def derobertis(Sv, bgn, thr):
"""
Mask Sv values when lower than background noise by a user-defined
threshold, following:
De Robertis and Higginbottom (2007) ‘A post-processing technique to
estimate the signal-to-noise ratio and remove echosounder background
noise’, ICES Journal of Marine Science, 64: 1282–1291.
Args:
Sv (float): 2D array with Sv data to be masked (dB)
background (float): 2 array with background noise data (dB)
thr (int): threshold value (dB)
Returns:
bool: 2D array mask (Sv < background = True)
"""
# subtract background noise
Svclean = tf.log(tf.lin(Sv) - tf.lin(bgn))
# signal to noise ratio
s2n = Svclean - bgn
# mask where Sv is less than background noise by a user-defined threshold
mask1 = np.ma.masked_less(s2n, thr).mask
mask2 = np.ma.masked_less(tf.lin(Sv) - tf.lin(bgn), 0).mask
mask = mask1 | mask2
return mask
def ariza(Sv, r, offset=20, thr=(-40, -35), m=20, n=50):
"""
Mask attenuated pings by looking at seabed breaches.
Ariza (in progress).
"""
# get ping array
p = np.arange(len(Sv[0]))
# set to NaN shallow waters and data below the Sv threshold
Sv_ = Sv.copy()
Sv_[0:np.nanargmin(abs(r - offset)), :] = np.nan
Sv_[Sv_ < -thr[0]] = np.nan
# bin Sv
# TODO: update to 'twod' and 'full' funtions
Sv_bnd, r_bnd, p_bnd = bin2d(Sv_, r, p, m, n, operation='mean')[0:3]
Sv_bnd = bin2dback(Sv_bnd, r_bnd, p_bnd, r, p)
# label binned Sv data features
Sv_lbl = label(~np.isnan(Sv_bnd))
labels = np.unique(Sv_lbl)
labels = np.delete(labels, np.where(labels == 0))
# list the median values for each Sv feature
val = []
for lbl in labels:
val.append(log(np.nanmedian(lin(Sv_bnd[Sv_lbl == lbl]))))
# keep the feature with a median above the Sv threshold (~seabed)
# and set the rest of the array to NaN
if val:
if np.nanmax(val) > thr[1]:
labels = labels[val != np.nanmax(val)]
for lbl in labels:
Sv_bnd[Sv_lbl == lbl] = np.nan
else:
Sv_bnd[:] = np.nan
else:
Sv_bnd[:] = np.nan
# remove everything in the original Sv array that is not seabed
Sv_sb = Sv.copy()
Sv_sb[np.isnan(Sv_bnd)] = np.nan
# compute the percentile 90th for each ping, at the range at which
# the seabed is supposed to be.
seabed_percentile = log(np.nanpercentile(lin(Sv_sb), 95, axis=0))
# get mask where this value falls bellow a Sv threshold (seabed breaches)
mask = seabed_percentile < thr[0]
mask = np.tile(mask, [len(Sv), 1])
return mask
def maxSv(Sv, r, r0=10, r1=1000, roff=0, thr=(-40, -60)):
"""
Initially detects the seabed as the ping sample with the strongest Sv value,
as long as it exceeds a dB threshold. Then it searchs up along the ping
until Sv falls below a secondary (lower) dB threshold, where the final
seabed is set.
Args:
Sv (float): 2D Sv array (dB).
r (float): 1D range array (m).
r0 (int): minimum range below which the search will be performed (m).
r1 (int): maximum range above which the search will be performed (m).
roff (int): seabed range offset (m).
thr (tuple): 2 integers with 1st and 2nd Sv threshold (dB).
Returns:
bool: 2D array with seabed mask.
"""
# get offset and range indexes
roff = np.nanargmin(abs(r-roff))
r0 = np.nanargmin(abs(r - r0))
r1 = np.nanargmin(abs(r - r1))
# get indexes for maximum Sv along every ping,
idx = np.int64(np.zeros(Sv.shape[1]))
idx[~np.isnan(Sv).all(axis=0)] = np.nanargmax(
Sv[r0:r1, ~np.isnan(Sv).all(axis=0)], axis=0) + r0
# indexes with maximum Sv < main threshold are discarded (=0)
maxSv = Sv[idx, range(len(idx))]
maxSv[np.isnan(maxSv)] = -999
idx[maxSv < thr[0]] = 0
# mask seabed, proceed only with acepted seabed indexes (!=0)
idx = idx
mask = np.zeros(Sv.shape, dtype=bool)
for j, i in enumerate(idx):
if i!=0:
# decrease indexes until Sv mean falls below the 2nd threshold
if np.isnan(Sv[i-5:i, j]).all():
Svmean = thr[1]+1
else:
Svmean = log(np.nanmean(lin(Sv[i-5:i, j])))
while (Svmean>thr[1]) & (i>=5):
i -= 1
# subtract range offset & mask all the way down
i -= roff
if i<0:
i = 0
mask[i:, j] = True
return mask
def ryan(Sv, r, m, n, thr, excludeabove=250, operation='percentile15'):
"""
Mask transient noise as in:
Ryan et al. (2015) ‘Reducing bias due to noise and attenuation in
open-ocean echo integration data’, ICES Journal of Marine Science,
72: 2482–2493.
This mask is based on the assumption that Sv values which exceed the median
value in a surrounding region of m metres by n pings must be due to
transient noise. Sv values are removed if exceed a threshold. Masking is
excluded above 250 m by default to avoid the removal of aggregated biota.
Args:
Sv (float): 2D numpy array with Sv data to be masked (dB)
r (float): 1D numpy array with range data (m)
m (int): height of surrounding region (m)
n (int): width of surrounding region (pings)
threshold (int): user-defined threshold for comparisons (dB)
excludeabove (int): range above which masking is excluded (m)
operation (str): type of average operation:
'mean'
'percentileXX'
'median'
'mode'
Returns:
bool: 2D numpy array mask (transient noise = True)
"""
# offsets for i and j indexes
ioff = np.argmin(abs(r - m))
joff = n
# preclude processing above a user-defined range
r0 = np.argmin(abs(r - excludeabove))
# mask if Sv sample greater than averaged block
# TODO: find out a faster method. The iteration below is too slow.
mask = np.ones(Sv.shape, dtype=bool)
mask[0:r0, :] = False
for i in range(r0, len(Sv)):
for j in range(len(Sv[0])):
# proceed only if enough room for setting the block
if (i - ioff >= 0) & (i + ioff < len(Sv)) & (j - joff >= 0) & (
j + joff < len(Sv[0])):
sample = Sv[i, j]
block = log(
np.nanpercentile(
lin(Sv[i - ioff:i + ioff, j - joff:j + joff]),
int(operation[-2:])))
mask[i, j] = sample - block > thr
return mask
def ryan(Sv, r, r0, r1, n, thr, start=0):
"""
Locate attenuated signal and create a mask following the attenuated signal
filter as in:
Ryan et al. (2015) ‘Reducing bias due to noise and attenuation in
open-ocean echo integration data’, ICES Journal of Marine Science,
72: 2482–2493.
Scattering Layers (SLs) are continuous high signal-to-noise regions with
low inter-ping variability. But attenuated pings create gaps within SLs.
attenuation attenuation ping evaluated
______ V _______________________ V ____________.....V.....____________
| | scattering layer | | . block . |
______| |_______________________| |____________...........____________|
The filter takes advantage of differences with preceding and subsequent
pings to detect and mask attenuation. A comparison is made ping by ping
with respect to a block of the reference layer. The entire ping is masked
if the ping median is less than the block median by a user-defined
threshold value.
Args:
Sv (float): 2D array with Sv data to be masked (dB).
r (float): 1D array with range data (m).
r0 (int): upper limit of SL (m).
r1 (int): lower limit of SL (m).
n (int): number of preceding & subsequent pings defining the block.
thr (int): user-defined threshold value (dB).
start (int): ping index to start processing.
Returns:
list: 2D boolean array with AS mask and 2D boolean array with mask
indicating where AS detection was unfeasible.
"""
# raise errors if wrong arguments
if r0 > r1:
raise Exception('Minimum range has to be shorter than maximum range')
# return empty mask if searching range is outside the echosounder range
if (r0 > r[-1]) or (r1 < r[0]):
mask = np.zeros_like(Sv, dtype=bool)
mask_ = np.zeros_like(Sv, dtype=bool)
return mask, mask_
# turn layer boundaries into arrays with length = Sv.shape[1]
r0 = np.ones(Sv.shape[1]) * r0
r1 = np.ones(Sv.shape[1]) * r1
# start masking process
mask_ = np.zeros(Sv.shape, dtype=bool)
mask = np.zeros(Sv.shape, dtype=bool)
for j in range(start, len(Sv[0])):
# find indexes for upper and lower SL limits
up = np.argmin(abs(r - r0[j]))
lw = np.argmin(abs(r - r1[j]))
# TODO: now indexes are the same at every loop, but future
# versions will have layer boundaries with variable range
# (need to implement mask_layer.py beforehand!)
# mask where AS evaluation is unfeasible (e.g. edge issues, all-NANs)
if (j - n < 0) | (j + n > len(Sv[0]) - 1) | np.all(
np.isnan(Sv[up:lw, j])):
mask_[:, j] = True
# compare ping and block medians otherwise & mask ping if too different
else:
pingmedian = log(np.nanmedian(lin(Sv[up:lw, j])))
blockmedian = log(np.nanmedian(lin(Sv[up:lw, (j - n):(j + n)])))
if (pingmedian - blockmedian) < thr:
mask[:, j] = True
return [mask[:, start:], mask_[:, start:]]
def experimental(Sv, r,
r0=10, r1=1000, roff=0, thr=(-30,-70), ns=150, nd=3):
"""
Mask Sv above a threshold to get a potential seabed mask. Then, the mask is
dilated to fill seabed breaches, and small objects are removed to prevent
masking high Sv features that are not seabed (e.g. fish schools or spikes).
Once this is done, the mask is built up until Sv falls below a 2nd
threshold, Finally, the mask is extended all the way down.
Args:
Sv (float): 2D Sv array (dB).
r (float): 1D range array (m).
r0 (int): minimum range below which the search will be performed (m).
r1 (int): maximum range above which the search will be performed (m).
roff (int): seabed range offset (m).
thr (tuple): 2 integers with 1st and 2nd Sv threshold (dB).
ns (int): maximum number of samples for an object to be removed.
nd (int): number of dilations performed to the seabed mask.
Returns:
bool: 2D array with seabed mask.
"""
# get indexes for range offset and range limits
roff = np.nanargmin(abs(r - roff))
r0 = np.nanargmin(abs(r - r0))
r1 = np.nanargmin(abs(r - r1)) + 1
# mask Sv above the first Sv threshold
mask = Sv[r0:r1, :] > thr[0]
maskabove = np.zeros((r0, mask.shape[1]), dtype =bool)
maskbelow = np.zeros((len(r) - r1, mask.shape[1]), dtype=bool)
mask = np.r_[maskabove, mask, maskbelow]
# remove small to prevent other high Sv features to be masked as seabed
# (e.g fish schools, impulse noise not properly masked. etc)
mask = remove_small_objects(mask, ns)
# dilate mask to fill seabed breaches
# (e.g. attenuated pings or gaps from previous masking)
kernel = np.ones((3,5))
mask = cv2.dilate(np.uint8(mask), kernel, iterations=nd)
mask = np.array(mask, dtype = 'bool')
# proceed with the following only if seabed was detected
idx = np.argmax(mask, axis=0)
for j, i in enumerate(idx):
if i != 0:
# rise up seabed until Sv falls below the 2nd threshold
while (log(np.nanmean(lin(Sv[i-5:i, j]))) > thr[1]) & (i>=5):
i -= 1
# subtract range offset & mask all the way down
i -= roff
if i<0:
i = 0
mask[i:, j] = True
# # dilate again to ensure not leaving seabed behind
# kernel = np.ones((3,3))
# mask = cv2.dilate(np.uint8(mask), kernel, iterations = 2)
# mask = np.array(mask, dtype = 'bool')
return mask
def blackwell_mod(Sv, theta, phi, r, r0=10, r1=1000, tSv=-75, ttheta=702,
tphi=282, wtheta=28 , wphi=52,
rlog=None, tpi=None, freq=None, rank=50):
"""
Detects and mask seabed using the split-beam angle and Sv, based in
"Blackwell et al (2019), Aliased seabed detection in fisheries acoustic
data". Complete article here: https://arxiv.org/abs/1904.10736
This is a modified version from the original algorithm. It includes extra
arguments to evaluate whether aliased seabed items can occur, given the
true seabed detection range, and the possibility of tuning the percentile's
rank.
Args:
Sv (float): 2D numpy array with Sv data (dB)
theta (float): 2D numpy array with the along-ship angle (degrees)
phi (float): 2D numpy array with the athwart-ship angle (degrees)
r (float): 1D range array (m)
r0 (int): minimum range below which the search will be performed (m)
r1 (int): maximum range above which the search will be performed (m)
tSv (float): Sv threshold above which seabed is pre-selected (dB)
ttheta (int): Theta threshold above which seabed is pre-selected (dB)
tphi (int): Phi threshold above which seabed is pre-selected (dB)
wtheta (int): window's size for mean square operation in Theta field
wphi (int): window's size for mean square operation in Phi field
rlog (float): Maximum logging range of the echosounder (m)
tpi (float): Transmit pulse interval, or ping rate (s)
freq (int): frequecy (kHz)
rank (int): Rank for percentile operation: [0, 100]
Returns:
bool: 2D array with seabed mask
"""
# raise errors if wrong arguments
if r0>r1:
raise Exception('Minimum range has to be shorter than maximum range')
# return empty mask if searching range is outside the echosounder range
if (r0>r[-1]) or (r1<r[0]):
return np.zeros_like(Sv, dtype=bool)
# delimit the analysis within user-defined range limits
i0 = np.nanargmin(abs(r - r0))
i1 = np.nanargmin(abs(r - r1)) + 1
Svchunk = Sv [i0:i1, :]
thetachunk = theta[i0:i1, :]
phichunk = phi [i0:i1, :]
# get blur kernels with theta & phi width dimensions
ktheta = np.ones((wtheta, wtheta))/wtheta**2
kphi = np.ones((wphi , wphi ))/wphi **2
# perform mean square convolution and mask if above theta & phi thresholds
thetamaskchunk = convolve2d(thetachunk, ktheta, 'same',
boundary='symm')**2 > ttheta
phimaskchunk = convolve2d(phichunk, kphi, 'same',
boundary='symm')**2 > tphi
anglemaskchunk = thetamaskchunk | phimaskchunk
# remove aliased seabed items when estimated True seabed can not be
# detected below the logging range
if (rlog is not None) and (tpi is not None) and (freq is not None):
items = label(anglemaskchunk)
item_labels = np.unique(label(anglemaskchunk))[1:]
for il in item_labels:
item = items==il
ritem = np.nanmean(r[i0:i1][np.where(item)[0]])
rseabed = aliased2seabed(ritem , rlog, tpi, freq)
if rseabed==[]:
anglemaskchunk[item] = False
anglemaskchunk = anglemaskchunk & (Svchunk>tSv)
# if aliased seabed, mask Sv above the Sv median of angle-masked regions
if anglemaskchunk.any():
Svmedian_anglemasked = log(
np.nanpercentile(lin(Svchunk[anglemaskchunk]), rank))
if np.isnan(Svmedian_anglemasked):
Svmedian_anglemasked = np.inf
if Svmedian_anglemasked < tSv:
Svmedian_anglemasked = tSv
Svmaskchunk = Svchunk > Svmedian_anglemasked
# label connected items in Sv mask
items = nd.label(Svmaskchunk, nd.generate_binary_structure(2,2))[0]
# get items intercepted by angle mask (likely, the seabed)
intercepted = list(set(items[anglemaskchunk]))
if 0 in intercepted:
intercepted.remove(intercepted==0)
# combine angle-intercepted items in a single mask
maskchunk = np.zeros(Svchunk.shape, dtype=bool)
for i in intercepted:
maskchunk = maskchunk | (items==i)
# add data above r0 and below r1 (removed in first step)
above = np.zeros((i0, maskchunk.shape[1]), dtype=bool)
below = np.zeros((len(r) - i1, maskchunk.shape[1]), dtype=bool)
mask = np.r_[above, maskchunk, below]
# return empty mask if aliased-seabed was not detected in Theta & Phi
else:
mask = np.zeros_like(Sv, dtype=bool)
return mask
def blackwell(Sv, theta, phi, r,
r0=10, r1=1000,
tSv=-75, ttheta=702, tphi=282,
wtheta=28 , wphi=52):
"""
Detects and mask seabed using the split-beam angle and Sv, based in
"Blackwell et al (2019), Aliased seabed detection in fisheries acoustic
data". Complete article here: https://arxiv.org/abs/1904.10736
Args:
Sv (float): 2D numpy array with Sv data (dB)
theta (float): 2D numpy array with the along-ship angle (degrees)
phi (float): 2D numpy array with the athwart-ship angle (degrees)
r (float): 1D range array (m)
r0 (int): minimum range below which the search will be performed (m)
r1 (int): maximum range above which the search will be performed (m)
tSv (float): Sv threshold above which seabed is pre-selected (dB)
ttheta (int): Theta threshold above which seabed is pre-selected (dB)
tphi (int): Phi threshold above which seabed is pre-selected (dB)
wtheta (int): window's size for mean square operation in Theta field
wphi (int): window's size for mean square operation in Phi field
Returns:
bool: 2D array with seabed mask
"""
# delimit the analysis within user-defined range limits
r0 = np.nanargmin(abs(r - r0))
r1 = np.nanargmin(abs(r - r1)) + 1
Svchunk = Sv[r0:r1, :]
thetachunk = theta[r0:r1, :]
phichunk = phi[r0:r1, :]
# get blur kernels with theta & phi width dimensions
ktheta = np.ones((wtheta, wtheta))/wtheta**2
kphi = np.ones((wphi , wphi ))/wphi **2
# perform mean square convolution and mask if above theta & phi thresholds
thetamaskchunk = convolve2d(thetachunk, ktheta, 'same',
boundary='symm')**2 > ttheta
phimaskchunk = convolve2d(phichunk, kphi, 'same',
boundary='symm')**2 > tphi
anglemaskchunk = thetamaskchunk | phimaskchunk
# if aliased seabed, mask Sv above the Sv median of angle-masked regions
if anglemaskchunk.any():
Svmedian_anglemasked = log(np.nanmedian(lin(Svchunk[anglemaskchunk])))
if np.isnan(Svmedian_anglemasked):
Svmedian_anglemasked = np.inf
if Svmedian_anglemasked < tSv:
Svmedian_anglemasked = tSv
Svmaskchunk = Svchunk > Svmedian_anglemasked
# label connected items in Sv mask
items = nd.label(Svmaskchunk, nd.generate_binary_structure(2,2))[0]
# get items intercepted by angle mask (likely, the seabed)
intercepted = list(set(items[anglemaskchunk]))
if 0 in intercepted:
intercepted.remove(intercepted==0)
# combine angle-intercepted items in a single mask
maskchunk = np.zeros(Svchunk.shape, dtype=bool)
for i in intercepted:
maskchunk = maskchunk | (items==i)
# add data above r0 and below r1 (removed in first step)
above = np.zeros((r0, maskchunk.shape[1]), dtype=bool)
below = np.zeros((len(r) - r1, maskchunk.shape[1]), dtype=bool)
mask = np.r_[above, maskchunk, below]
anglemask = np.r_[above, anglemaskchunk, below] # TODO remove
# return empty mask if aliased-seabed was not detected in Theta & Phi
else:
mask = np.zeros_like(Sv, dtype=bool)
return mask, anglemask
def twod(data, idim, jdim, irvals, jrvals, log=False, operation='mean'):
"""
Resample down an array along the two dimensions, i and j.
Args:
data (float): 2D array with data to be resampled.
idim (float): i vertical dimension.
jdim (float): j horizontal dimension.
irvals (float): i resampling intervals for i vertical dimension.
jrvals (float): j resampling intervals for j horizontal dimension.
log (bool ): if True, data is considered logarithmic and it will
be converted to linear during calculations.
operation(str): type of resampling operation. Accepts "mean" or "sum".
Returns:
float: 2D resampled data array
float: 1D resampled i vertical dimension
float: 1D resampled j horizontal dimension
float: 2D array with percentage of valid samples included on each
resampled cell.
"""
# check for appropiate inputs
if len(irvals) < 2:
raise Exception('length of i resampling intervals must be >2')
if len(jrvals) < 2:
raise Exception('length of j resampling intervals must be >2')
if len(data) != len(idim):
raise Exception('data height and idim length must be the same')
if len(data[0]) != len(jdim):
raise Exception('data width and jdim length must be the same')
for irval, jrval in zip(irvals, jrvals):
if (irval < idim[0]) | (idim[-1] < irval):
raise Exception('i resampling intervals must be within idim range')
if (jrval < jdim[0]) | (jdim[-1] < jrval):
raise Exception('j resampling intervals must be within jdim range')
# convert data to linear, if logarithmic
if log is True:
data = tf.lin(data)
# get i/j axes from i/j dimensions and i/j intervals
iax = np.arange(len(idim))
jax = np.arange(len(jdim))
iaxrs = dim2ax(idim, iax, irvals)
jaxrs = dim2ax(jdim, jax, jrvals)
# declare new array to allocate resampled values, and new array to
# alllocate the percentage of values used for resampling
datar = np.zeros((len(iaxrs) - 1, len(jaxrs) - 1)) * np.nan
percentage = np.zeros((len(iaxrs) - 1, len(jaxrs) - 1)) * np.nan
# iterate along-range
for i in range(len(iaxrs) - 1):
# get i indexes to locate the samples for the binning operation
idx0 = np.where(iax - iaxrs[i + 0] <= 0)[0][-1]
idx1 = np.where(iaxrs[i + 1] - iax > 0)[0][-1]
idx = np.arange(idx0, idx1 + 1)
# get i weights as the sum of the sample proportions taken
iweight0 = 1 - abs(iax[idx0] - iaxrs[i + 0])
iweight1 = abs(iax[idx1] - iaxrs[i + 1])
if len(idx) > 1:
iweights = np.r_[iweight0, np.ones(len(idx) - 2), iweight1]
else:
iweights = np.array([iweight0 - 1 + iweight1])
# iterate along-pings
for j in range(len(jaxrs) - 1):
# get j indexes to locate the samples for the binning operation
jdx0 = np.where(jax - jaxrs[j + 0] <= 0)[0][-1]
jdx1 = np.where(jaxrs[j + 1] - jax > 0)[0][-1]
jdx = np.arange(jdx0, jdx1 + 1)
# get j weights as the sum of the sample proportions taken
jweight0 = 1 - abs(jax[jdx0] - jaxrs[j + 0])
jweight1 = abs(jax[jdx1] - jaxrs[j + 1])
if len(jdx) > 1:
jweights = np.r_[jweight0, np.ones(len(jdx) - 2), jweight1]
else:
jweights = np.array([jweight0 - 1 + jweight1])
# get data and weight 2D matrices for the binning operation
d = data[idx[0]:idx[-1] + 1, jdx[0]:jdx[-1] + 1]
w = np.multiply.outer(iweights, jweights)
# if d is an all-NAN array, return NAN as the weighted operation
# and zero as the percentage of valid numbers used for binning
if np.isnan(d).all():
datar[i, j] = np.nan
percentage[i, j] = 0
#compute weighted operation & percentage of valid numbers otherwise
else:
w_ = w.copy()
w_[np.isnan(d)] = np.nan
if operation == 'mean':
datar[i, j] = np.nansum(d * w_) / np.nansum(w_)
elif operation == 'sum':
datar[i, j] = np.nansum(d * w_)
else:
raise Exception('Operation not recognised')
percentage[i, j] = np.nansum(w_) / np.nansum(w) * 100
# convert back to logarithmic, if data was logarithmic
if log is True:
datar = tf.log(datar)
# get resampled dimensions from resampling intervals
idimr = irvals[:-1]
jdimr = jrvals[:-1]
# return data
return datar, idimr, jdimr, percentage
def oned(data, dim, rvals, axis, log=False, operation='mean'):
"""
Resample down an array along i or j dimension.
Args:
data (float) : 2D array with data to be resampled.
dim (float) : original dimension.
rvals (float) : resampling dimension intervals.
axis (int ) : resampling axis (0= i vertical ax, 1= j horizontal ax).
log (bool ) : if True, data is considered logarithmic and it will
be converted to linear during the calculations.
operation(str): type of resampling operation. Accepts "mean" or "sum".
Returns:
float: 2D resampled data array
float: 1D resampled array corresponding to either i or j dimension.
float: 2D array with percentage of valid samples included on each
resampled cell.
"""
# check if appropiate axis input
if axis > 1:
raise Exception('axis must be 0 or 1')
# check if appropiate resampled dimension
if len(rvals) < 2:
raise Exception('length of resampling intervals must be >2')
# check if intervals are within the dimension range of values
for rval in rvals:
if (rval < dim[0]) | (dim[-1] < rval):
raise Exception('resampling intervals must be within dim range')
# convert data to linear, if logarithmic
if log is True:
data = tf.lin(data)
# get axis from dimension
ax = np.arange(len(dim))
axrs = dim2ax(dim, ax, rvals)
# proceed along i dimension
if axis == 0:
iax = ax
iaxrs = axrs
# check data and axis match
if len(data) != len(iax):
raise Exception('data height and i dimension length must be equal')
# declare new array to allocate resampled values, and new array to
# alllocate the percentage of values used for resampling
datar = np.zeros((len(iaxrs) - 1, len(data[0]))) * np.nan
percentage = np.zeros((len(iaxrs) - 1, len(data[0]))) * np.nan
# iterate along i dimension
for i in range(len(iaxrs) - 1):
# get i indexes to locate the samples for the resampling operation
idx0 = np.where(iax - iaxrs[i + 0] <= 0)[0][-1]
idx1 = np.where(iaxrs[i + 1] - iax > 0)[0][-1]
idx = np.arange(idx0, idx1 + 1)
# get i weights as the sum of the proportions of samples taken
iweight0 = 1 - abs(iax[idx0] - iaxrs[i + 0])
iweight1 = abs(iax[idx1] - iaxrs[i + 1])
if len(idx) > 1:
iweights = np.r_[iweight0, np.ones(len(idx) - 2), iweight1]
else:
iweights = np.array([iweight0 - 1 + iweight1])
# get data and weight 2D matrices for the resampling operation
d = data[idx[0]:idx[-1] + 1, :]
w = np.multiply.outer(iweights, np.ones(len(data[0])))
# if d is an all-NAN array, return NAN as the weighted operation
# and zero as the percentage of valid numbers used for binning
if np.isnan(d).all():
datar[i, :] = np.nan
percentage[i, :] = 0
# compute weighted operation and percentage valid numbers otherwise
else:
w_ = w.copy()
w_[np.isnan(d)] = np.nan
if operation == 'mean':
datar[i, :] = np.nansum(d * w_, axis=0) / np.nansum(w_,
axis=0)
elif operation == 'sum':
datar[i, :] = np.nansum(d * w_, axis=0)
else:
raise Exception('Operation not recognised')
percentage[i, :] = np.nansum(w_, axis=0) / np.nansum(
w, axis=0) * 100
# convert back to logarithmic, if data was logarithmic
if log is True:
datar = tf.log(datar)
# get resampled dimension from resampling interval
dimr = rvals
# return data
return datar, dimr, percentage
# proceed along j dimension
if axis == 1:
jax = ax
jaxrs = axrs
# check data and axis match
if len(data[0]) != len(jax):
raise Exception('data width and j dimension lenght must be equal')
# declare new array to allocate resampled values, and new array to
# alllocate the percentage of values used for resampling
datar = np.zeros((len(data), len(jaxrs) - 1)) * np.nan
percentage = np.zeros((len(data), len(jaxrs) - 1)) * np.nan
# iterate along j dimension
for j in range(len(jaxrs) - 1):
# get j indexes to locate the samples for the resampling operation
jdx0 = np.where(jax - jaxrs[j + 0] <= 0)[0][-1]
jdx1 = np.where(jaxrs[j + 1] - jax > 0)[0][-1]
jdx = np.arange(jdx0, jdx1 + 1)
# get j weights as the sum of the proportions of samples taken
jweight0 = 1 - abs(jax[jdx0] - jaxrs[j + 0])
jweight1 = abs(jax[jdx1] - jaxrs[j + 1])
if len(jdx) > 1:
jweights = np.r_[jweight0, np.ones(len(jdx) - 2), jweight1]
else:
jweights = np.array([jweight0 - 1 + jweight1])
# get data and weight 2D matrices for the resampling operation
d = data[:, jdx[0]:jdx[-1] + 1]
w = np.multiply.outer(np.ones(len(data)), jweights)
# if d is an all-NAN array, return NAN as the weighted operation
# and zero as the percentage of valid numbers used for resampling
if np.isnan(d).all():
datar[:, j] = np.nan
percentage[:, j] = 0
# compute weighted operation and percentage valid numbers otherwise
else:
w_ = w.copy()
w_[np.isnan(d)] = np.nan
if operation == 'mean':
datar[:, j] = np.nansum(d * w_, axis=1) / np.nansum(w_,
axis=1)
elif operation == 'sum':
datar[:, j] = np.nansum(d * w_, axis=1)
else:
raise Exception('Operation not recognised')
percentage[:, j] = np.nansum(w_, axis=1) / np.nansum(
w, axis=1) * 100
# convert back to logarithmic, if data was logarithmic
if log is True:
datar = tf.log(datar)
# get resampled dimension from resampling intervals
dimr = rvals
# return data
return datar, dimr, percentage
# Wang's algorithm.
Sv120clean = mIN.wang(Sv120,
thr=(-70, -40),
erode=[(3, 3)],
dilate=[(7, 7)],
median=[(7, 7)])[0]
#------------------------------------------------------------------------------
# Remove data outside the range 20-250 m
m120rg = mRG.outside(Sv120clean, r120, 20, 250)
Sv120clean[m120rg] = np.nan
#------------------------------------------------------------------------------
# Convolute Sv prior to mask swarms, with 3x3 moving window
k = np.ones((3, 3)) / 3**2
Sv120cvv = tf.log(convolve2d(tf.lin(Sv120clean), k, 'same', boundary='symm'))
#------------------------------------------------------------------------------
# Mask swarms using Weill's algorithm
m120sh = mSH.weill(Sv120cvv,
thr=-70,
maxvgap=15,
maxhgap=0,
minhlen=3,
minvlen=15)[0]
#------------------------------------------------------------------------------
# Figures
plt.figure(figsize=(8, 5))
plt.subplots_adjust(left=0.08, right=0.91, bottom=0.08, top=0.95, wspace=0.08)
gs = gridspec.GridSpec(1, 3, width_ratios=[1, 1, .05])
def wang(Sv, thr=(-70,-40), erode=[(3,3)], dilate=[(5,5),(7,7)],
median=[(7,7)]):
"""
Clean impulse noise from Sv data following the method decribed by:
Wang et al. (2015) ’A noise removal algorithm for acoustic data with
strong interference based on post-processing techniques’, CCAMLR
SG-ASAM: 15/02.
This algorithm runs different cycles of erosion, dilation, and median
filtering to clean impulse noise from Sv. Note that this function
returns a clean/corrected Sv array, instead of a boolean array indicating
the occurrence of impulse noise.
Args:
Sv (float) : 2D numpy array with Sv data (dB).
thr (int/float): 2-element tupple with bottom/top Sv thresholds (dB)
erode (int) : list of 2-element tupples indicating the window's
size for each erosion cycle.
dilate (int) : list of 2-element tupples indicating the window's
size for each dilation cycle.
median (int) : list of 2-element tupples indicating the window's
size for each median filter cycle.
Returns:
float : 2D array with clean Sv data.
bool : 2D array with mask indicating valid clean Sv data.
"""
# set weak noise and strong interference as vacant samples (-999)
Sv_thresholded = Sv.copy()
Sv_thresholded[(Sv<thr[0])|(Sv>thr[1])] = -999
# remaining weak interferences will take neighbouring vacant values
# by running erosion cycles
Sv_eroded = Sv.copy()
for e in erode:
Sv_eroded = erosion(Sv_thresholded, np.ones(e))
# the last step might have turned interferences inside biology into vacant
# samples, this is solved by running dilation cycles
Sv_dilated = Sv_eroded.copy()
for d in dilate:
Sv_dilated = dilation(Sv_dilated, np.ones(d))
# dilation has modified the Sv value of biological features, so these are
# now corrected to corresponding Sv values before the erosion/dilation
Sv_corrected1 = Sv_dilated.copy()
mask_bio = (Sv_dilated>=thr[0]) & (Sv_dilated<thr[1])
Sv_corrected1[mask_bio] = Sv_thresholded[mask_bio]
# compute median convolution in Sv corrected array
Sv_median = Sv_corrected1.copy()
for m in median:
Sv_median = tf.log(medianf(tf.lin(Sv_median), footprint=np.ones(m)))
# any vacant sample inside biological features will be corrected with
# the median of corresponding neighbouring samples
Sv_corrected2 = Sv_corrected1.copy()
mask_bio = (Sv>=thr[0]) & (Sv<thr[1])
mask_vacant = Sv_corrected1==-999
Sv_corrected2[mask_vacant&mask_bio] = Sv_median[mask_vacant&mask_bio]
# get mask indicating edges, where swarms analysis couldn't be performed
mask_ = np.ones_like(Sv_corrected2, dtype=bool)
idx = int((max([e[0], d[0]])-1)/2)
jdx = int((max([e[1], d[1]])-1)/2)
mask_[idx:-idx, jdx:-jdx] = False
return Sv_corrected2, mask_
def get_fmean(fdo, res, logtransformed=True):
"""
Computes mean profiles for every variable defined within the Functional
data object.
Args:
fdo (object): Functional data object
res (float): depth resolution of mean profile (m)
Returns:
dict: Mean profile results, including the following keys:
- 'x' : 1d- array with range dimension (m)
- 'y' : 2d-array with Sv profiles (dB)
- 'ymean' : 1d-array with Sv mean profile (dB)
- 'yplus' : 1d-array with effect of adding PC variance (dB)
- 'yminus': 1d-array with effect of subtracting PC variance (dB)
"""
# iterate through dimension variables
data = {}
ndim = len(fdo.dim_names)
for k in range(ndim):
# create single-dimension fdobjs
fdo_k = copy.deepcopy(fdo)
if ndim > 1:
fdo_k.coefficients = fdo_k.coefficients[:, :, k]
fdo_k.dim_names = [fdo_k.dim_names[k]]
# evaluate y at x positions in every single-dimension fdobj
xrange = fdo_k.domain_range[0]
x = np.arange(xrange[0], xrange[1] + res, res)
y = fdo_k.evaluate(x)[:, :, 0].T
x = x + fdo_k.domain_depth[k][0]
# compute 'y' mean, confidence intervals, and standard deviations
if logtransformed:
ymean = log(np.mean(lin(y), axis=1))
yminus = log(np.quantile(lin(y), .050, axis=1))
yplus = log(np.quantile(lin(y), .950, axis=1))
ystd = y * np.nan
else:
ymean = np.mean(y, axis=1)
yminus = np.quantile(y, .050, axis=1)
yplus = np.quantile(y, .950, axis=1)
ystd = np.std(y, axis=1)
# store data
data.update({
fdo_k.dim_names[0]: {
'x': x,
'y': y,
'ymean': ymean,
'yplus': yplus,
'yminus': yminus,
'ystd': ystd
}
})
# return data
return data
def fielding(Sv, r, r0, r1, n, thr, roff, jumps=5, maxts=-35, start=0):
"""
Mask transient noise with method proposed by Fielding et al (unpub.).
A comparison is made ping by ping with respect to a block in a reference
layer set at far range, where transient noise mostly occurs. If the ping
median is greater than the block median by a user-defined threshold, the
ping will be masked all the way up, until transient noise dissapears, or
until it gets the minimum range allowed by the user.
transient transient ping
noise noise evaluated
| | |
______ | _______________________ | ____________.....V.....____________
||| far range interval ||| . block . |
_____|||||_____________________|||||___________...........____________|
When transient noise is detected, comparisons start to be made in the same
ping but moving vertically every x meters (jumps). Pings with transient
noise will be masked up to where the ping is similar to the block according
with a secondary threshold or until it gets the exclusion range depth.
Args:
Sv (float): 2D numpy array with Sv data to be masked (dB).
r (float): 1D numpy array with range data (m).
r0 (int ): range below which transient noise is evaluated (m).
r1 (int ): range above which transient noise is evaluated (m).
n (int ): n of preceeding & subsequent pings defining the block.
thr (int ): user-defined threshold for side-comparisons (dB).
roff (int ): range above which masking is excluded (m).
maxts (int ): max transient noise permited, prevents to interpret
seabed as transient noise (dB).
jumps (int ): height of vertical steps (m).
start (int ): ping index to start processing.
Returns:
list: 2D boolean array with TN mask and 2D boolean array with mask
indicating where TN detection was unfeasible.
"""
# raise errors if wrong arguments
if r0 > r1:
raise Exception('Minimum range has to be shorter than maximum range')
# return empty mask if searching range is outside the echosounder range
if (r0 > r[-1]) or (r1 < r[0]):
mask = np.zeros_like(Sv, dtype=bool)
mask_ = np.zeros_like(Sv, dtype=bool)
return mask, mask_
# get upper and lower range indexes
up = np.argmin(abs(r - r0))
lw = np.argmin(abs(r - r1))
# get minimum range index admitted for processing
rmin = np.argmin(abs(r - roff))
# get scaling factor index
sf = np.argmin(abs(r - jumps))
# start masking process
mask_ = np.zeros(Sv.shape, dtype=bool)
mask = np.zeros(Sv.shape, dtype=bool)
for j in range(start, len(Sv[0])):
# mask where TN evaluation is unfeasible (e.g. edge issues, all-NANs)
if (j - n < 0) | (j + n > len(Sv[0]) - 1) | np.all(
np.isnan(Sv[up:lw, j])):
mask_[:, j] = True
# evaluate ping and block averages otherwise
else:
pingmedian = log(np.nanmedian(lin(Sv[up:lw, j])))
pingp75 = log(np.nanpercentile(lin(Sv[up:lw, j]), 75))
blockmedian = log(np.nanmedian(lin(Sv[up:lw, j - n:j + n])))
# if ping median below 'maxts' permited, and above enough from the
# block median, mask all the way up until noise dissapears
if (pingp75 < maxts) & ((pingmedian - blockmedian) > thr[0]):
r0, r1 = up - sf, up
while r0 > rmin:
pingmedian = log(np.nanmedian(lin(Sv[r0:r1, j])))
blockmedian = log(np.nanmedian(lin(Sv[r0:r1,
j - n:j + n])))
r0, r1 = r0 - sf, r1 - sf
if (pingmedian - blockmedian) < thr[1]:
break
mask[r0:, j] = True
return [mask[:, start:], mask_[:, start:]]