# Make some travel time data and add noise
seed = 0 # Set the random seed so that points are the same every time
src_loc = utils.random_points(area, 80, seed=seed)
rec_loc = utils.circular_points(area, 30, random=True, seed=seed)
srcs, recs = utils.connect_points(src_loc, rec_loc)
tts = ttime2d.straight(model, 'vp', srcs, recs)
tts, error = utils.contaminate(tts, 0.02, percent=True, return_stddev=True,
seed=seed)
# Make the mesh
mesh = SquareMesh(area, shape)
# and run the inversion
misfit = srtomo.SRTomo(tts, srcs, recs, mesh)
regularization = Smoothness2D(mesh.shape)
# Will use the l-curve criterion to find the best regularization parameter
tomo = LCurve(misfit, regularization,
[10 ** i for i in np.arange(0, 10, 1)], jobs=8).fit()
mesh.addprop('vp', tomo.estimate_)
# Plot the L-curve annd print the regularization parameter estimated
mpl.figure()
mpl.title('L-curve: triangle marks the best solution')
tomo.plot_lcurve()
print "Estimated regularization parameter: %g" % (tomo.regul_param_)
# Calculate and print the standard deviation of the residuals
# Should be close to the data error if the inversion was able to fit the data
residuals = tomo.residuals()
print "Assumed error: %g" % (error)
print "Standard deviation of residuals: %g" % (np.std(residuals))
mpl.figure(figsize=(14, 5))
rec_loc = utils.circular_points(area, 30, random=True, seed=seed)
srcs, recs = utils.connect_points(src_loc, rec_loc)
tts = ttime2d.straight(model, 'vp', srcs, recs)
tts, error = utils.contaminate(tts,
0.02,
percent=True,
return_stddev=True,
seed=seed)
# Make the mesh
mesh = SquareMesh(area, shape)
# and run the inversion
misfit = srtomo.SRTomo(tts, srcs, recs, mesh)
regularization = Smoothness2D(mesh.shape)
# Will use the l-curve criterion to find the best regularization parameter
tomo = LCurve(misfit,
regularization, [10**i for i in np.arange(0, 10, 1)],
jobs=8).fit()
mesh.addprop('vp', tomo.estimate_)
# Plot the L-curve annd print the regularization parameter estimated
mpl.figure()
mpl.title('L-curve: triangle marks the best solution')
tomo.plot_lcurve()
print "Estimated regularization parameter: %g" % (tomo.regul_param_)
# Calculate and print the standard deviation of the residuals
# Should be close to the data error if the inversion was able to fit the data
residuals = tomo.residuals()
print "Assumed error: %g" % (error)
print "Standard deviation of residuals: %g" % (np.std(residuals))
# Make synthetic data
inc, dec = -60, 23
props = {'magnetization': 10}
model = [mesher.Prism(-500, 500, -1000, 1000, 500, 4000, props)]
shape = (25, 25)
x, y, z = gridder.regular([-5000, 5000, -5000, 5000], shape, z=0)
tf = utils.contaminate(prism.tf(x, y, z, model, inc, dec), 5, seed=0)
# Setup the layer
layer = mesher.PointGrid([-7000, 7000, -7000, 7000], 700, (50, 50))
# Estimate the magnetization intensity
# Need to apply regularization so that won't try to fit the error as well
misfit = EQLTotalField(x, y, z, tf, inc, dec, layer)
regul = Damping(layer.size)
# Use an L-curve analysis to find the best regularization parameter
solver = LCurve(misfit, regul, [10 ** i for i in range(-30, -15)]).fit()
residuals = solver.residuals()
layer.addprop('magnetization', solver.estimate_)
print "Residuals:"
print "mean:", residuals.mean()
print "stddev:", residuals.std()
# Now I can forward model the layer at the south pole and check against the
# true solution of the prism
tfpole = prism.tf(x, y, z, model, -90, 0)
tfreduced = sphere.tf(x, y, z, layer, -90, 0)
mpl.figure()
mpl.suptitle('L-curve')
mpl.title("Estimated regularization parameter: %g" % (solver.regul_param_))
solver.plot_lcurve()
area = [-5000, 5000, -5000, 5000]
x1, y1, z1 = gridder.scatter(area, 80, z=0, seed=0)
gz = utils.contaminate(prism.gz(x1, y1, z1, model), 0.1, seed=0)
x2, y2, z2 = gridder.regular(area, (10, 50), z=-200)
gzz = utils.contaminate(prism.gzz(x2, y2, z2, model), 5, seed=0)
# Setup the layer
layer = mesher.PointGrid([-6000, 6000, -6000, 6000], 500, (50, 50))
# and the inversion
# Apply a scaling factor to make both portions of the misfit the same order of
# magnitude
scale = np.linalg.norm(gz)**2 / np.linalg.norm(gzz)**2
misfit = (EQLGravity(x1, y1, z1, gz, layer) +
scale * EQLGravity(x2, y2, z2, gzz, layer, field='gzz'))
regul = Smoothness2D(layer.shape)
# Use an L-curve analysis to find the best regularization parameter
solver = LCurve(misfit, regul, [10**i for i in range(-30, -20)]).fit()
layer.addprop('density', solver.estimate_)
# Now I can forward model gz using my layer to produce an integrated map in a
# much denser region
shape = (50, 50)
x, y, z = gridder.regular(area, shape, z=0)
gz_layer = sphere.gz(x, y, z, layer)
gz_true = prism.gz(x, y, z, model)
mpl.figure()
mpl.suptitle('L-curve')
mpl.title("Estimated regularization parameter: %g" % (solver.regul_param_))
solver.plot_lcurve()
mpl.grid()
# Make some travel time data and add noise
seed = 0 # Set the random seed so that points are the same everythime
src_loc = utils.random_points(area, 80, seed=seed)
rec_loc = utils.circular_points(area, 30, random=True, seed=seed)
srcs, recs = utils.connect_points(src_loc, rec_loc)
tts = ttime2d.straight(model, 'vp', srcs, recs)
tts, error = utils.contaminate(tts, 0.02, percent=True, return_stddev=True,
seed=seed)
# Make the mesh
mesh = SquareMesh(area, shape)
# and run the inversion
misfit = srtomo.SRTomo(tts, srcs, recs, mesh)
regularization = TotalVariation2D(10**-10, mesh.shape)
tomo = misfit + regularization
tomo = LCurve(misfit, regularization,
[10**i for i in np.arange(-3, 3, 0.5)], jobs=8)
# Since Total Variation is a non-linear function, then the tomography becomes
# non-linear. So we need to configure fit to use the Levemberg-Marquardt
# algorithm, a gradient descent method, that requires an initial estimate
tomo.config('levmarq', initial=0.00001*np.ones(mesh.size)).fit()
mesh.addprop('vp', tomo.estimate_)
# Plot the L-curve annd print the regularization parameter estimated
mpl.figure()
mpl.title('L-curve: triangle marks the best solution')
tomo.plot_lcurve()
print "Estimated regularization parameter: %g" % (tomo.regul_param_)
# Calculate and print the standard deviation of the residuals
# Should be close to the data error if the inversion was able to fit the data
residuals = tomo.residuals()
area = [-5000, 5000, -5000, 5000]
x1, y1, z1 = gridder.scatter(area, 80, z=0, seed=0)
gz = utils.contaminate(prism.gz(x1, y1, z1, model), 0.1, seed=0)
x2, y2, z2 = gridder.regular(area, (10, 50), z=-200)
gzz = utils.contaminate(prism.gzz(x2, y2, z2, model), 5, seed=0)
# Setup the layer
layer = mesher.PointGrid([-6000, 6000, -6000, 6000], 500, (50, 50))
# and the inversion
# Apply a scaling factor to make both portions of the misfit the same order of
# magnitude
scale = np.linalg.norm(gz) ** 2 / np.linalg.norm(gzz) ** 2
misfit = (EQLGravity(x1, y1, z1, gz, layer)
+ scale * EQLGravity(x2, y2, z2, gzz, layer, field='gzz'))
regul = Smoothness2D(layer.shape)
# Use an L-curve analysis to find the best regularization parameter
solver = LCurve(misfit, regul, [10 ** i for i in range(-30, -20)]).fit()
layer.addprop('density', solver.estimate_)
# Now I can forward model gz using my layer to produce an integrated map in a
# much denser region
shape = (50, 50)
x, y, z = gridder.regular(area, shape, z=0)
gz_layer = sphere.gz(x, y, z, layer)
gz_true = prism.gz(x, y, z, model)
mpl.figure()
mpl.suptitle('L-curve')
mpl.title("Estimated regularization parameter: %g" % (solver.regul_param_))
solver.plot_lcurve()
mpl.grid()
# Reverse x because vertices must be clockwise.
xs = np.linspace(0, 100000, 100)[::-1]
depths = (-1e-15*(xs - 50000)**4 + 8000 -
3000*np.exp(-(xs - 70000)**2/(10000**2)))
depths -= depths.min() # Reduce depths to zero
props = {'density': -300}
model = Polygon(np.transpose([xs, depths]), props)
x = np.linspace(0, 100000, 100)
z = -100*np.ones_like(x)
data = utils.contaminate(talwani.gz(x, z, [model]), 0.5, seed=0)
# Make the solver and run the inversion
misfit = PolygonalBasinGravity(x, z, data, 50, props, top=0)
regul = Smoothness1D(misfit.nparams)
# Use an L-curve analysis to find the best regularization parameter
lc = LCurve(misfit, regul, [10**i for i in np.arange(-10, -5, 0.5)], jobs=4)
initial = 3000*np.ones(misfit.nparams)
lc.config('levmarq', initial=initial).fit()
mpl.figure()
mpl.subplot(2, 2, 1)
mpl.plot(x, data, 'ok', label='observed')
mpl.plot(x, lc.predicted(), '-r', linewidth=2, label='predicted')
mpl.legend()
ax = mpl.subplot(2, 2, 3)
mpl.polygon(model, fill='gray', alpha=0.5)
mpl.polygon(lc.estimate_, style='o-r')
ax.invert_yaxis()
mpl.subplot(1, 2, 2)
mpl.title('L-curve')
lc.plot_lcurve()
