Phase 2 · does the weather predict nocturnal migration?

A boosted-tree model (LightGBM) estimating the nightly bird density (VID, birds/km²) of each radar from the weather of the night (wind and temperature on the 925, 850 and 700 hectopascal pressure levels, that is at roughly 750, 1,500 and 3,000 m above sea level, which is where the birds fly, plus the 10 and 100 m wind, temperature, humidity, pressure, precipitation and cloud cover at the surface; archive of operational analyses and forecasts from Open-Meteo, available from 2021), the day of the year, the radar's local climatology and its position. The wind is split into a tailwind and a crosswind component relative to the season's migratory heading, and the 24 h changes of pressure and temperature, which mark frontal passages, are added. The target is cube-root transformed. Only nights with coverage ≥ 60 % inside the migration windows (15 Feb-31 May, 15 Aug-30 Nov).

Two honest validations: (1) leave a whole year out and predict it with the rest (does it work in a new year?); (2) leave a whole radar out (does it work in a place it was not trained on? — that is the case of a city with no radar nearby, and also that of the new Spanish radars, which only have one season).

Every fold only sees what would have been available. The radar's local climatology is computed from its own observations, so handing it to the model as is would leak the answer. Leaving a year out, it is recomputed with the remaining years (moving window of ±7 days); leaving a radar out it disappears from the features, because in a city without a radar it does not exist. That is why the per-radar figures are lower than the annual ones: the model is left with the weather, the calendar and the geographic position alone.

How an alert is decided. A heavy-passage night is one above the local historical 90th percentile of its season, that is one night in ten. A mean model shrinks the predictions towards the centre and almost never crosses that absolute value, so the alert is decided by a second model, trained to classify that event directly, and the threshold is calibrated on the distribution of its predictions: the alert fires in the top decile of the forecast nights of that radar and season. In operation that distribution comes from running the model over ten years of reanalysis at the point of interest, with no need for a radar in the city. Since as many alerts are issued as there are heavy-passage nights, the captured percentage is directly comparable with the 10 % that chance would give. The area under the curve summarises the ability to separate those nights without depending on any threshold (0.5 = chance).

What the model uses

Relative gain of each feature training on everything, that is in the scenario with the local climatology available.

Radar-by-radar validation

Observed against predicted

Examples: the new Spanish radars and Porto

How much does the wind at flight altitude add?

The same nights and the same folds, changing only the feature set: with the wind at 925, 850 and 700 hectopascals (roughly 750, 1,500 and 3,000 m above sea level) against only the 10 and 100 m wind. The row that matters for the service is «radar out»: a city where there is no radar to train with.

What this means for the pilot

• The four renewed Spanish radars are the hard case. Madrid, Barcelona, Cáceres and Málaga give an area under the curve of 0.55-0.67 and capture 8%-21% of the heavy-passage nights. They beat chance (10 %) by little more than a factor of two, and with a single season of data every alert rests on very few nights. That is not yet enough for a reliable alert in those cities.
• The best results are from the Azores (area under the curve 0.85-0.97, 57%-79% captured): oceanic islands where the passage is very episodic and depends almost only on the weather. They must be read with caution, because they are few nights and do not represent the continent.
• On the continent the median area under the curve is 0.76 and the best radars reach 0.88 with 56% captured. There the service does work: they are almost all French, plus Porto.
• The operational conclusion is that the alert works where the radar sees well and there are several seasons of archive. In Spain it is better to wait for the 2027 season and, above all, to resolve the profile truncated at six layers and the insect contamination before publishing alerts.

Tables

Validation

Feature importance

Limitations

• The wind in the flight layer comes from the archive of operational forecasts, which only goes back to 2021. That shortens the usable archive and leaves out the first years of the French series, which are the long ones. The comparison with the surface model measures what is gained in exchange.
• Without the speed from vol2bird, insects cannot be separated from birds by airspeed; in autumn and in the south part of the VID is insect. The local climatology absorbs part of the bias, not all of it.
• Only the 49 radars with at least 100 nights get their own validation; the rest contribute to the training but are not validated separately (6 radars do not reach 20 nights).
• The «R² climatology» column of the per-radar folds compares against the observed climatology of that same radar, which the model could not use: it is a deliberately generous benchmark.
• Years with few radars give a very negative R² (the model gets the ranking of the nights right but not the absolute level of a radar it has barely seen). Spearman and the area under the curve are the metrics to trust; the R² only makes sense next to the climatology one, in the same column.
• The prediction uses the analysis of the night itself, not a forecast several days ahead: it is the upper bound of what will be achieved in operation.