- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper trains a conditional diffusion generator on paired reanalysis atmospheric states and satellite precipitation estimates, then conditions it on forecasts from frozen weather models to produce daily precipitation. A multiscale recurrent encoder represents atmospheric evolution, and cross-attention guides denoising. The evaluation covers global and regional forecasts, architecture comparisons, and a station-based rainfall case study.
Critical strengths
Separating atmospheric forecasting from precipitation supervision has clear practical value. In “Cross-Backbone Reuse and Fine-Tuning Diagnostic,” the paper states that different forecast models at the same resolution use the same generator checkpoint and variable interface; the main evaluation also includes both data-driven and numerical forecasts. This supports the claimed reusability more directly than an improvement shown with a particular forecasting backbone.
Critical weaknesses
The method motivates cross-attention through a “nonlocal spatiotemporal alignment challenge” and argues that precipitation features can selectively attend to atmospheric conditions. In the configuration comparison, attention fusion outperforms gated fusion, and the complete model outperforms variants with modules removed. These results establish that the design is useful, but changing the fusion operator also changes how features interact and how flexibly the model can represent their relation. The comparison does not establish whether the alternative has comparable cross-location interaction ability, and it does not examine whether attention actually uses upstream or surrounding atmospheric locations. The performance difference therefore cannot distinguish the proposed spatial alignment mechanism from the benefit of a more flexible fusion operator, limiting attribution of a central contribution.
The station case study uses a heavy-rainfall event to support improved local accumulations and event detection, and its figure displays local maxima. Yet the reported critical success indices cover only relatively low rainfall thresholds. The appendix adds station maps at other times within the same event but does not report intensity errors in the high-rainfall range or missed and false heavy-rainfall events. Better detection at low thresholds could result from more accurate rain-area coverage even if the magnitudes of local peaks remain biased. Thus, the station evidence does not directly establish how much the method improves heavy rainfall accumulation, and visually sharper peaks should not be equated with more accurate heavy-rainfall forecasts.
Questions for authors
Could the authors provide a fusion comparison with comparable cross-location interaction range and model capacity, or analyze which atmospheric locations the attention actually uses, to clarify how much of the gain comes from the proposed spatial alignment mechanism?
Could the authors use the existing station observations to report intensity errors in the high-rainfall range and missed and false heavy-rainfall events, so that the apparent improvement in visualized peaks can be checked against actual heavy-rainfall forecast skill?
7 comments
Response for Reviewer fF5r (2/3)
Table R2b. Gridded CSI at 50 mm.
| Model | Day 1 | Day 2 | Day 3 | Day 4 | Day 5 | Day 6 | Day 7 | Day 8 | Day 9 | Day 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| GraphCast | 0.2408 | 0.1744 | 0.1433 | 0.1210 | 0.1000 | 0.0877 | 0.0755 | 0.0660 | 0.0578 | 0.0519 |
| HRES | 0.2213 | 0.1762 | 0.1551 | 0.1389 | 0.1260 | 0.1128 | 0.1016 | 0.0868 | 0.0774 | 0.0692 |
| GenCast | 0.2351 | 0.1642 | 0.1589 | 0.1464 | 0.1385 | 0.1274 | 0.1143 | 0.0932 | 0.0869 | 0.0542 |
| ENS Mean | 0.1748 | 0.1021 | 0.0695 | 0.0452 | 0.0251 | 0.0121 | 0.0059 | 0.0029 | 0.0017 | 0.0010 |
| GraphCast+Ours | 0.3432 | 0.2943 | 0.2480 | 0.2060 | 0.1756 | 0.1552 | 0.1332 | 0.1173 | 0.1058 | 0.0941 |
| HRES+Ours | 0.3178 | 0.2783 | 0.2563 | 0.2364 | 0.2152 | 0.1948 | 0.1736 | 0.1555 | 0.1410 | 0.1291 |
| Pangu+Ours | 0.3073 | 0.2442 | 0.2192 | 0.1968 | 0.1750 | 0.1457 | 0.1205 | 0.1006 | 0.0854 | 0.0746 |
Table R2c. Gridded FSS at 25 mm. Higher is better.
| Model | Day 1 | Day 2 | Day 3 | Day 4 | Day 5 | Day 6 | Day 7 | Day 8 | Day 9 | Day 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| GraphCast | 0.7303 | 0.6586 | 0.6066 | 0.5518 | 0.4958 | 0.4415 | 0.3885 | 0.3435 | 0.3042 | 0.2695 |
| HRES | 0.7387 | 0.6715 | 0.6279 | 0.5806 | 0.5174 | 0.4437 | 0.3723 | 0.3068 | 0.2635 | 0.2239 |
| ENS Mean | 0.7382 | 0.6436 | 0.5764 | 0.5002 | 0.4191 | 0.3310 | 0.2435 | 0.1734 | 0.1226 | 0.0820 |
| GraphCast+Ours | 0.7975 | 0.7380 | 0.6664 | 0.6111 | 0.5649 | 0.5108 | 0.4571 | 0.4043 | 0.3623 | 0.3270 |
| HRES+Ours | 0.7656 | 0.7155 | 0.6756 | 0.6332 | 0.5877 | 0.5355 | 0.4794 | 0.4250 | 0.3777 | 0.3397 |
| Pangu+Ours | 0.7568 | 0.7007 | 0.6562 | 0.5901 | 0.5260 | 0.4755 | 0.4140 | 0.3648 | 0.3172 | 0.2776 |
Table R2d. Gridded FSS at 50 mm.
| Model | Day 1 | Day 2 | Day 3 | Day 4 | Day 5 | Day 6 | Day 7 | Day 8 | Day 9 | Day 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| GraphCast | 0.5389 | 0.4284 | 0.3616 | 0.3085 | 0.2508 | 0.2156 | 0.1730 | 0.1518 | 0.1275 | 0.1065 |
| HRES | 0.5342 | 0.4561 | 0.4042 | 0.3623 | 0.3257 | 0.2827 | 0.2372 | 0.2007 | 0.1677 | 0.1418 |
| ENS Mean | 0.4664 | 0.3364 | 0.2486 | 0.1698 | 0.0946 | 0.0417 | 0.0149 | 0.0053 | 0.0027 | 0.0017 |
| GraphCast+Ours | 0.5975 | 0.4969 | 0.4407 | 0.3970 | 0.3502 | 0.2966 | 0.2470 | 0.2079 | 0.1780 | 0.1492 |
| HRES+Ours | 0.6050 | 0.5430 | 0.4994 | 0.4582 | 0.4145 | 0.3678 | 0.3159 | 0.2714 | 0.2328 | 0.2058 |
| Pangu+Ours | 0.6483 | 0.5735 | 0.4856 | 0.4036 | 0.3297 | 0.2633 | 0.2142 | 0.1641 | 0.1324 | 0.1059 |
Across the 2,288-station network, RMSE, CSI, and FSS improve at 25 mm across all ten lead times under GraphCast and HRES conditioning. (Tables R3a–R3e).
Table R3a. Station RMSE for station-based evaluation.
| Model | Day 1 | Day 2 | Day 3 | Day 4 | Day 5 | Day 6 | Day 7 | Day 8 | Day 9 | Day 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| GraphCast | 7.1325 | 7.2531 | 7.4342 | 7.7262 | 8.0099 | 8.3415 | 8.6037 | 8.8955 | 9.2586 | 9.6167 |
| HRES | 7.1702 | 7.4379 | 7.7896 | 8.2153 | 8.4028 | 8.8524 | 9.0474 | 9.2005 | 9.6900 | 9.9643 |
| GraphCast+Ours | 6.3910 | 6.5930 | 6.8636 | 7.2502 | 7.5642 | 7.7212 | 7.8475 | 8.2619 | 8.7115 | 9.1270 |
| HRES+Ours | 6.4641 | 6.9026 | 7.3608 | 7.8014 | 8.0678 | 7.8486 | 8.2241 | 8.4546 | 8.8972 | 9.7440 |
| Pangu+Ours | 6.4627 | 6.8088 | 7.1220 | 7.3774 | 7.5366 | 8.5163 | 8.7467 | 8.9275 | 9.3291 | 9.3852 |
Table R3b. Station CSI at 25 mm. Higher is better.
| Model | Day 1 | Day 2 | Day 3 | Day 4 | Day 5 | Day 6 | Day 7 | Day 8 | Day 9 | Day 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| GraphCast | 0.2298 | 0.1981 | 0.1915 | 0.1697 | 0.1530 | 0.1326 | 0.1223 | 0.1055 | 0.0809 | 0.0707 |
| HRES | 0.2215 | 0.2183 | 0.1988 | 0.1730 | 0.1529 | 0.1195 | 0.1074 | 0.1067 | 0.0852 | 0.0807 |
| GraphCast+Ours | 0.3074 | 0.2912 | 0.2597 | 0.2263 | 0.1972 | 0.1736 | 0.1438 | 0.1319 | 0.0971 | 0.0812 |
| HRES+Ours | 0.3153 | 0.2877 | 0.2561 | 0.2199 | 0.1945 | 0.1545 | 0.1370 | 0.1260 | 0.1017 | 0.0864 |
| Pangu+Ours | 0.2967 | 0.2608 | 0.2207 | 0.1757 | 0.1600 | 0.1308 | 0.1157 | 0.1041 | 0.0845 | 0.0680 |
Table R3c. Station CSI at 50 mm.
| Model | Day 1 | Day 2 | Day 3 | Day 4 | Day 5 | Day 6 | Day 7 | Day 8 | Day 9 | Day 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| GraphCast | 0.0844 | 0.0914 | 0.0844 | 0.0688 | 0.0605 | 0.0527 | 0.0407 | 0.0338 | 0.0302 | 0.0257 |
| HRES | 0.0807 | 0.0757 | 0.0740 | 0.0608 | 0.0404 | 0.0381 | 0.0267 | 0.0242 | 0.0231 | 0.0226 |
| GraphCast+Ours | 0.1123 | 0.1124 | 0.0961 | 0.0674 | 0.0630 | 0.0516 | 0.0506 | 0.0425 | 0.0284 | 0.0312 |
| HRES+Ours | 0.1004 | 0.0996 | 0.0796 | 0.0629 | 0.0618 | 0.0467 | 0.0372 | 0.0355 | 0.0329 | 0.0294 |
| Pangu+Ours | 0.0973 | 0.0990 | 0.0671 | 0.0639 | 0.0502 | 0.0390 | 0.0373 | 0.0332 | 0.0259 | 0.0275 |
Table R3d. Station FSS at 25 mm. Higher is better.
| Model | Day 1 | Day 2 | Day 3 | Day 4 | Day 5 | Day 6 | Day 7 | Day 8 | Day 9 | Day 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| GraphCast | 0.3737 | 0.3307 | 0.3214 | 0.2902 | 0.2654 | 0.2342 | 0.2179 | 0.1909 | 0.1497 | 0.1321 |
| HRES | 0.3627 | 0.3584 | 0.3317 | 0.2950 | 0.2652 | 0.2135 | 0.1940 | 0.1928 | 0.1570 | 0.1493 |
| GraphCast + Ours | 0.4702 | 0.4511 | 0.4123 | 0.3691 | 0.3294 | 0.2958 | 0.2514 | 0.2331 | 0.1770 | 0.1502 |
| HRES + Ours | 0.4794 | 0.4468 | 0.4078 | 0.3605 | 0.3257 | 0.2676 | 0.2410 | 0.2238 | 0.1846 | 0.1591 |
| Pangu + Ours | 0.4576 | 0.4137 | 0.3616 | 0.2989 | 0.2759 | 0.2313 | 0.2074 | 0.1886 | 0.1558 | 0.1273 |
Table R3e. Station FSS at 50 mm
| Model | Day 1 | Day 2 | Day 3 | Day 4 | Day 5 | Day 6 | Day 7 | Day 8 | Day 9 | Day 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| GraphCast | 0.1557 | 0.1675 | 0.1557 | 0.1287 | 0.1141 | 0.1001 | 0.0782 | 0.0654 | 0.0586 | 0.0501 |
| HRES | 0.1493 | 0.1407 | 0.1378 | 0.1146 | 0.0777 | 0.0734 | 0.0520 | 0.0473 | 0.0452 | 0.0442 |
| GraphCast + Ours | 0.2019 | 0.2021 | 0.1753 | 0.1263 | 0.1185 | 0.0981 | 0.0963 | 0.0815 | 0.0552 | 0.0605 |
| HRES + Ours | 0.1825 | 0.1812 | 0.1475 | 0.1184 | 0.1164 | 0.0892 | 0.0717 | 0.0686 | 0.0637 | 0.0571 |
| Pangu + Ours | 0.1773 | 0.1802 | 0.1258 | 0.1201 | 0.0956 | 0.0751 | 0.0719 | 0.0643 | 0.0505 | 0.0535 |
Response for Reviewer fF5r (3/3)
Tables R4a–R4d report the requested miss rates and false alarm ratios. At 25 mm/day, both decrease across all ten lead times for GraphCast and HRES.
Table R4a. Station Miss rate at 25 mm. Lower is better.
| Model | Day 1 | Day 2 | Day 3 | Day 4 | Day 5 | Day 6 | Day 7 | Day 8 | Day 9 | Day 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| GraphCast | 0.6960 | 0.7370 | 0.7333 | 0.7549 | 0.7774 | 0.7906 | 0.7909 | 0.8195 | 0.8669 | 0.8737 |
| HRES | 0.7083 | 0.6902 | 0.7087 | 0.7411 | 0.7660 | 0.8163 | 0.8284 | 0.8325 | 0.8621 | 0.8662 |
| GraphCast+Ours | 0.6080 | 0.6141 | 0.6549 | 0.6929 | 0.7278 | 0.7580 | 0.7906 | 0.8096 | 0.8576 | 0.8751 |
| HRES+Ours | 0.5822 | 0.5855 | 0.6175 | 0.6676 | 0.7027 | 0.7566 | 0.7836 | 0.8049 | 0.8384 | 0.8570 |
| Pangu+Ours | 0.5720 | 0.6189 | 0.6926 | 0.7642 | 0.7749 | 0.8143 | 0.8346 | 0.8468 | 0.8705 | 0.8960 |
Table R4b. Station False alarm ratio at 25 mm. Lower is better.
| Model | Day 1 | Day 2 | Day 3 | Day 4 | Day 5 | Day 6 | Day 7 | Day 8 | Day 9 | Day 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| GraphCast | 0.5151 | 0.5547 | 0.5957 | 0.6445 | 0.6714 | 0.7342 | 0.7726 | 0.7974 | 0.8288 | 0.8616 |
| HRES | 0.5205 | 0.5750 | 0.6150 | 0.6574 | 0.6937 | 0.7451 | 0.7768 | 0.7727 | 0.8176 | 0.8310 |
| GraphCast+Ours | 0.4126 | 0.4571 | 0.4879 | 0.5377 | 0.5830 | 0.6193 | 0.6854 | 0.6998 | 0.7663 | 0.8117 |
| HRES+Ours | 0.4377 | 0.5152 | 0.5633 | 0.6063 | 0.6401 | 0.7026 | 0.7281 | 0.7376 | 0.7846 | 0.8207 |
| Pangu+Ours | 0.5084 | 0.5474 | 0.5610 | 0.5919 | 0.6439 | 0.6936 | 0.7222 | 0.7546 | 0.8045 | 0.8358 |
Table R4c. Station Miss rate at 50 mm.
| Model | Day 1 | Day 2 | Day 3 | Day 4 | Day 5 | Day 6 | Day 7 | Day 8 | Day 9 | Day 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| GraphCast | 0.9019 | 0.8945 | 0.9013 | 0.9198 | 0.9275 | 0.9324 | 0.9497 | 0.9582 | 0.9599 | 0.9659 |
| HRES | 0.9077 | 0.9101 | 0.9048 | 0.9207 | 0.9515 | 0.9528 | 0.9679 | 0.9699 | 0.9696 | 0.9697 |
| GraphCast+Ours | 0.8777 | 0.8693 | 0.8789 | 0.9138 | 0.9173 | 0.9314 | 0.9260 | 0.9423 | 0.9641 | 0.9585 |
| HRES+Ours | 0.8898 | 0.8891 | 0.9107 | 0.9286 | 0.9238 | 0.9458 | 0.9550 | 0.9566 | 0.9577 | 0.9570 |
| Pangu+Ours | 0.8868 | 0.8789 | 0.9170 | 0.9189 | 0.9333 | 0.9478 | 0.9521 | 0.9566 | 0.9650 | 0.9597 |
Table R4d. Station False alarm ratio at 50 mm.
| Model | Day 1 | Day 2 | Day 3 | Day 4 | Day 5 | Day 6 | Day 7 | Day 8 | Day 9 | Day 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| GraphCast | 0.6242 | 0.5935 | 0.6327 | 0.6748 | 0.7320 | 0.8077 | 0.8246 | 0.8498 | 0.8903 | 0.9063 |
| HRES | 0.6081 | 0.6762 | 0.7511 | 0.7929 | 0.8047 | 0.8347 | 0.8636 | 0.8901 | 0.9120 | 0.9186 |
| GraphCast+Ours | 0.4225 | 0.5560 | 0.6826 | 0.7635 | 0.7907 | 0.8279 | 0.8624 | 0.8605 | 0.8793 | 0.8880 |
| HRES+Ours | 0.4700 | 0.5066 | 0.5769 | 0.6561 | 0.7542 | 0.7487 | 0.8249 | 0.8372 | 0.8705 | 0.9149 |
| Pangu+Ours | 0.5915 | 0.6478 | 0.7400 | 0.7682 | 0.8316 | 0.8659 | 0.8554 | 0.8757 | 0.9096 | 0.9205 |
Response for Reviewer fF5r (1/3)
We thank the reviewer for the constructive feedback and will incorporate the additional fusion ablations, attention visualizations, and gridded and station-based heavy-rainfall evaluations presented below into the revised manuscript.
W1 & Q1. Disentangling Spatial Alignment from Fusion Flexibility.
We added ablations to assess whether cross-attention provides greater gains than alternative fusion operators.
All three variants access the same module parameter size (12M) and the same global grid of 15 × 30 atmospheric tokens, providing comparable cross-location interaction coverage.
Across Days 1–10, ConvGRU + attention achieves higher CSI25 and FSS25 than both alternatives at every lead time under both backbones.
Table R1a. CSI at 25 mm/24 h, Days 1–10. Higher is better.
Table R1b. FSS at 25 mm/24 h, Days 1–10. Higher is better.
The comparison with global-mixing CNN supports input-dependent nonlocal fusion under comparable spatial coverage and parameter budgets. The comparison with channel-stacked CNN supports the effectiveness of ConvGRU-based temporal encoding.
W1 & Q1. Which atmospheric locations the attention actually uses?
To examine the spatial behavior of cross-attention, we visualize the atmospheric key locations attended to by rain-core queries in two precipitation events over South China and Pakistan in Figure 15 and 16 in revised manuscript. We will upload the revised revised manuscript as soon as possible.
W2 & Q2. Heavy-Rainfall Evaluation Using Gridded and Station Observations.
Thanks for the suggestion. We extend heavy-rainfall verification to 25 and 50 mm/day over Days 1–10, using gridded IMERG and the 2,288-station network.
In the gridded evaluation, our framework improves both CSI and FSS at every lead time under GraphCast, HRES, and Pangu conditioning (Tables R2a–R2d).
Table R2a. Gridded CSI at 25 mm. Higher is better.