Sonde Predict

What it does

Map, 3D and profile

The flight on an OpenStreetMap map, in 3D over the terrain, or as an altitude profile with the winds it met. Predictions close to you are shown in bold.

The best winds for the Nordic sky

Inside the Nordic area it averages MET Norway's 2.5 km MEPS model, on its own 65 levels, with ECMWF's global forecast — better than either alone. ECMWF covers the rest of the world.

Tested against real flights

Checked against FMI soundings and against 45 Jokioinen flights that hunters' receivers followed to the ground. The balloon defaults come from hundreds of real flights.

A landing area, not just a point

Fifty varied flights, each with a wind error as large as the forecasts' real errors, show where the sonde will come down. Checked on real flights: about 95 % land inside the 95 % area.

Every station you follow

Batches cover several launch sites with their own launch times and their own balloons, days ahead, and can repeat on their own while the program is open.

Out in the field

Export to GPS and phone apps as GPX, run it from the command line, and read what it is doing line by line in the activity log. It tells you when a new version is out.

A day of predictions on the map with landing areas — A day of launches on the map, each with its landing area.

The profile of three launches: the same climb, very different winds — a 41 m/s jet stream on the first.

The activity log while a batch runs — The activity log says which forecast it uses and what it downloads.

How close does it get?

We predicted 45 real flights from Jokioinen (8–22 October 2026) with the winds available at launch time and compared them with where SondeHub's receivers last heard each sonde.

Balloon settings	Typical landing error	Worst 10 %
Before: generic values	13.7 km	27 km
Now: the station's recent flights, measured ascent and descent	8.8 km	17 km
If each balloon's burst were known in advance	6.5 km	12 km

Balloons from one station vary (burst altitudes 31–36 km at Jokioinen), and the forecast is never perfect, so the landing area matters as much as the point. Other stations were checked the same way: Tallinn 5 km, Stavanger 6 km typical error; stations with slow parachute descents drift longer and are harder (Greifswald about 22 km), and their landing areas are larger to match. The full comparison — against soundings, by forecast lead time, per wind model and per station — is in the source.

95 %of real landings inside the 95 % landing area, once forecast errors are included

−13 %slower climb at 9–18 km, as real balloons do; now part of the flight model

22stations with descents measured on their own flights; 320 launch sites with typical values from SondeHub

An old program, revived

Sonde Predict started in 2016 as a small Python 2 program for one question: is the sonde from Jokioinen worth chasing today? It downloaded GFS GRIB files from NOAA, ran them through wgrib2, and showed the flight in Google Earth. It worked for years, until Python 2, the data server and Google Earth all moved on, and only a zip file was left.

In 2026 it was rebuilt from that zip: the same idea and the same workflow, now in C++ and Qt, on Windows and Linux, with its own map and 3D view, better winds and a flight model tuned on real flights.

2016 Sonde Predict, up to v3: Python 2.7, wxPython, GFS GRIB, Google Earth.
2022 Last build of v3.
2026 Sonde Predict 5: rewritten from the recovered source; map, 3D, profile, Windows and Linux.
2026 Sonde Predict 5.1: MEPS + ECMWF winds, a flight model and landing area checked against real flights, multi-site batches.

Get it

Windows installer Windows zip Linux AppImage User manual (PDF)

On Linux, make the AppImage executable (chmod +x) and run it. Coming from v3? File → Import Sonde Predict v3 settings brings your locations, balloon and batch along.