Source code for clustering

#!/usr/bin/python
"""
Functions to cluster seismograms by a range of constraints.

Copyright 2015 Calum Chamberlain

This file is part of EQcorrscan.

    EQcorrscan is free software: you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
    the Free Software Foundation, either version 3 of the License, or
    (at your option) any later version.

    EQcorrscan is distributed in the hope that it will be useful,
    but WITHOUT ANY WARRANTY; without even the implied warranty of
    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
    GNU General Public License for more details.

    You should have received a copy of the GNU General Public License
    along with EQcorrscan.  If not, see <http://www.gnu.org/licenses/>.

"""

# I want to make sure that the lags for those that I cluster are similar as well
# as making sure that they are similarly correlated - will cluster based on
# cross-channel correlation sum
import numpy as np
import warnings

[docs]def cross_chan_coherence(st1, st2, i=0): """ Function to determine the cross-channel coherancy between two streams of multichannel seismic data. :type st1: obspy Stream :param st1: Stream one :type st2: obspy Stream :param st2: Stream two :type i: int :param i: index used for parallel async processing, returned unaltered :returns: cross channel coherence, float - normalized by number of channels, if i, returns tuple of (cccoh, i) where i is int, as intput. """ from eqcorrscan.core.match_filter import normxcorr2 cccoh=0.0 kchan=0 for tr in st1: tr1=tr.data # Assume you only have one waveform for each channel tr2=st2.select(station=tr.stats.station, \ channel=tr.stats.channel) if tr2: cccoh+=normxcorr2(tr1,tr2[0].data)[0][0] kchan+=1 if kchan: cccoh=cccoh/kchan return (cccoh, i) else: warnings.warn('No matching channels') return (0, i)
[docs]def distance_matrix(stream_list, cores=1): """ Function to compute the distance matrix for all templates - will give distance as 1-abs(cccoh), e.g. a well correlated pair of templates will have small distances, and an equally well correlated reverse image will have the same distance as apositively correlated image - this is an issue :type stream_list: List of obspy.Streams :param tream_list: List of the streams to compute the distance matrix for :type core: int :param cores: Number of cores to parallel process using, defaults to 1. :returns: ndarray - distance matrix """ from multiprocessing import Pool # Initialize square matrix dist_mat=np.array([np.array([0.0]*len(stream_list))]*len(stream_list)) for i, master in enumerate(stream_list): # Start a parallel processing pool pool=Pool(processes=cores) # Parallel processing results=[pool.apply_async(cross_chan_coherence, args=(master,\ stream_list[j], j))\ for j in range(len(stream_list))] pool.close() # Extract the results when they are done dist_list=[p.get() for p in results] # Close and join all the processes back to the master process pool.join() # Sort the results by the input j dist_list.sort(key=lambda tup:tup[1]) # Sort the list into the dist_mat structure for j in range(i,len(stream_list)): if i==j: dist_mat[i,j]=0.0 else: dist_mat[i,j]=1-dist_list[j][0] # Reshape the distance matrix for i in range(1,len(stream_list)): for j in range(i): dist_mat[i,j]=dist_mat.T[i,j] return dist_mat
[docs]def cluster(stream_list, show=True, corr_thresh=0.3, save_corrmat=False,\ cores='all', debug=1): """ Function to take a set of templates and cluster them, will return groups as lists of streams. Clustering is done by computing the cross-channel correlation sum of each stream in stream_list with every other stream in the list. Scipy.cluster.hierachy functions are then used to compute the complete distance matrix, where distance is 1 minus the normalised cross-correlation sum such that larger distances are less similar events. Groups are then created by clustering the distance matrix at distances less than 1 - corr_thresh. Will compute the distance matrix in parallel, using all available cores :type stream_list: List of Obspy.Stream :param stream_list: List of templates to compute clustering for :type show: bool :param show: plot linkage on screen if True, defaults to True :type corr_thresh: float :param corr_thresh: Cross-channel correlation threshold for grouping :type save_corrmat: bool :param save_corrmat: If True will save the distance matrix to dist_mat.npy :type cores: int :param cores: numebr of cores to use when computing the distance matrix,\ defaults to 'all' which will work out how many cpus are available\ and hog them. :type debug: int :param debug: Level of debugging from 1-5, higher is more output, currently\ only level 1 implimented. :returns: List of groups with each group a list of streams making up\ that group. """ from scipy.spatial.distance import squareform from scipy.cluster.hierarchy import linkage, dendrogram, fcluster import matplotlib.pyplot as plt from multiprocessing import cpu_count if cores=='all': num_cores=cpu_count() else: num_cores=cores # Compute the distance matrix if debug >= 1: print 'Computing the distance matrix using '+str(num_cores)+' cores' dist_mat=distance_matrix(stream_list, cores=num_cores) if save_corrmat: np.save('dist_mat.npy', dist_mat) if debug >= 1: print 'Saved the distance matrix as dist_mat.npy' dist_vec=squareform(dist_mat) # plt.matshow(dist_mat, aspect='auto', origin='lower', cmap=pylab.cm.YlGnB) if debug >= 1: print 'Computing linkage' Z = linkage(dist_vec) if show: if debug >= 1: print 'Plotting the dendrogram' D = dendrogram(Z, color_threshold = 1 - corr_thresh,\ distance_sort='ascending') plt.show() # Get the indeces of the groups if debug >= 1: print 'Clustering' indeces = fcluster(Z, 1 - corr_thresh, 'distance') group_ids=list(set(indeces)) # Unique list of group ids if debug >= 1: print 'Found '+str(len(group_ids))+' groups' # Convert to tuple of (group id, stream id) indeces=[(indeces[i], i) for i in xrange(len(indeces))] # Sort by group id indeces.sort(key=lambda tup:tup[0]) groups=[] if debug >= 1: print 'Extracting and grouping' for group_id in group_ids: group=[] for ind in indeces: if ind[0] == group_id: group.append(stream_list[ind[1]]) elif ind[0] > group_id: # Because we have sorted by group id, when the index is greater # than the group_id we can break the inner loop. # Patch applied by CJC 05/11/2015 groups.append(group) break return groups
[docs]def group_delays(stream_list): """ Function to group template waveforms according to their delays :type stream_list: List of obspy.Stream :param stream_list: List of the waveforms you want to group :returns: List of List of obspy.Streams where each initial list is a group\ with the same delays """ groups=[] group_delays=[] group_chans=[] # Sort templates by number of channels stream_list=[(st, len(st)) for st in stream_list] stream_list.sort(key=lambda tup:tup[1]) stream_list=[st[0] for st in stream_list] for i in xrange(1,len(stream_list)): print 'Working on waveform '+str(i)+' of '+str(len(stream_list)) # Calculate the delays starttimes=[] chans=[] for tr in stream_list[i]: starttimes.append(tr.stats.starttime) chans.append((tr.stats.station, tr.stats.channel)) # This delays calculation will be an issue if we have changes in channels delays=[starttimes[m]-min(starttimes) for m in xrange(len(starttimes))] delays=[round(d,2) for d in delays] if len(groups)==0: groups.append([stream_list[i]]) group_delays.append(delays) group_chans.append(chans) else: j=0 match=False while not match: kmatch=0 # Find the set of shared stations and channels shared_chans=[] shared_delays_slave=[] shared_delays_master=[] k=0 for chan in chans: if chan in group_chans[j]: shared_chans.append(chan) shared_delays_slave.append(delays[k]) shared_delays_master.append(group_delays[j][group_chans[j].index(chan)]) k+=1 # Normalize master and slave delay times shared_delays_slave=[delay-min(shared_delays_slave)\ for delay in shared_delays_slave] shared_delays_master=[delay-min(shared_delays_master)\ for delay in shared_delays_master] for k in xrange(len(shared_chans)): # Check if the channel and delay match another group if shared_delays_slave[k]==shared_delays_master[k]: kmatch+=1 # increase the match index if kmatch==len(shared_chans): # If all the channels match, add it to the group groups[j].append(stream_list[i]) match=True elif j<len(groups)-1: j+=1 else: # Create a new group and break the loop groups.append([stream_list[i]]) group_delays.append(delays) group_chans.append(chans) match=True # Use this to break the loop return groups
[docs]def SVD(stream_list): """ Function to compute the SVD of a number of templates and return the singular vectors and singular values of the templates. :type stream_list: List of Obspy.Stream :param stream_list: List of the templates to be analysed :return: SVector(list of ndarray), SValues(list) for each channel, \ Uvalues(list of ndarray) for each channel, \ stachans, List of String (station.channel) .. rubric:: Note It is recommended that you align the data before computing the SVD, e.g., the P-arrival on all templates for the same channel should appear at the same time in the trace. """ # Convert templates into ndarrays for each channel # First find all unique channels: stachans=[] for st in stream_list: for tr in st: stachans.append(tr.stats.station+'.'+tr.stats.channel) stachans=list(set(stachans)) # Initialize a list for the output matrices, one matrix per-channel SValues=[] SVectors=[] Uvectors=[] for stachan in stachans: chan_mat=[stream_list[i].select(station=stachan.split('.')[0], \ channel=stachan.split('.')[1])[0].data \ for i in range(len(stream_list)) if \ len(stream_list[i].select(station=stachan.split('.')[0], \ channel=stachan.split('.')[1])) != 0] # chan_mat=[chan_mat[i]/np.max(chan_mat[i]) for i in xrange(len(chan_mat))] chan_mat=np.asarray(chan_mat) print chan_mat.shape U, s, V = np.linalg.svd(chan_mat, full_matrices=False) SValues.append(s) SVectors.append(V) Uvectors.append(U) return SVectors, SValues, Uvectors, stachans
[docs]def empirical_SVD(stream_list, linear=True): """ Empirical subspace detector generation function. Takes a list of templates and computes the stack as the first order subspace detector, and the differential of this as the second order subspace detector following the emprical subspace method of Barrett & Beroza, 2014 - SRL. :type stream_list: list of stream :param stream_list: list of template streams to compute the subspace detectors\ from :type linear: Bool :param linear: Set to true by default to compute the linear stack as the\ first subspace vector, False will use the phase-weighted stack as the\ first subspace vector. :returns: list of two streams """ from eqcorrscan.utils import stacking if linear: first_subspace=stacking.linstack(stream_list) second_subspace=first_subspace.copy() for i in range(len(second_subspace)): second_subspace[i].data=np.diff(second_subspace[i].data) second_subspace[i].stats.starttime+=0.5*second_subspace[i].stats.delta return [first_subspace, second_subspace]
[docs]def SVD_2_stream(SVectors, stachans, k, sampling_rate): """ Function to convert the singular vectors output by SVD to streams, one for each singular vector level, for all channels. :type SVectors: List of np.ndarray :param SVectors: Singular vectors :type stachans: List of Strings :param stachans: List of station.channel Strings :type k: int :param k: Number of streams to return = number of SV's to include :type sampling_rate: float :param sampling_rate: Sampling rate in Hz :returns: SVstreams, List of Obspy.Stream, with SVStreams[0] being composed of the highest rank singular vectors. """ from obspy import Stream, Trace SVstreams=[] for i in range(k): SVstream=[] for j, stachan in enumerate(stachans): SVstream.append(Trace(SVectors[j][i], \ header={'station': stachan.split('.')[0], 'channel': stachan.split('.')[1], 'sampling_rate': sampling_rate})) SVstreams.append(Stream(SVstream)) return SVstreams
[docs]def corr_cluster(trace_list, thresh=0.9): """ Group traces based on correlations above threshold with the stack - will run twice, once with a lower threshold, then again with your threshold to remove large outliers :type trace_list: List of :class:obspy.Trace :param trace_list: Traces to compute similarity between :type thresh: float :param thrsh: Correlation threshold between -1-1 :returns: np.ndarray of bool """ from eqcorrscan.utils import stacking from obspy import Stream from core.match_filter import normxcorr2 stack=stacking.linstack([Stream(tr) for tr in trace_list])[0] output=np.array([False]*len(trace_list)) group1=[] for i, tr in enumerate(trace_list): if normxcorr2(tr.data,stack.data)[0][0] > 0.6: output[i]=True group1.append(tr) stack=stacking.linstack([Stream(tr) for tr in group1])[0] group2=[] for i, tr in enumerate(trace_list): if normxcorr2(tr.data,stack.data)[0][0] > thresh: group2.append(tr) output[i]=True else: output[i]=False return output
[docs]def extract_detections(detections, templates, contbase_list, extract_len=90.0, \ outdir=None, extract_Z=True, additional_stations=[]): """ Function to extract the waveforms associated with each detection in a list of detections for the template, template. Waveforms will be returned as a list of obspy.Streams containing segments of extract_len. They will also be saved if outdir is set. The default is unset. The default extract_len is 90 seconds per channel. :type detections: List tuple of of :class: datetime.datetime, string :param detections: List of datetime objects, and their associated template\ name :type templates: List of tuple of string and :class: obspy.Stream :param templates: A list of the tuples of the template name and the template\ Stream used to detect detections. :type contbase_list: List of tuple of string :param contbase_list: List of tuples of the form ['path', 'type', 'network']\ Where path is the path to the continuous database, type is\ the directory structure, which can be either Yyyyy/Rjjj.01,\ which is the standard IRIS Year, julian day structure, or,\ yyyymmdd which is a single directory for every day. :type extract_len: float :param extract_len: Length to extract around the detection (will be equally\ cut around the detection time) in seconds. Default is 90.0. :type outdir: Bool or String :param outdir: Default is None, with None set, no files will be saved,\ if set each detection will be saved into this directory with files\ named according to the detection time, NOT than the waveform\ start time. Detections will be saved into template subdirectories. :type extract_Z: Bool :param extract_Z: Set to True to also extract Z channels for detections\ delays will be the same as horizontal channels, only applies if\ only horizontal channels were used in the template. :type additional_stations: List of tuple :param additional_stations: List of stations, chanels and networks to also\ extract data for using an average delay. :returns: List of :class: obspy.Stream """ from obspy import read, UTCDateTime, Stream from eqcorrscan.utils import pre_processing import datetime as dt import os from joblib import Parallel, delayed # Sort the template according to starttimes, needed so that stachan[i] # corresponds to delays[i] all_delays=[] # List of tuples of template name, delays all_stachans=[] for template in templates: templatestream=template[1].sort(['starttime']) stachans=[(tr.stats.station,tr.stats.channel,tr.stats.network) \ for tr in templatestream] mintime=templatestream[0].stats.starttime delays=[tr.stats.starttime-mintime for tr in templatestream] all_delays.append((template[0], delays)) all_stachans.append((template[0], stachans)) # Sort the detections and group by day detections.sort() detection_days=[detection[0].date() for detection in detections] detection_days=list(set(detection_days)) detection_days.sort() # Initialize output list detection_wavefiles=[] # Also include Z channels when extracting detections if extract_Z: new_all_stachans=[] new_all_delays=[] t=0 for template in all_stachans: stachans=template[1] delays=all_delays[t][1] new_stachans=[] new_delays=[] j=0 for i, stachan in enumerate(stachans): if j==1: new_stachans.append((stachan[0], stachan[1][0]+'Z',\ stachan[2])) new_delays.append(delays[i]) new_stachans.append(stachan) new_delays.append(delays[i]) j=0 else: new_stachans.append(stachan) new_delays.append(delays[i]) j+=1 new_all_stachans.append((template[0], new_stachans)) new_all_delays.append((template[0], new_delays)) t+=1 all_delays=new_all_delays all_stachans=new_all_stachans if not len(additional_stations)==0: t=0 for template in all_stachans: av_delay=np.mean(all_delays[t][1]) for sta in additional_stations: if not sta in template[1]: template[1].append(sta) all_delays[t][1].append(av_delay) t+=1 # Loop through the days for detection_day in detection_days: print 'Working on detections for day: '+str(detection_day) stachans=list(set([stachans[1] for stachans in all_stachans][0])) # List of all unique stachans - read in all data for stachan in stachans: contbase=[base for base in contbase_list\ if base[2]==stachan[2]][0] if contbase[1]=='yyyymmdd': dayfile=detection_day.strftime('%Y%m%d')+'/*'+stachan[0]+\ '.'+stachan[1][0]+'?'+stachan[1][-1]+'.*' elif contbase[1]=='Yyyyy/Rjjj.01': dayfile=detection_day.strftime('Y%Y/R%j.01')+'/'+stachan[0]+\ '.*.'+stachan[1][0]+'?'+stachan[1][-1]+'.'+detection_day.strftime('%Y.%j') if not 'st' in locals(): try: st=read(contbase[0]+'/'+dayfile) except: print 'No data for '+contbase[0]+'/'+dayfile else: try: st+=read(contbase[0]+'/'+dayfile) except: print 'No data for '+contbase[0]+'/'+dayfile st.merge(fill_value='interpolate') # We now have a stream of day long data, we should process it! # st=Parallel(n_jobs=3)(delayed(pre_processing.dayproc)(tr, lowcut,\ # highcut,\ # filter_order,\ # samp_rate,\ # debug, detection_day)\ # for tr in st) # st=Stream(st) # for tr in st: # Convert to int32 for STEIM2 format # tr.data=tr.data.astype(np.int32) day_detections=[detection for detection in detections\ if detection[0].date() == detection_day] for detection in day_detections: template=detection[1] t_stachans=[stachans[1] for stachans in all_stachans \ if stachans[0] == template][0] t_delays=[delays[1] for delays in all_delays\ if delays[0] == template][0] print 'Cutting for detections at: '+detection[0].strftime('%Y/%m/%d %H:%M:%S') detect_wav=st.copy() for tr in detect_wav: tr.trim(starttime=UTCDateTime(detection[0])-extract_len/2,\ endtime=UTCDateTime(detection[0])+extract_len/2) if outdir: if not os.path.isdir(outdir+'/'+template): os.makedirs(outdir+'/'+template) detect_wav.write(outdir+'/'+template+'/'+\ detection[0].strftime('%Y-%m-%d_%H-%M-%S')+\ '.ms', format='MSEED', encoding='STEIM2') # '.ms', format='MSEED', encoding='STEIM2') print 'Written file: '+outdir+'/'+template+'/'+\ detection[0].strftime('%Y-%m-%d_%H-%M-%S')+'.ms' if not outdir: detection_wavefiles.append(detect_wav) del detect_wav del st if outdir: detection_wavefiles=[] if not outdir: return detection_wavefiles else: return
[docs]def space_time_cluster(detections, t_thresh, d_thresh): """ Function to cluster detections in space and time, use to seperate repeaters from other events :type detections: List :param detections: List of tuple of tuple of location (lat, lon, depth (km)),\ and time as a datetime object :type t_thresh: float :param t_thresh: Maximum inter-event time threshold in seconds :type d_thresh: float :param d_thresh: Maximum inter-event distance in km :returns: List of tuple (detections, clustered) and list of indeces of\ clustered detections """ from eqcorrscan.utils.mag_calc import dist_calc import datetime as dt # Ensure they are sorted by time first, not that we need it. detections.sort(key=lambda tup:tup[1]) clustered=[] clustered_indeces=[] for master_ind, master in enumerate(detections): keep=False for slave in detections: if not master==slave and\ abs((master[1] - slave[1]).total_seconds()) <= t_thresh and \ dist_calc(master[0], slave[0]) <= d_thresh: # If the slave events is close in time and space to the master # keep it and break out of the loop. keep=True break if keep: clustered.append(master) clustered_indeces.append(master_ind) return clustered, clustered_indeces
[docs]def re_thresh_csv(path, old_thresh, new_thresh, chan_thresh): """ Function to remove detections by changing the threshold, can only be done to remove detection by increasing threshold, threshold lowering will have no affect. :type path: Str :param path: Path to the .csv detection file :type old_thresh: float :param old_thresh: Old threshold MAD multiplier :type new_thresh: float :param new_thresh: New threhsold MAD multiplier :type chan_thresh: int :param chan_thresh: Minimum number of channels for a detection returns: List of detections """ f=open(path,'r') old_thresh=float(old_thresh) new_thresh=float(new_thresh) # Be nice, ensure that the thresholds are float detections=[] detections_in=0 detections_out=0 for line in f: if not line.split(', ')[0]=='template' and len(line) >2: detections_in+=1 if abs(float(line.split(', ')[3])) >=\ (new_thresh/old_thresh)*float(line.split(', ')[2]) and\ int(line.split(', ')[4]) >= chan_thresh: detections_out+=1 detections.append(line.split(', ')) print 'Read in '+str(detections_in)+' detections' print 'Left with '+str(detections_out)+' detections' return detections