# -*- coding: utf-8 -*-
# ------------------------------------------------------------------
# Filename: konnoohmachismoothing.py
# Purpose: Small module to smooth spectra with the so called Konno & Ohmachi
# method.
# Author: Lion Krischer
# Email: krischer@geophysik.uni-muenchen.de
# License: GPLv2
#
# Copyright (C) 2011 Lion Krischer
# --------------------------------------------------------------------
"""
Functions to smooth spectra with the so called Konno & Ohmachi method.
:copyright:
The ObsPy Development Team (devs@obspy.org)
:license:
GNU Lesser General Public License, Version 3
(https://www.gnu.org/copyleft/lesser.html)
"""
from __future__ import (absolute_import, division, print_function,
unicode_literals)
from future.builtins import * # NOQA
import warnings
import numpy as np
[docs]def konno_ohmachi_smoothing_window(frequencies, center_frequency,
bandwidth=40.0, normalize=False):
"""
Returns the Konno & Ohmachi Smoothing window for every frequency in
frequencies.
Returns the smoothing window around the center frequency with one value per
input frequency defined as follows (see [Konno1998]_)::
[sin(b * log_10(f/f_c)) / (b * log_10(f/f_c)]^4
b = bandwidth
f = frequency
f_c = center frequency
The bandwidth of the smoothing function is constant on a logarithmic scale.
A small value will lead to a strong smoothing, while a large value of will
lead to a low smoothing of the Fourier spectra.
The default (and generally used) value for the bandwidth is 40. (From the
`Geopsy documentation <http://www.geopsy.org>`_)
All parameters need to be positive. This is not checked due to performance
reasons and therefore any negative parameters might have unexpected
results.
:type frequencies: :class:`numpy.ndarray` (float32 or float64)
:param frequencies:
All frequencies for which the smoothing window will be returned.
:type center_frequency: float
:param center_frequency:
The frequency around which the smoothing is performed. Must be greater
or equal to 0.
:type bandwidth: float
:param bandwidth:
Determines the width of the smoothing peak. Lower values result in a
broader peak. Must be greater than 0. Defaults to 40.
:type normalize: bool, optional
:param normalize:
The Konno-Ohmachi smoothing window is normalized on a logarithmic
scale. Set this parameter to True to normalize it on a normal scale.
Default to False.
"""
if frequencies.dtype != np.float32 and frequencies.dtype != np.float64:
msg = 'frequencies needs to have a dtype of float32/64.'
raise ValueError(msg)
# If the center_frequency is 0 return an array with zero everywhere except
# at zero.
if center_frequency == 0:
smoothing_window = np.zeros(len(frequencies), dtype=frequencies.dtype)
smoothing_window[frequencies == 0.0] = 1.0
return smoothing_window
# Disable div by zero errors and return zero instead
with np.errstate(divide='ignore', invalid='ignore'):
# Calculate the bandwidth*log10(f/f_c)
smoothing_window = bandwidth * np.log10(frequencies / center_frequency)
# Just the Konno-Ohmachi formulae.
smoothing_window[:] = (
np.sin(smoothing_window) / smoothing_window) ** 4
# Check if the center frequency is exactly part of the provided
# frequencies. This will result in a division by 0. The limit of f->f_c is
# one.
smoothing_window[frequencies == center_frequency] = 1.0
# Also a frequency of zero will result in a logarithm of -inf. The limit of
# f->0 with f_c!=0 is zero.
smoothing_window[frequencies == 0.0] = 0.0
# Normalize to one if wished.
if normalize:
smoothing_window /= smoothing_window.sum()
return smoothing_window
[docs]def calculate_smoothing_matrix(frequencies, bandwidth=40.0, normalize=False):
"""
Calculates a len(frequencies) x len(frequencies) matrix with the Konno &
Ohmachi window for each frequency as the center frequency.
Any spectrum with the same frequency bins as this matrix can later be
smoothed by using
:func:`~obspy.signal.konnoohmachismoothing.apply_smoothing_matrix`.
This also works for many spectra stored in one large matrix and is even
more efficient.
This makes it very efficient for smoothing the same spectra again and again
but it comes with a high memory consumption for larger frequency arrays!
:type frequencies: :class:`numpy.ndarray` (float32 or float64)
:param frequencies:
The input frequencies.
:type bandwidth: float
:param bandwidth:
Determines the width of the smoothing peak. Lower values result in a
broader peak. Must be greater than 0. Defaults to 40.
:type normalize: bool, optional
:param normalize:
The Konno-Ohmachi smoothing window is normalized on a logarithmic
scale. Set this parameter to True to normalize it on a normal scale.
Default to False.
"""
# Create matrix to be filled with smoothing entries.
sm_matrix = np.empty((len(frequencies), len(frequencies)),
frequencies.dtype)
for _i, freq in enumerate(frequencies):
sm_matrix[_i, :] = konno_ohmachi_smoothing_window(
frequencies, freq, bandwidth, normalize=normalize)
return sm_matrix
[docs]def apply_smoothing_matrix(spectra, smoothing_matrix, count=1):
"""
Smooths a matrix containing one spectra per row with the Konno-Ohmachi
smoothing window, using a smoothing matrix pre-computed through the
:func:`~obspy.signal.konnoohmachismoothing.calculate_smoothing_matrix`
function.
This function is useful if one needs to smooth the same type of spectrum
(same shape) through different function calls.
All spectra need to have frequency bins corresponding to the same
frequencies.
"""
if spectra.dtype not in (np.float32, np.float64):
msg = '`spectra` needs to have a dtype of float32/64.'
raise ValueError(msg)
new_spec = np.dot(spectra, smoothing_matrix)
# Eventually apply more than once.
for _i in range(count - 1):
new_spec = np.dot(new_spec, smoothing_matrix)
return new_spec
[docs]def konno_ohmachi_smoothing(spectra, frequencies, bandwidth=40, count=1,
enforce_no_matrix=False, max_memory_usage=512,
normalize=False):
"""
Smooths a matrix containing one spectra per row with the Konno-Ohmachi
smoothing window.
All spectra need to have frequency bins corresponding to the same
frequencies.
This method first will estimate the memory usage and then either use a fast
and memory intensive method or a slow one with a better memory usage.
:type spectra: :class:`numpy.ndarray` (float32 or float64)
:param spectra:
One or more spectra per row. If more than one the first spectrum has to
be accessible via spectra[0], the next via spectra[1], ...
:type frequencies: :class:`numpy.ndarray` (float32 or float64)
:param frequencies:
Contains the frequencies for the spectra.
:type bandwidth: float
:param bandwidth:
Determines the width of the smoothing peak. Lower values result in a
broader peak. Must be greater than 0. Defaults to 40.
:type count: int, optional
:param count:
How often the apply the filter. For very noisy spectra it is useful to
apply is more than once. Defaults to 1.
:type enforce_no_matrix: bool, optional
:param enforce_no_matrix:
An efficient but memory intensive matrix-multiplication algorithm is
used in case more than one spectra is to be smoothed or one spectrum is
to be smoothed more than once if enough memory is available. This flag
disables the matrix algorithm altogether. Defaults to False
:type max_memory_usage: int, optional
:param max_memory_usage:
Set the maximum amount of extra memory in MB for this method. Decides
whether or not the matrix multiplication method is used. Defaults to
512 MB.
:type normalize: bool, optional
:param normalize:
The Konno-Ohmachi smoothing window is normalized on a logarithmic
scale. Set this parameter to True to normalize it on a normal scale.
Default to False.
"""
if spectra.dtype not in (np.float32, np.float64):
msg = '`spectra` needs to have a dtype of float32/64.'
raise ValueError(msg)
if frequencies.dtype not in (np.float32, np.float64):
msg = '`frequencies` needs to have a dtype of float32/64.'
raise ValueError(msg)
# Spectra and frequencies should have the same dtype.
if frequencies.dtype != spectra.dtype:
frequencies = np.require(frequencies, np.float64)
spectra = np.require(spectra, np.float64)
msg = '`frequencies` and `spectra` should have the same dtype. It ' + \
'will be changed to np.float64 for both.'
warnings.warn(msg)
# Check the dtype to get the correct size.
if frequencies.dtype == np.float32:
size = 4.0
elif frequencies.dtype == np.float64:
size = 8.0
# Calculate the approximate usage needs for the smoothing matrix algorithm.
length = len(frequencies)
approx_mem_usage = (length * length + 2 * len(spectra) + length) * \
size / 1048576.0
# If smaller than the allowed maximum memory consumption build a smoothing
# matrix and apply to each spectrum. Also only use when more then one
# spectrum is to be smoothed.
if enforce_no_matrix is False and (len(spectra.shape) > 1 or count > 1) \
and approx_mem_usage < max_memory_usage:
smoothing_matrix = calculate_smoothing_matrix(
frequencies, bandwidth, normalize=normalize)
return apply_smoothing_matrix(spectra, smoothing_matrix, count=count)
# Otherwise just calculate the smoothing window every time and apply it.
else:
new_spec = np.empty(spectra.shape, spectra.dtype)
# Separate case for just one spectrum.
if len(new_spec.shape) == 1:
for _i in range(len(frequencies)):
window = konno_ohmachi_smoothing_window(
frequencies, frequencies[_i], bandwidth,
normalize=normalize)
new_spec[_i] = (window * spectra).sum()
# Reuse smoothing window if more than one spectrum.
else:
for _i in range(len(frequencies)):
window = konno_ohmachi_smoothing_window(
frequencies, frequencies[_i], bandwidth,
normalize=normalize)
for _j, spec in enumerate(spectra):
new_spec[_j, _i] = (window * spec).sum()
# Eventually apply more than once.
for _i in range(count - 1):
new_spec = konno_ohmachi_smoothing(
new_spec, frequencies, bandwidth, enforce_no_matrix=True,
normalize=normalize)
return new_spec