Original title: Distilling deep optical flow stereo methods to retrieve dense three-dimensional wind fields
Thomas J. Vandal, Dong L. Wu, James L. Carr, Derek J. Posselt +4 · September 2026
Weather models rely on tracking wind speeds and wind heights across the globe. Standard satellite methods track features slowly, cover few points, and depend on weather predictions to estimate height. This creates a circular loop that causes inaccurate heights. Stereo satellites can calculate heights geometrically from different viewpoints, but they require overlapping views and need huge computational power. To fix this, researchers used deep optical flowA computer vision technique that estimates motion between image frames, used to track features in satellite imagery., an artificial intelligence method that tracks motion between image frames. They trained a stereo teacher model and then distilled its knowledge into a student model that uses just one satellite. This student model calculates winds globally across full-disk images. Tests show stereo windsWind estimates that use geometric parallax from multiple satellite views to accurately determine height. beat operational tracking for three water vapor bands (6.2, 6.9, and 7.3 μm), but degraded in the long-wave infrared 11.2 μm band.
You're trying to figure out how high a kite is flying from the ground. From one spot, it's impossible to tell whether it's high up and far away or lower and closer to you. But if you have a friend standing somewhere else, you can both look and use the difference in your viewpoints to calculate its exact height. Here, satellites do the same with clouds to measure wind, and this work lets a single satellite mimic that two-view calculation.
This improves weather forecasting by providing more accurate wind data from a single satellite. It eliminates a circular dependency that made height estimates unreliable, helping meteorologists predict storms and climate patterns more effectively.
The abstract notes that performance degrades in the long-wave infrared band at 11.2 μm compared to other bands. It also focuses on geostationary satellites, so applicability to other orbit types is not addressed. The validation is against specific datasets, which may not cover all atmospheric conditions.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Distilling deep optical flow stereo methods to retrieve dense three-dimensional wind fields — Thomas J. Vandal, Dong L. Wu, James L. Carr et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.