'''
:copyright:
The SeisMIC development team (makus@gfz-potsdam.de).
:license:
EUROPEAN UNION PUBLIC LICENCE v. 1.2
(https://joinup.ec.europa.eu/collection/eupl/eupl-text-eupl-12)
:author:
Peter Makus (makus@gfz-potsdam.de)
Created: Thursday, 3rd June 2021 04:15:57 pm
Last Modified: Friday, 29th September 2023 11:06:52 am
'''
from copy import deepcopy
import json
import logging
import os
from typing import Generator, List, Tuple
import warnings
import yaml
import fnmatch
from glob import glob
from mpi4py import MPI
import numpy as np
from obspy import UTCDateTime
from tqdm import tqdm
from seismic.db.corr_hdf5 import CorrelationDataBase
from seismic.monitor.dv import DV, read_dv
from seismic.monitor.wfc import WFC
from seismic.utils.miic_utils import log_lvl
[docs]class Monitor(object):
def __init__(self, options: dict | str):
"""
Object that handles the computation of seismic velocity changes.
This will access correlations that have been computed previously with
the :class:`~seismic.correlate.correlate.Correlator` class.
Takes the parameter file as input (either as yaml or as dict).
:param options: Parameters for the monitor. Usually, they come in form
of a *yaml* file. If this parameter is a str, it will be
interpreted to be the path to said *yaml* file.
:type options: dict or str
"""
if isinstance(options, str):
with open(options) as file:
options = yaml.load(file, Loader=yaml.FullLoader)
self.options = options
self.starttimes, self.endtimes = self._starttimes_list()
self.proj_dir = options['proj_dir']
self.outdir = os.path.join(
options['proj_dir'], options['dv']['subdir'])
self.indir = os.path.join(
options['proj_dir'], options['co']['subdir']
)
try:
self.save_comps_separately = options['save_comps_separately']
except KeyError:
# If it's not specified it's probably an old params.yaml
self.save_comps_separately = False
# init MPI
self.comm = MPI.COMM_WORLD
self.psize = self.comm.Get_size()
self.rank = self.comm.Get_rank()
# directories:
logdir = os.path.join(self.proj_dir, options['log_subdir'])
if self.rank == 0:
os.makedirs(self.outdir, exist_ok=True)
os.makedirs(logdir, exist_ok=True)
# Logging - rank dependent
loglvl = log_lvl[self.options['log_level']]
if self.rank == 0:
tstr = UTCDateTime.now().strftime('%Y-%m-%d-%H:%M')
else:
tstr = None
tstr = self.comm.bcast(tstr, root=0)
# replace colon in tstr
tstr = tstr.replace(':', '-')
self.logger = logging.getLogger(
"seismic.monitor.Monitor%s" % str(self.rank).zfill(3))
self.logger.setLevel(loglvl)
# also catch the warnings
logging.captureWarnings(True)
warnlog = logging.getLogger('py.warnings')
fh = logging.FileHandler(os.path.join(
logdir, f'monitor{tstr}rank{str(self.rank).zfill(3)}'))
fh.setLevel(loglvl)
self.logger.addHandler(fh)
warnlog.addHandler(fh)
fmt = logging.Formatter(
fmt='%(asctime)s - %(levelname)s - %(message)s')
fh.setFormatter(fmt)
consoleHandler = logging.StreamHandler()
consoleHandler.setFormatter(fmt)
self.logger.addHandler(consoleHandler)
# Write the options dictionary to the log file
if self.rank == 0:
opt_dump = deepcopy(options)
# json cannot write the UTCDateTime objects that might be in here
for step in opt_dump['co']['preProcessing']:
if 'stream_mask_at_utc' in step['function']:
startsstr = [
t.format_fissures() for t in step['args']['starts']]
step['args']['starts'] = startsstr
if 'ends' in step['args']:
endsstr = [
t.format_fissures() for t in step['args']['ends']]
step['args']['ends'] = endsstr
with open(os.path.join(
logdir, 'params%s.txt' % tstr), 'w') as file:
file.write(json.dumps(opt_dump, indent=1))
# Find available stations and network
self.netlist, self.statlist, self.infiles = \
self._find_available_corrs()
def _starttimes_list(self) -> Tuple[np.ndarray, np.ndarray]:
"""
Returns numpy array of starttimes and endtimes, for which each data
point of velocity change should be computed.
:return: A numpy array holding the starttimes and endtimes of the
computing time windows in seconds (as UTC timestamps).
:rtype: Tuple[np.ndarray, np.ndarray]
..seealso:: For complete description,
:func:`~seismic.monitor.monitor.make_time_list`
"""
starttimes, endtimes = make_time_list(
self.options['dv']['start_date'], self.options['dv']['end_date'],
self.options['dv']['date_inc'], self.options['dv']['win_len'])
return starttimes, endtimes
def _find_available_corrs(self) -> Tuple[List[str], List[str], List[str]]:
"""
1. Finds the files of available correlations.
2. Out of those files, it will extract the ones that are requested.
:return: list of network combination codes, list of station combination
codes, and list of the file paths to their corresponding
correlation files.
:rtype: Tuple[List[str], List[str], List[str]]
"""
netlist, statlist, infiles = corr_find_filter(
self.indir, **self.options)
self.logger.info(
'Found correlation data for the following station '
+ 'and network combinations %s' % str(
['{n}.{s}'.format(n=n, s=s) for n, s in zip(
netlist, statlist)]))
return netlist, statlist, infiles
[docs] def compute_velocity_change(
self, corr_file: str, tag: str, network: str, station: str,
channel: str, ref_trcs: np.ndarray = None):
"""
Computes the velocity change for a cross (or auto) correlation
time-series. This process is executed "per station combination".
Meaning, several processes of this method can be started via MPI
and the method
:meth:`~seismic.monitor.monitor.Monitor.compute_velocity_change_bulk`,
which is the function that should rather be acessed than this one.
Data will be written into the defined dv folder and plotted if this
option was previously set to True.
:param corr_file: File to read the correlation data from.
:type corr_file: str
:param tag: Tag of the data in the file (almost always `'subdivision'`)
:type tag: str
:param network: Network combination code
:type network: str
:param station: Station Combination Code
:type station: str
:param channel: Channel combination code
:type channel: str
:param ref_trcs: Feed in on or several custom reference traces.
If None the program will determine a reference trace from
the chosen method in the config. Defaults to None
:type ref_trcs: np.ndarray, optional
"""
self.logger.info(
f'Computing velocity change for file: {corr_file} and channel: '
f' {channel}.')
with CorrelationDataBase(corr_file, mode='r') as cdb:
# get the corrstream containing all the corrdata for this combi
cst = cdb.get_data(network, station, channel, tag)
lts = cdb.get_corr_options()['corr_args']['lengthToSave']
tw_start = self.options['dv']['tw_start']
if self.options['dv']['compute_tt']:
if not hasattr(cst[0].stats, 'dist'):
warnings.warn(
f'{network}.{station} does not include distance. '
'SeisMIC will assume an interstation distance of 0.')
else:
# Assume flat earth to include topography
# Note that dist is in km and elevation information in m
d = np.sqrt(
cst[0].stats.dist**2
+ ((cst[0].stats.stel-cst[0].stats.evel)/1000)**2)
tt = round(
d/self.options['dv']['rayleigh_wave_velocity'], 0)
tw_start += tt
self.logger.info(
f'Computed travel time for {network}.{station} is '
f'{tt} s. The assumed direct-line distance was {d} km.')
if lts < tw_start + self.options['dv']['tw_len']:
reqtw = tw_start + self.options[
'dv']['tw_len']
raise ValueError(
'Requested lapse time window (time window start + time window'
f' length = {reqtw} contains segments after the Correlation'
'Function.'
' When computing the correlations make sure to set an '
f'appropriate value for lengthToSave. The value was {lts}.'
f' The direct-line distance between the stations is {d} km.'
)
if 'preprocessing' in self.options['dv']:
for func in self.options['dv']['preprocessing']:
# This one goes on the CorrStream
if func['function'] in ['pop_at_utcs', 'select_time']:
f = cst.__getattribute__(func['function'])
cst = f(**func['args'])
cb = cst.create_corr_bulk(
network=network, station=station, channel=channel, inplace=True)
# for possible rest bits
del cst
# Do the actual processing:
cb.normalize(normtype='absmax')
# That is were the stacking is happening
cb.resample(self.starttimes, self.endtimes)
cb.filter(
(self.options['dv']['freq_min'], self.options['dv']['freq_max']))
# Preprocessing on the correlation bulk
if 'preprocessing' in self.options['dv']:
for func in self.options['dv']['preprocessing']:
if func['function'] in ['pop_at_utcs', 'select_time']:
continue
f = cb.__getattribute__(func['function'])
cb = f(**func['args'])
# Retain a copy of the stats
stats_copy = deepcopy(cb.stats)
if self.options['dv']['tw_len'] is None:
trim0 = cb.stats.start_lag
trim1 = cb.stats.end_lag
cbt = cb
else:
trim0 = -(
tw_start+self.options['dv']['tw_len'])
trim1 = (
tw_start+self.options['dv']['tw_len'])
cbt = cb.copy().trim(trim0, trim1)
if cbt.data.shape[1] <= 20:
raise ValueError('CorrBulk extremely short.')
# 15.11.22
# Do trimming after stretching of the reference trace to avoid
# edges having an influence on dv/v estimate
if ref_trcs is None:
tr = cb.extract_multi_trace(**self.options['dv']['dt_ref'])
else:
tr = ref_trcs
# Compute time window
tw = [np.arange(
tw_start*cbt.stats['sampling_rate'],
trim1*cbt.stats['sampling_rate'], 1)]
# 2023/09/29
# sim_mat is not required in this operation
dv = cbt.stretch(
ref_trc=tr, return_sim_mat=False,
stretch_steps=self.options['dv']['stretch_steps'],
stretch_range=self.options['dv']['stretch_range'],
tw=tw, sides=self.options['dv']['sides'],
ref_tr_trim=(trim0, trim1), ref_tr_stats=stats_copy,
processing=self.options['dv'])
ccb = cb.correct_stretch(dv)
# extract the final reference trace (mean excluding very different
# traces)
if ref_trcs is None:
tr = ccb.extract_multi_trace(**self.options['dv']['dt_ref'])
ccb.trim(trim0, trim1)
# obtain an improved time shift measurement
dv = cbt.stretch(
ref_trc=tr, return_sim_mat=self.options['dv']['return_sim_mat'],
stretch_steps=self.options['dv']['stretch_steps'],
stretch_range=self.options['dv']['stretch_range'],
tw=tw, sides=self.options['dv']['sides'],
ref_tr_trim=(trim0, trim1), ref_tr_stats=stats_copy,
processing=self.options['dv'])
# Postprocessing on the dv object
if 'postprocessing' in self.options['dv']:
for func in self.options['dv']['postprocessing']:
f = dv.__getattribute__(func['function'])
dv = f(**func['args'])
outf = os.path.join(
self.outdir, 'DV-%s.%s.%s' % (network, station, channel))
dv.save(outf)
if self.options['dv']['plot_vel_change']:
fname = '%s_%s_%s' % (
dv.stats.network, dv.stats.station, dv.stats.channel)
savedir = os.path.join(
self.options['proj_dir'], self.options['fig_subdir'])
dv.plot(
save_dir=savedir, figure_file_name=fname,
normalize_simmat=True, sim_mat_Clim=[-1, 1])
[docs] def compute_velocity_change_bulk(self):
"""
Compute the velocity change for all correlations using MPI.
The parameters for the correlation defined in the *yaml* file will be
used.
This function will just call
:meth:`~seismic.monitor.monitor.Monitor.compute_velocity_change`
several times.
"""
tag = 'subdivision'
# get number of available channel combis
if self.rank == 0:
plist = []
for f, n, s in zip(self.infiles, self.netlist, self.statlist):
with CorrelationDataBase(f, mode='r') as cdb:
ch = cdb.get_available_channels(
tag, n, s)
plist.extend([f, n, s, c] for c in ch)
else:
plist = None
plist = self.comm.bcast(plist, root=0)
pmap = np.arange(len(plist))*self.psize/len(plist)
pmap = pmap.astype(np.int32)
ind = pmap == self.rank
ind = np.arange(len(plist), dtype=int)[ind]
# Assign a task to each rank
for ii in tqdm(ind):
corr_file, net, stat, cha = plist[ii]
try:
self.compute_velocity_change(
corr_file, tag, net, stat, cha)
except KeyError:
self.logger.exception(
f'No correlation data found for {net}.{stat}.{cha} with '
+ f'tag {tag} in file {corr_file}.'
)
except Exception as e:
self.logger.exception(f'{e} for file {corr_file}.')
[docs] def compute_components_average(self, method: str = 'AutoComponents'):
"""
Averages the Similarity matrix of different velocity changes in
the whole dv folder. Based upon those values, new dvs and correlations
are computed.
:param method: Which components should be averaged? Can be
'StationWide' if all component-combinations for one station should
be averaged, 'AutoComponents' if only the dv results of
autocorrelations are to be averaged, 'CrossComponents' if you wish
to average dvs from the same station but only intercomponent
corrrelations, or 'CrossStations' if cross-station correlations
should be averaged (same station combination and all component
combinations).
Defaults to 'AutoComponents'
:type method: str, optional
:raises ValueError: For Unknown combination methods.
"""
av_methods = (
'autocomponents', 'crosscomponents', 'stationwide',
'crossstations', 'betweenstations', 'betweencomponents')
if method.lower() not in av_methods:
raise ValueError('Averaging method not in %s.' % str(av_methods))
infiles = glob(os.path.join(self.outdir, '*.npz'))
if method.lower() == 'autocomponents':
ch = 'av-auto'
elif method.lower() in ('betweencomponents', 'crosscomponents'):
ch = 'av-xc'
else:
ch = 'av'
while len(infiles):
# Find files belonging to same station
pat = '.'.join(infiles[0].split('.')[:-2]) + '*'
filtfil = fnmatch.filter(infiles, pat)
fffil = [] # Second filter
for f in filtfil:
if ch in f.split('.'):
# Is already computed for this station
# So let's remove all of them from the initial list
for ff in filtfil:
try:
infiles.remove(ff)
except ValueError:
# Has already been removed by one of the other
# check
pass
filtfil.clear()
fffil.clear()
self.logger.debug('Skipping already averaged dv...%s' % f)
break
elif 'av' in f:
# computed by another averaging method
infiles.remove(f)
continue
components = f.split('.')[-2].split('-')
stations = f.split('.')[-3].split('-')
if method.lower() == 'autocomponents':
# Remove those from combined channels
if components[0] != components[1] \
or stations[0] != stations[1]:
infiles.remove(f)
continue
elif method.lower() in (
'betweencomponents', 'crosscomponents'):
# Remove those from equal channels
components = f.split('.')[-2].split('-')
if components[0] == components[1] \
or stations[0] != stations[1]:
infiles.remove(f)
continue
elif method.lower() == 'stationwide':
if stations[0] != stations[1]:
infiles.remove(f)
continue
elif method.lower() in ('betweenstations', 'crossstations'):
if stations[0] == stations[1]:
infiles.remove(f)
continue
fffil.append(f)
dvs = []
for f in fffil:
try:
dvs.append(read_dv(f))
except Exception:
self.logger.exception(
'An unexpected error has ocurred '
+ 'while reading file %s.' % f)
# Remove so they are not processed again
infiles.remove(f)
if not len(dvs):
continue
elif len(dvs) == 1:
self.logger.warn(
'Only component %s found for station %s.%s... Skipping.'
% (dvs[0].stats.channel, dvs[0].stats.network,
dvs[0].stats.station))
continue
dv_av = average_components(dvs)
outf = os.path.join(
self.outdir, 'DV-%s.%s.%s' % (
dv_av.stats.network, dv_av.stats.station, ch))
dv_av.save(outf)
if self.options['dv']['plot_vel_change']:
# plot if desired
fname = '%s_%s_%s' % (
dv_av.stats.network, dv_av.stats.station, ch)
savedir = os.path.join(
self.options['proj_dir'], self.options['fig_subdir'])
dv_av.plot(
save_dir=savedir, figure_file_name=fname,
normalize_simmat=True, sim_mat_Clim=[-1, 1])
[docs]def make_time_list(
start_date: str, end_date: str, date_inc: int, win_len: int) -> Tuple[
np.ndarray, np.ndarray]:
"""
Returns numpy array of starttimes and endtimes, for which each data point
of velocity change should be computed.
:param start_date: Date (including time) to start monitoring
:type start_date: str
:param end_date: Last date (including time) of monitoring
:type end_date: str
:param date_inc: Increment in seconds between each datapoint
:type date_inc: int
:param win_len: Window Length, for which each data point is computed.
:type win_len: int
:return: A numpy array holding the starttimes and endtimes of the computing
time windows in seconds (as UTC timestamps)
:rtype: Tuple[ np.ndarray, np.ndarray]
..note::
see `obspy's documentation
<https://docs.obspy.org/packages/autogen/obspy.core.utcdatetime.UTCDateTime.html>`
for compatible input strings.
"""
if date_inc <= 0 or win_len <= 0:
raise ValueError(
'Negative increments or window length are not allowed.')
start = UTCDateTime(start_date).timestamp
end = UTCDateTime(end_date).timestamp
if start >= end:
raise ValueError('Start Date should be earlier than end date.')
inc = date_inc
starttimes = np.arange(start, end, inc)
endtimes = starttimes + win_len
return starttimes, endtimes
[docs]def corr_find_filter(
indir: str, net: dict, **kwargs) -> Tuple[
List[str], List[str], List[str]]:
"""
1. Finds the files of available correlations.
2. Out of those files, it will extract the ones that are requested.
:param indir: Directory that the correlations are saved in.
:type indir: str
:param net: Network dictionary from *yaml* file.
:type net: dict
:return: list of network combination codes, list of station combination
codes, and list of the file paths to their corresponding
correlation files.
:rtype: Tuple[List[str], List[str], List[str]]
"""
netlist = []
statlist = []
infiles = []
for f in glob(os.path.join(indir, '*-*.*-*.h5')):
matched = False
f = os.path.basename(f)
split = f.split('.')
# Find the files that should actually be processed
nsplit = split[0].split('-')
ssplit = split[1].split('-')
if isinstance(net['network'], str) and not fnmatch.filter(
nsplit, net['network']) == nsplit:
continue
elif isinstance(net['network'], list):
# account for the case of wildcards in the network list
for n in net['network']:
if fnmatch.filter(nsplit, n) == nsplit:
matched = True
break
if not matched:
continue
if isinstance(net['station'], str) and not fnmatch.filter(
ssplit, net['station']) == ssplit:
continue
elif isinstance(net['station'], list):
# account for the case of wildcards in the station list
for s in net['station']:
if fnmatch.filter(ssplit, s) == ssplit:
matched = True
break
if not matched:
continue
netlist.append(split[0])
statlist.append(split[1])
infiles.append(os.path.join(indir, f))
return netlist, statlist, infiles
[docs]def correct_dv_shift(
dv0: DV, dv1: DV, method: str = 'mean',
n_overlap: int = 0) -> Tuple[DV, DV]:
"""
Shift dv0 with respect to dv1, so that the same times will have the same
value.
.. note::
in-place operation.
:param dv0: DV object to be shifted
:type dv0: DV
:param dv1: DV object to compare to
:type dv1: DV
:param method: either mean or median, defaults to 'mean'. If mean is chosen
the points will be weighted by the correlation coefficient.
:type method: str, optional
:param n_overlap: Number of points to compare to beyond the
already overlapping point. Makes sense if, e.g., one station replaces
the other, defaults to 0.
:type n_overlap: int, optional
:raises ValueError: If the two dvs only have non-nan value more than
n_overlap apart.
:return: The two dv object. Only the first one is modified. Modification
also happens in-place.
:rtype: Tuple[DV, DV]
"""
# Times when both measured
both_avail = np.all(np.vstack((dv0.avail, dv1.avail)), axis=0)
if np.any(both_avail):
# Unless one of them become less than 0 or more than length
start = np.where(both_avail)[0].min() - n_overlap
stop = np.where(both_avail)[0].max() + n_overlap
ii = jj = 0
elif (
np.array(dv0.stats.starttime)[dv0.avail][0]
> np.array(dv1.stats.starttime)[dv1.avail][-1]):
# dv0 comes after dv1:
ii = np.where(
np.array(dv0.stats.starttime) == np.array(
dv0.stats.starttime)[dv0.avail][0])[0][0]
stop = ii + n_overlap
jj = np.where(
np.array(dv1.stats.starttime) == np.array(
dv1.stats.starttime)[dv1.avail][-1])[0][0]
start = jj - n_overlap
else:
# dv1 comes after dv0:
ii = np.where(
np.array(dv1.stats.starttime) == np.array(
dv1.stats.starttime)[dv1.avail][0])[0][0]
stop = ii + n_overlap
jj = np.where(
np.array(dv0.stats.starttime) == np.array(
dv0.stats.starttime)[dv0.avail][-1])[0][0]
start = jj - n_overlap
if ii-jj > n_overlap:
raise ValueError(
'The gap in active times of the two DVs is larger than '
'n_overlap, i.e., '
f'{n_overlap*(dv0.stats.starttime[1]-dv0.stats.starttime[0])/3600}'
'hours. The result would not make sense.')
if stop >= len(dv0.value):
stop = len(dv0.value) - 1
if start < 0:
start = 0
dv0c = dv0.corr[start:stop]
dv0v = dv0.value[start:stop]
dv1c = dv1.corr[start:stop]
dv1v = dv1.value[start:stop]
if method == 'mean':
shift = np.nanmean(dv0v*dv0c)/np.nanmean(dv0c)\
- np.nanmean(dv1v*dv1c)/np.nanmean(dv1c)
elif method == 'median':
shift = np.nanmedian(
dv0.value[start:stop])-np.nanmedian(dv1.value[start:stop])
else:
raise ValueError('Method has to be either mean or median')
roll = int(round(-shift/(dv0.second_axis[1]-dv0.second_axis[0])))
dv0.sim_mat = np.roll(dv0.sim_mat, (roll, 0))
dv0.value = dv0.second_axis[
np.nanargmax(np.nan_to_num(dv0.sim_mat), axis=1)]
return dv0, dv1
[docs]def correct_time_shift_several(dvs: List[DV], method: str, n_overlap: int):
"""
Suppose you have a number of dv/v time series from the same location,
but station codes changed several times. To stitch these together,
this code will iteratively shift each progressive time-series to the
shift of the end of its precessor.
.. note::
The shifting correction happens in-place.
:param dvs: List of dvs to be shifted
:type dvs: List[DV]
:param method: method to be used to measure the shift. Can be 'mean' or
'median'. If 'mean' is chosen, the points will be weighted by their
respective correlation coefficient.
:type method: str
:param n_overlap: amount of points beyond the overlap to use to compute
the shift.
:type n_overlap: int
"""
# If three different time-series are provided, find the one that
# spans the middle time
avl = np.array([
np.mean(np.array([
t.timestamp for t in dv.stats.corr_start])[
dv.avail]) for dv in dvs])
avl_u = np.sort(np.unique(avl))
# Loop over each unique start and shift and correct them iteratively
for k, avstart in enumerate(avl_u):
if k+1 == len(avl_u):
# everything is shifted
break
ii_ref = np.concatenate(np.where(avl == avstart))[0]
ii_corr = np.concatenate(np.where(avl == avl_u[k+1]))
dv_r = dvs[ii_ref]
dv_corr = [dvs[ii] for ii in ii_corr]
for dv in dv_corr:
try:
# The correction should happen in-place
correct_dv_shift(
dv, dv_r, method=method,
n_overlap=n_overlap)
except ValueError as e:
warnings.warn(
f'{e} for {dv.stats.id} and reference dv '
f'{dv_r.stats.id}.')
[docs]def average_components_mem_save(
dvs: Generator[DV, None, None], save_scatter: bool = True) -> DV:
"""
Averages the Similariy matrix of the DV objects. Based on those,
it computes a new dv value and a new correlation value. Less memory intense
but slower than :func:`average_components`. Should be used if you run
out of memory using the former. Uses the sum of squares rule for the
standard deviation.
:param dvs: Iterator over dvs from the different components to compute an
average from. Note that it is possible to use almost anything as
input as long as the similarity matrices of the dvs have the same shape
:type dvs: Iterator[class:`~seismic.monitor.dv.DV`]
:param save_scatter: Saves the scattering of old values and correlation
coefficients to later provide a statistical measure. Defaults to True.
:type save_scatter: bool, optional
:raises TypeError: for DVs that were computed with different methods
:return: A single dv with an averaged similarity matrix.
:rtype: DV
"""
statsl = []
corrs = []
stretches = []
for ii, dv in enumerate(dvs):
if ii == 0:
sim_mat_sum = np.zeros_like(dv.sim_mat)
n_stat = np.zeros_like(dv.corr)
stats = deepcopy(dv.stats)
strvec = dv.second_axis
value_type = dv.value_type
method = dv.method
dv_processing = dv.dv_processing
if dv.method != method:
raise TypeError('DV has to be computed with the same method.')
if 'av' in dv.stats.channel+dv.stats.network+dv.stats.station:
warnings.warn('Averaging of averaged dvs not allowed. Skipping dv')
continue
if dv.sim_mat.shape != sim_mat_sum.shape or any(
dv.second_axis != strvec):
warnings.warn(
'The shapes of the similarity matrices of the input DVs '
+ 'vary. Make sure to compute the dvs with the same parameters'
+ ' (i.e., start & end dates, date-inc, stretch increment, '
+ 'and stretch steps.\n\nThis dv will be skipped'
)
continue
sim_mat_sum += np.nan_to_num(dv.sim_mat)
n_stat[~np.isnan(dv.value)] += 1
if save_scatter:
corrs.append(dv.corr)
stretches.append(dv.value)
statsl.append(dv.stats)
# Now inf where it was nan before
av_sim_mat = (sim_mat_sum.T/n_stat).T
av_sim_mat[np.isinf(av_sim_mat)] = np.nan
# Now we would have to recompute the dv value and corr value
iimax = np.nanargmax(np.nan_to_num(av_sim_mat), axis=1)
corr = np.nanmax(av_sim_mat, axis=1)
dt = strvec[iimax]
dt[np.isnan(corr)] = np.nan
if save_scatter:
# use sum of square for std
corrs = np.array(corrs)
stretches = np.array(stretches)
else:
corrs = None
stretches = None
if not all(np.array([st.channel for st in statsl]) == stats.channel):
stats['channel'] = 'av'
if not all(np.array([st.station for st in statsl]) == stats.station):
stats['station'] = 'av'
if not all(np.array([st.network for st in statsl]) == stats.network):
stats['network'] = 'av'
dvout = DV(
corr, dt, value_type, av_sim_mat, strvec,
method, stats, stretches=stretches, corrs=corrs,
n_stat=n_stat, dv_processing=dv_processing)
return dvout
[docs]def average_components(
dvs: List[DV], save_scatter: bool = True, correct_shift: bool = False,
correct_shift_method: str = 'mean',
correct_shift_overlap: int = 0) -> DV:
"""
Averages the Similariy matrix of the DV objects. Based on those,
it computes a new dv value and a new correlation value.
:param dvs: List of dvs from the different components to compute an
average from. Note that it is possible to use almost anything as
input as long as the similarity matrices of the dvs have the same shape
:type dvs: List[class:`~seismic.monitor.dv.DV`]
:param save_scatter: Saves the scattering of old values and correlation
coefficients to later provide a statistical measure. Defaults to True.
:type save_scatter: bool, optional
:param correct_shift: Shift the dvs with respect to each other. This is
relevant if not all stations are active all the time and the references
are not describing the same state. This only works if stations were
active almost simultaneously at least for a bit
(i.e., corr-start[n+n_overlap]-corr-start[n])
:raises TypeError: for DVs that were computed with different methods
:return: A single dv with an averaged similarity matrix.
:rtype: DV
.. seealso::
If you should get an `OutOfMemoryError` consider using
:func:`average_components_mem_save`.
"""
dv_use = []
for dv in dvs:
if dv.method != dvs[0].method:
raise TypeError('DV has to be computed with the same method.')
if 'av' in dv.stats.channel+dv.stats.network+dv.stats.station:
warnings.warn('Averaging of averaged dvs not allowed. Skipping dv')
continue
if dv.sim_mat.shape != dvs[0].sim_mat.shape or any(
dv.second_axis != dvs[0].second_axis):
warnings.warn(
'The shapes of the similarity matrices of the input DVs '
+ 'vary. Make sure to compute the dvs with the same parameters'
+ ' (i.e., start & end dates, date-inc, stretch increment, '
+ 'and stretch steps.\n\nThis dv will be skipped'
)
continue
dv_use.append(dv)
# Correct shift
if correct_shift:
correct_time_shift_several(
dv_use, method=correct_shift_method,
n_overlap=correct_shift_overlap)
av_sim_mat = np.nanmean([dv.sim_mat for dv in dv_use], axis=0)
# Now we would have to recompute the dv value and corr value
iimax = np.nanargmax(np.nan_to_num(av_sim_mat), axis=1)
corr = np.nanmax(av_sim_mat, axis=1)
strvec = dv_use[0].second_axis
dt = strvec[iimax]
dt[np.isnan(corr)] = np.nan
if save_scatter:
stretches = np.array([dv.value for dv in dv_use])
corrs = np.array([dv.corr for dv in dv_use])
else:
stretches = None
corrs = None
# Number of stations per corr_start
n_stat = np.array(
[dv.avail*np.ones_like(dv.corr, dtype=int) for dv in dv_use])
n_stat = np.sum(n_stat, axis=0)
stats = deepcopy(dv_use[0].stats)
if not all(np.array([dv.stats.channel for dv in dv_use]) == stats.channel):
stats['channel'] = 'av'
if not all(np.array([dv.stats.station for dv in dv_use]) == stats.station):
stats['station'] = 'av'
if not all(np.array([dv.stats.network for dv in dv_use]) == stats.network):
stats['network'] = 'av'
dvout = DV(
corr, dt, dv_use[0].value_type, av_sim_mat, strvec,
dv_use[0].method, stats, stretches=stretches, corrs=corrs,
n_stat=n_stat, dv_processing=dv_use[0].dv_processing)
return dvout
[docs]def average_dvs_by_coords(
dvs: List[DV], lat: Tuple[float, float], lon: Tuple[float, float],
el: Tuple[float, float] = (-1e6, 1e6), return_std: bool = False) -> DV:
"""
Averages the Similariy matrix of the DV objects if they are inside of the
queried coordionates. Based on those,
it computes a new dv value and a new correlation value.
:param dvs: List of dvs from the different components to compute an
average from. Note that it is possible to use almost anything as
input as long as the similarity matrices of the dvs have the same shape
:type dvs: List[class:`~seismic.monitor.dv.DV`]
:param lat: Minimum and maximum latitude of the stations(s). Note that
for cross-correlations both stations must be inside of the window.
:type lat: Tuple[float, float]
:param lon: Minimum and maximum longitude of the stations(s). Note that
for cross-correlations both stations must be inside of the window.
:type lon: Tuple[float, float]
:param el: Minimum and maximum elevation of the stations(s) in m. Note that
for cross-correlations both stations must be inside of the window.
Defaults to (-1e6, 1e6) - i.e., no filter.
:type el: Tuple[float, float], optional
:param return_std: Save the standard deviation of the similarity
matrices into the dv object. Defaults to False.
:type return_std: bool, optional
:raises TypeError: for DVs that were computed with different methods
:return: A single dv with an averaged similarity matrix.
:rtype: DV
"""
dv_filt = []
for dv in dvs:
s = dv.stats
if s.channel == 'av':
warnings.warn('Averaging of averaged dvs not allowed. Skipping dv')
continue
if all((
lat[0] <= s.stla <= lat[1], lat[0] <= s.evla <= lat[1],
lon[0] <= s.stlo <= lon[1], lon[0] <= s.evlo <= lon[1],
el[0] <= s.stel <= el[1], el[0] <= s.evel <= el[1])):
dv_filt.append(dv)
if not len(dv_filt):
raise ValueError(
'No DVs within geographical constraints found.')
av_dv = average_components(dv_filt, return_std)
s = av_dv.stats
# adapt header
s['network'] = s['station'] = 'geoav'
s['stel'] = s['evel'] = el
s['stlo'] = s['evlo'] = lon
s['stla'] = s['evla'] = lat
return av_dv
[docs]def average_components_wfc(wfcs: List[WFC]) -> WFC:
"""
Averages the Similariy matrix of the three DV objects. Based on those,
it computes a new dv value and a new correlation value.
:param dvs: List of dvs from the different components to compute an
average from. Note that it is possible to use almost anything as
input (also from different stations). However, at the time,
the function requires the input to be of the same shape
:type dvs: List[class:`~seismic.monitor.dv.DV`]
:raises TypeError: for DVs that were computed with different methods
:return: A single dv with an averaged similarity matrix
:rtype: DV
"""
stats = deepcopy(wfcs[0].stats)
wfcp = wfcs[0].wfc_processing
stats['channel'] = 'av'
for k in ['tw_start', 'tw_len', 'freq_min', 'freq_max']:
if any((wfcp[k] != wfc.wfc_processing[k] for wfc in wfcs)):
raise ValueError('%s is not allowed to differ.' % k)
meandict = {}
for k in wfcs[0].keys():
meandict[k] = np.nanmean([wfc[k] for wfc in wfcs], axis=0)
wfcout = WFC(meandict, stats, wfcp)
return wfcout
