At 10:22 AM on October 14, 2025, a wildfire ignited in the Bastrop County cedar forest east of Austin. By 10:37 AM — just fifteen minutes later — Trace AQ's smoke detection system had flagged the emerging plume and issued an automated alert to our subscribed users in the affected forecast zone. By noon, the smoke was visible from downtown Austin. Our users had nearly two hours of advance warning.
That fifteen-minute detection latency is not an accident. It is the result of careful engineering around the GOES-R satellite constellation and a real-time processing pipeline that we've spent over two years building and refining. In this article, we're pulling back the curtain on how it works.
GOES-R: The Geostationary Backbone
The GOES-R series — currently GOES-16 (East) and GOES-18 (West) — are operated by NOAA and represent the most capable geostationary weather satellite constellation ever deployed over the Americas. The Advanced Baseline Imager (ABI) instrument aboard each satellite captures imagery in 16 spectral bands at spatial resolutions between 500 meters and 2 kilometers, with a full-disk scan every 10–15 minutes and mesoscale sector scans every 30–60 seconds during active events.
For smoke detection, the key bands are in the mid-infrared and visible spectrum. Smoke particles scatter and absorb incoming solar radiation in ways that create distinctive spectral signatures compared to clear air, cloud, and other atmospheric constituents. The ABI's 0.64 µm visible channel and 2.25 µm shortwave infrared channel are particularly valuable for differentiating smoke from clouds — a non-trivial challenge, since both appear as bright, optically thick features in visible imagery.
Our Ingestion Architecture
NOAA distributes GOES-R data through the NOAA Open Data Dissemination (NODD) program via Amazon S3. New ABI scan files appear in the bucket within minutes of satellite downlink completion. Our ingestion system subscribes to S3 event notifications and processes each new file within seconds of it appearing.
The raw NetCDF files require several preprocessing steps before smoke analysis can begin. We apply the standard ABI Level 1b to Level 2 calibration pipeline to convert raw counts to physical radiances and reflectances. We then reproject the data from the native fixed-grid projection to a standard geographic coordinate system at 1 km resolution over our coverage area.
The entire preprocessing pipeline runs on a dedicated GPU cluster. Each ABI scan covering the continental United States processes in under 90 seconds, well within our target of beginning smoke analysis before the next scan arrives.
Smoke Detection Algorithm
Our smoke detection uses a two-stage approach. The first stage applies a physics-based spectral test: we compute the difference between specific ABI band pairs that are known to behave differently for smoke versus cloud versus clear sky. This spectral test produces a preliminary smoke probability map with roughly 80% precision at the pixel level.
The second stage applies a convolutional neural network trained on three years of manually labeled GOES-R imagery. The neural network takes the spectral test output plus raw multi-band imagery as input and produces a refined smoke probability with spatial coherence — it understands that smoke plumes have characteristic shapes and transport patterns, not random pixel configurations. This second stage pushes precision above 94% while maintaining a recall rate over 91% against our validation dataset.
The output of this detection pipeline is a smoke probability raster and a set of detected plume objects with estimated centroid locations, extents, and optical depths. These become inputs to our atmospheric transport model, which forecasts where the smoke will go over the next 24–96 hours.
Integration with the Forecast Model
Smoke detection tells us where smoke currently is. The forecast model tells us where it will be and at what ground-level concentration. Bridging these two requires estimating the vertical distribution of smoke — how high it is in the atmosphere determines how it will be transported.
We use LIDAR data from CALIPSO satellite overpasses and ground-based AERONET sun photometers to calibrate our smoke layer height estimates from GOES-R optical depth retrievals. This calibration allows us to assign a plausible vertical profile to each detected smoke plume object, which then gets ingested into the atmospheric transport model.
The full pipeline — from satellite scan to updated AQI forecast — completes in under 15 minutes for routine smoke events and within 20 minutes for complex multi-source events where plume attribution requires additional processing. This latency is competitive with or better than any other operational smoke forecasting system we're aware of.
Night and Low-Visibility Operations
One significant challenge with visible and near-infrared smoke detection is that it doesn't work well at night. Wildfires obviously don't stop at sunset, and smoke transported overnight can be poorly characterized without daytime satellite observations.
For nighttime operations, we shift weight toward our atmospheric transport model (which runs continuously regardless of satellite availability) and supplement with VIIRS imagery from the Suomi-NPP and NOAA-20 polar-orbiting satellites. While polar orbiters provide lower temporal resolution than GOES-R, their nighttime-capable low-light imager provides useful fire detection and limited smoke characterization during gaps in geostationary coverage.
We're also evaluating integration with ground-based camera networks operated by ALERTCalifornia and similar state programs, which provide high temporal resolution fire detection under line-of-sight conditions. This network complements satellite observations well in mountainous terrain where view angles create sensor blind spots.
Building this infrastructure has been one of the most technically demanding parts of creating Trace AQ. We're committed to continuing to invest in our satellite data capabilities and to sharing what we learn with the broader environmental monitoring community.
