You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
Gesa 8a9884b6ce fixed arrivals from xml picks 3 weeks ago
cpt removed bar plot figure from commit 6 months ago
examples added stacking kwargs from config 1 year ago
helper_functions comment out example use 2 months ago
src fixed arrivals from xml picks 3 weeks ago
README.md upd readme, example values and comments for config file 3 weeks ago
setup.py fixed stacking functions for S and P, name for setup.py 1 year ago

README.md

CLUSTY

Clusty is a toolbox for the clustering of earthquakes based on waveform similarity abserved across a network of seismic stations.

Input: Seismic waveforms, station metadata, (optional: picks)

This maual is still in preparation, please send us an email if you need further help or have suggestions: gesap@gfz-potsdam.de or pniemz@gfz-potsdam.de

prerequisits:

Basic commands to run clusty:

A full run can be started using:

  clusty --config CONFIG_FILE --run --log_level LOGLEVEL

log_level argument can be: DEBUG, INFO, ERROR, WARNING

However, we recommend to run the tool bit by bit, using the following options:

--cc to compute cross correlations, --netsim to compute the network similarity, --eps_analysis to obtain insight into dbscan parameter settings, --cluster to cluster the earthquakes based on the precomputed network similarity, --plot_results to obtain result plots, --merge_freq_results to merge clustering results obtaine in different frequency ranges or --export_stacked_traces to export stacked waveforms for each cluster.

Input:

  • catalog file in pyrocko format
  • station file in pyrocko format
  • picks (optional) in pyrocko format or xml
  • waveform data - one subdirectory per event with NETWORK.STATION in waveform-file (see helper_functions/data_download.py for fdsn download example to this format or contact us for help in converting into this format from continous data)

Example configuration file:

A basic config file is created by running the command clusty --init. Settings need to be adjusted afterwards.

Values given here indicate those values that we used for the study of the aftershock sequence of the Zakynthos Oct. 2018 Mw 6.9 event.

  --- !clusty.config.clusty_config
settings:

- !clusty.config.GeneralSettings
  n_workers: 1
  work_dir: ./
  catalog_file: path/to/catalog
  waveform_dir: path/to/waveforms
  station_file: path/to/stationfile
  station_subset:  # min. and max. distance between stations and events [km]
    maxdist_stations: 200.0
    mindist_stations: 50.0

- !clusty.config.cc_comp_settings
  bp: [3.0, 0.05, 0.2]  # order, corner highpass [Hz], corner lowpass [Hz]
  downsample_to: 0.1  # [s]
  #pick_dir: ''  # optional
  phase: [R, L]
  components: [HHZ, HHE, HHN]
  compute_arrivals: true
  vmodel: path/to/velocity-model
  use_precalc_ccs: false  # boolean value to indicate if cross-correlations are already computed
  snr_calc: true  # boolean value to indicate weather SNRs should be computed for each trace
  snr_thresh: 2.0  # minimum SNR
  max_dist: 30.0  # maximum inter event distance [km]
  debug_mode: false  # opens an interactive waveform browser to check time window and filter settings, 
                     # for testing only
  debug_mode_S: false   # same for second phase

- !clusty.config.network_similarity_settings
  get_station_weights: false  # should a station weighting be applied? in case of uneven station distribution
  method_similarity_computation: trimmed_mean_ccs  # other methods: median, mean, max., product, weighted_sum_c_diff (see Petersen & Niemz et al. 2020)
  use_precalc_net_sim: false  # boolean value to indicate whether network similarity matrix is already computed
  trimm_cut: 0.3  # parameter for trimmed mean method - cut off percentage of worst stations
  apply_cc_station_thresh: true  # should a cross-correlation based threshold be used (see Petersen & Niemz et al. 2020)
  cc_thresh: 0.7  # min. cc to be met at ```min_n_stats``` stations covering an azimuthal range of at least ```az_thresh``` deg to consider an event for clustering
  min_n_stats: 5 
  az_thresh: 60  # [deg]
  combine_components: true  # combine all components (or separate results?)
  weights_components: [0.4, 0.3, 0.3]  # weightings for components, here HHZ,HHN,HHE

- !clusty.config.clustering_settings
  method: dbscan
  dbscan_eps: [0.08, 0.12, 0.14, 0.16, 0.18, 0.2]  # range of eps values to get started, use smaller steps in a smaller range after finding a rough best value
  wf_plot: [] # add tuple with (eps, minpts) if you want wf plots
  wf_plot_stats: []  # add net.stat if waveform plots should be returned
  export_stacked_traces: false  # set to true if stacked traces are needed (slow) 
  debug_stacking: false

values used in Zakynthos study:

same settings for all methods:

  • SNR > 2.0
  • ccmin of 0.7 met at above 5 stations covering > 60 deg az.
  • HHZ (0.4), HHE (0.3), HHN (0.3)
  • MinPts = 5 (tested 3-8)
  • Eps range tested: 0.03 - 0.30

method: TRIMMED MEAN

  • trim: 0.3 (tested also 0.1 and 0.2)
  • 0.05-0.2 Hz --> eps = 0.13
  • 0.02-0.15 Hz --> eps = 0.15
  • 0.1 - 0.5 Hz --> eps = 0.24
  • 0.2 - 1 Hz --> eps = 0.26

method: Weighted sum, weighting based on difference between first and second cc-function max.

  • MinPts = 5
  • 0.02-0.15 Hz --> eps = 0.17

method: products

  • MinPts = 5
  • 0.02-0.15 Hz --> eps = 0.20

method: mean cc of all stations (no trimming)

  • MinPts = 5
  • 0.02-0.15 Hz --> eps = 0.15

method: median cc of all stations

  • MinPts = 5
  • 0.02-0.15 Hz --> eps = 0.15

method: max. cc of all stations

  • MinPts = 5
  • 0.02-0.15 Hz --> eps = 0.05
  • note that this method only requires a large cc value at a single station. Results are therefore not represetative for entire mechanism...

METHODS

(1) Stacking Methods

  • max_cc:

    • maximum cc of all stations for one event pair used as network similarity --> only for testing, time window and filter selection etc.
  • median_ccs (Scott and Ater, 1993):

    • can be used as a proxy for network similarity, but problem if wide magnitude range: for smaller events only the closest stations can record event --> use median of all that pass snr ratio
  • weighted_sum_c_diff:

    • sum of cc at all stations; weighted by difference of first to second cc-maxima
  • product and product_combPS (Stuermer et al. 2011):

    • single phase: combine stations as n-th root of product of cc at all (n) stations
    • P and S: multiplication of P and S products, then 2nth- root
  • mean and trimmed mean (Maurer Deichmann 1995):

    • trimmed: lowest values removed before mean calculation, k percent of stations removed before net sim is computed as mean of remaining stations. ‘k must be determined by trial and error’.
  • combine P and S:

    • mean of P and S network similarity after using above methods for computing P and S net sim independently. P and S can be weighted. waveform differences can be more distinct on S phases.
  • stacking in different freuqncy bands (Souberster 2017)

(2) Station weighting methods

  • based on az. station position; only implemented for stachking method “weighted_sum_c_diff”

(3) DBScan

  • using implementation in scikit-learn with our precalculated distance matrice: Scikit-learn: Machine Learning in Python, Pedregosa et al., JMLR 12, pp. 2825-2830, 2011. https://scikit-learn.org/stable/modules/clustering.html#dbscan

  • DBSCAN (Ester et al., 1996): Clusters can have any shape, based on densities. The number of clusters are not predefined. Two samples belong to one cluster, if their distance is less than eps. core points are objects, that have at least min_samples within the distance eps. A point p is directly reachable from a point q, if it is within distance eps, and reachable from point q if there is a path connecting p and q via directly connected points. Points, that do not lie within distance eps of any other point do not belong to any cluster (outliers/ noise). All points within a cluster are density connected and any point that is density-reachable to any point of the cluster is part of the cluster.