fig.subplots_adjust(wspace=0.45)
i=0
for (ax1,ax2,ax3,ax4), (_,files) in zip(m_axs, subset.iterrows()):
print("file = ",files['shot'])
input_data = ReadSegyio(segyfile=files['shot'],
keep_hdrs=[],drop_hdrs=[],
gather_id="FieldRecord",verbose=2)
dt = input_data.sample_rate/1000
# Extract numpy array with trace values and transpose so
# col=time and row=trace
d = input_data.data["gather"][0].T
# Apply simple processing: tpow gain to balance amplitudes upto time=maxt
d = gains.tpow(d,dt,tpow=0.25,tmin=0,maxt=2)
# 95th percentile gain to clip large outliers
d = gains.perc(d,95)
tmax = (input_data.n_samples-1)*dt
d = gains.standardize(d)
ax1.imshow(d,cmap='Greys',aspect='auto',extent=[0,DX*input_data.n_traces,
1000*dt*input_data.n_samples,0])
if i==0: ax1.set_title('data')
if i==len(subset)-1: ax1.set_xlabel('Offset (m)')
ax1.set_ylabel('Time (ms)')
# Convert shot gather to phase velocity vs. frequency panel
fv_abs = np.squeeze(d_to_fv(d,dt=dt,side=1,f_abs=0).numpy())
fvimg1 = ax2.imshow(fv_abs,cmap='jet',aspect='auto',extent=[0,FMAX,VMIN,VMAX],
vmin=-3,vmax=3)
if i==0: ax2.set_title('phase velocity spectrum')
if i==len(subset)-1: ax2.set_xlabel('Frequency (Hz)')
def get_seismic_and_vpvs(data_file, model_file):
"""
Used in training data loader to create input and two outputs
Reads data and model SEGY from path strings and returns Tensorflow objects
Model: SEGY file consisting of Vp, Vs and density. This functions extracts
only vp and vs and converts to float32 tensors
Data: SEGY file. This function extract trace values and applies
pre-processing to balance amplitudes, low-cut filter. Then the shot gather
is transformed from t-x to v-f (freq vs phase velocity)
The phase velocity vs. frequency panel is resized to IMSZ[0] x IMSZ[1]
Data augmentation is commented out. Perhaps data augmentation is best applied
in the time domain
Relies on global parameter IMSZ and boolean AUGMENT!
Parameters:
data_file Full path to SEGY format seismic data
model_filename Full path to SEGY format model
Output
fv (input_1) phase velocity (col) vs. frequency (row) panel (tf.float32)
vp (vp_output) 1D array of velocity values (tf.float32)
vs (vs_output) 1D array of velocity values (tf.float32)
"""
# Process filenames
# Convert to numpy and then convert bytestreams to ASCII
data_file = data_file.numpy().decode('ASCII')
model_file = model_file.numpy().decode('ASCII')
# For debugging
#print("data_file: {}, model_file: {}".format(data_file, model_file))
# Read SEGY model file with SEGYIO
model1d = ReadSegyio(segyfile=model_file,
keep_hdrs=[],
drop_hdrs=[],
gather_id="FieldRecord",
verbose=0)
# Extract numpy array with trace values
m = model1d.data["gather"][0].T
# Extract Vs information
vp = m[:IZMAX, 0]
vs = m[:IZMAX, 1]
# Convert to tensorflow Tensor dtype
vp = tf.convert_to_tensor(vp)
vs = tf.convert_to_tensor(vs)
# Read SEGY data file with SEGYIO
seismic = ReadSegyio(segyfile=data_file,
keep_hdrs=[],
drop_hdrs=[],
gather_id="FieldRecord",
verbose=0)
dt = seismic.sample_rate / 1000
# Extract numpy array with trace values and transpose so
# col=time and row=trace
d = seismic.data["gather"][0].T
# Apply simple processing: tpow gain to balance amplitudes upto time=maxt
d = gains.tpow(d, dt, tpow=0.25, tmin=0, maxt=2)
# Apply a high-pass filter
#d = np.apply_along_axis(lambda m: butter_highpass_filter(m,order=6,
# lowcut=6,fs=1/dt),
# axis=0, arr=d)
# 95th percentile gain to clip large outliers
d = gains.perc(d, 95)
# Convert shot gather to phase velocity vs. frequency panel
fv = d_to_fv(d, dt=dt, side=1)
#if AUGMENT:
# # Randomly change amplitudes (bulk changes)
# Careful: random_brightness changes the mean
# fv = tf.image.random_brightness(fv, 1, seed=42)
# fv = tf.image.random_contrast( fv, 0.75, 2, seed=42)
# Resize image, avoid tf.image.resize
# Bi-linear interpolation is fine, it preserve the amplitude ranges fairly
# well if we turn off the antialias filter. More advanced schemes lessen
# the dynamic range
fv = tf.image.resize(fv, [IMSZ[0], IMSZ[1]],
method='bilinear',
antialias=False)
# Force to specified dtypes to reduce memory requirements. Note that
# tf.image.resize always outputs float32. These have to be consistent
# with the dtypes specified in the Tout option in the tf.py_function during
# loading with the parallel mapping function
fv = tf.cast(fv, dtype=tf.float32)
vp = tf.cast(vp, dtype=tf.float32)
vs = tf.cast(vs, dtype=tf.float32)
# Return dictionary so when can specify named input/output in the
# Tensorflow model
return ({"input_1": fv}, {"vp_output": vp, "vs_output": vs})
def get_seismic(data_file):
"""
Used in data loader with seismic only (in case no velocity model is
available. Reads data SEGY file from path strings and returns Tensorflow
object
Data: SEGY file. This function extract trace values and applies
pre-processing to balance amplitudes, low-cut filter. Then the shot gather
is transformed from t-x to v-f (freq vs phase velocity)
The phase velocity vs. frequency panel is resized to IMSZ[0] x IMSZ[1]
Relies on global parameter IMSZ !
Parameters:
data_file Full path to SEGY format seismic data
Output
fv phase velocity (col) vs. frequency (row) panel (tf.float32)
"""
# Process filename
# Convert to numpy and then convert bytestreams to ASCII
data_file = data_file.numpy().decode('ASCII')
# Read SEGY data file with SEGYIO
seismic = ReadSegyio(segyfile=data_file,
keep_hdrs=[],
drop_hdrs=[],
gather_id="FieldRecord",
verbose=0)
dt = seismic.sample_rate / 1000
# Extract numpy array with trace values and transpose so
# col=time and row=trace
d = seismic.data["gather"][0].T
# Apply simple processing: tpow gain to balance amplitudes upto time=maxt
d = gains.tpow(d, seismic.sample_rate / 1000, tpow=0.25, tmin=0, maxt=2)
# Apply a high-pass filter
#d = np.apply_along_axis(lambda m: butter_highpass_filter(m,order=6,
# lowcut=6,fs=1/dt),
# axis=0, arr=d)
# 95th percentile gain to clip large outliers
d = gains.perc(d, 95)
# Convert shot gather to phase velocity vs. frequency panel
fv = d_to_fv(d, dt=dt, side=OFFSIDE)
# Resize image, avoid tf.image.resize
# Bi-linear interpolation is fine, it preserve the amplitude ranges fairly
# well if we turn off the antialias filter. More advanced schemes lessen
# the dynamic range
fv = tf.image.resize(fv, [IMSZ[0], IMSZ[1]],
method='bilinear',
antialias=False)
# Force to specified dtypes to reduce memory requirements. Note that
# tf.image.resize always outputs float32. These have to be consistent
# with the dtypes specified in the Tout option in the tf.py_function during
# loading with the parallel mapping function
fv = tf.cast(fv, dtype=tf.float32)
return fv
#for (ax1,ax2,ax3), (_,c_row) in zip(m_axs, model_2b.sample(sample_rows).iterrows()):
for (ax1, ax2, ax3), (_, c_row) in zip(m_axs, subset.iterrows()):
input_data = ReadSegyio(segyfile=c_row['shot'],
keep_hdrs=[],
drop_hdrs=[],
gather_id="FieldRecord",
verbose=2)
input_model = ReadSegyio(segyfile=c_row['model'],
keep_hdrs=[],
drop_hdrs=[],
gather_id="FieldRecord",
verbose=2)
dt = input_data.sample_rate / 1000
d = input_data.data["gather"][0].T
d = gains.tpow(d, dt, tpow=0.25, tmin=0, maxt=2)
d = gains.perc(d, 95)
tmax = (input_data.n_samples - 1) * dt
d = gains.standardize(d)
fv = np.squeeze(d_to_fv(d, dt=dt, side=1).numpy())
m = input_model.data["gather"][0].T
vs1d = m[:IZMAX, 1] # 0=Vp, 1=Vs, 2=rho
vp1d = m[:IZMAX, 0] # 0=Vp, 1=Vs, 2=rho
ax1.imshow(d,
cmap='Greys',
aspect='auto',
extent=[
0, DX * input_data.n_traces,
1000 * dt * input_data.n_samples, 0
])