Sampling rate: all inputs resampled to 100 Hz (incl. 200 Hz stations) — verify anti-aliasing & downstream effects

[seismoseo]

Summary

PocketQuake (via the korea-cluster-relocation engine) resamples every input trace to a fixed target_sampling_hz = 100 Hz before AI picking, regardless of the station's native sampling rate. Stations recording at 200 Hz (or any non-100 Hz rate) are therefore down-sampled to 100 Hz. Filing this for later verification — recording the behavior now, not asserting a bug yet.

Where (code refs)

• pipeline/config.py:181 — target_sampling_hz: float = 100.0 (single global default)
• pipeline/core/picking.py:133 — pick_event_pnplus(model, comp_files, sampling_rate=cfg.target_sampling_hz)
• pipeline/core/eqnet_backend.py:138 — SeismicTraceIterableDataset(..., sampling_rate=sampling_rate, ...); the EQNet / PhaseNet+ data loader performs the actual resample of each trace to that rate.
• (separate path) pipeline/core/xcorr.py:83 — cross-correlation interpolates to its own interp_hz (lanczos, a=20), independent of target_sampling_hz.

So all picking input is normalized to 100 Hz inside the EQNet loader; a 200 Hz station's data is decimated 2× there.

Why it may matter (to check later)

• Anti-aliasing. 200 → 100 Hz halves Nyquist (100 → 50 Hz). Need to confirm EQNet's resampling applies a proper anti-alias low-pass before decimation; otherwise energy >50 Hz can alias into the 1–40 Hz pick band.
• First-motion polarity & amplitude. Both are read from the (resampled) traces for the focal-mechanism / magnitude stages. Down-sampling can blunt the P first sample and bias polarity/amplitude.
• Pick timing. 100 Hz = 10 ms samples; for 200 Hz native data this discards sub-10 ms timing (xcorr re-interpolates separately, so dt.cc may be unaffected — but dt.ct from picks could be).
• Configurability. target_sampling_hz is one global value — decide whether it should be configurable or matched per-network.

To verify

• Does EQNet SeismicTraceIterableDataset anti-alias on down-sample (vs naive interpolation)?
• Compare picks / polarity / amplitude for a 200 Hz station at native vs forced-100 Hz.
• Confirm xcorr interp_hz vs pick target_sampling_hz consistency for dt.cc / dt.ct.

[seismoseo]

Verification progress — the mechanism is now CONFIRMED: EQNet's loader resamples with trace.interpolate(sampling_rate, method="linear") (eqnet/data/seismic_trace.py, the '## interpolate to 100 Hz' block) — plain linear interpolation, no anti-alias low-pass before decimation. So for 200 Hz stations, energy above the new 50 Hz Nyquist can alias into the pick band; the highpass applied afterwards does not remove it. Remaining to quantify (keeping this open): the empirical size of the effect on picks/polarity/amplitude for a 200 Hz station at native vs forced-100 Hz — most pick energy sits at 1–40 Hz, so the impact may be small, but that's exactly what the comparison should establish. Also confirmed: xcorr is unaffected (it re-interpolates from the gathered SAC with its own lanczos a=20 path); dt.ct from picks is the exposed product. A candidate minimal fix once quantified: lowpass at 0.4×target_sampling_hz before handing traces to the loader.