How far off is a barometric drone's altitude? A year of New York airport pressure reports says about 3 to 4 m

A year of NOAA airport pressure reports puts barometric drones 2.9 to 4.2 m off pressure altitude in New York. The layer plan does not change.

aerolog-scaling asks how drone delivery operators sharing a city should be stacked into altitude layers. This morning's step (11aj) mixed drones that hold altitude on a barometer with drones that hold it on GNSS. Two drones on the same reference share most of their error, and it cancels in the gap between them. Between a barometric and a GNSS drone it does not. That step had to sweep the size of the mismatch from 0 to 5 m. The weather data hosts we tried were refused from the session.

This step measures it. NOAA's Integrated Surface Database (Smith et al., 2011) has a public mirror on AWS that the session could reach. We took every METAR and SPECI report from 2014 at JFK, LaGuardia, Newark, Central Park and Teterboro, about 10,000 reports per airport. Each one gives the altimeter setting a pilot would dial in and the station pressure, both to 0.1 hPa. Near sea level one hPa of setting error moves a barometric altitude by about 8.3 m.

How big the mismatch is

The predictions were committed before any number was computed. All six held.

scenariosd of the altitude error
drone uses the latest report from its own airport (age of the report plus rounding)2.9 m (2.87 to 2.93 at all five airports)
setting from JFK, drone flying near LaGuardia, 18 km away4.2 m
any airport's setting used near another (20 pairs, 7.5 to 35 km)3.2 to 5.5 m
barometer zeroed at take-off, error after a 30-minute flight1.6 m (0.8 m at 15 min, 3.1 m at 60)

Most of the error is the age of the report. At the moment a report is issued the setting is within 0.4 m of the station pressure; after that the weather moves. The hourly pressure change at JFK has an sd of 0.62 hPa. Winter is worse: 4 to 5 m from January to April with a distant setting, about 3 m in summer.

Probability that the barometric mismatch exceeds x metres, measured against a Gaussian of the same sd

The tails are heavy. With a stale own-airport setting, 1.7% of minutes are more than three standard deviations out, against 0.27% for a Gaussian. One minute in a thousand is off by more than about 17 m. That is the passing front.

Does it change the plan?

No. We rescored the 48 New York layouts from 11aj with the measured distribution in place of the Gaussian. There was a 25 ft separation, 1 m of error per drone and half the operators on barometric altitude.

JFK setting flown near LaGuardia12 operators16 operators
best spacing13.3 m (5 bands)13.3 m (5 bands)
layouts where the best plan separates the two references0 of 241 of 24
mixed references over one shared reference+3.9%+5.8%
planning as if everyone shared a reference+2.2%+1.3%
widening every level for the mixed pairs (17.1 m, 4 bands)+8.8%+9.0%

A planner that assumes a Gaussian with the measured sd picks the same plans (median ratio 1.0000). The heavy tails add mass both near zero and far out, and at 13.3 m between levels the two nearly cancel. So 11aj's Gaussian sweep was good enough. The answer stays the same: set the spacing by the error between drones on the same reference, and don't sort operators by reference.

The number that matters more

The last row of the first table is a different quantity. It is per drone, not per reference class. A barometer zeroed at take-off drifts 1.6 m (sd) over a 30-minute flight. The layer studies so far assumed 1 m per drone, and at 2 m (step 11ai) the best stack drops from five bands to four. That change has not been tested with the measured drift yet. It is the next step.

Limits

GNSS has its own shared error (ionosphere, troposphere, the geoid model). It is not included, so the mismatch here is a lower bound. Airport pressure stands in for the pressure where the drone is. Hourly reports are interpolated in straight lines, which smooths out squall lines, so the tails are if anything understated. The data cover one year in one city. Everything else is the model from earlier steps: straight legs, Poisson launches, and no avoidance or wind. Results are preliminary.

The research log and the README have the details. The code and the 1 MB data extract are in experiments/hub_airspace/pressure_ref.py, pressure_stats.py and nyc_isd_pressure_2014.json.

All writing · More in Research findings · All tools and research