From Nowcasting to Forecasting: Adapting a Reanalysis-Trained
Mikko Partio, Leila Hieta, Ossi Laine
cs.LG · physics.ao-ph
Quick verdict
CloudCast v2: reanalysis pretraining + flow matching extends satellite-initialized cloud forecasts to 12 h; honest ablations, single-region one-year test.
Summary
The paper introduces CloudCast v2, a Swin Transformer-based U-shaped model producing 12-hour cloud-cover forecasts on a 5 km grid over Scandinavia, initialized from satellite-derived NWCSAF effective cloudiness. Training proceeds in four stages: one-step and autoregressive pre-training on 1984–2020 CERRA reanalysis; deterministic fine-tuning on 2020–2024 NWCSAF data with MEPS NWP fields as dynamic forcings; and a final conditional flow matching (CFM) stage using a clean-target parameterization. On the held-out 2025 test year (n = 2,855 initializations), CloudCast v2 cuts mean absolute error by roughly 10% versus CloudCast v1 (0.213 vs 0.238) and overtakes v1 in fractions skill score after roughly 2–6 hours depending on cloudiness category. Notably, the deterministic no-CFM ablation outperforms the full model on MAE, ACC and MSESS, while CFM raises neighborhood-scale FSS in every category — a sharpness-versus-pixel-accuracy trade-off the authors report candidly. On the top-quintile dynamic subset (n = 570), both variants retain positive skill against Lagrangian persistence. Main limitations: a single domain and single test year, verification only against the satellite product that is also the training target, no significance tests or uncertainty quantification for metric differences, unspecified CFM sampling details, and unarchived NWCSAF-derived datasets, although code and most inputs are public.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- CloudCast v2 reduces mean absolute error by 10% relative to CloudCast v1 over the 1–12 h range (MAE 0.213 vs 0.238, Table 4).
- CloudCast v2 overtakes CloudCast v1 in fractions skill score after approximately 2–6 hours depending on cloudiness category (Section 3.3).
- The CFM adaptation stage accounts for most of CloudCast v2's long-lead FSS advantage, at the cost of slightly worse MAE, ACC and MSESS than the no-CFM ablation (Table 4).
- On the top-quintile dynamic subset (n = 570), CloudCast v2 retains positive MSESS against Lagrangian persistence throughout 12 h and beats CloudCast v1 in most cases (96.3% at t+5).
- Observation-initialized ML forecasts can usefully extend beyond the 1–3 h nowcasting range when conditioned on NWP atmospheric fields, complementing rather than replacing NWP.
Strengths
Unusually honest ablation reporting
Table 4 shows the no-CFM deterministic variant beats the flagship CFM model on MAE (0.203 vs 0.213), ACC at t+12 (0.589 vs 0.483) and MSESS at t+12 (0.600 vs 0.479); Section 3.5 and the Discussion explicitly frame CFM's benefit as FSS-only. This candor strengthens trust in the spatial-skill claims.
Multi-perspective verification design
Section 2.6/3 combines pixel-wise bias/MAE (Fig 5), ACC/MSESS with per-month-hour climatologies (Fig 6), category-wise FSS over 25–110 km neighborhoods (Fig 7), and a dynamic-subset skill score (top 20%, n = 570) against a purpose-built Färnebäck optical-flow Lagrangian persistence baseline — going well beyond single-metric claims.
Correct handling of the NWP variable mismatch
Transparent failure case
Compute and protocol transparency
Weaknesses
No statistical significance or uncertainty quantification
Single domain, single test year
Verification target equals the fine-tuning target
Generative sampler underspecified
Thin baseline coverage in the spatial metric
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
CloudCast v2 (Partio, Hieta & Laine, Finnish Meteorological Institute) targets the operational gap between satellite-based nowcasting — skillful for roughly 1–3 h — and NWP forecasts that may misrepresent the observed initial cloud state. The model produces 12-hour cloud-cover forecasts on a 5 km Scandinavia grid (535×475), initialized from NWCSAF effective cloudiness and conditioned on MEPS atmospheric fields. Central claims: (1) a 10% MAE reduction versus CloudCast v1 over 1–12 h; (2) FSS crossover ahead of v1 after roughly 2–6 h by category; (3) observation-initialized ML forecasting that retains satellite spatial detail beyond the nowcasting range. The claims are matched to evidence with unusual candor — the paper openly reports that its own no-CFM ablation beats the flagship model on most pixel-wise and anomaly metrics.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is CloudCast v2?
A machine-learning model from the Finnish Meteorological Institute that forecasts cloud cover up to 12 hours ahead over Scandinavia on a 5 km grid. It is initialized from satellite-derived cloud observations (NWCSAF effective cloudiness), pre-trained on the CERRA reanalysis to learn cloud evolution, and adapted to observations using conditional flow matching, with MEPS numerical weather prediction fields supplied as inputs at each forecast lead time.
How much better is CloudCast v2 than its predecessor CloudCast v1?
On a held-out 2025 test year it cuts mean absolute error by about 10% versus CloudCast v1 (0.213 vs 0.238 averaged over 1–12 h lead times) and exceeds v1's spatial fractions skill score after roughly 2–6 hours depending on cloudiness. Notably, an internal ablation without the flow-matching stage scores even better on pixel-wise and anomaly metrics; flow matching mainly improves spatial structure sharpness.
Can CloudCast v2 replace numerical weather prediction?
No. It depends on MEPS NWP forecasts for its atmospheric input fields, so it cannot issue a forecast before those are available. The authors position it as a fast add-on (about 90 seconds on one H100 GPU) that blends the latest satellite-observed cloud field with existing NWP guidance, not as a replacement for the NWP production cycle.
Is the code or data for CloudCast v2 available?
Source code is public on GitHub (fmidev/cloudcast). CERRA reanalysis and MEPS forecast archives are openly available, but the NWCSAF effective-cloudiness data and the processed training/evaluation datasets are not publicly archived due to volume and distribution constraints, and trained model weights are not mentioned as released.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are extracted only as axis labels and captions; curve values (exact FSS/ACC/MSESS trajectories in Figs 5–8 and the case-study imagery in Figs 9–10) could not be read, so figure-level judgments rest on the text and Table 4 numbers.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.