"""**Module for 2D interpolation over a rectangular mesh**
This module:
* provides piecewise constant (nearest neighbour) and bilinear interpolation
* is fast (based on numpy vector operations)
* depends only on numpy
* guarantees that interpolated values never exceed the four nearest neighbours
* handles missing values in domain sensibly using NaN
* is unit tested with a range of common and corner cases
See end of this file for documentation of the mathematical derivation used.
"""
__author__ = 'Ole Nielsen <ole.moller.nielsen@gmail.com>'
__revision__ = '$Format:%H$'
__date__ = '01/11/2010'
__license__ = "GPL"
__copyright__ = 'Copyright 2012, Australia Indonesia Facility for '
__copyright__ += 'Disaster Reduction'
import logging
import numpy
from safe.common.exceptions import BoundsError, InaSAFEError
LOGGER = logging.getLogger('InaSAFE')
# pylint: disable=W0105
[docs]def interpolate2d(x, y, Z, points, mode='linear', bounds_error=False):
"""Fundamental 2D interpolation routine
Args:
* x: 1D array of x-coordinates of the mesh on which to interpolate
* y: 1D array of y-coordinates of the mesh on which to interpolate
* Z: 2D array of values for each x, y pair
* points: Nx2 array of coordinates where interpolated values are sought
* mode: Determines the interpolation order. Options are
* 'constant' - piecewise constant nearest neighbour interpolation
* 'linear' - bilinear interpolation using the four
nearest neighbours (default)
* bounds_error: Boolean flag. If True (default) a BoundsError exception
will be raised when interpolated values are requested
outside the domain of the input data. If False, nan
is returned for those values
Returns:
* 1D array with same length as points with interpolated values
Raises: Exception, BoundsError (see note about bounds_error)
Notes:
Input coordinates x and y are assumed to be monotonically increasing,
but need not be equidistantly spaced. No such assumption regarding
ordering of points is made.
Z is assumed to have dimension M x N, where M = len(x) and N = len(y).
In other words it is assumed that the x values follow the first
(vertical) axis downwards and y values the second (horizontal) axis
from left to right.
If this routine is to be used for interpolation of raster grids where
data is typically organised with longitudes (x) going from left to
right and latitudes (y) from left to right then user
interpolate_raster in this module
"""
# Input checks
x, y, Z, xi, eta = check_inputs(x, y, Z, points, mode, bounds_error)
# Identify elements that are outside interpolation domain or NaN
outside = (xi < x[0]) + (eta < y[0]) + (xi > x[-1]) + (eta > y[-1])
outside += numpy.isnan(xi) + numpy.isnan(eta)
inside = -outside
xi = xi[inside]
eta = eta[inside]
# Find upper neighbours for each interpolation point
idx = numpy.searchsorted(x, xi, side='left')
idy = numpy.searchsorted(y, eta, side='left')
# Internal check (index == 0 is OK)
msg = ('Interpolation point outside domain. This should never happen. '
'Please email Ole.Moller.Nielsen@gmail.com')
if len(idx) > 0:
if not max(idx) < len(x):
raise InaSAFEError(msg)
if len(idy) > 0:
if not max(idy) < len(y):
raise InaSAFEError(msg)
# Get the four neighbours for each interpolation point
x0 = x[idx - 1]
x1 = x[idx]
y0 = y[idy - 1]
y1 = y[idy]
z00 = Z[idx - 1, idy - 1]
z01 = Z[idx - 1, idy]
z10 = Z[idx, idy - 1]
z11 = Z[idx, idy]
# Coefficients for weighting between lower and upper bounds
oldset = numpy.seterr(invalid='ignore') # Suppress warnings
alpha = (xi - x0) / (x1 - x0)
beta = (eta - y0) / (y1 - y0)
numpy.seterr(**oldset) # Restore
if mode == 'linear':
# Bilinear interpolation formula
dx = z10 - z00
dy = z01 - z00
z = z00 + alpha * dx + beta * dy + alpha * beta * (z11 - dx - dy - z00)
else:
# Piecewise constant (as verified in input_check)
# Set up masks for the quadrants
left = alpha < 0.5
right = -left
lower = beta < 0.5
upper = -lower
lower_left = lower * left
lower_right = lower * right
upper_left = upper * left
# Initialise result array with all elements set to upper right
z = z11
# Then set the other quadrants
z[lower_left] = z00[lower_left]
z[lower_right] = z10[lower_right]
z[upper_left] = z01[upper_left]
# Self test
if len(z) > 0:
mz = numpy.nanmax(z)
mZ = numpy.nanmax(Z)
msg = ('Internal check failed. Max interpolated value %.15f '
'exceeds max grid value %.15f ' % (mz, mZ))
if not(numpy.isnan(mz) or numpy.isnan(mZ)):
if not mz <= mZ:
raise InaSAFEError(msg)
# Populate result with interpolated values for points inside domain
# and NaN for values outside
r = numpy.zeros(len(points))
r[inside] = z
r[outside] = numpy.nan
return r
[docs]def interpolate_raster(x, y, Z, points, mode='linear', bounds_error=False):
"""2D interpolation of raster data
It is assumed that data is organised in matrix Z as latitudes from
bottom up along the first dimension and longitudes from west to east
along the second dimension.
Further it is assumed that x is the vector of longitudes and y the
vector of latitudes.
See interpolate2d for details of the interpolation routine
"""
# Flip matrix Z up-down to interpret latitudes ordered from south to north
Z = numpy.flipud(Z)
# Transpose Z to have y coordinates along the first axis and x coordinates
# along the second axis
Z = Z.transpose()
# Call underlying interpolation routine and return
res = interpolate2d(x, y, Z, points, mode=mode, bounds_error=bounds_error)
return res
"""
Bilinear interpolation is based on the standard 1D linear interpolation
formula:
Given points (x0, y0) and (x1, x0) and a value of x where x0 <= x <= x1,
the linearly interpolated value y at x is given as
alpha*(y1-y0) + y0
or
alpha*y1 + (1-alpha)*y0 (1)
where alpha = (x-x0)/(x1-x0) (1a)
2D bilinear interpolation aims at obtaining an interpolated value z at a point
(x,y) which lies inside a square formed by points (x0, y0), (x1, y0),
(x0, y1) and (x1, y1) for which values z00, z10, z01 and z11 are known.
This obtained be first applying equation (1) twice in in the x-direction
to obtain interpolated points q0 and q1 for (x, y0) and (x, y1), respectively.
q0 = alpha*z10 + (1-alpha)*z00 (2)
and
q1 = alpha*z11 + (1-alpha)*z01 (3)
Then using equation (1) in the y-direction on the results from (2) and (3)
z = beta*q1 + (1-beta)*q0 (4)
where beta = (y-y0)/(y1-y0) (4a)
Substituting (2) and (3) into (4) yields
z = alpha*beta*z11 + beta*z01 - alpha*beta*z01 +
alpha*z10 + z00 - alpha*z00 - alpha*beta*z10 - beta*z00 +
alpha*beta*z00
= alpha*beta*(z11 - z01 - z10 + z00) +
alpha*(z10 - z00) + beta*(z01 - z00) + z00
which can be further simplified to
z = alpha*beta*(z11 - dx - dy - z00) + alpha*dx + beta*dy + z00 (5)
where
dx = z10 - z00
dy = z01 - z00
Equation (5) is what is implemented in the function interpolate2d above.
Piecewise constant interpolation can be implemented using the same coefficients
(1a) and (4a) that are used for bilinear interpolation as they are a measure of
the relative distance to the left and lower neigbours. A value of 0 will pick
the left or lower bound whereas a value of 1 will pick the right or higher
bound. Hence z can be assigned to its nearest neigbour as follows
| z00 alpha < 0.5 and beta < 0.5 # lower left corner
|
| z10 alpha >= 0.5 and beta < 0.5 # lower right corner
z = |
| z01 alpha < 0.5 and beta >= 0.5 # upper left corner
|
| z11 alpha >= 0.5 and beta >= 0.5 # upper right corner
"""
# pylint: enable=W0105