acceptodds
Under review as a conference paper at ICLR 2027

Yushi: Unified Precipitation Estimation and Nowcasting from Direct Observations

Abstract

Precipitation nowcasting is critical for hydrological forecasting and weather-related hazard warning, while most learning-based methods assume that spatially complete precipitation fields are already available as model inputs or prediction targets. In practice, such fields are usually derived from radar through quantitative precipitation estimation (QPE), interpolation, or multi-source merging rather than directly observed. In this paper, we propose Yushi, a unified framework for precipitation estimation and nowcasting directly from radar reflectivity and sparse automatic weather station (AWS) observations. Yushi first learns a gauge-constrained representation in which radar provides dense spatial structure while AWS measurements impose quantitative surface-precipitation constraints. Future precipitation states are then modeled in this latent space using causal conditional flow matching, and an intensity-aware balanced loss transfers precipitation-intensity information to latent flow optimization to emphasize heavy rainfall. Experiments over the Yangtze River Delta region show that Yushi consistently outperforms representative nowcasting baselines across various deterministic and spatially pooled metrics, with particularly clear improvements for intense precipitation. The results demonstrate the potential of directly learning and forecasting precipitation states from complementary primary observations without requiring a pre-constructed gridded precipitation analysis.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Deep Learning, Precipitation Nowcasting, Flow Matching

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_b1k1OpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose Yushi, a model that unifies quantitative precipitation estimation (QPE) and nowcasting from radar reflectivity and sparse weather station measurements. It learns a gauge-constrained latent representation (GCR) and uses causal conditional flow matching (CCFM) to forecast future precipitation, with an intensity aware loss to improve heavy rainfall predictions. Results on the Yangtze River Delta show improvements over the baselines considered.

Critical strengths
  1. Strong motivation and interesting approach. Unifying QPE and nowcasting by learning a latent representation directly from radar and station observations is an interesting idea. Forecasting in this space using flow matching, with joint prediction of future frames rather than autoregressive generation, is an appealing design. The intensity-aware loss is also well motivated.
  2. Clear presentation and promising results. The paper is well written, easy to follow, and the different components are clearly explained. The reported improvements are positive, and the ablation studies and held-out station experiments are appreciated.
Critical weaknesses
  1. Limited evaluation and unclear fairness of comparisons. The experiments are restricted to a custom dataset from a single region, with only four months used for testing. Evaluation on an established benchmark such as SEVIR would strengthen the paper, particularly given the relatively low CSI values in Table 1. More importantly, Yushi uses radar and station observations, while the baselines only use radar-derived precipitation. This makes it difficult to isolate the architectural improvements from the benefits of additional information. The ablation without GCR in Table 4 performs similarly to the baselines, reinforcing this concern. The held-out station experiment is useful but does not fully address it, since Yushi still benefits from station observations unavailable to the baselines.
  2. Missing comparisons and incomplete evaluation. Relevant work such as MetNet-3 and From Drops to Grid is not discussed. Since GCR is trained separately and frozen, it would be valuable to compare it against stronger reconstruction methods, including those based on Neural Processes, rather than only interpolation and Z-R conversion. Additionally, despite using a generative forecasting model, no probabilistic evaluation of CRPS, calibration, or ensemble diversity is reported, limiting the assessment of CCFM.
Questions for authors
  1. Could the authors clarify the evaluation targets in Table 1 and evaluate Yushi under matched input conditions? How does its performance change with different levels of station availability, including radar-only inputs?
  2. How are the stochastic predictions converted into deterministic forecasts for evaluation? Could the authors provide probabilistic metrics to demonstrate the benefits of flow matching?
  3. Since From Drops to Grid also estimates dense precipitation fields directly from radar and sparse station observations, could the authors clarify how GCR differs from this approach and provide a comparison of their reconstruction performance?
  4. Since station observations are used both as inputs and reconstruction targets, how do the authors prevent GCR from simply reproducing station values while relying on the Z-R relationship elsewhere? Could random station masking help?
4 comments

We sincerely thank the reviewer for the generous appreciation and positive assessment of our paper! Your suggestions are very valuable to us. We have revised the manuscript accordingly and explain the specific changes below. We sincerely appreciate your help in making the contribution and scope of our work clearer!

W1: Evaluation is limited to a single region and four test months.

We agree that evaluation across additional regions would provide stronger evidence of geographical generalization. Public benchmarks such as SEVIR, HKO-7, and MeteoNet do not contain colocated AWS/gauge observations and therefore cannot directly evaluate the radar-AWS formulation proposed in this work. To strengthen the present evaluation, we have added stronger deterministic and generative baselines, holdout-station experiments, QPE comparisons, additional ablations, and probabilistic ensemble evaluation.

The test period is temporally separated from the training and validation periods, rather than being randomly sampled from the same time range. We now more clearly describe this temporal split in the manuscript. We also acknowledge cross-region evaluation as a limitation and plan to release our radar-AWS dataset to support future research on this setting.

W2: It is difficult to separate the benefit of AWS observations from architectural improvements.

We agree that the contribution of additional AWS information should be distinguished from the contribution of the proposed architecture. We address this from three complementary perspectives. First, we add RN-Net as a radar+AWS baseline, and Yushi outperforms it in the evaluation as below.

ModelCSI-MHSS-MBIAS-MCSI 1mmCSI 2mmCSI 4mmCSI-4POOL 1mmCSI-4POOL 2mmCSI-4POOL 4mmCSI-16POOL 1mmCSI-16POOL 2mmCSI-16POOL 4mm
RN-Net0.08550.15110.17730.15780.07660.02210.15460.07480.02110.14520.07350.0215
Yushi0.10790.19210.77980.16100.10520.05750.16570.11090.06250.24920.19410.1269

Second, we include controlled Yushi variants in the table below. Yushi w/o GCR replaces the learned gauge-constrained representation with a compressed radar-derived precipitation background, and Yushi w/o GCR (radar+AWS) uses radar and AWS observations for direct precipitation forecasting without the proposed GCR representation. Comparing these variants with the full model separates the benefit of AWS information from the benefit of the proposed representation learning strategy.

ModelCSI-MHSS-MBIAS-MCSI 1mmCSI 2mmCSI 4mmCSI-4POOL 1mmCSI-4POOL 2mmCSI-4POOL 4mmCSI-16POOL 1mmCSI-16POOL 2mmCSI-16POOL 4mm
Yushi0.10790.19210.77980.16100.10520.05750.16570.11090.06250.24920.19410.1269
Yushi w/o GCR0.07740.14280.57660.10580.07700.04950.11170.08170.05320.19460.14840.1009
Yushi w/o GCR (radar+AWS)0.09030.16200.48430.15200.08390.03490.15440.08640.03660.20020.12990.0661

Third, we now evaluate Yushi under different levels of AWS availability. As shown in the following table, reducing the available stations from 100% to 80% causes only marginal performance degradation, while further reducing them to 60% leads to a more pronounced decline.

ModelCSI-MHSS-MBIAS-MCSI 1mmCSI 2mmCSI 4mmCSI-4POOL 1mmCSI-4POOL 2mmCSI-4POOL 4mmCSI-16POOL 1mmCSI-16POOL 2mmCSI-16POOL 4mm
Yushi-100%0.10910.19400.79120.16170.10770.05780.16640.11300.06250.24980.19550.1269
Yushi-80%0.10790.19210.77980.16100.10520.05750.16570.11090.06250.24920.19410.1269
Yushi-60%0.08740.15850.55370.13410.08460.04340.13860.08960.04710.22110.16600.1007

These results distinguish the value of gauge information from the value of the proposed representation learning strategy: additional AWS observations are beneficial, as expected, but Yushi remains robust when station availability is reduced, and the full GCR-based model consistently performs better than variants that use the same observation sources without the proposed representation.

Q1: How does Yushi perform under matched-input conditions and different levels of station availability?

As mentioned earlier, we have added an explicit station-availability experiment to address this question. In addition to the full setting using all available stations, we randomly retain only 80% and 60% of the AWS observations while keeping the remaining model configuration unchanged. The results demonstrate that Yushi is robust to moderate reductions in station availability. Using 80% of the stations yields performance very close to the full setting: CSI-M changes only from 0.1091 to 0.1079, while HSS-M changes from 0.1940 to 0.1921. The heavy-precipitation performance is also largely preserved; for example, CSI-16POOL at 4 mm remains 0.1269 in both settings. When station availability is further reduced to 60%, performance decreases more noticeably, but the model still retains useful forecasting skill.

Q2: How are stochastic predictions converted to deterministic forecasts, and can the probabilistic benefit of flow matching be demonstrated?

Excellent question! We have added probabilistic ensemble evaluation and an ensemble-size study. For probabilistic evaluation, we report CRPS and fair-CRPS directly on the raw ensemble predictions. For deterministic metrics such as CSI, HSS, and BIAS, we additionally use Probability-Matched Mean (PMM) to obtain a deterministic ensemble composite while better preserving heavy-precipitation amplitudes than a simple ensemble mean.

ModelCSI-MHSS-MBIAS-MCSI 1mmCSI 2mmCSI 4mmCSI-4POOL 1mmCSI-4POOL 2mmCSI-4POOL 4mmCSI-16POOL 1mmCSI-16POOL 2mmCSI-16POOL 4mm
Prediff0.04410.08030.58050.06650.04140.02430.07070.04390.02560.13970.09260.0548
Prediff-PMM (k=8)0.07940.14290.66670.10640.07710.05480.10970.08050.05690.17080.13420.0919
FlowCast0.05600.10310.34510.07650.05690.03450.08150.06050.03700.15820.12020.0754
FlowCast-PMM (k=8)0.08710.15720.37200.11060.08890.06170.11440.09240.06440.17000.14040.0937
Yushi0.10790.19210.77980.16100.10520.05750.16570.11090.06250.24920.19410.1269
Yushi-PMM (k=8)0.14460.25020.77930.19720.14320.09330.20040.14820.09880.26270.21870.1578

We further evaluate multiple ensemble sizes (K). Performance improves as the ensemble size increases and tends to saturate beyond approximately K=8~12, while the inference cost continues to grow. We therefore use K=8 in the main deterministic ensemble results as a practical accuracy-cost trade-off. This additional evaluation directly demonstrates the probabilistic capability of the CCFM formulation rather than assessing it only through deterministic scores.

ModelCSI-MHSS-MBIAS-MCSI 1mmCSI 2mmCSI 4mmCSI-4POOL 1mmCSI-4POOL 2mmCSI-4POOL 4mmCSI-16POOL 1mmCSI-16POOL 2mmCSI-16POOL 4mmCRPSFair-CRPS
Yushi (K=1)0.10790.19210.77980.16100.10520.05750.16570.11090.06250.24920.19410.1269--
Yushi (K=2)0.12180.21460.77170.17420.11850.07260.17840.12410.07780.25390.20270.14140.04300.0330
Yushi (K=4)0.13560.23640.77710.18710.13380.08590.19060.13890.09140.25830.21340.15350.03800.0330
Yushi (K=8)0.14460.25020.77930.19720.14320.09330.20040.14820.09880.26270.21870.15780.03550.0330
Yushi (K=12)0.14770.25490.77960.20080.14640.09610.20380.15120.10140.26410.21950.15850.03470.0330

W4/Q3: How does GCR differ from From Drops to Grid, and how does their reconstruction performance compare?

We have added From Drops to Grid, MSMN, and Kriging interpolation to the QPE comparison in the following table.

MethodMSE 1mmMSE 2mmMSE 4mmCORR 1mmCORR 2mmCORR 4mm
Bilinear interpolation from AWS2.41495.168112.81090.55400.45740.3489
Kriging interpolation from AWS2.36675.103612.75880.56770.46460.3561
MSMN1.94833.97029.00580.65430.54310.4143
From Drops to Grid2.23245.130213.9450.65750.53510.3689
Z-R conversion from radar reflectivity2.97795.386510.28920.57200.50820.3805
Yushi (decoding from GCR)2.02543.85278.31880.67200.59590.4570

The main difference is functional. From Drops to Grid and MSMN are designed primarily to reconstruct dense precipitation fields from observations. GCR must satisfy an additional requirement: its latent representation must both decode into a quantitatively constrained precipitation field and serve as the state space for subsequent generative nowcasting. Despite this additional latent-space requirement, GCR achieves competitive or better QPE performance on the holdout stations. This indicates that the representation remains effective for precipitation estimation while simultaneously supporting future-state prediction.

W3/Q4: How do you prevent GCR from simply copying station values and reverting to Z-R elsewhere? Could random station masking help?

We agree that this is an important concern, and we now provide two complementary experiments. First, we use a fixed 20% station holdout set whose observations are completely excluded from both training and test-time inputs and are used exclusively for evaluation. Therefore, performance at these locations cannot result from directly copying or memorizing their gauge values. As shown in above table, GCR still achieves lower MSE and higher CORR than the Z-R background and the other reconstruction baselines on these unseen stations.

Second, we explicitly evaluate the effect of random gauge masking during GCR training. Without masking, reconstruction performance deteriorates substantially. For example, at the 1-mm threshold, MSE increases from 2.0254 to 2.9779 and CORR decreases from 0.6720 to 0.5720. The masked model achieves the best results across all reported MSE and CORR metrics. This indicates that random station masking improves generalization to unobserved gauge locations and reduces dependence on any fixed station configuration.

MethodMSE 1mmMSE 2mmMSE 4mmCORR 1mmCORR 2mmCORR 4mm
Yushi (decoding from GCR) w/o mask2.97794.14468.66440.57200.57610.4453
Yushi (decoding from GCR)2.02543.85278.31880.67200.59590.4570

Together, the holdout-station and masking experiments provide evidence that GCR learns transferable radar--gauge relationships rather than simply memorizing observed station values and falling back to the Z-R relation elsewhere.

Thank you for the detailed response and the additional experiments. However, my main concerns remain, and the manuscript has not yet been updated to reflect the rebuttal, so I keep my score.

W1/Q1: How Yushi's forecasting model compares to the baselines without station data remains unanswered. The only radar-only variant, Yushi w/o GCR, was already in Table 4 and performs below ConvGRU, Earthformer and PhyDNet. With 60% of the stations, Yushi is already similar to the radar-only baselines and falls below them at 4 mm. The gains therefore appear to come mainly from the station data. Additionally, the evaluation is still limited to a single region.

Q2: CRPS is reported only for Yushi, not for the baselines. PMM should be described and discussed in the main paper.

Q3: The QPE comparison uses only MSE and CORR, without precipitation-specific metrics such as CSI, HSS, BIAS or pooled CSI.

Q4: The holdout evaluation is appreciated, but random station masking is not described in the paper. Also, the 1 mm "w/o mask" values are identical to the Z-R conversion row.

Reviewer_cxqNOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents Yushi for two-hour nowcasting from radar and sparse AWS observations and the GCR part of Yushi could also implement the quantitative precipitation estimation. It learns a gauge-constrained latent representation, predicts future states with causal conditional flow matching, and emphasizes heavy rainfall through an intensity-weighted loss.

Critical strengths

(1) The study assembles 38,447 radar–AWS samples for training, validation, and testing. (2) The reported results improve CSI and HSS, with event-frequency BIAS closer to 1. These gains suggest potential to reduce under-forecasting. (3) The abalation study shows that each component is useful.

Critical weaknesses

Main weaknesses: (1) Unfair comparison with baseline models. As the authors stated in Section 4.2, the baselines do not use AWS data as input, so we strongly suggest that the authors conduct an ablation study: evaluating Yushi under the same conditions as the other baselines, using only radar data to ensure fairness. The original results could be retained to demonstrate the effectiveness of AWS data integration. (2) The selection of baselines. In my opinion, the chosen baselines are spatiotemporal prediction methods for image sequences or radar echo map nowcasting. As the authors stated, they cannot accept raw AWS data. But why not compare Yushi with models that can accept AWS as input, like RN-Net [1]? (3) Additional comparisons with baselines need to be made. (a) In Table 1, please consider comparing Yushi with PreDiff [2] (a latent-space diffusion method for precipitation nowcasting) and AlphaPre [3] (a deterministic precipitation nowcasting model). (b) In Table 3, please consider comparing with the Kriging interpolation method, which is always used for interpolating meteorological variables. Besides, GCR should also be compared with other deep-learning-based methods, like [4] and [5].

Minor weaknesses: (1) Parameter selection analysis and declaration: (a) Training loss of GCR: the selection of α and β should be analyzed. Why use 0.1:50? (b) AvgPool in IAB: why use 8×8 AvgPool? How about 4×4 and 16×16? (c) Ensemble evaluation: the CRPS/fair CRPS performance of FlowCast and Yushi should be reported for different ensemble sizes (K). The number of ensemble members required for CRPS to converge should also be reported. (2) Visualization of precipitation nowcasting results (Figures 3, 5, 6, and 7): Is 4 mm only a display limit, or are values clipped? Please specify whether thresholds use ≥ or >. (3) In Figures 3, 5, 6, and 7, GCR’s results are reported together with other models’ prediction results. Because GCR’s results are QPE reconstructed from radar and AWS data, they are reconstructions, not predictions. Therefore, I think GCR’s visualizations should be listed independently and compared with those of the other methods shown in Table 3.

References [1] RN-Net: A Deep Learning Approach to 0–2 Hour Rainfall Nowcasting Based on Radar and Automatic Weather Station Data. DOI: 10.3390/s21061981. [2] PreDiff: Precipitation Nowcasting with Latent Diffusion Models. NeurIPS 2023. DOI: 10.52202/075280-3439. [3] AlphaPre: Amplitude-Phase Disentanglement Model for Precipitation Nowcasting. CVPR 2025. DOI: 10.1109/CVPR52734.2025.01662. [4] A Deep Learning Multimodal Method for Precipitation Estimation. DOI: 10.3390/rs13163278. [5] Radar–Gauge Merging of Hourly Rainfall Based on the Multisource Spatial Merging Network Model. DOI: 10.1175/JTECH-D-23-0110.1.

Questions for authors

Please address the main concerns in the Weaknesses section above.

3 comments

We sincerely thank the reviewer for the generous appreciation and positive assessment of our paper! Your suggestions are very valuable to us. We have revised the manuscript accordingly and explain the specific changes below. We sincerely appreciate your help in making the contribution and scope of our work clearer!

W1: The comparison is unfair because Yushi uses AWS while most baselines do not.

We agree that it is important to distinguish improvements caused by the proposed architecture from improvements caused by access to additional AWS observations. We therefore add two types of controlled comparisons.

First, we include RN-Net, which accepts both radar and AWS observations, as a multimodal baseline. Yushi still outperforms RN-Net in both the full-station and holdout-station evaluations, as shown in the table below.

ModelCSI-MHSS-MBIAS-MCSI 1mmCSI 2mmCSI 4mmCSI-4POOL 1mmCSI-4POOL 2mmCSI-4POOL 4mmCSI-16POOL 1mmCSI-16POOL 2mmCSI-16POOL 4mm
RN-Net0.08550.15110.17730.15780.07660.02210.15460.07480.02110.14520.07350.0215
Yushi0.10790.19210.77980.16100.10520.05750.16570.11090.06250.24920.19410.1269

Second, we include controlled Yushi variants in the table below. Yushi w/o GCR replaces the learned gauge-constrained representation with a compressed radar-derived precipitation background, and Yushi w/o GCR (radar+AWS) uses radar and AWS observations for direct precipitation forecasting without the proposed GCR representation. Comparing these variants with the full model separates the benefit of AWS information from the benefit of the proposed representation learning strategy. We have revised the text to distinguish these two evaluation goals: matched-input comparisons are used to assess architectural improvements, while the full radar+AWS setting reflects the intended operational use of Yushi, where all available observations are exploited to maximize forecast accuracy.

ModelCSI-MHSS-MBIAS-MCSI 1mmCSI 2mmCSI 4mmCSI-4POOL 1mmCSI-4POOL 2mmCSI-4POOL 4mmCSI-16POOL 1mmCSI-16POOL 2mmCSI-16POOL 4mm
Yushi0.10790.19210.77980.16100.10520.05750.16570.11090.06250.24920.19410.1269
Yushi w/o GCR0.07740.14280.57660.10580.07700.04950.11170.08170.05320.19460.14840.1009
Yushi w/o GCR (radar+AWS)0.09030.16200.48430.15200.08390.03490.15440.08640.03660.20020.12990.0661

W2: Why not compare with methods that can ingest AWS, such as RN-Net?

Following this suggestion, we have added RN-Net as a radar-AWS baseline. As mentioned earlier, Yushi outperforms RN-Net in the evaluation.

W3: Additional forecasting and QPE baselines are needed.

We have expanded both sets of comparisons. For precipitation nowcasting, we add PreDiff and AlphaPre in the evaluation below, covering latent diffusion and deterministic precipitation forecasting approaches in addition to the original baselines.

ModelCSI-MHSS-MBIAS-MCSI 1mmCSI 2mmCSI 4mmCSI-4POOL 1mmCSI-4POOL 2mmCSI-4POOL 4mmCSI-16POOL 1mmCSI-16POOL 2mmCSI-16POOL 4mm
Prediff0.04410.08030.58050.06650.04140.02430.07070.04390.02560.13970.09260.0548
AlphaPre0.08380.15140.24840.11580.08570.04980.11710.08600.04930.14520.10260.0589
Yushi0.10790.19210.77980.16100.10520.05750.16570.11090.06250.24920.19410.1269

For QPE, the following table now includes Kriging interpolation, MSMN, and From Drops to Grid, in addition to bilinear interpolation and radar Z-R conversion. It should be noted that MSMN and From Drops to Grid were originally developed for precipitation reconstruction rather than latent generative nowcasting. We therefore evaluate them specifically in the QPE experiment, where their functionality is directly comparable to the decoding capability of GCR.

MethodMSE 1mmMSE 2mmMSE 4mmCORR 1mmCORR 2mmCORR 4mm
Bilinear interpolation from AWS2.41495.168112.81090.55400.45740.3489
Kriging interpolation from AWS2.36675.103612.75880.56770.46460.3561
MSMN1.94833.97029.00580.65430.54310.4143
From Drops to Grid2.23245.130213.9450.65750.53510.3689
Z-R conversion from radar reflectivity2.97795.386510.28920.57200.50820.3805
Yushi (decoding from GCR)2.02543.85278.31880.67200.59590.4570

W4(a): Why are the GCR loss weights set to α=0.1\alpha=0.1 and β=50\beta=50?

The two loss terms play different roles in GCR. The radar-derived precipitation background provides spatially dense but potentially biased guidance, whereas AWS measurements provide direct surface precipitation observations and therefore serve as the primary quantitative constraint. We consequently assign a substantially larger weight to the AWS term.

To further examine this choice, we add a sensitivity experiment by keeping α=0.1\alpha=0.1 fixed and reducing the AWS weight from β=50\beta=50 to β=10\beta=10. As shown in the following table, reducing β\beta degrades the reconstruction performance across all precipitation thresholds. For example, at the 4-mm threshold, MSE increases from 8.3188 to 9.9487 and CORR decreases from 0.4570 to 0.4251. This supports the use of a stronger gauge constraint and confirms that the performance gain does not simply come from reproducing the radar-derived Z-R background.

MethodMSE 1mmMSE 2mmMSE 4mmCORR 1mmCORR 2mmCORR 4mm
Yushi (decoding from GCR) (α=0.1,β=10\alpha=0.1, \beta=10)2.62304.91669.94870.61820.55610.4251
Yushi (decoding from GCR) (α=0.1,β=50\alpha=0.1, \beta=50)2.02543.85278.31880.67200.59590.4570

W4(b): Why use $8 \times 8 $ AvgPool in IAB? What about 4×44 \times 4 and 16×1616 \times 16?

The pooling operation maps precipitation intensity from the physical precipitation field to the spatial resolution of the latent representation. In GCR, the original spatial dimension of 461 is first padded to 464 and then downsampled by a factor of 8, resulting in a latent spatial dimension of 58. Accordingly, we use an $ 8 \times $ pooling factor here to establish a direct spatial correspondence between precipitation intensity in the physical space and locations in the latent space—a prerequisite for calculating the loss. We apologize for the confusion caused by the lack of clarity in our previous description.

W4(c): Probabilistic evaluation and ensemble-size analysis should be provided.

Excellent suggestion! We have added probabilistic ensemble evaluation and an ensemble-size study. For probabilistic evaluation, we report CRPS and fair-CRPS directly on the raw ensemble predictions. For deterministic metrics such as CSI, HSS, and BIAS, we additionally use Probability-Matched Mean (PMM) to obtain a deterministic ensemble composite while better preserving heavy-precipitation amplitudes than a simple ensemble mean.

ModelCSI-MHSS-MBIAS-MCSI 1mmCSI 2mmCSI 4mmCSI-4POOL 1mmCSI-4POOL 2mmCSI-4POOL 4mmCSI-16POOL 1mmCSI-16POOL 2mmCSI-16POOL 4mm
Prediff0.04410.08030.58050.06650.04140.02430.07070.04390.02560.13970.09260.0548
Prediff-PMM (k=8)0.07940.14290.66670.10640.07710.05480.10970.08050.05690.17080.13420.0919
FlowCast0.05600.10310.34510.07650.05690.03450.08150.06050.03700.15820.12020.0754
FlowCast-PMM (k=8)0.08710.15720.37200.11060.08890.06170.11440.09240.06440.17000.14040.0937
Yushi0.10790.19210.77980.16100.10520.05750.16570.11090.06250.24920.19410.1269
Yushi-PMM (k=8)0.14460.25020.77930.19720.14320.09330.20040.14820.09880.26270.21870.1578

We further evaluate multiple ensemble sizes (K). Performance improves as the ensemble size increases and tends to saturate beyond approximately K=8~12, while the inference cost continues to grow. We therefore use K=8 in the main deterministic ensemble results as a practical accuracy-cost trade-off. This additional evaluation directly demonstrates the probabilistic capability of the CCFM formulation rather than assessing it only through deterministic scores.

ModelCSI-MHSS-MBIAS-MCSI 1mmCSI 2mmCSI 4mmCSI-4POOL 1mmCSI-4POOL 2mmCSI-4POOL 4mmCSI-16POOL 1mmCSI-16POOL 2mmCSI-16POOL 4mmCRPSFair-CRPS
Yushi (K=1)0.10790.19210.77980.16100.10520.05750.16570.11090.06250.24920.19410.1269--
Yushi (K=2)0.12180.21460.77170.17420.11850.07260.17840.12410.07780.25390.20270.14140.04300.0330
Yushi (K=4)0.13560.23640.77710.18710.13380.08590.19060.13890.09140.25830.21340.15350.03800.0330
Yushi (K=8)0.14460.25020.77930.19720.14320.09330.20040.14820.09880.26270.21870.15780.03550.0330
Yushi (K=12)0.14770.25490.77960.20080.14640.09610.20380.15120.10140.26410.21950.15850.03470.0330

W5: Are values above 4 mm clipped? Do the threshold metrics use \textgreater\textgreater or ≥\geq?

The 4-mm value is only the upper limit of the visualization color scale. Values above 4 mm are saturated visually but are not clipped in the quantitative evaluation. All threshold-based metrics use precipitation values ≥\geq the specified threshold. We have clarified both points in the revised manuscript. For reference, 4 mm per 5 min corresponds to an hourly rate of 48 mm/h, so this visualization range highlights the spatial structure of heavy precipitation.

W6: GCR reconstruction should not be mixed with forecasting visualizations.

We agree. GCR produces a QPE reconstruction rather than a future forecast, and presenting it together with forecasting outputs can cause confusion. Therefore, we have separated the visualizations. The visualization of the forecast results remains unchanged, while for QPE, we have added a comparison between the GCR reconstruction results and other baselines. The figure can be found in the supplementary material.

Reviewer_ko6sOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose a method for precipitation estimation and nowcasting from the radar and sparse automatic weather stations (AWS) observations. To this end, they formulate a gauge-constrained representation learner and causal conditional flow matching model.

Critical strengths

The paper is sound, and proposes a valid method for precipitation forecasting.

Critical weaknesses

Limited evaluation: I am afraid one use case of precipitation nowcasting, Yangtze River Delta region dataset, does not look like it is enough to justify that the method works convincingly well. I think incorporating other data sources from diverse regions, or using standard benchmark datasets, would be critical to make the paper well-supported.

Contributions: it seems like the authors give a number of disjoint contributions, as opposed to tackling one large gap in precipitation forecasting. They target three deficiencies of the architecture: (1) tying the precipitation reconstruction problem to the latent space, offering simultaneous solution to reconstruction and nowcasting (2) gauge-constrained representation learning (3) precipitation intensity-aware latent flow matching. I would imagine that these three questions would need to be decoupled, and analysed separately, comparing with the existing approaches and showing that the proposed strategy works better.

Questions for authors

I think addressing the critical weaknesses by justifying limited evaluation concern, as well as a concern that even if there were more evaluations, there are three different contributions which, I believe, need to be demonstrated separately. Intensity-aware latent flow matching should demonstrate the ability to work on extreme precipitation; gauge-constrained representation learning helps learn representations more accurately; simultaneous reconstruction and nowcasting can help improve accuracy. Probably, if confirmed rigorously, these three questions, each on their own, may be a subject of a work each; now it seems like for all the good ideas that the authors have, there is not enough evidence to support them.

3 comments

We sincerely thank the reviewer for the generous appreciation and positive assessment of our paper! Your suggestions are very valuable to us. We have revised the manuscript accordingly and explain the specific changes below. We sincerely appreciate your help in making the contribution and scope of our work clearer!

W1: Evaluation is limited to one regional dataset.

We agree that cross-region evaluation would further strengthen the evidence for generalization. However, widely used public nowcasting benchmarks such as SEVIR, HKO-7, and MeteoNet do not provide colocated gauge/AWS observations, which are required to evaluate the central radar-gauge formulation of Yushi. Therefore, these datasets cannot directly reproduce our observation setting. We also explicitly acknowledge cross-region validation as an important limitation. We plan to release the radar-AWS dataset used in this study to facilitate future evaluation of observation-to-precipitation methods on this setting.

W2/Q1: Intensity-aware latent flow matching should demonstrate its effectiveness for extreme precipitation.

Thank you for this suggestion. We agree that the contribution of the intensity-aware objective should be evaluated specifically for heavy precipitation rather than only through overall forecasting metrics. We therefore analyze the IAB loss directly through the ablation and hyper-parameter experiments in the table below.

ModelCSI-MHSS-MBIAS-MCSI 1mmCSI 2mmCSI 4mmCSI-4POOL 1mmCSI-4POOL 2mmCSI-4POOL 4mmCSI-16POOL 1mmCSI-16POOL 2mmCSI-16POOL 4mm
Yushi0.10790.19210.77980.16100.10520.05750.16570.11090.06250.24920.19410.1269
Yushi w/o IAB0.10190.18260.45210.15130.09750.05680.15470.10200.06120.22470.16970.1137
Yushi (γ=0.5\gamma=0.5)0.10260.18370.51720.15260.09820.05690.15680.10330.06130.23620.17830.1174
Yushi (γ=1.5\gamma=1.5)0.10420.18631.12240.15520.10120.05640.16080.10740.06160.25610.19870.1298

Removing IAB (γ=0\gamma=0) consistently degrades the forecasting performance, while increasing γ\gamma places progressively greater emphasis on intense-precipitation regions. In particular, the pooled CSI and BIAS increase as γ\gamma increases, showing that the intensity-aware weighting encourages a more aggressive response to heavy precipitation. When γ\gamma becomes too large, BIAS can exceed 1, indicating over-forecasting. These results demonstrate that IAB explicitly controls the model's emphasis on rare high-intensity precipitation rather than merely improving average performance.

W2/Q2: Gauge-constrained representation learning should be shown to learn more accurate precipitation representations.

We agree. We have strengthened the standalone evaluation of GCR by comparing its decoded precipitation fields against both conventional and learning-based QPE approaches.As shown in the table below, we now include bilinear interpolation, Kriging interpolation, radar Z-R conversion, MSMN, and From Drops to Grid. All methods are evaluated on the same holdout stations that are never provided to the model during training or test-time inference. GCR achieves lower estimation error and higher correlation, particularly under heavier precipitation thresholds.

MethodMSE 1mmMSE 2mmMSE 4mmCORR 1mmCORR 2mmCORR 4mm
Bilinear interpolation from AWS2.41495.168112.81090.55400.45740.3489
Kriging interpolation from AWS2.36675.103612.75880.56770.46460.3561
MSMN1.94833.97029.00580.65430.54310.4143
From Drops to Grid2.23245.130213.9450.65750.53510.3689
Z-R conversion from radar reflectivity2.97795.386510.28920.57200.50820.3805
Yushi (decoding from GCR)2.02543.85278.31880.67200.59590.4570

These results show that GCR is not simply reproducing gauge values or relying on the Z-R background elsewhere. Instead, it combines the dense spatial information from radar with quantitative constraints from sparse gauges to infer precipitation at unseen locations. A further distinction from reconstruction-only approaches such as From Drops to Grid and MSMN is that GCR is designed not only to decode a dense precipitation field but also to provide the latent state used by the subsequent generative nowcasting model. Thus, the same learned representation supports both QPE and forecasting.

W2/Q3: The benefit of unifying precipitation estimation and nowcasting should be demonstrated separately.

Thank you for this comment. We have reorganized the ablation analysis to isolate the roles of the main components. Specifically, the following table contains the following controlled variants:

• Yushi w/o GCR: replaces the learned gauge-constrained representation with a compressed radar-derived precipitation background.

• Yushi w/o GCR (radar+AWS): uses radar and AWS observations for direct precipitation forecasting without the proposed GCR representation.

• Yushi w/o CCFM: keeps GCR fixed but replaces the proposed CCFM forecasting model with FlowCast.

• Yushi w/o IAB: uses the standard CFM objective with γ=0\gamma=0.

ModelCSI-MHSS-MBIAS-MCSI 1mmCSI 2mmCSI 4mmCSI-4POOL 1mmCSI-4POOL 2mmCSI-4POOL 4mmCSI-16POOL 1mmCSI-16POOL 2mmCSI-16POOL 4mm
Yushi0.10790.19210.77980.16100.10520.05750.16570.11090.06250.24920.19410.1269
Yushi w/o GCR0.07740.14280.57660.10580.07700.04950.11170.08170.05320.19460.14840.1009
Yushi w/o GCR (radar+AWS)0.09030.16200.48430.15200.08390.03490.15440.08640.03660.20020.12990.0661
Yushi w/o CCFM0.06430.11560.88810.09770.06070.03460.10210.06440.03750.18250.13380.0849
Yushi w/o IAB0.10190.18260.45210.15130.09750.05680.15470.10200.06120.22470.16970.1137

The full Yushi model consistently outperforms these variants, showing that the gain does not arise from a single component alone. More importantly, the experiments separately support the three design choices: GCR improves the precipitation representation, CCFM improves forecasting in this shared latent space, and IAB improves the treatment of heavy precipitation.

In addition, we have incorporated further baseline models based on the feedback from other reviewers and evaluated the ensemble forecasting performance of our approach. The detailed results can be found in the table below.

ModelCSI-MHSS-MBIAS-MCSI 1mmCSI 2mmCSI 4mmCSI-4POOL 1mmCSI-4POOL 2mmCSI-4POOL 4mmCSI-16POOL 1mmCSI-16POOL 2mmCSI-16POOL 4mm
ConvGRU0.07880.14340.22280.10710.08050.04890.10840.08060.04880.13390.09530.0580
SimVP0.06480.11950.16320.09040.06310.04080.09100.06350.04110.11230.07840.0499
PhyDNet0.08100.14660.22640.11680.08090.04530.11810.08120.04540.14380.09620.0536
Earthformer0.08000.14500.25220.11140.08200.04660.11280.08240.04640.13920.10050.0565
Prediff0.04410.08030.58050.06650.04140.02430.07070.04390.02560.13970.09260.0548
Prediff-PMM (k=8)0.07940.14290.66670.10640.07710.05480.10970.08050.05690.17080.13420.0919
FlowCast0.05600.10310.34510.07650.05690.03450.08150.06050.03700.15820.12020.0754
FlowCast-PMM (k=8)0.08710.15720.37200.11060.08890.06170.11440.09240.06440.17000.14040.0937
RN-Net0.08550.15110.17730.15780.07660.02210.15460.07480.02110.14520.07350.0215
AlphaPre0.08380.15140.24840.11580.08570.04980.11710.08600.04930.14520.10260.0589
Yushi0.10790.19210.77980.16100.10520.05750.16570.11090.06250.24920.19410.1269
Yushi-PMM (k=8)0.14460.25020.77930.19720.14320.09330.20040.14820.09880.26270.21870.1578

Many thanks for taking your time to address the concerns, and especially the experiments which partially alleviate the concerns.

Two comments that are outstanding:

  • it is not clear how this has been incorporated into the paper
  • the comment about disjoint contributions sounds critical. The authors consider three questions: (1) tying the precipitation reconstruction problem to the latent space, offering simultaneous solution to reconstruction and nowcasting (2) gauge-constrained representation learning (3) precipitation intensity-aware latent flow matching. I cannot see any obvious answer to why these three questions the authors discuss would need to be inextricably coupled.

Clarification on these points would be very valuable 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.