Reputation: 1337

How to perform bilinear interpolation in Python

I would like to perform blinear interpolation using python.
Example gps point for which I want to interpolate height is:

B = 54.4786674627
L = 17.0470721369

using four adjacent points with known coordinates and height values:

n = [(54.5, 17.041667, 31.993), (54.5, 17.083333, 31.911), (54.458333, 17.041667, 31.945), (54.458333, 17.083333, 31.866)]

and here's my primitive attempt:

import math
z00 = n[0][2]
z01 = n[1][2]
z10 = n[2][2]
z11 = n[3][2]
c = 0.016667 #grid spacing
x0 = 56 #latitude of origin of grid
y0 = 13 #longitude of origin of grid
i = math.floor((L-y0)/c)
j = math.floor((B-x0)/c)
t = (B - x0)/c - j
z0 = (1-t)*z00 + t*z10
z1 = (1-t)*z01 + t*z11
s = (L-y0)/c - i
z = (1-s)*z0 + s*z1

where z0 and z1

z01  z0  z11

     z
z00  z1   z10

I get 31.964 but from other software I get 31.961.
Is my script correct?
Can You provide another approach?

2022 Edit:
I would like to thank everyone who, even more than a decade after publication of this question, gives new answers to it.

Upvotes: 33

Answers (9)

BBSysDyn

Reputation: 4601

A solution here shows bilinear interpolation, I applied his method here. But my adaption of this method, simply computing a proximity to all four corners from Euclidian distance and using that as simple weighted average works much better (my adaption is in the same link).

import numpy as np

def func(x, y):
    s1 = 0.2; x1 = 36.5; y1 = 32.5
    s2 = 0.4; x2 = 36.1; y2 = 32.8
    g1 = np.exp( -4 *np.log(2) * ((x-x1)**2+(y-y1)**2) / s1**2)
    g2 = np.exp( -2 *np.log(2) * ((x-x2)**2+(y-y2)**2) / s2**2)    
    return g1 + g2 

D = 20
x = np.linspace(36,37,D)
y = np.linspace(32,33,D)
xx,yy = np.meshgrid(x,y)
zz = func(xx,yy)

def find_corners(xi,yi):
    idx1 = np.searchsorted(x, xi, side="left")
    idx2 = np.searchsorted(y, yi, side="left")
    cs = [(idx2,idx1),(idx2-1,idx1),(idx2,idx1-1),(idx2-1,idx1-1)]
    return cs

def cdist(p1,p2):    
    distances = np.linalg.norm(p1 - p2, axis=1)
    return distances

def cell_interp(x, y, points):
    a = np.array([x,y]).reshape(-1,2)
    b = np.array(points)[:,:2]
    ds = cdist(a,b)
    ds = ds / np.sum(ds)
    ds = 1. - ds
    c = np.array(points)[:,2]
    iz = np.sum(c * ds) / np.sum(ds)
    return iz

def grid_interp(intx,inty):
    cs = find_corners(intx,inty)

    i,j = cs[0][0],cs[0][1]
    i,j = cs[1][0],cs[1][1]
    i,j = cs[2][0],cs[2][1]
    i,j = cs[3][0],cs[3][1]
    
    i0,j0 = cs[0][0],cs[0][1]
    i1,j1 = cs[1][0],cs[1][1]
    i2,j2 = cs[2][0],cs[2][1]
    i3,j3 = cs[3][0],cs[3][1]
    
    introw = [(xx[i0,j0],yy[i0,j0],zz[i0,j0]),
              (xx[i1,j1],yy[i1,j1],zz[i1,j1]),
              (xx[i2,j2],yy[i2,j2],zz[i2,j2]),
              (xx[i3,j3],yy[i3,j3],zz[i3,j3])]
    return cell_interp(intx,inty,introw)

x2 = np.linspace(36.0001,36.9999,D*2)
y2 = np.linspace(32.0001,32.9999,D*2)
xx2,yy2 = np.meshgrid(x2,y2)
zz2 = func(xx2,yy2)

grid_interp_vec = np.vectorize(grid_interp,otypes=[np.float])
zz2_grid = grid_interp_vec(xx2,yy2)
print (np.mean(np.square(zz2-zz2_grid)))

Upvotes: 0

Khalil Al Hooti

Reputation: 4516

This is the same solution as defined here but applied to some function and compared with interp2d available in Scipy. We use numba library to make the interpolation function even faster than Scipy implementation.

import numpy as np
from scipy.interpolate import interp2d
import matplotlib.pyplot as plt

from numba import jit, prange
@jit(nopython=True, fastmath=True, nogil=True, cache=True, parallel=True)
def bilinear_interpolation(x_in, y_in, f_in, x_out, y_out):
    f_out = np.zeros((y_out.size, x_out.size))
    
    for i in prange(f_out.shape[1]):
        idx = np.searchsorted(x_in, x_out[i])
        
        x1 = x_in[idx-1]
        x2 = x_in[idx]
        x = x_out[i]
        
        for j in prange(f_out.shape[0]):
            idy = np.searchsorted(y_in, y_out[j])
            y1 = y_in[idy-1]
            y2 = y_in[idy]
            y = y_out[j]

            
            f11 = f_in[idy-1, idx-1]
            f21 = f_in[idy-1, idx]
            f12 = f_in[idy, idx-1]
            f22 = f_in[idy, idx]
            

            
            f_out[j, i] = ((f11 * (x2 - x) * (y2 - y) +
                            f21 * (x - x1) * (y2 - y) +
                            f12 * (x2 - x) * (y - y1) +
                            f22 * (x - x1) * (y - y1)) /
                           ((x2 - x1) * (y2 - y1)))
    
    return f_out

We make it quite a large interpolation array to assess the performance of each method.

The sample function is,

$z(x, y)=\sin \left(\frac{\pi x}{2}\right) e^{y / 2}$

x = np.linspace(0, 4, 13)
y = np.array([0, 2, 3, 3.5, 3.75, 3.875, 3.9375, 4])
X, Y = np.meshgrid(x, y)
Z = np.sin(np.pi*X/2) * np.exp(Y/2)

x2 = np.linspace(0, 4, 1000)
y2 = np.linspace(0, 4, 1000)
Z2 = bilinear_interpolation(x, y, Z, x2, y2)

fun = interp2d(x, y, Z, kind='linear')
Z3 = fun(x2, y2)


fig, ax = plt.subplots(nrows=1, ncols=3, figsize=(10, 6))
ax[0].pcolormesh(X, Y, Z, shading='auto')
ax[0].set_title("Original function")
X2, Y2 = np.meshgrid(x2, y2)
ax[1].pcolormesh(X2, Y2, Z2, shading='auto')
ax[1].set_title("bilinear interpolation")

ax[2].pcolormesh(X2, Y2, Z3, shading='auto')
ax[2].set_title("Scipy bilinear function")

plt.show()

Performance Test

Python without numba library

bilinear_interpolation function, in this case, is the same as numba version except that we change prange with python normal range in the for loop, and remove function decorator jit

%timeit bilinear_interpolation(x, y, Z, x2, y2)

Gives 7.15 s ± 107 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

Python with numba numba

%timeit bilinear_interpolation(x, y, Z, x2, y2)

Gives 2.65 ms ± 70.5 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

Scipy implementation

%%timeit
f = interp2d(x, y, Z, kind='linear')
Z2 = f(x2, y2)

Gives 6.63 ms ± 145 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

Performance tests are performed on 'Intel(R) Core(TM) i7-8700K CPU @ 3.70GHz'

Upvotes: 2

Shubham Shaswat

Reputation: 1310

A numpy implementation of based on this formula:

def bilinear_interpolation(x,y,x_,y_,val):

    a = 1 /((x_[1] - x_[0]) * (y_[1] - y_[0]))
    xx = np.array([[x_[1]-x],[x-x_[0]]],dtype='float32')
    f = np.array(val).reshape(2,2)
    yy = np.array([[y_[1]-y],[y-y_[0]]],dtype='float32')
    b = np.matmul(f,yy)

    return a * np.matmul(xx.T, b)

Input: Here,x_ is list of [x0,x1] and y_ is list of [y0,y1]

bilinear_interpolation(x=54.4786674627,
                       y=17.0470721369,
                       x_=[54.458333,54.5],
                       y_=[17.041667,17.083333],
                       val=[31.993,31.911,31.945,31.866])

Output:

array([[31.95912739]])

Upvotes: 4

Valeriu Radu

Reputation: 1

I suggest the following solution:

def bilinear_interpolation(x, y, z01, z11, z00, z10):

    def linear_interpolation(x, z0, z1):
        return z0 * x + z1 * (1 - x)

    return linear_interpolation(y, linear_interpolation(x, z01, z11),
                                   linear_interpolation(x, z00, z10))

Upvotes: 0

pmav99

Reputation: 2057

Inspired from here, I came up with the following snippet. The API is optimized for reusing a lot of times the same table:

from bisect import bisect_left

class BilinearInterpolation(object):
    """ Bilinear interpolation. """
    def __init__(self, x_index, y_index, values):
        self.x_index = x_index
        self.y_index = y_index
        self.values = values

    def __call__(self, x, y):
        # local lookups
        x_index, y_index, values = self.x_index, self.y_index, self.values

        i = bisect_left(x_index, x) - 1
        j = bisect_left(y_index, y) - 1

        x1, x2 = x_index[i:i + 2]
        y1, y2 = y_index[j:j + 2]
        z11, z12 = values[j][i:i + 2]
        z21, z22 = values[j + 1][i:i + 2]

        return (z11 * (x2 - x) * (y2 - y) +
                z21 * (x - x1) * (y2 - y) +
                z12 * (x2 - x) * (y - y1) +
                z22 * (x - x1) * (y - y1)) / ((x2 - x1) * (y2 - y1))

You can use it like this:

table = BilinearInterpolation(
    x_index=(54.458333, 54.5), 
    y_index=(17.041667, 17.083333), 
    values=((31.945, 31.866), (31.993, 31.911))
)

print(table(54.4786674627, 17.0470721369))
# 31.957986883136307

This version has no error checking and you will run into trouble if you try to use it at the boundaries of the indexes (or beyond). For the full version of the code, including error checking and optional extrapolation, look here.

Upvotes: 7

dgorissen

Reputation: 6305

You can also refer to the interp function in matplotlib.

Upvotes: 3

Raymond Hettinger

Reputation: 226416

Here's a reusable function you can use. It includes doctests and data validation:

def bilinear_interpolation(x, y, points):
    '''Interpolate (x,y) from values associated with four points.

    The four points are a list of four triplets:  (x, y, value).
    The four points can be in any order.  They should form a rectangle.

        >>> bilinear_interpolation(12, 5.5,
        ...                        [(10, 4, 100),
        ...                         (20, 4, 200),
        ...                         (10, 6, 150),
        ...                         (20, 6, 300)])
        165.0

    '''
    # See formula at:  http://en.wikipedia.org/wiki/Bilinear_interpolation

    points = sorted(points)               # order points by x, then by y
    (x1, y1, q11), (_x1, y2, q12), (x2, _y1, q21), (_x2, _y2, q22) = points

    if x1 != _x1 or x2 != _x2 or y1 != _y1 or y2 != _y2:
        raise ValueError('points do not form a rectangle')
    if not x1 <= x <= x2 or not y1 <= y <= y2:
        raise ValueError('(x, y) not within the rectangle')

    return (q11 * (x2 - x) * (y2 - y) +
            q21 * (x - x1) * (y2 - y) +
            q12 * (x2 - x) * (y - y1) +
            q22 * (x - x1) * (y - y1)
           ) / ((x2 - x1) * (y2 - y1) + 0.0)

You can run test code by adding:

if __name__ == '__main__':
    import doctest
    doctest.testmod()

Running the interpolation on your dataset produces:

>>> n = [(54.5, 17.041667, 31.993),
         (54.5, 17.083333, 31.911),
         (54.458333, 17.041667, 31.945),
         (54.458333, 17.083333, 31.866),
    ]
>>> bilinear_interpolation(54.4786674627, 17.0470721369, n)
31.95798688313631

Upvotes: 57

jterrace

Reputation: 67083

Not sure if this helps much, but I get a different value when doing linear interpolation using scipy:

>>> import numpy as np
>>> from scipy.interpolate import griddata
>>> n = np.array([(54.5, 17.041667, 31.993),
                  (54.5, 17.083333, 31.911),
                  (54.458333, 17.041667, 31.945),
                  (54.458333, 17.083333, 31.866)])
>>> griddata(n[:,0:2], n[:,2], [(54.4786674627, 17.0470721369)], method='linear')
array([ 31.95817681])

Upvotes: 10

machine yearning

Reputation: 10119

I think the point of doing a floor function is that usually you're looking to interpolate a value whose coordinate lies between two discrete coordinates. However you seem to have the actual real coordinate values of the closest points already, which makes it simple math.

z00 = n[0][2]
z01 = n[1][2]
z10 = n[2][2]
z11 = n[3][2]

# Let's assume L is your x-coordinate and B is the Y-coordinate

dx = n[2][0] - n[0][0] # The x-gap between your sample points
dy = n[1][1] - n[0][1] # The Y-gap between your sample points

dx1 = (L - n[0][0]) / dx # How close is your point to the left?
dx2 = 1 - dx1              # How close is your point to the right?
dy1 = (B - n[0][1]) / dy # How close is your point to the bottom?
dy2 = 1 - dy1              # How close is your point to the top?

left = (z00 * dy1) + (z01 * dy2)   # First interpolate along the y-axis
right = (z10 * dy1) + (z11 * dy2)

z = (left * dx1) + (right * dx2)   # Then along the x-axis

There might be a bit of erroneous logic in translating from your example, but the gist of it is you can weight each point based on how much closer it is to the interpolation goal point than its other neighbors.