Skip to content

Tune the detection

A first search with the defaults tells you a lot about your data. This guide shows how to check the results and which settings to change for too few detections, too many false detections or imprecise locations.

Check the results

  • Waveforms and picks: open the run in Snuffler and compare the modeled and picked arrivals with the waveforms of a few detections.

    qseek snuffler my-search/ --show-observed --show-modelled --show-semblance

  • Detection function: --show-semblance adds the detection function, the maximum semblance over all nodes. Clear peaks above a flat noise level mean the stack works.

  • Quality columns: csv/detections.csv has the semblance, the number of picks n_picks, the travel time residuals rms and the location uncertainty of every detection. Plot them to find a threshold that separates events from noise. See the run directory.

Too few detections

  • Threshold: the default detection_threshold "MAD" is 10 times the median absolute deviation of the detection function in each window. Set a fixed value, below the semblance of the weakest events you see, to detect more.
  • Image function: try pre-trained weights that fit your data, e.g. trained on a similar region or instrument type, or another model. See image functions.
  • Frequency band: adjust the bandpass of the pre-processing to the frequencies of your events, e.g. higher frequencies for microseismicity.
  • Minimum stations: lower min_stations for small networks.

Too many false detections

  • Threshold: raise the detection_threshold, or set a fixed value above the semblance of the noise.
  • Picks: filter the detections by n_picks. Noise rarely produces consistent picks at many stations.
  • Distance weighting: with the default distance weights, each node relies on its closest stations. Noise at a single close station can dominate small networks; increase required_closest_stations.

Imprecise locations

  • Velocity model: the travel times need a velocity model that fits your region. A 1D model of the region is better than a constant velocity, a 3D model better still in complex geology. See ray tracers.
  • Station corrections: extract station corrections from a first run and search again. They correct what the velocity model misses, see station corrections.
  • Resolution: increase n_levels to the search volume for smaller nodes, and keep node_interpolation on.
  • Boundary: detections at the border of the volume are ignored. If the seismicity reaches the border, enlarge the volume.

Repeated detections of one event

detection_blinding sets the minimum time between two detections, 1 s by default. Raise it if a long or complex event is detected twice.