seismoai-qc 0.2.0

QC module for seismic trace detection — part of seismoai

seismoai-qc

Quality Control module for seismic trace detection — part of the seismoai pipeline.

Built on real data from the Forge 2D Survey (2017): 166 SGY files, 27,722 total traces.

Install

pip install seismoai-qc

Usage

from seismoai_qc import detect_dead_traces, detect_noisy_traces, qc_report
import numpy as np

# traces: 2D numpy array (167, 4001) — output of seismoai_io
report = qc_report(traces)
print(report['label'].value_counts())
# good     21373
# dead      6183
# noisy      166

# Pass to seismoai_model:
labels = report['label']

Functions

Function	What it does
detect_dead_traces(traces)	Flags traces with std_dev < 1e-4 (no signal)
detect_noisy_traces(traces)	Flags traces with max_amp > 50 (spike up to 758)
qc_report(traces)	Full DataFrame with labels for seismoai_model

Real Data Stats

Analyzed across all 166 SGY files (27,722 total traces):

Category	Count	Percentage
Good traces	21,373	77.1%
Dead traces	6,183	22.3%
Noisy traces	166	0.6%
Total	27,722	100%

Why are some traces dead?

Dead traces (std_dev < 0.0001) occur at far offsets where seismic source energy does not reach the receiver. In Forge 2D Survey, 22.3% traces are dead — mostly far-offset receivers.

Why do some traces reach 758?

Normal traces stay below 15 (99th percentile = 10.71). Noisy traces spike up to 758.22 due to electrical interference or instrument malfunction during acquisition. Only 0.6% traces are noisy.

Thresholds (derived from real data)

# Dead threshold
std_dev < 0.0001
# Dead trace std range: 0.00000021 to 0.0001
# Live trace std range: > 0.001 (clear gap)

# Noisy threshold  
max_amp > 50.0
# Normal 99th percentile: 10.71
# Noisy minimum: 200+ (all 166 noisy traces exceed 200)
# Worst spike: 758.22

How Pair 4 Uses This

from seismoai_io import load_sgy
from seismoai_qc import qc_report

# Load traces
traces = load_sgy("file.sgy")

# Get QC labels
report = qc_report(traces)

# Hand off to seismoai_model
labels = report['label']   # 'good', 'dead', 'noisy'
features = traces[report['is_dead'] == False]  # remove dead

Run Tests

pytest tests/ -v

Real Dataset Test

python test_real_data.py

Reflection

We built the seismoai_qc module which detects dead and noisy traces in seismic data from the Forge 2D Survey. Dead traces have std_dev below 0.0001 — out of 27,722 total traces, 6,183 (22.3%) were dead because far-offset receivers did not receive enough source energy. Noisy traces have amplitudes above 50, spiking as high as 758, caused by electrical interference — we found 166 noisy traces (0.6%). We derived these thresholds by analyzing the real dataset rather than guessing. Our qc_report() produces labeled output that Pair 4 will use directly to train their noise classifier on 21,373 good traces.