Case study 1 · October 2026
Machine-learning earthquake picking on a Raspberry Shake
What we did
With a fellow seismology enthusiast, we ran the open-source easyQuake toolkit on one day of data (Oct 21, 2026) from our Raspberry Shake station AM.R1CFD in High Valley, Idaho, together with nearby public network stations.
- easyQuake ran three machine-learning phase pickers: PhaseNet, EQTransformer and GPD.
- Four small local events that day were then picked and reviewed station by station, with P and S arrivals checked on review plots and graded for quality.
- The reviewed picks were used to locate the events and estimate local magnitude.
- Finally, each ML picker’s output was compared against the reviewed picks.
What the data showed
- Three events located close together roughly 85 km west of R1CFD, toward Hells Canyon; one located about 124 km to the east.
- Estimated local magnitudes ranged from about M1.0 to M1.7.
- Against the reviewed picks, EQTransformer matched 31 picks (median difference 0.07 s) and GPD matched 45 (median 0.20 s). PhaseNet’s run produced no picks that matched in this comparison.
- Takeaway: ML pickers are a fast first pass, but human review still matters, especially for tiny events on a single-component hobby station.
Skills practiced: seismic data handling, phase picking and QC, earthquake location basics, Python data tools. Preliminary, informal analysis; not an official catalog.