Can human seismologists verify machine-learning detected earthquakes?

Lingsen Meng , Hui Huang, Jinzhi Ma, Yang Ma

Routine seismic network catalogues are reviewed event by event by analysts, whereas machine-learning catalogues, which contain five to fifty times more events, leave most of their detections unexamined. Whether those detections are earthquakes can nevertheless be measured, because each one can be checked blind against its own waveforms. Human seismologists returned 4,800 blind verdicts on 1,200 detections from six catalogues, and confirmed 60 to 98% of what they were willing to judge. A score fitted to seven quantities a catalogue already records reproduces that judgement and transfers to a sequence it was not trained on, so the standard the panel set can be applied automatically to millions of events nobody can review by hand, and we suggest releasing machine catalogues in tiers by it. The score also separates the catalogues geometrically: at Ridgecrest the high-scored events trace the conjugate strike-slip fault network while the low-scored events fill the same volume diffusely, and beneath northeast Japan the high-scored events resolve both planes of the double seismic zone where the low-scored events blur them. False detections are therefore concentrated among the low-scored events. Quantifying what a seismologist will endorse keeps human judgement inside the scientific record rather than upstream of it.