Eta Aquariids 2026 Peak Tonight Will Be Measured by 1,100 Volunteer Pi Cameras Running Open-Source ML

The Eta Aquariid meteor shower reaches peak activity overnight Wednesday into Thursday, May 6 to May 7, 2026. The forecasted zenithal hourly rate is 60. Southern Hemisphere observers in dark-sky conditions are expected to see 40 to 60 meteors per hour. Northern Hemisphere observers above about 40 degrees north latitude will see less than 10, capped both by the low radiant elevation and by tonight's waning gibbous moon, which will rise before the radiant clears the horizon and wash out the dim end of the population.

These numbers are appearing in every general-news outlet covering the shower. Almost none of those outlets are explaining where the numbers actually come from. They come out of two operational sensor networks that have quietly become the most productive astronomical instruments of their class anywhere on Earth: the Global Meteor Network and Cameras for Allsky Meteor Surveillance (CAMS). Both run on consumer-grade hardware. Both run open-source ML pipelines nightly. Both are mostly volunteer-staffed. Together they produce more than a million precision meteor orbits per year.

This is a story about the technology, not the meteor shower. The shower happens whether we look or not. The instrumentation is the recent thing.

The hardware: $300 per node, deployed by individuals

A standard GMN station is documented in detail on the operator wiki and follows a tight specification: a Raspberry Pi 4 single-board computer, a low-light Sony IMX291 sensor in an IP66-rated weatherproof housing, a 4mm or 6mm CS-mount lens, an SSD for video buffering, a Wi-Fi connection, and a roof, balcony, or garden mount with a clear sky view. The all-in bill of materials at 2026 prices is approximately $270.

The unit cost matters. CAMS, which is a few years older, was designed around higher-end CCD cameras and dedicated PCs at a per-station cost in the $5,000 to $8,000 range. CAMS has scaled to roughly 60 to 80 active stations across the United States, Australia, and the Benelux countries. GMN, with its lower per-station cost and Pi-based reference build, has scaled to over 1,100 active stations across more than 50 countries. The two networks complement each other. CAMS has deeper per-station data and a longer baseline of observations. GMN has wider sky coverage and finer geographic granularity.

The unit-cost difference is what made the GMN model work. At under $300 per station, the procurement decision moves from institutional to individual. An amateur astronomer or a school astronomy club can put two stations on rooftops over a weekend without filing a budget request. That dynamic — the same dynamic that took home weather stations and consumer-grade GNSS receivers from niche to ubiquitous — is what densifies the network.

The software: three classifiers stacked at the edge

The reference software, called RMS (Raspberry Meteor Software), is open source and runs entirely on each Pi. It does not require a network connection during overnight recording. The processing flow is classical edge ML: capture, threshold, classify locally, upload sparingly.

Each Pi records video at 25 frames per second after sunset, uses difference imaging to extract bright transient pixels, then runs three classifiers in cascading order to filter the candidate events down to high-confidence meteors:

A random forest binary classifier on hand-crafted features (track length, angular velocity, photometric profile, frame coverage, color ratios) discards roughly 95% of candidates as obvious non-meteors.
A convolutional neural network scores the survivors using stacked image frames. This stage catches the harder cases — particularly satellite glints, which can fool the random forest on the photometric axis but have characteristic visual signatures.
A long short-term memory network predicts whether the light-curve profile of each event is consistent with atmospheric ablation. Real meteors have a sharp rise, peak, and decay structure on a sub-second to few-second timescale. Aircraft and strobing satellites do not.

The compute footprint of all three stages combined runs on a Pi 4 overnight in the time available between sunrise and the operator checking their morning email. Inference is GPU-free. The trained weights are versioned in the open-source repository and updated periodically as new false-positive classes are identified by operator review.

What gets uploaded each morning is a small bundle per high-confidence detection: a short clip, the extracted track, calibration metadata, the photometric profile. A typical clear-night station uploads under 200 megabytes. The bandwidth budget is trivial.

The math: triangulation, orbit recovery, parent-body association

A single camera sees a meteor as a streak in a 2D plane. To recover the 3D atmospheric trajectory and the heliocentric orbit, you need at least two stations with overlapping fields of view, separated by a baseline of typically 50 to 200 kilometers. The central back end ingests detections from every station each morning, indexes by NTP-synchronized timestamp, and looks for spatially and temporally coincident detections. When two or more stations agree on a meteor, the back end solves a non-linear least-squares fit for the atmospheric trajectory, then propagates backward to recover the heliocentric orbit.

If the recovered orbit is sufficiently similar to a known parent body's orbit (using the standard Drummond similarity criterion), the meteor is associated with that body's debris stream. The Eta Aquariid associations to comet 1P/Halley are unambiguous. The orbit is highly inclined, retrograde, and characteristic enough that misclassification is rare.

This is the analytical layer that turns a network of independent cameras into an instrument. No single station can recover an orbit. The network produces tens of thousands per year.

The result: real-time rate forecasting, ML-discovered comets

Two products flow out of this architecture. The first is the rate forecast that anchors every shower-coverage news article. The model combines multi-year orbit data with atmospheric corrections for radiant elevation, limiting magnitude, population index, and lunar phase to predict zenithal hourly rates as a function of time and observer latitude. The peak rate of 60 you are reading about for tonight is the output of that model. Tomorrow morning's network observations will be compared against the prediction and used to refine the model for next year.

The second product is more remarkable. CAMS pioneered ML-based shower identification by clustering trajectories in the orbital catalog without prior knowledge of parent bodies. Over the last decade this approach has produced more than three dozen new meteor shower confirmations, and critically, has flagged candidate parent body orbits that were subsequently confirmed by optical and radar follow-up. ML systems on volunteer hardware have been finding comets that the world's largest optical surveys had not yet imaged.

This is "AI for science" working at production scale, on consumer hardware, with open data and open code, on a budget that would not buy a single H100 today.

Why this matters beyond astronomy

The architectural pattern — cheap edge devices, on-device cascade classifiers, sparse uplink, central triangulation, open data — is directly transferable to bioacoustic monitoring, atmospheric chemistry, distributed seismology, agricultural pest surveillance, and a half-dozen other domains where the data flywheel is gated by hardware cost rather than algorithmic novelty. The astronomy community got there first because they had decades of ground-truth catalogs and an active amateur community.

For a deeper analysis of the architecture and its lessons for enterprise distributed observability, see today's CrashBytes deep-dive on edge AI and citizen meteor networks. We also have a running prediction on ML-driven comet discoveries through 2027 that the catalog growth makes increasingly testable.

How this compares to other distributed sensor networks

The closest peer to GMN in 2026 is the eBird citizen-science platform run by the Cornell Lab of Ornithology, which has roughly 900,000 active observers contributing distributed bird-sighting data and uses ML classifiers (Merlin, BirdNET) to validate audio submissions. eBird is larger by participant count but produces a fundamentally different data shape — categorical occurrence reports rather than precision physical measurements. Both are open-data projects. Both have produced peer- reviewed science. The architectural lessons run in both directions.

Other adjacent examples include the Raspberry Shake seismic network (~1,200 personal seismographs uploading to a central back end), the PurpleAir air-quality network (~30,000 nodes), the AllskyCams meteor network (older, smaller than GMN, focused on US coverage), and the SkySentinel fireball detection network. None of these matches GMN's combination of ML pipeline maturity, data flywheel, and open-source methodology. Some — Raspberry Shake especially — have nearly identical architectural patterns but lag in classifier sophistication.

The lesson is that the architectural pattern is general, the implementations vary in maturity, and meteor astronomy currently leads the field on the ML pipeline axis specifically. The next 18 to 24 months are likely to see the GMN/CAMS reference architecture adopted with attribution by adjacent volunteer-data domains. Some of that adoption is already happening at the operator level, with individual hobbyists running multiple sensor types on the same Pi infrastructure.

What enterprise infrastructure teams can take away

Three patterns from GMN are immediately portable to enterprise distributed observability and edge ML deployments. First, cascade classifiers staged in increasing-cost order — random forest, then CNN, then LSTM — applied to candidate events at the edge, dramatically reduce upstream bandwidth and central compute load. Most enterprise trace pipelines do not stage their classifiers this way. They should.

Second, schema stability and a stateless central back end allow new clients to join the network without per-client capacity planning or explicit enrollment workflows. The GMN back end has no inventory of stations. Stations push detections in the right format and become part of the network. The schema does the work that configuration management would otherwise do.

Third, unit economics that fit a hobbyist budget produce networks that grow without procurement friction. The same dynamic applies inside enterprises: agents that deploy in an afternoon, with no quarterly capacity review, will spread; agents that require governance approval will stall at the pilot. The cost of the deployment friction is often larger than the cost of the deployed software.

What to expect tomorrow morning

By Thursday morning, GMN and CAMS stations across both hemispheres will have flat-fielded their frames, run the classifiers, and uploaded their detection bundles. The back ends will have triangulated, fit orbits, and associated to Halley. Public dashboards will show the observed rate versus the prediction. The catalog will have grown by tens of thousands of trajectories. Some of those will likely be rare flagged candidates queued for parent-body follow-up by professional surveys.

This is what edge AI looks like when it has been allowed to grow without a procurement cycle. It is operational, it is reproducible, and it is five years ahead of most "AI everywhere" enterprise pitches. Tonight is a good night to look up. Tomorrow is a good morning to look at the public data.

The viewing window tonight

For readers in the Southern Hemisphere or the tropics, the optimal viewing window opens roughly two hours before astronomical twilight on the morning of May 7 and extends until the sky brightens. The radiant is in the eastern sky in Aquarius, climbing as the night progresses. The waning gibbous moon will be present but the brighter Eta Aquariids — which tend toward magnitude zero or brighter and often leave persistent trains because of their high entry velocity — will punch through the moonlight without trouble. No optical equipment is required. A dark sky and a clear horizon are the only real prerequisites. Observers in the higher Northern latitudes should manage expectations: the radiant elevation does not exceed roughly forty degrees before sunrise, and the rate caps in single digits regardless of moon phase. South of about thirty degrees north, rates climb rapidly with declining latitude.