- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper combines automated airport observations, aviation weather reports, regional numerical forecast fields, and airport metadata. It defines precipitation and thunderstorm-state and flight-condition classification at several forecast horizons, then evaluates a reference model alongside learned and operational forecast baselines using overall classification, rare-event ranking, and categorical decision measures.
Critical strengths
Separating rare-event ranking from decisions provides a useful experimental view. In the main results table, some learned models retain a Thunder ranking signal but never predict Thunder as their final class, while the operational forecast baseline predicts adverse events at a markedly different frequency. Reporting these behaviors together reveals differences that an aggregate classification measure alone would hide.
Critical weaknesses
In “Anchor Construction and Causal Alignment,” a GRIB record is eligible when its initialization time is no later than the forecast anchor; the adjoining data description calls this nominal-cycle causal selection. Initialization time need not coincide with the time at which the forecast file was actually published or received, so this rule alone does not establish that the selected field was available at the anchor. For TAF, by contrast, the paper explicitly uses the later of issue and archive-arrival times to establish availability. If a GRIB field became available after the anchor, results using that modality could incorporate information unavailable when the prediction was made. The present description does not rule this out, leaving the benchmark's forecast-time causality and the associated performance interpretation uncertain.
In “Input Adaptation for Learnable Baselines,” the neural time-series baselines reduce each GRIB channel around an airport to a spatial mean, and the appendix explicitly notes that the grid's internal spatial arrangement is not retained. MISTNet instead keeps spatial grid tokens and aggregates them conditionally. The ablations show that regional input and fusion within MISTNet matter, but they cannot distinguish whether the gap in the main comparison comes from the fusion design or from spatial information discarded before the baselines receive it. Claims about the reference model's advantage over other learned methods therefore need to be confined to the present input adaptations.
Questions for authors
Could the authors clarify how the actual publication or archive-arrival times of the GRIB fields relate to each forecast anchor, and verify whether the causal protocol and main results hold when records are selected by actual availability time?
Could the authors provide a learned baseline that retains the GRIB spatial arrangement under comparable input and training conditions, or clarify how the main model comparison should be interpreted given the current spatial-mean adaptation?