def test_eql_boost_large_region():
"""
Check if EQLHarmonicBoost works on a large region
The iterative process ignores the effect of sources on far observation
points. If the region is very large, this error should be diminished.
"""
# Define a squared region
region = (-1000e3, 1000e3, -1000e3, 1000e3)
# Build synthetic point masses
points = vd.grid_coordinates(region=region,
shape=(6, 6),
extra_coords=-1e3)
masses = vd.datasets.CheckerBoard(amplitude=1e13,
region=region).predict(points)
# Define a set of observation points
coordinates = vd.grid_coordinates(region=region,
shape=(40, 40),
extra_coords=0)
# Get synthetic data
data = hm.point_mass_gravity(coordinates, points, masses, field="g_z")
# The interpolation should be sufficiently accurate on the data points
eql = EQLHarmonicBoost(window_size=100e3)
eql.fit(coordinates, data)
assert mean_squared_error(
data, eql.predict(coordinates)) < 1e-5 * vd.maxabs(data)
# Gridding onto a denser grid should be reasonably accurate when compared
# to synthetic values
grid = vd.grid_coordinates(region=region, shape=(60, 60), extra_coords=0)
true = hm.point_mass_gravity(grid, points, masses, field="g_z")
assert mean_squared_error(true, eql.predict(grid)) < 1e-3 * vd.maxabs(data)
def test_eql_boost_single_window():
"""
Check if EQLHarmonicBoost works with a single window that covers the whole region
"""
# Define a squared region
region = (-3e3, -1e3, 5e3, 7e3)
# Build synthetic point masses
points = vd.grid_coordinates(region=region,
shape=(6, 6),
extra_coords=-1e3)
masses = vd.datasets.CheckerBoard(amplitude=1e13,
region=region).predict(points)
# Define a set of observation points
coordinates = vd.grid_coordinates(region=region,
shape=(40, 40),
extra_coords=0)
# Get synthetic data
data = hm.point_mass_gravity(coordinates, points, masses, field="g_z")
# The interpolation should be perfect on the data points
eql = EQLHarmonicBoost(window_size=region[1] - region[0])
eql.fit(coordinates, data)
npt.assert_allclose(data, eql.predict(coordinates), rtol=1e-5)
# Gridding onto a denser grid should be reasonably accurate when compared
# to synthetic values
grid = vd.grid_coordinates(region=region, shape=(60, 60), extra_coords=0)
true = hm.point_mass_gravity(grid, points, masses, field="g_z")
npt.assert_allclose(true, eql.predict(grid), rtol=1e-3)
def test_eql_boost_random_state():
"""
Check if EQLHarmonicBoost produces same result by setting random_state
"""
region = (-3e3, -1e3, 5e3, 7e3)
# Build synthetic point masses
points = vd.grid_coordinates(region=region,
shape=(6, 6),
extra_coords=-1e3)
masses = vd.datasets.CheckerBoard(amplitude=1e13,
region=region).predict(points)
# Define a set of observation points
coordinates = vd.grid_coordinates(region=region,
shape=(20, 20),
extra_coords=0)
# Get synthetic data
data = hm.point_mass_gravity(coordinates, points, masses, field="g_z")
# Initialize two EQLHarmonicBoost with the same random_state
eql_a = EQLHarmonicBoost(window_size=500, random_state=0)
eql_a.fit(coordinates, data)
eql_b = EQLHarmonicBoost(window_size=500, random_state=0)
eql_b.fit(coordinates, data)
# Check if fitted coefficients are the same
npt.assert_allclose(eql_a.coefs_, eql_b.coefs_)
def test_eql_boost_warm_start():
"""
Check if EQLHarmonicBoost can be fitted with warm_start
"""
region = (-3e3, -1e3, 5e3, 7e3)
# Build synthetic point masses
points = vd.grid_coordinates(region=region,
shape=(6, 6),
extra_coords=-1e3)
masses = vd.datasets.CheckerBoard(amplitude=1e13,
region=region).predict(points)
# Define a set of observation points
coordinates = vd.grid_coordinates(region=region,
shape=(20, 20),
extra_coords=0)
# Get synthetic data
data = hm.point_mass_gravity(coordinates, points, masses, field="g_z")
# Check if refitting with warm_start=True changes its coefficients
eql = EQLHarmonicBoost(window_size=500, warm_start=True)
eql.fit(coordinates, data)
coefs = eql.coefs_.copy()
eql.fit(coordinates, data)
assert not np.allclose(coefs, eql.coefs_)
# Check if refitting with warm_start=False doesn't change its coefficients
# (need to set random_state, otherwise coefficients might be different due
# to another random shuffling of the windows).
eql = EQLHarmonicBoost(window_size=500, warm_start=False, random_state=0)
eql.fit(coordinates, data)
coefs = eql.coefs_.copy()
eql.fit(coordinates, data)
npt.assert_allclose(coefs, eql.coefs_)
def test_eql_boost_warm_start_new_coords():
"""
Check if sources are not changed when warm_start is True
If the gridder is already fitted (and has warm_start=True), the location of
the sources must not be modified, even if the new fitting process is
carried out with another set of observation points.
"""
region = (-3e3, -1e3, 5e3, 7e3)
# Build synthetic point masses
points = vd.grid_coordinates(region=region,
shape=(6, 6),
extra_coords=-1e3)
masses = vd.datasets.CheckerBoard(amplitude=1e13,
region=region).predict(points)
# Define a set of observation points
coordinates = vd.grid_coordinates(region=region,
shape=(20, 20),
extra_coords=0)
# Get synthetic data
data = hm.point_mass_gravity(coordinates, points, masses, field="g_z")
# Capture the location of sources created by EQLHarmonicBoost on the first fit
eql = EQLHarmonicBoost(window_size=500, warm_start=True)
eql.fit(coordinates, data)
sources = tuple(c.copy() for c in eql.points_)
# Fit the gridder with a new set of observation data
coordinates_new = (coordinates[0] + 100, coordinates[1] - 100,
coordinates[2])
data_new = hm.point_mass_gravity(coordinates_new,
points,
masses,
field="g_z")
eql.fit(coordinates_new, data_new)
# Check if sources remain on the same location
for i in range(3):
npt.assert_allclose(sources[i], eql.points_[i])
def test_eql_boost_custom_points():
"""
Check EQLHarmonicBoost with custom points
Check if the iterative gridder works well with a custom set of sources.
By default, the gridder puts one source beneath each data point, therefore
the indices of data and sources windows are identical. When passing
a custom set of sources, these indices may differ.
"""
region = (-3e3, -1e3, 5e3, 7e3)
# Build synthetic point masses
points = vd.grid_coordinates(region=region,
shape=(6, 6),
extra_coords=-1e3)
masses = vd.datasets.CheckerBoard(amplitude=1e13,
region=region).predict(points)
# Define a set of observation points
coordinates = vd.grid_coordinates(region=region,
shape=(20, 20),
extra_coords=0)
# Get synthetic data
data = hm.point_mass_gravity(coordinates, points, masses, field="g_z")
# Define a set of block-averaged equivalent sources
sources = block_averaged_sources(coordinates,
100,
depth_type="relative_depth",
depth=500)
# Fit EQLHarmonicBoost with the block-averaged sources
eql = EQLHarmonicBoost(window_size=500, points=sources)
eql.fit(coordinates, data)
# Check if sources are located on the same points
for i in range(3):
npt.assert_allclose(sources[i], eql.points_[i])
# The interpolation should be sufficiently accurate on the data points
assert mean_squared_error(
data, eql.predict(coordinates)) < 1e-3 * vd.maxabs(data)
northing = [7e3, 13e3]
# The vertical coordinate is positive upward so negative numbers represent depth
upward = [-0.5e3, -1e3]
points = [easting, northing, upward]
# We're using "negative" masses to represent a "mass deficit" since we assume
# measurements are gravity disturbances, not actual gravity values.
masses = [3e11, -10e11]
# Define computation points on a grid at 500m above the ground
coordinates = vd.grid_coordinates(
region=[0, 20e3, 0, 20e3], shape=(100, 100), extra_coords=500
)
# Compute the downward component of the gravitational acceleration
gravity = hm.point_mass_gravity(
coordinates, points, masses, field="g_z", coordinate_system="cartesian"
)
print(gravity)
# Plot the results on a map
fig, ax = plt.subplots(figsize=(8, 6))
ax.set_aspect("equal")
# Get the maximum absolute value so we can center the colorbar on zero
maxabs = vd.maxabs(gravity)
img = ax.contourf(
*coordinates[:2], gravity, 60, vmin=-maxabs, vmax=maxabs, cmap="seismic"
)
plt.colorbar(img, ax=ax, pad=0.04, shrink=0.73, label="mGal")
# Plot the point mass locations
ax.plot(easting, northing, "oy")
ax.set_title("Gravitational acceleration (downward)")