailed precipitation, particularly for extremes, so baselines like ConvCNP are not properly adapted for this task. A unified evaluation using held-out stations, all-pixel errors, CSI/POD/FAR, CRPS, calibration, and appropriate predictive distributions is needed.
  2. Insufficiently competitive baselines and unclear SDE contribution. The paper overlaps substantially with “From Drops to Grid: Noise-Aware Spatio-Temporal Neural Process for Rainfall Estimation”. However, it does not compare against it or other recent spatio-temporal Neural Processes. Existing baselines are mostly older architectures, and learned baselines use different training objectives despite receiving comparable inputs. The latent SDE is a promising contribution, particularly for predicting during gauge outages, but its benefits are not sufficiently established and highlighted. The paper would benefit from focusing more on this contribution, with stronger comparisons against modern temporal models and deterministic alternatives, rather than emphasizing broad benchmark improvements.
Questions for authors
  1. How do the main results change when evaluating exclusively on held-out stations and all radar pixels, including false-positive rainfall? Could the authors provide a unified comparison using also CSI/POD/FAR and CRPS?
  2. Could the authors compare against DropsToGrid adapted to satellite inputs and other recent spatio-temporal Neural Processes, ensuring comparable inputs, training objectives, and evaluation protocols?
  3. What specific advantages do stochastic latent transitions provide over deterministic alternatives or temporal conditioning on historical station observations?
Reviewer_kw5LOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Neural Stochastic Process (NSP), a framework for refining satellite precipitation estimates using sparse rain gauge observations. The method combines a set-conditioned spatial encoder with a Neural SDE-based latent transition regularizer to incorporate temporal information during training while maintaining single-hour inference. The authors also introduce QPEBench, a multi-year benchmark integrating satellite precipitation, gauge observations, elevation, and radar references over the contiguous United States. Extensive experiments against 17 baselines demonstrate improvements in deterministic precipitation estimation, while additional ablation, gauge-outage, and transfer experiments evaluate the contributions of temporal regularization and learned latent transitions.

Critical strengths
  1. The paper provides extensive comparisons with conventional statistical methods and operational gauge-corrected satellite precipitation products(in Tab1), demonstrating the practical value of the proposed framework beyond comparisons with DL baselines.

  2. The ablation studies are thorough, particularly the experiments examining temporal regularization and learned latent transitions under gauge-outage conditions. These analyses provide useful evidence of how temporal information can improve spatial precipitation refinement, even without explicit temporal forecasting during standard inference.

Critical weaknesses
  1. Insufficient justification for the Neural SDE formulation.

Although the extensive ablation studies demonstrate the benefits of temporal regularization, the improvements over deterministic or simplified stochastic transitions remain modest. In particular, Table 3 shows relatively small performance differences, while Table 4 shows only marginal improvements over constant diffusion, especially considering the reported standard deviations. Furthermore, the model uses a one-step Euler-discretized transition without SDE integration during standard inference, making the necessity of a continuous-time stochastic formulation unclear.

  1. Limited evidence of the practical benefits of stochastic temporal modeling.

The primary task is single-hour precipitation refinement rather than forecasting. While temporal regularization improves spatial accuracy, the authors do not directly demonstrate that it improves the temporal consistency of consecutive refined precipitation fields. Moreover, NSP does not consistently outperform competing methods in probabilistic evaluation. Consequently, the practical advantages of stochastic temporal dynamics remain insufficiently established.

  1. Limited comparison with recent gauge-conditioned precipitation refinement methods.

Despite the extensive baseline comparisons, the paper insufficiently discusses recent approaches that fuse sparse gauge observations with satellite or gridded precipitation data (such as PRIMER(Sun et al, 2026), QCGS(Kim et al, 2026), BCFS(Shen et al, 2026)) methods. Although these approaches differ in their inputs and supervision settings, a clearer discussion of their methodological differences and practical trade-offs would help establish the novelty and advantages of NSP.

Questions for authors
  1. Does Neural SDE-based temporal regularization improve the temporal consistency of consecutive refined precipitation fields? An evaluation of temporal gradients or accumulated precipitation errors relative to a model without the Neural SDE would help clarify this contribution.

  2. Although Appendix provides extensive gauge-density sensitivity analyses, the experiments primarily rely on randomly subsampled station networks. How well do these findings generalize to geographically clustered or systematically sparse gauge networks, particularly in mountainous regions where precipitation patterns are highly heterogeneous?

Reviewer_SQG4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies hourly refinement of satellite precipitation estimates (GSMaP) using sparse rain-gauge observations. The proposed Neural Stochastic Process (NSP) encodes satellite precipitation, elevation, and a sparse context-gauge map with its mask into a diagonal Gaussian posterior over a 2D spatial latent field ztz_t, and a decoder predicts a log-space residual correction to the satellite estimate together with a heteroscedastic observation variance. During training, consecutive hours are coupled by a transition regularizer: a Neural SDE is discretized with a single Euler-Maruyama step (Δt=1\Delta t=1 hour), and the analytic Gaussian KL between the resulting transition pθ(zt+1∣zt)p_\theta(z_{t+1}\mid z_t) and the independently encoded posterior qϕ(zt+1∣xt+1)q_\phi(z_{t+1}\mid x_{t+1}) is added to the objective, so no latent trajectory is simulated and inference remains a single encoder-decoder pass per hour. The authors also introduce QPEBench, a CONUS benchmark of 43,756 hourly samples (2021-2025) on a 0.1-degree grid that aligns GSMaP, elevation, and Synoptic gauge observations with MRMS radar used only as an evaluation reference, with three chronological folds and six deterministic metrics. Against 17 baselines and the operational GSMaP GC product, NSP reports the lowest rainy-area RMSE (2.818 vs. 2.942 for GSMaP GC) and the highest FSS, and the paper additionally presents a Kyushu regional experiment, transfer to excluded regions and to MSWEP inputs without retraining, temporal-pairing controls, and gauge-outage experiments using latent SDE rollout.

Critical strengths
  1. Benchmark contribution. QPEBench is a systematic benchmark for satellite precipitation refinement from sparse rain gauges. It provides chronological year-based splits, explicit evaluation masks, and a clear distinction between point observations (gauges) and an areal reference (radar), which makes it highly reusable in this area. The design choice of excluding radar from model inputs and training targets and using it only as an evaluation reference is also well justified.

  2. Broad experimental comparison. The paper compares against a wide range of baselines and also against operational products, namely GSMaP GC, Daymet, and MSWEP. It further examines generalization from multiple angles, including a Kyushu regional experiment, transfer to regions excluded from training, and transfer to a different precipitation product as input.

  3. Practical design. Because the transition KL is computed analytically, training does not require simulating SDE trajectories, and inference requires only a single encoder-decoder pass per hour (about 2.6 ms), which makes the method easy to deploy in operational settings. The auxiliary mode that rolls out the latent state with the learned transition during temporary gauge outages is also practically useful.

Critical weaknesses
  1. Limited methodological novelty. The KL regularization with a single-step Euler Gaussian transition is not substantially different from existing dynamical VAE approaches, and since the continuous-time SDE structure is not actually exploited, the "Neural SDE" framing seems overstated.

  2. Small performance gains. The improvement on the main metric is only about 4%, and several metrics are effectively tied with baselines, so the claim of being the best on all metrics is overstated.

  3. Inconsistent effect of the temporal regularizer. The contribution of Ltrans\mathcal{L}_{\mathrm{trans}} differs considerably between Table 3 and the multi-seed controlled experiments in Appendix G, and the absolute performance also varies substantially with the training configuration.

  4. Fairness of baseline comparisons. The learned baselines lack a satellite-residual structure, so a large portion of the performance gap appears to come from the residual parameterization rather than from the proposed method.

  5. Limitations in critical regimes. There is no improvement for extreme precipitation (above 25 mm/h), and the method also falls behind some baselines on probabilistic metrics (CRPS).

Questions for authors
  1. Why does the contribution of Ltrans\mathcal{L}_{\mathrm{trans}} differ so much between Table 3 and Appendix G? Which of the two do you consider to better reflect the actual effect of the temporal regularizer?

  2. If the learned baselines (e.g., U-Net and ConvCNP) were given the same satellite-residual output and trained with the same NLL objective, how much of the performance gap in the main results table would remain?

  3. What practical advantage does the continuous-time SDE formulation provide over a simple discrete Gaussian transition network?

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

Related papers

Discussion (0)

Sign in to comment.