- 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
- 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.
- 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
- 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.
- 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
- 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?
- How are the stochastic predictions converted into deterministic forecasts for evaluation? Could the authors provide probabilistic metrics to demonstrate the benefits of flow matching?
- 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?
- 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
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.
| Model | CSI-M | HSS-M | BIAS-M | CSI 1mm | CSI 2mm | CSI 4mm | CSI-4POOL 1mm | CSI-4POOL 2mm | CSI-4POOL 4mm | CSI-16POOL 1mm | CSI-16POOL 2mm | CSI-16POOL 4mm |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Prediff | 0.0441 | 0.0803 | 0.5805 | 0.0665 | 0.0414 | 0.0243 | 0.0707 | 0.0439 | 0.0256 | 0.1397 | 0.0926 | 0.0548 |
| Prediff-PMM (k=8) | 0.0794 | 0.1429 | 0.6667 | 0.1064 | 0.0771 | 0.0548 | 0.1097 | 0.0805 | 0.0569 | 0.1708 | 0.1342 | 0.0919 |
| FlowCast | 0.0560 | 0.1031 | 0.3451 | 0.0765 | 0.0569 | 0.0345 | 0.0815 | 0.0605 | 0.0370 | 0.1582 | 0.1202 | 0.0754 |
| FlowCast-PMM (k=8) | 0.0871 | 0.1572 | 0.3720 | 0.1106 | 0.0889 | 0.0617 | 0.1144 | 0.0924 | 0.0644 | 0.1700 | 0.1404 | 0.0937 |
| Yushi | 0.1079 | 0.1921 | 0.7798 | 0.1610 | 0.1052 | 0.0575 | 0.1657 | 0.1109 | 0.0625 | 0.2492 | 0.1941 | 0.1269 |
| Yushi-PMM (k=8) | 0.1446 | 0.2502 | 0.7793 | 0.1972 | 0.1432 | 0.0933 | 0.2004 | 0.1482 | 0.0988 | 0.2627 | 0.2187 | 0.1578 |
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.
| Model | CSI-M | HSS-M | BIAS-M | CSI 1mm | CSI 2mm | CSI 4mm | CSI-4POOL 1mm | CSI-4POOL 2mm | CSI-4POOL 4mm | CSI-16POOL 1mm | CSI-16POOL 2mm | CSI-16POOL 4mm | CRPS | Fair-CRPS |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Yushi (K=1) | 0.1079 | 0.1921 | 0.7798 | 0.1610 | 0.1052 | 0.0575 | 0.1657 | 0.1109 | 0.0625 | 0.2492 | 0.1941 | 0.1269 | - | - |
| Yushi (K=2) | 0.1218 | 0.2146 | 0.7717 | 0.1742 | 0.1185 | 0.0726 | 0.1784 | 0.1241 | 0.0778 | 0.2539 | 0.2027 | 0.1414 | 0.0430 | 0.0330 |
| Yushi (K=4) | 0.1356 | 0.2364 | 0.7771 | 0.1871 | 0.1338 | 0.0859 | 0.1906 | 0.1389 | 0.0914 | 0.2583 | 0.2134 | 0.1535 | 0.0380 | 0.0330 |
| Yushi (K=8) | 0.1446 | 0.2502 | 0.7793 | 0.1972 | 0.1432 | 0.0933 | 0.2004 | 0.1482 | 0.0988 | 0.2627 | 0.2187 | 0.1578 | 0.0355 | 0.0330 |
| Yushi (K=12) | 0.1477 | 0.2549 | 0.7796 | 0.2008 | 0.1464 | 0.0961 | 0.2038 | 0.1512 | 0.1014 | 0.2641 | 0.2195 | 0.1585 | 0.0347 | 0.0330 |
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.
| Method | MSE 1mm | MSE 2mm | MSE 4mm | CORR 1mm | CORR 2mm | CORR 4mm |
|---|---|---|---|---|---|---|
| Bilinear interpolation from AWS | 2.4149 | 5.1681 | 12.8109 | 0.5540 | 0.4574 | 0.3489 |
| Kriging interpolation from AWS | 2.3667 | 5.1036 | 12.7588 | 0.5677 | 0.4646 | 0.3561 |
| MSMN | 1.9483 | 3.9702 | 9.0058 | 0.6543 | 0.5431 | 0.4143 |
| From Drops to Grid | 2.2324 | 5.1302 | 13.945 | 0.6575 | 0.5351 | 0.3689 |
| Z-R conversion from radar reflectivity | 2.9779 | 5.3865 | 10.2892 | 0.5720 | 0.5082 | 0.3805 |
| Yushi (decoding from GCR) | 2.0254 | 3.8527 | 8.3188 | 0.6720 | 0.5959 | 0.4570 |
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.
| Method | MSE 1mm | MSE 2mm | MSE 4mm | CORR 1mm | CORR 2mm | CORR 4mm |
|---|---|---|---|---|---|---|
| Yushi (decoding from GCR) w/o mask | 2.9779 | 4.1446 | 8.6644 | 0.5720 | 0.5761 | 0.4453 |
| Yushi (decoding from GCR) | 2.0254 | 3.8527 | 8.3188 | 0.6720 | 0.5959 | 0.4570 |
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.
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.
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.
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.
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.