Interactive demo · MetDetect

See a meteorite fall in the radar

When a meteorite fall's debris drifts to the ground it briefly shows up in NEXRAD Doppler weather radar. Fireball networks tell us when and where to look; MetDetect answers which radar echoes are the falling debris. Each event below shows every candidate the detector surfaces, with the ones clearing its 10%-false-alarm operating point marked; false positives are left in, and every run uses the current pipeline. Everything renders in your browser.

Event …

How to read the score. Each event's score comes from the current detector, trained on real no-fall radar windows and evaluated leave-one-event-out (the event's own data is never in its training set). A detection means the score clears a threshold set at a 10% false-alarm rate on those no-fall windows (leave-one-out ROC AUC 0.865 against real nulls). Calibration is preliminary — the pre-registered held-out evaluation is still pending. Two hard cases are kept honestly: Hamburg's top candidate is the recovered debris but scores just below threshold in winter weather, and on the Washington coast nothing clears threshold — the real debris ranks second, a hair behind a rain cluster.

What am I looking at? Each dot is a radar “gate” inside a candidate the detector flagged at its operating point, colored by reflectivity. Real falls appear as an elongated streak — fragments strung out along the wind as they descend. The 3D view shows the same gates at their true altitudes. The radar site is marked ✳; ◆ marks the candidate's centroid. Rings around dots mark agreement with the manually identified truth.

This is a preloaded demo of precomputed results (the site is served as static pages; the desktop app can't run here). A live companion tool — upload a fireball, search the radar on demand — is planned. See the MetDetect and research pages.

How MetDetect works

A window of raw NEXRAD Level-II radar (~10⁶ gates) is reduced, stage by stage, to a handful of candidate debris clusters, and a null-trained detector scores each against real no-fall radar behavior.

MetDetect data-reduction cascade flowchart — The data-reduction cascade: radar volumes → polar cube → prefilters → KDE anomaly weighting → weighted clustering → physics gates → supervised pruning → candidates → null-trained detector.

The MetDetect desktop application — The MetDetect desktop app — radar loading, cascade runs, and gate-level inspection. This page shows its outputs; the tool itself runs on the desktop.

Osceola fall: height vs range cross-section of the debris column — What the detector sees vertically: the Osceola debris column descending through the radar beams (diagnostic panel from the June 2026 analysis run; scores on this page use the current detector).