The atmosphere’s three-dimensional wind field is the single most impactful variable for weather forecasting and the least observed. Operational wind products from geostationary satellites produce roughly 5,000 vectors per hemisphere per hour, with height assignments that carry errors contributing up to 70% of total wind uncertainty. At the same time, these GEO winds provide key information to data assimilation systems driving numerical weather prediction. 

We’ve developed a system that produces over 90 million height-tagged wind observations across six atmospheric levels from a single pair of geostationary satellites, updated every 10 minutes in real time. That’s 87× more vectors than a full GFS grid, each at 2 km resolution with a geometric height accurate to better than one kilometer. Accounting for the resolution difference, the information density per unit area is roughly 3,000× greater than global forecast models.

How it works

When two geostationary satellites observe the same cloud from different longitudes, the cloud appears at slightly different positions in each image. This parallax shift is proportional to cloud height and the angular separation between the satellites. By measuring both temporal cloud motion and cross-satellite parallax simultaneously, we jointly solve for height and wind at every cloudy pixel.

Our system combines two existing technologies in a new way. WindFlow, a RAFT-based deep optical flow model developed at NASA by Vandal et al. for atmospheric feature tracking, produces dense per-pixel displacement fields between satellite image pairs. A five-state geometric solver, following the stereo wind retrieval framework of Carr et al. (2020), decomposes those displacements into cloud-top height, pointing offsets, and two wind components through weighted least squares.

The key advance is density and speed. Traditional stereo wind methods use template matching at discrete sites, producing ~5,000 vectors per scene. By replacing template matching with deep optical flow, we retrieve winds at every pixel where trackable features exist, over 15 million vectors per spectral channel, with roughly 52% of the full disk yielding valid retrievals. And best of all, this can be run every 10 minutes. 

Multi-level winds from six channels

We run the retrieval independently on six ABI infrared channels, each sensitive to a different atmospheric layer. The lower-tropospheric channels, C14 window IR (11.2 µm) and C12 ozone (9.6 µm), resolve boundary layer flow and low-level cloud motion at 3 km median height, while C10 lower water vapor (7.3 µm) bridges to the mid-troposphere at 7.8 km. The upper-level channels, C09 mid-level water vapor (6.9 µm) at 9.1 km, C08 upper water vapor (6.2 µm) at 10.1 km, and C04 cirrus (1.38 µm) at 7.4 km, capture jet stream flow and high ice clouds. Each channel produces 12–13 million quality-controlled vectors, totaling over 90 million across all six bands.

The figure below shows all six channels across the full disk at 23:00 UTC on March 10 — 6 PM Central time, as the supercell tracked through Illinois and Indiana. The storm’s anvil is visible in the upper-level panels as a concentrated region of high cloud over the central US, with the broader synoptic flow, Gulf of Mexico moisture feed, upper-level westerlies, subtropical easterlies, resolved simultaneously.

Zooming in to the storm scale reveals how each spectral band sees the supercell differently. In C14 and C12, the low-level inflow from the south feeds into the storm base at 2–3 km, with the massive anvil shield appearing as a bright white canopy overhead. In the upper-level channels, the anvil outflow streams NE at 10–14 km, the storm’s exhaust racing away from the updraft core. The height coloring reveals the three-dimensional shape of the storm: low-level green inflow beneath the anvil, mid-level yellow within it, and upper-level orange and red at the overshooting top where the updraft punches through the tropopause.

Vertical structure of a tornadic supercell

Extracting winds along a SW-to-NE transect through the supercell reveals the storm’s internal structure in cross-section. The transect passes through the 6-inch hail report at Kankakee and the EF3 tornado location, with wind barbs colored by along-track wind component.

All six channels converge at 15 km over the updraft core between Kankakee and the tornado — where the storm is optically thick through the entire troposphere. The anvil outflow races NE at 20–30 m/s at 10–14 km. On the NE flank, the channels separate dramatically as the anvil thins: C14 and C12 plunge from 15 km to below 4 km over 200 km, tracing the complete outer boundary of the storm from the tropopause down to the boundary layer. That 13 km height drop, and the accompanying wind speed transition from 45 m/s outflow to the undisturbed environment, captures the full vertical extent of the convective system in a single observation.

The vertical stratification is a direct consequence of the channel physics. The water vapor channels peak in the upper troposphere (8–11 km), while the window IR and ozone channels see primarily lower-tropospheric clouds (2–3 km). Combined, they provide continuous vertical sampling of the wind field from a single observation cycle.

Validation

We validated against the independent stereo wind product of Carr et al. For Band 14 (11.2 µm) on GOES full-disk imagery:

• Height RMSE: 874 m with +17 m bias and 0.949 correlation
• Wind RMSVD: 4.51 m/s (u-RMSE 3.54, v-RMSE 2.78 m/s)
• 19,095 collocated points passing quality control

This compares favorably to NOAA operational AMVs, which show ~5 m/s mean vector difference against rawinsondes.

Real-time processing

The complete six-channel full-disk retrieval processes in 10 minutes, matching the ABI observation cadence. This was demonstrated on live GOES-19/GOES-18 data on March 15, 2026, and again on the March 10 EF5 tornado outbreak just five days earlier. The system runs on current hardware with no modifications to the satellite constellation.

What’s next

These dense stereo wind retrievals provide the observational foundation for several capabilities that do not currently exist.

We are generating a multi-year archive of stereo wind labels to train a model that predicts dense, multi-level winds from single-satellite imagery alone, deployable on any geostationary satellite worldwide without requiring stereo overlap at inference. The spectral similarity of current geostationary imagers, ABI (GOES), AHI (Himawari), FCI (Meteosat Third Generation), makes this transfer feasible with no new hardware.

Combined with in-situ observations from radiosondes and long-duration balloon networks, microwave sounders from polar orbiters, and surface station networks, these stereo winds form the backbone for a continuously updated three-dimensional analysis of the global wind field. We look forward to sharing detailed results in the coming months.

Vandal, Thomas J., et al. “Dense feature tracking of atmospheric winds with deep optical flow.” proceedings of the 28th ACM SIGKDD Conference on Knowledge discovery and data mining. 2022.

Carr, James L., et al. “GEO–GEO stereo-tracking of atmospheric motion vectors (AMVs) from the geostationary ring.” Remote Sensing 12.22 (2020): 3779.