- Rating
- 3(+1)
- Confidence
- 2(−1)
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes Earth-JEPA, a multi-resolution Joint-Embedding Predictive Architecture that fuses FY-4B infrared brightness temperatures, ERA5, and static geography into a 4-scale latent pyramid, using full attention at the coarsest scale and axial attention at finer scales to forecast future latent states supervised by an EMA target encoder via a State-Transition Loss.
Critical strengths
The problem is well motivated and the reporting is candid, evaluating across regional nowcasting, global medium-range forecasting, and cross-sensor transfer while honestly showing in Table 3 that Pangu-Weather and GraphCast reach 0.979 and 0.984 in 24-hour U850 ACC compared to 0.888 for E-JEPA, and noting that E-JEPA trails RainPro-2R on 5 of 6 SEVIR metrics. The low-cost training recipe and public benchmark have practical value, providing a backbone pretrained in 29.7 GPU-hours and a Drop-ERA5 setting where RMSE rises by only 4.1%, which is useful when reanalysis data is unavailable in real time.
Critical weaknesses
The headline accuracy and efficiency claim rests on a mischaracterized baseline and mixed compute accounting, as ClimaX's 0.754 ACC comes from an adapted regime of 10,000 rows, 40 epochs, and a 1.5-degree grid while WeatherGFM at 0.508 falls below persistence at 0.524, indicating poorly tuned baselines; moreover, the 12.8% compute figure counts only E-JEPA's 29.7 GPU-hour pretraining cost whereas the 0.888 ACC comes from the fine-tuned model costing 138.2 GPU-hours, which is 59.5% of ClimaX's 232.3 GPU-hours, while the frozen model matching 29.7 GPU-hours achieves 0.829 ACC for a 9.9% gain.
The transfer comparisons lack controls needed to attribute gains to pretraining, omitting a same-architecture supervised-from-scratch baseline on WeatherBench2 and failing to explain how a backbone pretrained on 4 km regional patches is adapted across a 40x scale mismatch to a global 1.5-degree grid.
The ablations do not support the claimed role of the design components, as the Supervised control in Table 5 is more accurate with a TBB RMSE of 5.32 K versus 5.39 K and ERA5 nRMSE of 0.0886 versus 0.0904, while tube masking, FPN, and axial attention each change probe error by less than 0.01%, dropping the State-Transition Loss changes TBB RMSE by 0.06%, and removing EMA raises TBB RMSE by only 0.75% despite a 92% drop in effective rank.
Questions for authors
Please clarify for every baseline in Tables 1 to 3 what data it was pretrained on and whether it is frozen or fine-tuned, and report WeatherBench2 results for E-JEPA's architecture trained from scratch alongside fine-tuned E-JEPA versus baselines under the same total 138.2 GPU-hours.
Given that ERA5 arrives with delay in operations, how do the main results change under realistic latency, and since ERA5 is hourly, how are the +15, +30, and +45 minute future targets constructed and is surface-pressure forecasting on interpolated reanalysis meaningful?
How are the 121x240 global 1.5-degree WeatherBench2 fields fed into encoders pretrained on 128x128 4 km regional patches in terms of patch size, geography conditioning, and positional handling?
6 comments
Response to Reviewer n4Hc (Part 2 / 3)
W3: The ablations do not support the claimed role of the design components, as the Supervised control in Table 5 is more accurate with a TBB RMSE of 5.32 K versus 5.39 K and ERA5 nRMSE of 0.0886 versus 0.0904, while tube masking, FPN, and axial attention each change probe error by less than 0.01%, dropping the State-Transition Loss changes TBB RMSE by 0.06%, and removing EMA raises TBB RMSE by only 0.75% despite a 92% drop in effective rank.
A3: Thank you for this sharp observation. We agree that effective rank and short-horizon regional probes alone do not prove downstream forecasting gains. In Section 4.4, Table 5, and Appendix Tables 15–17, we separated the regional reconstruction control from the WeatherBench2 target-task control and added downstream WeatherBench2 (24-h and 72-h) objective ablations:
Table R3-1: WeatherBench2 objective ablations and target-task supervision.
| Variant or control | 24-h nRMSE ↓ | 72-h nRMSE ↓ | Mean nRMSE ↓ | 24-h mean ACC ↑ | 72-h mean ACC ↑ |
|---|---|---|---|---|---|
| w/o EMA | 0.1550 | 0.2669 | 0.21095 | 0.8074 | 0.4016 |
| w/o state-transition loss | 0.1156 | 0.2114 | 0.16345 | 0.8961 | 0.6497 |
| Supervised (from scratch) | 0.0983 | 0.1805 | 0.13936 | 0.9251 | 0.7524 |
| Earth-JEPA (frozen) | 0.1027 | 0.1898 | 0.14623 | 0.9192 | 0.7245 |
- Clear downstream gains: While uncollapsed fine scales compensate over 60-min regional horizons, 72-h global rollouts rely on synoptic-scale dynamics: removing State-Transition Loss increases WeatherBench2 mean nRMSE by about 12% (), and removing EMA increases it by about 44% ().
- Component & control roles: Masking, FPN, and axial attention provide missing-data robustness, cross-scale fusion, and efficiency at no cost to representation quality (Section 4.4). Meanwhile, both the regional reconstruction control (5.32 K vs. 5.39 K) and the WB2 supervised control confirm that frozen Earth-JEPA approaches end-to-end target supervision while learning an 18.13x more compact, task-reusable latent target.
Q1: Please clarify for every baseline in Tables 1 to 3 what data it was pretrained on and whether it is frozen or fine-tuned, and report WeatherBench2 results for E-JEPA's architecture trained from scratch alongside fine-tuned E-JEPA versus baselines under the same total 138.2 GPU-hours.
A4: Thank you for requesting these provenance and control details. We have clarified every baseline in Appendix C.3 and Appendix D.2: ClimaX (CMIP6/ERA5 pretraining) and WeatherGFM (multimodal weather pretraining) adapt frozen backbones; Earthformer is pretrained on EarthCastBench and adapted frozen; Earthformer Sup. and CasCast Sup. are task-supervised on precipitation; Pangu-Weather and GraphCast are trained end-to-end on global 0.25° ERA5. Across baselines and controls, training follows a unified 30-epoch budget with 5-epoch validation early stopping, and Table 12 reports the actual measured GPU-hours:
Table R3-2: WeatherBench2 controls and baselines with complete GPU-hour costs.
| Method | Pretraining | Adaptation | Total | 24-h U850 ACC ↑ | 72-h U850 ACC ↑ |
|---|---|---|---|---|---|
| ClimaX (frozen) | 232.3 | 26.6 | 258.9 | 0.754 | 0.308 |
| Earthformer (frozen) | 72.2 | 31.2 | 103.4 | 0.765 | 0.329 |
| WeatherGFM (frozen) | 192.9 | 0.3 | 193.2 | 0.766 | 0.346 |
| E-JEPA (supervised, from scratch) | -- | 32.6 | 32.6 | 0.9219 | 0.7414 |
| E-JEPA (frozen) | 29.7 | 29.3 | 59.0 | 0.9119 | 0.7064 |
| E-JEPA (fine-tuned) | 29.7 | 40.2 | 69.9 | 0.916 | 0.719 |
Frozen E-JEPA uses 59.0 GPU-hours in total (and fine-tuned 69.9 GPU-hours, well below the earlier budget), while the pretrained backbone (29.7 h) can be reused across tasks.
Response to Reviewer n4Hc (Part 3 / 3)
Q2: Given that ERA5 arrives with delay in operations, how do the main results change under realistic latency, and since ERA5 is hourly, how are the +15, +30, and +45 minute future targets constructed and is surface-pressure forecasting on interpolated reanalysis meaningful?
A5: Thank you for raising this practical question. We have summarized the Drop-ERA5 comparison in Section 5 and added the full comparison in Appendix B (Table 10), alongside the alignment details in Appendix B.2:
Table R3-3: Effect of ERA5 input availability on EarthCastBench (matching Appendix Table 10).
| Method | Rain RMSE ↓ | Rain MAE ↓ | CSI ↑ | MSL RMSE ↓ | MSL MAE ↓ | SP RMSE ↓ | SP MAE ↓ |
|---|---|---|---|---|---|---|---|
| ClimaX | 0.8082 | 0.2651 | 0.7571 | 0.5785 | 0.4370 | 0.5571 | 0.4287 |
| Earthformer | 0.8169 | 0.2729 | 0.7515 | 0.5672 | 0.4383 | 0.5578 | 0.4365 |
| WeatherGFM | 0.7534 | 0.2515 | 0.7693 | 0.5687 | 0.4298 | 0.5449 | 0.4203 |
| E-JEPA (Drop-ERA5) | 0.7179 | 0.2427 | 0.7676 | 0.5502 | 0.4157 | 0.5365 | 0.4121 |
| E-JEPA | 0.6776 | 0.2250 | 0.7818 | 0.5040 | 0.3796 | 0.5003 | 0.3800 |
Under Drop-ERA5 (all ERA5 inputs removed at inference), Earth-JEPA stays close to its full-input performance and still achieves lower precipitation and pressure RMSE than every full-input baseline. Sub-hourly targets are linearly interpolated between adjacent hourly ERA5 analyses (Appendix B.2), applied identically across all models to evaluate agreement with aligned reanalysis references.
Q3: How are the 121x240 global 1.5-degree WeatherBench2 fields fed into encoders pretrained on 128x128 4 km regional patches in terms of patch size, geography conditioning, and positional handling?
A6: Thank you for asking for these implementation details. As detailed in Section 4.2 and Appendix C.3, each 121x240 global field is bilinearly resampled to the 128x128 backbone grid. The 4x4 patch embedding, 8x8 Swin windows, and relative-position bias tables remain unchanged. Global orography, latitude, and longitude fill the geographic slots (with unavailable land-cover, solar-zenith, and 9 TBB slots zeroed, while 4 time encodings are retained). The decoder applies circular longitude padding and replicated polar padding back to 121x240.
Additional Clarification & Core Advantages
"...honestly showing in Table 3 that Pangu-Weather and GraphCast reach 0.979 and 0.984 in 24-hour U850 ACC compared to 0.888 for E-JEPA, and noting that E-JEPA trails RainPro-2R on 5 of 6 SEVIR metrics."
We report these candidly to position Earth-JEPA as a resource-efficient, reusable foundation model trained on 8x A800 GPUs (3.40 TB archive) rather than a native 0.25° global supervised model (where Pangu/GraphCast use ~961x/321x our cost). On SEVIR, after unifying frozen transfer, E-JEPA outperforms RainPro-2R by about 12% on CSI-p16 and occupies the top two positions on 5 of 6 deterministic metrics (Table 4).
Table R3-4: Summary of core advantages and supporting evidence.
| Core Advantage | Key Result | Location |
|---|---|---|
| Efficiency | WB2 cost: 29.7 + 29.3 GPU-h (E-JEPA pretrain + frozen adaptation), 0.10% of Pangu-Weather’s A800-equivalent cost | Fig. 1, Table 6 |
| Transferability | SEVIR performance: beats previous supervised SOTA RainPro-2R (ICLR 2026) on 5/6 metrics, with +12.0% CSI-p16 | Table 4, Sec. 4.3 |
| Accuracy | WB2 performance vs. ClimaX: +47% mean ACC, 39% lower mean MAE, 38% lower mean RMSE | Tables 3, 13, 14 |
| Scalability (preliminary) | WB2 72-h mean ACC: 0.678 → 0.725 → 0.749 (50M → 113M → 239M, frozen E-JEPA) | Table 3, Fig. 5 |
We have uploaded the anonymous supplementary material with our model definitions and evaluation code, and will release the full training code, checkpoints, and EarthCastBench upon acceptance. Given the unified cost-accuracy pairing, new supervised experiments, and downstream ablations, we respectfully hope our endeavors can help inform your reassessment of our work.
Response to Reviewer n4Hc (Part 1 / 3)
Thank you for the thorough and constructive review. We appreciate your recognition of the motivation, the low-cost training recipe, and the practical value of EarthCastBench.
In the revised manuscript, we corrected the cost–accuracy comparison and the WeatherGFM evaluation, added same-architecture supervised controls on WeatherBench2 and SEVIR, and clarified how we adapt a regionally pretrained backbone to global forecasting. We address each point below.
W1: The headline accuracy and efficiency claim rests on a mischaracterized baseline and mixed compute accounting, as ClimaX's 0.754 ACC comes from an adapted regime of 10,000 rows, 40 epochs, and a 1.5-degree grid while WeatherGFM at 0.508 falls below persistence at 0.524, indicating poorly tuned baselines; moreover, the 12.8% compute figure counts only E-JEPA's 29.7 GPU-hour pretraining cost whereas the 0.888 ACC comes from the fine-tuned model costing 138.2 GPU-hours, which is 59.5% of ClimaX's 232.3 GPU-hours, while the frozen model matching 29.7 GPU-hours achieves 0.829 ACC for a 9.9% gain.
A1: Thank you for this careful audit. We have corrected both the compute pairing and the WeatherGFM evaluation in the Abstract, Introduction, Figure 1, Section 4.2, Table 3, and Table 6:
W2: The transfer comparisons lack controls needed to attribute gains to pretraining, omitting a same-architecture supervised-from-scratch baseline on WeatherBench2 and failing to explain how a backbone pretrained on 4 km regional patches is adapted across a 40x scale mismatch to a global 1.5-degree grid.
A2: Thank you for pointing out these missing controls and interface details. We have added same-architecture supervised-from-scratch controls on both WeatherBench2 (Table 3) and SEVIR (Table 4), and detailed the regional-to-global interface in Section 4.2 and Appendix C.3: