Aurora anomaly detection

All-sky cameras photograph the night sky every few seconds, every night, through every aurora season. Nearly all of those frames are ordinary. The scientifically interesting ones (rare auroral structures and unusual dynamics) are buried in the noise at a rate no human reviewer can keep up with. Through UC Berkeley’s Data Discovery program, I joined the Space Sciences Laboratory to help models do the looking.

I engineered both supervised and unsupervised computer vision approaches for the problem: supervised models trained on labeled auroral forms, and unsupervised methods for surfacing the anomalies nobody had thought to label yet, since the most valuable frames are by definition the ones that don’t match existing categories.

The hard part is that “ordinary” won’t hold still. What counts as a normal sky drifts with season, sensor, moonlight, and solar activity, so a model that memorizes one winter’s normal flags everything the next. The models had to learn the sky’s habits before they could be trusted to notice its surprises. That, it turns out, is a pretty good description of research itself.