The experiment in one sentence
We trained on 2018, chose alert thresholds on 2019, and evaluated on held-out 2020 data—first using local weather, then adding three regional dust clues.
Held-out 2020 development results
| Metric | Local weather | + regional dust |
|---|---|---|
| F₂ | 0.4995 | 0.5213 |
| PR-AUC | 0.3481 | 0.3886 |
| Brier (lower is better) | 0.0979 | 0.0925 |
| Recall | 71.2% | 76.5% |
What does F₂ = 0.521 mean in real terms?
F₂ is a summary that rewards catching real dust onsets more than avoiding extra alerts. Here is the result without the statistical shorthand:
The improved model caught 21 additional onsets (303 instead of 282), while issuing 83 additional alerts. F₂ considers that a useful trade because the project gives extra importance to not missing dust.
What could that improvement mean at a larger scale?
The observed recall improvement is equivalent to approximately 5.3 additional station-onset detections per 100 actual onsets. Move the slider to explore an illustrative network.
This is scaling, not a forecast of how many unique storms occur. One regional storm may be recorded at several stations.
Saudi context: one long-term study reported a maximum of 203 station-recorded dust events across 27 Saudi stations in 2003. At our observed improvement, a network of that scale would correspond to roughly 11 additional detected station events. Definitions, stations, and years differ, so this is an illustration—not a national performance estimate. Source study.
Global context: the World Meteorological Organization reports that sand and dust storms affect about 330 million people in more than 150 countries, but it does not provide one standardized annual count of comparable storms. For that reason, the calculator scales by hypothetical station-level onsets rather than claiming a global number “saved.” WMO source.
Try a weather scenario
Unshown inputs remain at typical values. This illustrates model behavior; it is not a live warning.
Estimated onset probability
–
–
What should we conclude?
Regional reports improve ranking, recall, and probability error. That supports the idea that nearby dust contains information local weather misses. It does not prove the exact path of a plume or make this an operational forecast.
2020 was excluded from fitting and threshold selection, but later data-source comparisons also examined 2020. Treat these as development results requiring confirmation on a new period or region.
Probability challenge
Imagine you are issuing tomorrow's dust outlook. Read the plain-language situation and guess the percentage chance of a new dust onset. Then compare your answer with the weather-only model and the improved regional model.
You are guessing the model's estimated probability—not a guaranteed real-world frequency.
Your estimate
It sees local wind, humidity, rain and season—but no nearby dust reports.
It also sees whether dust exists nearby and whether its position agrees with the wind.
Why can the answers differ?
The older model asks mainly, “Could today's local weather support dust?” The improved model also asks, “Is dust already present in the surrounding region?” A windy day without regional dust and a moderately windy day with a nearby plume can therefore receive different estimates.