trigger.py

# -*- coding: utf-8 -*-
# -------------------------------------------------------------------
# Filename: trigger.py
#  Purpose: Python trigger/picker routines for seismology.
#   Author: Moritz Beyreuther, Tobias Megies
#    Email: moritz.beyreuther@geophysik.uni-muenchen.de
#
# Copyright (C) 2008-2012 Moritz Beyreuther, Tobias Megies
# -------------------------------------------------------------------
"""
Various routines related to triggering/picking

Module implementing the Recursive STA/LTA. Two versions, a fast ctypes one and
a bit slower python one. Furthermore, the classic and delayed STA/LTA, the
carl_sta_trig and the z_detect are implemented.
Also includes picking routines, routines for evaluation and visualization of
characteristic functions and a coincidence triggering routine.

.. seealso:: [Withers1998]_ (p. 98) and [Trnkoczy2012]_

:copyright:
    The ObsPy Development Team (devs@obspy.org)
:license:
    GNU Lesser General Public License, Version 3
    (https://www.gnu.org/copyleft/lesser.html)
"""
from collections import deque
import ctypes as C  # NOQA
import warnings

import numpy as np
import scipy

from obspy import UTCDateTime
from obspy.signal.cross_correlation import templates_max_similarity
from obspy.signal.headers import clibsignal, head_stalta_t


def recursive_sta_lta(a, nsta, nlta):
    """
    Recursive STA/LTA.

    Fast version written in C.

    :note: This version directly uses a C version via CTypes
    :type a: :class:`numpy.ndarray`, dtype=float64
    :param a: Seismic Trace, numpy.ndarray dtype float64
    :type nsta: int
    :param nsta: Length of short time average window in samples
    :type nlta: int
    :param nlta: Length of long time average window in samples
    :rtype: :class:`numpy.ndarray`, dtype=float64
    :return: Characteristic function of recursive STA/LTA

    .. seealso:: [Withers1998]_ (p. 98) and [Trnkoczy2012]_
    """
    # be nice and adapt type if necessary
    a = np.ascontiguousarray(a, np.float64)
    ndat = len(a)
    charfct = np.empty(ndat, dtype=np.float64)
    # do not use pointer here:
    clibsignal.recstalta(a, charfct, ndat, nsta, nlta)
    return charfct


def recursive_sta_lta_py(a, nsta, nlta):
    """
    Recursive STA/LTA written in Python.

    .. note::

        There exists a faster version of this trigger wrapped in C
        called :func:`~obspy.signal.trigger.recursive_sta_lta` in this module!

    :type a: NumPy :class:`~numpy.ndarray`
    :param a: Seismic Trace
    :type nsta: int
    :param nsta: Length of short time average window in samples
    :type nlta: int
    :param nlta: Length of long time average window in samples
    :rtype: NumPy :class:`~numpy.ndarray`
    :return: Characteristic function of recursive STA/LTA

    .. seealso:: [Withers1998]_ (p. 98) and [Trnkoczy2012]_
    """
    ndat = len(a)
    # compute the short time average (STA) and long time average (LTA)
    # given by Evans and Allen
    csta = 1. / nsta
    clta = 1. / nlta
    sta = 0.
    lta = np.finfo(0.0).tiny  # avoid zero division
    a = np.square(a)
    charfct = np.zeros(ndat, dtype=np.float64)
    icsta = 1 - csta
    iclta = 1 - clta
    for i in range(1, ndat):
        sta = csta * a[i] + icsta * sta
        lta = clta * a[i] + iclta * lta
        charfct[i] = sta / lta
    charfct[:nlta] = 0

    return charfct


def carl_sta_trig(a, nsta, nlta, ratio, quiet):
    """
    Computes the carlSTAtrig characteristic function.

    eta = star - (ratio * ltar) - abs(sta - lta) - quiet

    :type a: NumPy :class:`~numpy.ndarray`
    :param a: Seismic Trace
    :type nsta: int
    :param nsta: Length of short time average window in samples
    :type nlta: int
    :param nlta: Length of long time average window in samples
    :type ration: float
    :param ratio: as ratio gets smaller, carl_sta_trig gets more sensitive
    :type quiet: float
    :param quiet: as quiet gets smaller, carl_sta_trig gets more sensitive
    :rtype: NumPy :class:`~numpy.ndarray`
    :return: Characteristic function of CarlStaTrig
    """
    m = len(a)
    #
    sta = np.zeros(len(a), dtype=np.float64)
    lta = np.zeros(len(a), dtype=np.float64)
    star = np.zeros(len(a), dtype=np.float64)
    ltar = np.zeros(len(a), dtype=np.float64)
    pad_sta = np.zeros(nsta)
    pad_lta = np.zeros(nlta)  # avoid for 0 division 0/1=0
    #
    # compute the short time average (STA)
    for i in range(nsta):  # window size to smooth over
        sta += np.concatenate((pad_sta, a[i:m - nsta + i]))
    sta /= nsta
    #
    # compute the long time average (LTA), 8 sec average over sta
    for i in range(nlta):  # window size to smooth over
        lta += np.concatenate((pad_lta, sta[i:m - nlta + i]))
    lta /= nlta
    lta = np.concatenate((np.zeros(1), lta))[:m]  # XXX ???
    #
    # compute star, average of abs diff between trace and lta
    for i in range(nsta):  # window size to smooth over
        star += np.concatenate((pad_sta,
                               abs(a[i:m - nsta + i] - lta[i:m - nsta + i])))
    star /= nsta
    #
    # compute ltar, 8 sec average over star
    for i in range(nlta):  # window size to smooth over
        ltar += np.concatenate((pad_lta, star[i:m - nlta + i]))
    ltar /= nlta
    #
    eta = star - (ratio * ltar) - abs(sta - lta) - quiet
    eta[:nlta] = -1.0
    return eta


def classic_sta_lta(a, nsta, nlta):
    """
    Computes the standard STA/LTA from a given input array a. The length of
    the STA is given by nsta in samples, respectively is the length of the
    LTA given by nlta in samples.

    Fast version written in C.

    :type a: NumPy :class:`~numpy.ndarray`
    :param a: Seismic Trace
    :type nsta: int
    :param nsta: Length of short time average window in samples
    :type nlta: int
    :param nlta: Length of long time average window in samples
    :rtype: NumPy :class:`~numpy.ndarray`
    :return: Characteristic function of classic STA/LTA
    """
    data = a
    # initialize C struct / NumPy structured array
    head = np.empty(1, dtype=head_stalta_t)
    head[:] = (len(data), nsta, nlta)
    # ensure correct type and contiguous of data
    data = np.ascontiguousarray(data, dtype=np.float64)
    # all memory should be allocated by python
    charfct = np.empty(len(data), dtype=np.float64)
    # run and check the error-code
    errcode = clibsignal.stalta(head, data, charfct)
    if errcode != 0:
        raise Exception('ERROR %d stalta: len(data) < nlta' % errcode)
    return charfct


def classic_sta_lta_py(a, nsta, nlta):
    """
    Computes the standard STA/LTA from a given input array a. The length of
    the STA is given by nsta in samples, respectively is the length of the
    LTA given by nlta in samples. Written in Python.

    .. note::

        There exists a faster version of this trigger wrapped in C
        called :func:`~obspy.signal.trigger.classic_sta_lta` in this module!

    :type a: NumPy :class:`~numpy.ndarray`
    :param a: Seismic Trace
    :type nsta: int
    :param nsta: Length of short time average window in samples
    :type nlta: int
    :param nlta: Length of long time average window in samples
    :rtype: NumPy :class:`~numpy.ndarray`
    :return: Characteristic function of classic STA/LTA
    """
    # The cumulative sum can be exploited to calculate a moving average (the
    # cumsum function is quite efficient)
    sta = np.cumsum(a ** 2, dtype=np.float64)

    # Copy for LTA
    lta = sta.copy()

    # Compute the STA and the LTA
    sta[nsta:] = sta[nsta:] - sta[:-nsta]
    sta /= nsta
    lta[nlta:] = lta[nlta:] - lta[:-nlta]
    lta /= nlta

    # Pad zeros
    sta[:nlta - 1] = 0

    # Avoid division by zero by setting zero values to tiny float
    dtiny = np.finfo(0.0).tiny
    idx = lta < dtiny
    lta[idx] = dtiny

    return sta / lta


def delayed_sta_lta(a, nsta, nlta):
    """
    Delayed STA/LTA.

    :type a: NumPy :class:`~numpy.ndarray`
    :param a: Seismic Trace
    :type nsta: int
    :param nsta: Length of short time average window in samples
    :type nlta: int
    :param nlta: Length of long time average window in samples
    :rtype: NumPy :class:`~numpy.ndarray`
    :return: Characteristic function of delayed STA/LTA

    .. seealso:: [Withers1998]_ (p. 98) and [Trnkoczy2012]_
    """
    m = len(a)
    #
    # compute the short time average (STA) and long time average (LTA)
    # don't start for STA at nsta because it's muted later anyway
    sta = np.zeros(m, dtype=np.float64)
    lta = np.zeros(m, dtype=np.float64)
    for i in range(m):
        sta[i] = (a[i] ** 2 + a[i - nsta] ** 2) / nsta + sta[i - 1]
        lta[i] = (a[i - nsta - 1] ** 2 + a[i - nsta - nlta - 1] ** 2) / \
            nlta + lta[i - 1]
    sta[0:nlta + nsta + 50] = 0
    lta[0:nlta + nsta + 50] = 1  # avoid division by zero
    return sta / lta


def z_detect(a, nsta):
    """
    Z-detector.

    :param nsta: Window length in Samples.

    .. seealso:: [Withers1998]_, p. 99
    """
    # Z-detector given by Swindell and Snell (1977)
    # Standard Sta shifted by 1
    sta = np.cumsum(a ** 2, dtype=np.float64)
    sta[nsta + 1:] = sta[nsta:-1] - sta[:-nsta - 1]
    sta[nsta] = sta[nsta - 1]
    sta[:nsta] = 0
    a_mean = np.mean(sta)
    a_std = np.std(sta)
    _z = (sta - a_mean) / a_std
    return _z


def energy_ratio(a, nsta):
    r"""
    Energy ratio defined as

    .. math::
        \text{er}(i) = \frac{\sum_{j=i}^{i+L}{a_j^2}}{\sum_{j=i-L}^{i}{a_j^2}}

    where :math:`L` is ``nsta``.

    :type a: NumPy :class:`~numpy.ndarray`
    :param a: Seismic Trace
    :type nsta: int
    :param nsta: Length of the energy ratio window in samples. It's the same
                 length as ``nsta`` in the classical STA/LTA methods.
    :rtype: NumPy :class:`~numpy.ndarray`
    :return: Energy Ratio

    .. seealso:: [Han2009]_
    """
    sig_power = np.r_[0, np.cumsum(a ** 2, dtype=np.float64)]
    energy_diff = sig_power[nsta:] - sig_power[:-nsta]
    er = np.zeros(len(a), dtype=np.float64)
    np.divide(energy_diff[nsta:], energy_diff[:-nsta],
              where=energy_diff[:-nsta] != 0,
              out=er[nsta:-nsta + 1])
    return er


def modified_energy_ratio(a, nsta, power=3):
    r"""
    Improvement of the :func:`energy_ratio` that accounts for the signal
    itself:

    .. math::
        \text{mer}(i) = (\text{er}(i) * |a(i)|)^3

    where :math:`text{er}(i)` is the :func:`energy_ratio`.

    :type a: NumPy :class:`~numpy.ndarray`
    :param a: Seismic Trace
    :type nsta: int
    :param nsta: Length of energy ratio window in samples. It's the same length
                 as ``nsta`` in the classical STA/LTA methods.
    :type power: int
    :param power: The power exponent in the equation above. Default: 3
    :rtype: NumPy :class:`~numpy.ndarray`
    :return: Modified Energy Ratio
    """
    er = energy_ratio(a, nsta=nsta)
    mer = np.power(er * np.abs(a), power, out=er)
    return mer


def trigger_onset(charfct, thres1, thres2, max_len=9e99, max_len_delete=False):
    """
    Calculate trigger on and off times.

    Given thres1 and thres2 calculate trigger on and off times from
    characteristic function.

    This method is written in pure Python and gets slow as soon as there
    are more then 1e6 triggerings ("on" AND "off") in charfct --- normally
    this does not happen.

    :type charfct: NumPy :class:`~numpy.ndarray`
    :param charfct: Characteristic function of e.g. STA/LTA trigger
    :type thres1: float
    :param thres1: Value above which trigger (of characteristic function)
                   is activated (higher threshold)
    :type thres2: float
    :param thres2: Value below which trigger (of characteristic function)
        is deactivated (lower threshold)
    :type max_len: int
    :param max_len: Maximum length of triggered event in samples. A new
                    event will be triggered as soon as the signal reaches
                    again above thres1.
    :type max_len_delete: bool
    :param max_len_delete: Do not write events longer than max_len into
                           report file.
    :rtype: List
    :return: Nested List of trigger on and of times in samples
    """
    # 1) find indices of samples greater than threshold
    # 2) calculate trigger "of" times by the gap in trigger indices
    #    above the threshold i.e. the difference of two following indices
    #    in ind is greater than 1
    # 3) in principle the same as for "of" just add one to the index to get
    #    start times, this operation is not supported on the compact
    #    syntax
    # 4) as long as there is a on time greater than the actual of time find
    #    trigger on states which are greater than last of state an the
    #    corresponding of state which is greater than current on state
    # 5) if the signal stays above thres2 longer than max_len an event
    #    is triggered and following a new event can be triggered as soon as
    #    the signal is above thres1
    ind1 = np.where(charfct >= thres1)[0]
    if len(ind1) == 0:
        return []
    ind2 = np.where(charfct >= thres2)[0]
    #
    on = deque([ind1[0]])
    of = deque([-1])
    # determine the indices where charfct falls below off-threshold
    ind2_ = np.empty_like(ind2, dtype=bool)
    ind2_[:-1] = np.diff(ind2) > 1
    # last occurence is missed by the diff, add it manually
    ind2_[-1] = True
    of.extend(ind2[ind2_].tolist())
    on.extend(ind1[np.where(np.diff(ind1) > 1)[0] + 1].tolist())
    # include last pick if trigger is on or drop it
    if max_len_delete:
        # drop it
        of.extend([1e99])
        on.extend([on[-1]])
    else:
        # include it
        of.extend([ind2[-1]])

    # add last sample to ensure trigger gets switched off if ctf does not fall
    # below off-threshold before hitting the end
    of.append(len(charfct))
    #
    pick = []
    while on[-1] > of[0]:
        while on[0] <= of[0]:
            on.popleft()
        while of[0] < on[0]:
            of.popleft()
        if of[0] - on[0] > max_len:
            if max_len_delete:
                on.popleft()
                continue
            of.appendleft(on[0] + max_len)
        pick.append([on[0], of[0]])
    return np.array(pick, dtype=np.int64)


def pk_baer(reltrc, samp_int, tdownmax, tupevent, thr1, thr2, preset_len,
            p_dur, return_cf=False):
    """
    Wrapper for P-picker routine by M. Baer, Schweizer Erdbebendienst.

    :param reltrc: time series as numpy.ndarray float32 data, possibly filtered
    :param samp_int: number of samples per second
    :param tdownmax: if dtime exceeds tdownmax, the trigger is examined for
        validity
    :param tupevent: min nr of samples for itrm to be accepted as a pick
    :param thr1: threshold to trigger for pick (c.f. paper)
    :param thr2: threshold for updating sigma  (c.f. paper)
    :param preset_len: no of points taken for the estimation of variance of
        SF(t) on preset()
    :param p_dur: p_dur defines the time interval for which the maximum
        amplitude is evaluated Originally set to 6 secs
    :type return_cf: bool
    :param return_cf: If ``True``, also return the characteristic function.
    :return: (pptime, pfm [,cf]) pptime sample number of parrival;
        pfm direction of first motion (U or D), optionally also the
        characteristic function.

    .. note:: currently the first sample is not taken into account

    .. seealso:: [Baer1987]_
    """
    pptime = C.c_int()
    # c_chcar_p strings are immutable, use string_buffer for pointers
    pfm = C.create_string_buffer(b"     ", 5)
    # be nice and adapt type if necessary
    reltrc = np.ascontiguousarray(reltrc, np.float32)
    # Initiliaze CF array (MB)
    c_float_p = C.POINTER(C.c_float)
    cf_arr = np.zeros(len(reltrc) - 1, dtype=np.float32, order="C")
    cf_p = cf_arr.ctypes.data_as(c_float_p)
    # index in pk_mbaer.c starts with 1, 0 index is lost, length must be
    # one shorter
    args = (len(reltrc) - 1, C.byref(pptime), pfm, samp_int,
            tdownmax, tupevent, thr1, thr2, preset_len, p_dur, cf_p)
    errcode = clibsignal.ppick(reltrc, *args)
    if errcode != 0:
        raise MemoryError("Error in function ppick of mk_mbaer.c")
    # Switch cf_arr param (MB)
    # add the sample to the time which is not taken into account
    # pfm has to be decoded from byte to string
    if return_cf:
        return pptime.value + 1, pfm.value.decode('utf-8'), cf_arr
    else:
        return pptime.value + 1, pfm.value.decode('utf-8')


def aic_simple(a):
    r"""
    Simple Akaike Information Criterion [Maeda1985]_.

    It's computed directly from input data :math:`a` and defined as

    .. math::
        \text{AIC}(k) = k\log(\text{Var}(a_{1..k})) +
                        (N-k-1)\log(\text{Var}(a_{k+1..N}))

    which variance denoted as :math:`\text{Var}`.

    The true output is one data sample less. To make it convenient with other
    metrics in this module, where the output length is preserved, the last
    element is appended to the output: ``aic[-2] == aic[-1]``.

    :type a: :class:`numpy.ndarray` or :class:`list`
    :param a: Input time series
    :rtype: :class:`numpy.ndarray`
    :return: aic - Akaike Information Criterion array
    """
    n = len(a)
    if n <= 2:
        return np.zeros(n, dtype=np.float64)
    a = np.ascontiguousarray(a, np.float64)
    aic_res = np.empty(n, dtype=np.float64)
    clibsignal.aic_simple(aic_res, a, n)
    return aic_res


def ar_pick(a, b, c, samp_rate, f1, f2, lta_p, sta_p, lta_s, sta_s, m_p, m_s,
            l_p, l_s, s_pick=True):
    """
    Pick P and S arrivals with an AR-AIC + STA/LTA algorithm.

    The algorithm picks onset times using an Auto Regression - Akaike
    Information Criterion (AR-AIC) method. The detection intervals are
    successively narrowed down with the help of STA/LTA ratios as well as
    STA-LTA difference calculations. For details, please see [Akazawa2004]_.

    An important feature of this algorithm is that it requires comparatively
    little tweaking and site-specific settings and is thus applicable to large,
    diverse data sets.

    :type a: :class:`numpy.ndarray`
    :param a: Z signal the data.
    :type b: :class:`numpy.ndarray`
    :param b: N signal of the data.
    :type c: :class:`numpy.ndarray`
    :param c: E signal of the data.
    :type samp_rate: float
    :param samp_rate: Number of samples per second.
    :type f1: float
    :param f1: Frequency of the lower bandpass window.
    :type f2: float
    :param f2: Frequency of the upper .andpass window.
    :type lta_p: float
    :param lta_p: Length of LTA for the P arrival in seconds.
    :type sta_p: float
    :param sta_p: Length of STA for the P arrival in seconds.
    :type lta_s: float
    :param lta_s: Length of LTA for the S arrival in seconds.
    :type sta_s: float
    :param sta_s: Length of STA for the S arrival in seconds.
    :type m_p: int
    :param m_p: Number of AR coefficients for the P arrival.
    :type m_s: int
    :param m_s: Number of AR coefficients for the S arrival.
    :type l_p: float
    :param l_p: Length of variance window for the P arrival in seconds.
    :type l_s: float
    :param l_s: Length of variance window for the S arrival in seconds.
    :type s_pick: bool
    :param s_pick: If ``True``, also pick the S phase, otherwise only the P
        phase.
    :rtype: tuple
    :returns: A tuple with the P and the S arrival.
    """
    if not (len(a) == len(b) == len(c)):
        raise ValueError("All three data arrays must have the same length.")

    a = scipy.signal.detrend(a, type='linear')
    b = scipy.signal.detrend(b, type='linear')
    c = scipy.signal.detrend(c, type='linear')

    # be nice and adapt type if necessary
    a = np.require(a, dtype=np.float32, requirements=['C_CONTIGUOUS'])
    b = np.require(b, dtype=np.float32, requirements=['C_CONTIGUOUS'])
    c = np.require(c, dtype=np.float32, requirements=['C_CONTIGUOUS'])

    # scale amplitudes to avoid precision issues in case of low amplitudes
    # C code picks the horizontal component with larger amplitudes, so scale
    # horizontal components with a common scaling factor
    data_max = np.abs(a).max()
    if data_max < 100:
        a *= 1e6
        a /= data_max
    data_max = max(np.abs(b).max(), np.abs(c).max())
    if data_max < 100:
        for data in (b, c):
            data *= 1e6
            data /= data_max

    s_pick = C.c_int(s_pick)  # pick S phase also
    ptime = C.c_float()
    stime = C.c_float()
    args = (len(a), samp_rate, f1, f2,
            lta_p, sta_p, lta_s, sta_s, m_p, m_s, C.byref(ptime),
            C.byref(stime), l_p, l_s, s_pick)
    errcode = clibsignal.ar_picker(a, b, c, *args)
    if errcode != 0:
        bufs = ['buff1', 'buff1_s', 'buff2', 'buff3', 'buff4', 'buff4_s',
                'f_error', 'b_error', 'ar_f', 'ar_b', 'buf_sta', 'buf_lta',
                'extra_tr1', 'extra_tr2', 'extra_tr3']
        if errcode <= len(bufs):
            raise MemoryError('Unable to allocate %s!' % (bufs[errcode - 1]))
        raise Exception('Error during PAZ calculation!')
    return ptime.value, stime.value


def plot_trace(trace, cft):
    """
    Plot characteristic function of trigger along with waveform data.

    :type trace: :class:`~obspy.core.trace.Trace`
    :param trace: waveform data
    :type cft: :class:`numpy.ndarray`
    :param cft: characteristic function as returned by a trigger in
        :mod:`obspy.signal.trigger`
    :rtype: tuple
    :returns: Matplotlib figure instance and axes
    """
    import matplotlib.pyplot as plt
    df = trace.stats.sampling_rate
    npts = trace.stats.npts
    t = np.arange(npts, dtype=np.float32) / df
    fig, axes = plt.subplots(nrows=2, sharex=True)
    axes[0].plot(t, trace.data, 'k')
    axes[1].plot(t, cft, 'k')
    axes[1].set_xlabel("Time after %s [s]" % trace.stats.starttime.isoformat())
    fig.suptitle(trace.id)
    fig.canvas.draw()
    return fig, axes


def plot_trigger(trace, cft, thr_on, thr_off, show=True):
    """
    Plot characteristic function of trigger along with waveform data and
    trigger On/Off from given thresholds.

    :type trace: :class:`~obspy.core.trace.Trace`
    :param trace: waveform data
    :type cft: :class:`numpy.ndarray`
    :param cft: characteristic function as returned by a trigger in
        :mod:`obspy.signal.trigger`
    :type thr_on: float
    :param thr_on: threshold for switching trigger on
    :type thr_off: float
    :param thr_off: threshold for switching trigger off
    :type show: bool
    :param show: Do not call `plt.show()` at end of routine. That way,
        further modifications can be done to the figure before showing it.
    :rtype: tuple
    :returns: Matplotlib figure instance and axes
    """
    import matplotlib.pyplot as plt
    df = trace.stats.sampling_rate
    fig, axes = plot_trace(trace, cft)
    ax1, ax2 = axes
    on_off = np.array(trigger_onset(cft, thr_on, thr_off))
    ymin, ymax = ax1.get_ylim()
    try:
        ax1.vlines(on_off[:, 0] / df, ymin, ymax, color='r', lw=2,
                   label="Trigger On")
        ax1.vlines(on_off[:, 1] / df, ymin, ymax, color='b', lw=2,
                   label="Trigger Off")
        ax1.legend()
    except IndexError:
        pass
    ax2.axhline(thr_on, color='red', lw=1, ls='--')
    ax2.axhline(thr_off, color='blue', lw=1, ls='--')
    ax2.set_xlabel("Time after %s [s]" % trace.stats.starttime.isoformat())
    fig.suptitle(trace.id)
    fig.canvas.draw()
    if show:
        plt.show()
    return fig, axes


def coincidence_trigger(trigger_type, thr_on, thr_off, stream,
                        thr_coincidence_sum, trace_ids=None,
                        max_trigger_length=1e6, delete_long_trigger=False,
                        trigger_off_extension=0, details=False,
                        event_templates={}, similarity_threshold=0.7,
                        **options):
    """
    Perform a network coincidence trigger.

    The routine works in the following steps:
      * take every single trace in the stream
      * apply specified triggering routine (can be skipped to work on
        precomputed custom characteristic functions)
      * evaluate all single station triggering results
      * compile chronological overall list of all single station triggers
      * find overlapping single station triggers
      * calculate coincidence sum of every individual overlapping trigger
      * add to coincidence trigger list if it exceeds the given threshold
      * optional: if master event templates are provided, also check single
        station triggers individually and include any single station trigger if
        it exceeds the specified similarity threshold even if no other stations
        coincide with the trigger
      * return list of network coincidence triggers

    .. note::
        An example can be found in the
        `Trigger/Picker Tutorial
        <https://tutorial.obspy.org/code_snippets/trigger_tutorial.html>`_.

    .. note::
        Setting `trigger_type=None` precomputed characteristic functions can
        be provided.

    .. seealso:: [Withers1998]_ (p. 98) and [Trnkoczy2012]_

    :param trigger_type: String that specifies which trigger is applied (e.g.
        ``'recstalta'``). See e.g. :meth:`obspy.core.trace.Trace.trigger` for
        further details. If set to `None` no triggering routine is applied,
        i.e.  data in traces is supposed to be a precomputed characteristic
        function on which the trigger thresholds are evaluated.
    :type trigger_type: str or None
    :type thr_on: float
    :param thr_on: threshold for switching single station trigger on
    :type thr_off: float
    :param thr_off: threshold for switching single station trigger off
    :type stream: :class:`~obspy.core.stream.Stream`
    :param stream: Stream containing waveform data for all stations. These
        data are changed inplace, make a copy to keep the raw waveform data.
    :type thr_coincidence_sum: int or float
    :param thr_coincidence_sum: Threshold for coincidence sum. The network
        coincidence sum has to be at least equal to this value for a trigger to
        be included in the returned trigger list.
    :type trace_ids: list or dict, optional
    :param trace_ids: Trace IDs to be used in the network coincidence sum. A
        dictionary with trace IDs as keys and weights as values can
        be provided. If a list of trace IDs is provided, all
        weights are set to 1. The default of ``None`` uses all traces present
        in the provided stream. Waveform data with trace IDs not
        present in this list/dict are disregarded in the analysis.
    :type max_trigger_length: int or float
    :param max_trigger_length: Maximum single station trigger length (in
        seconds). ``delete_long_trigger`` controls what happens to single
        station triggers longer than this value.
    :type delete_long_trigger: bool, optional
    :param delete_long_trigger: If ``False`` (default), single station
        triggers are manually released at ``max_trigger_length``, although the
        characteristic function has not dropped below ``thr_off``. If set to
        ``True``, all single station triggers longer than
        ``max_trigger_length`` will be removed and are excluded from
        coincidence sum computation.
    :type trigger_off_extension: int or float, optional
    :param trigger_off_extension: Extends search window for next trigger
        on-time after last trigger off-time in coincidence sum computation.
    :type details: bool, optional
    :param details: If set to ``True`` the output coincidence triggers contain
        more detailed information: A list with the trace IDs (in addition to
        only the station names), as well as lists with single station
        characteristic function peak values and standard deviations in the
        triggering interval and mean values of both, relatively weighted like
        in the coincidence sum. These values can help to judge the reliability
        of the trigger.
    :param options: Necessary keyword arguments for the respective trigger
        that will be passed on. For example ``sta`` and ``lta`` for any STA/LTA
        variant (e.g. ``sta=3``, ``lta=10``).
        Arguments ``sta`` and ``lta`` (seconds) will be mapped to ``nsta``
        and ``nlta`` (samples) by multiplying with sampling rate of trace.
        (e.g. ``sta=3``, ``lta=10`` would call the trigger with 3 and 10
        seconds average, respectively)
    :param event_templates: Event templates to use in checking similarity of
        single station triggers against known events. Expected are streams with
        three traces for Z, N, E component. A dictionary is expected where for
        each station used in the trigger, a list of streams can be provided as
        the value to the network/station key (e.g. {"GR.FUR": [stream1,
        stream2]}). Templates are compared against the provided `stream`
        without the specified triggering routine (`trigger_type`) applied.
    :type event_templates: dict
    :param similarity_threshold: similarity threshold (0.0-1.0) at which a
        single station trigger gets included in the output network event
        trigger list. A common threshold can be set for all stations (float) or
        a dictionary mapping station names to float values for each station.
    :type similarity_threshold: float or dict
    :rtype: list
    :returns: List of event triggers sorted chronologically.
    """
    # if no trace ids are specified use all traces ids found in stream
    if trace_ids is None:
        trace_ids = [tr.id for tr in stream]
    # we always work with a dictionary with trace ids and their weights later
    if isinstance(trace_ids, list) or isinstance(trace_ids, tuple):
        trace_ids = dict.fromkeys(trace_ids, 1)
    # set up similarity thresholds as a dictionary if necessary
    if not isinstance(similarity_threshold, dict):
        similarity_threshold = dict.fromkeys(
            [tr.stats.station for tr in stream], similarity_threshold)

    # the single station triggering
    triggers = []
    # prepare kwargs for trigger_onset
    kwargs = {'max_len_delete': delete_long_trigger}
    for tr in stream:
        tr = tr.copy()
        if tr.id not in trace_ids:
            msg = "At least one trace's ID was not found in the " + \
                  "trace ID list and was disregarded (%s)" % tr.id
            warnings.warn(msg, UserWarning)
            continue
        if trigger_type is not None:
            tr.trigger(trigger_type, **options)
        kwargs['max_len'] = int(
            max_trigger_length * tr.stats.sampling_rate + 0.5)
        tmp_triggers = trigger_onset(tr.data, thr_on, thr_off, **kwargs)
        for on, off in tmp_triggers:
            try:
                cft_peak = tr.data[on:off].max()
                cft_std = tr.data[on:off].std()
            except ValueError:
                cft_peak = tr.data[on]
                cft_std = 0
            on = tr.stats.starttime + float(on) / tr.stats.sampling_rate
            off = tr.stats.starttime + float(off) / tr.stats.sampling_rate
            triggers.append((on.timestamp, off.timestamp, tr.id, cft_peak,
                             cft_std))
    triggers.sort()

    for i, (on, off, tr_id, cft_peak, cft_std) in enumerate(triggers):
        sta = tr_id.split(".")[1]
        templates = event_templates.get(sta)
        if templates:
            simil = templates_max_similarity(
                stream, UTCDateTime(on), templates)
        else:
            simil = None
        triggers[i] = (on, off, tr_id, cft_peak, cft_std, simil)

    # the coincidence triggering and coincidence sum computation
    coincidence_triggers = []
    last_off_time = 0.0
    while triggers != []:
        # remove first trigger from list and look for overlaps
        on, off, tr_id, cft_peak, cft_std, simil = triggers.pop(0)
        sta = tr_id.split(".")[1]
        event = {}
        event['time'] = UTCDateTime(on)
        event['stations'] = [tr_id.split(".")[1]]
        event['trace_ids'] = [tr_id]
        event['coincidence_sum'] = float(trace_ids[tr_id])
        event['similarity'] = {}
        if details:
            event['cft_peaks'] = [cft_peak]
            event['cft_stds'] = [cft_std]
        # evaluate maximum similarity for station if event templates were
        # provided
        if simil is not None:
            event['similarity'][sta] = simil
        # compile the list of stations that overlap with the current trigger
        for (tmp_on, tmp_off, tmp_tr_id, tmp_cft_peak, tmp_cft_std,
                tmp_simil) in triggers:
            tmp_sta = tmp_tr_id.split(".")[1]
            # skip retriggering of already present station in current
            # coincidence trigger
            if tmp_tr_id in event['trace_ids']:
                continue
            # check for overlapping trigger,
            # break if there is a gap in between the two triggers
            if tmp_on > off + trigger_off_extension:
                break
            event['stations'].append(tmp_sta)
            event['trace_ids'].append(tmp_tr_id)
            event['coincidence_sum'] += trace_ids[tmp_tr_id]
            if details:
                event['cft_peaks'].append(tmp_cft_peak)
                event['cft_stds'].append(tmp_cft_std)
            # allow sets of triggers that overlap only on subsets of all
            # stations (e.g. A overlaps with B and B overlaps w/ C => ABC)
            off = max(off, tmp_off)
            # evaluate maximum similarity for station if event templates were
            # provided
            if tmp_simil is not None:
                event['similarity'][tmp_sta] = tmp_simil
        # skip if both coincidence sum and similarity thresholds are not met
        if event['coincidence_sum'] < thr_coincidence_sum:
            if not event['similarity']:
                continue
            elif not any([val > similarity_threshold[_s]
                          for _s, val in event['similarity'].items()]):
                continue
        # skip coincidence trigger if it is just a subset of the previous
        # (determined by a shared off-time, this is a bit sloppy)
        if off <= last_off_time:
            continue
        event['duration'] = off - on
        if details:
            weights = np.array([trace_ids[i] for i in event['trace_ids']])
            weighted_values = np.array(event['cft_peaks']) * weights
            event['cft_peak_wmean'] = weighted_values.sum() / weights.sum()
            weighted_values = np.array(event['cft_stds']) * weights
            event['cft_std_wmean'] = \
                (np.array(event['cft_stds']) * weights).sum() / weights.sum()
        coincidence_triggers.append(event)
        last_off_time = off
    return coincidence_triggers