Seismic First-Break Picker

Automatic P-wave first-break picking for active-source (vertical-component) land seismic, designed to produce physically-consistent traveltimes ready for traveltime inversion (pyGIMLi .sgt) with no further human editing.

Trained on 33 manually-picked shot gathers and applied to the full 117-gather survey.

Approach

The pipeline is classical-first, learned-refined, physically-regularized:

1. SEG-Y reader (pick/segy_io.py) — parses Stryde source gathers, geometry from the dataset's custom trace-header bytes, honours the −100 ms recording delay.
2. Classical baseline (pick/picker.py) — STA/LTA + AIC + energy-ratio with a velocity-prior search window. Unbiased but limited to ~3 ms MAD (energy detectors lag the human's first-motion convention and occasionally cycle-skip).
3. Supervised picker (pick/ml_pick.py, pick/train_cv.py) — a 1-D U-Net (PhaseNet-style) that regresses an onset-probability bump, learning the human first-break convention from the 33 labelled gathers. Trained on Apple-Silicon GPU (MPS). An ensemble provides a disagreement signal for rejection.
4. Gating — "fewer but correct" (pick/gate.py) — a pick is kept only if it clears a confidence floor, the ensemble agrees, and it fits a robust local moveout trend.
5. Moveout regularization (pick/moveout.py) — per shot, a confidence-weighted LOESS + isotonic fit turns jittery independent picks into a smooth, monotonic, physically-consistent first-break curve (removes the ±1-sample grid jitter that arises from 1 ms sampling over 0.15 m receiver spacing), dropping off-curve picks.
6. Export (pick/dataset.py) — writes a pyGIMLi .sgt reusing the original point indices; physically-invalid picks (t ≤ 0) are removed.

Measured quality

Leave-shots-out cross-validation against the manual picks:

Metric	Full coverage	Conf-gated (50%)
Median residual (bias)	+0.16 ms	—
MAD	1.39 ms	1.05 ms
within ±2 ms	63%	72%

Label-free checks on the exported survey:

• Reciprocity (source↔receiver) median |Δt| = 1.68 ms (P90 5 ms).
• Moveout monotonic fraction = 0.96 for the auto picks (vs 0.66 for the raw manual picks) — the auto curves are smoother and more physical than the human's.

See out/fig_traveltimes.png, out/fig_gather_auto.png, out/fig_gather_manual.png.

Layout

pick/        pipeline (reader, picker, ML, gates, moveout, export, render, verify, diagnostics)
out/         exported picks (.sgt), QC csv, figures
stacked_segy_groups_of_3/   117 source-gather SEG-Y files
3D_Picking_1-33.sgt         manual first-break picks (training labels, pyGIMLi format)

Run

python -m venv .venv && .venv/bin/pip install -r requirements.txt
.venv/bin/python pick/train_cv.py     # cross-validated accuracy
.venv/bin/python pick/run_all.py      # train ensemble + pick all 117 -> out/picks_full_survey.sgt
.venv/bin/python pick/verify.py       # format + reciprocity + moveout QC
.venv/bin/python pick/render.py       # figures

Output for inversion: out/picks_full_survey.sgt (pyGIMLi TravelTimeManager).