Bilinear interpolation of point data on a raster in Python?

Question

I have a raster that I'd like to do some point interpolations with. Here is where I'm at:

from osgeo import gdal
from numpy import array

# Read raster
source = gdal.Open('my_raster.tif')
nx, ny = source.RasterXSize, source.RasterYSize
gt = source.GetGeoTransform()
band_array = source.GetRasterBand(1).ReadAsArray()
# Close raster
source = None

# Compute mid-point grid spacings
ax = array([gt[0] + ix*gt[1] + gt[1]/2.0 for ix in range(nx)])
ay = array([gt[3] + iy*gt[5] + gt[5]/2.0 for iy in range(ny)])

Up to now, I've tried SciPy's interp2d function:

from scipy import interpolate
bilinterp = interpolate.interp2d(ax, ay, band_array, kind='linear')

however I get a memory error on my 32-bit Windows system with a 317×301 raster:

Traceback (most recent call last):
  File "<interactive input>", line 1, in <module>
  File "C:\Python25\Lib\site-packages\scipy\interpolate\interpolate.py", line 125, in __init__
    self.tck = fitpack.bisplrep(self.x, self.y, self.z, kx=kx, ky=ky, s=0.)
  File "C:\Python25\Lib\site-packages\scipy\interpolate\fitpack.py", line 873, in bisplrep
tx,ty,nxest,nyest,wrk,lwrk1,lwrk2)
MemoryError

I'll admit, I have limited confidence in this SciPy function, as the bounds_error or fill_value parameters don't work as documented. I don't see why I should have a memory error, since my raster is 317×301, and the bilinear algorithm should not be difficult.

Has anyone come across a good bilinear interpolation algorithm, preferably in Python, possibly tailored with NumPy? Any hints or advice?

---

(Note: the nearest neighbor interpolation algorithm is easy cake:

from numpy import argmin, NAN

def nearest_neighbor(px, py, no_data=NAN):
    '''Nearest Neighbor point at (px, py) on band_array
    example: nearest_neighbor(2790501.920, 6338905.159)'''
    ix = int(round((px - (gt[0] + gt[1]/2.0))/gt[1]))
    iy = int(round((py - (gt[3] + gt[5]/2.0))/gt[5]))
    if (ix < 0) or (iy < 0) or (ix > nx - 1) or (iy > ny - 1):
        return no_data
    else:
        return band_array[iy, ix]

... but I much prefer bilinear interpolation methods)

Answer 1

I've translated the formula below (from Wikipedia) into Python-speak to yield the following algorithm, which appears to work.

from numpy import floor, NAN

def bilinear(px, py, no_data=NAN):
    '''Bilinear interpolated point at (px, py) on band_array
    example: bilinear(2790501.920, 6338905.159)'''
    ny, nx = band_array.shape
    # Half raster cell widths
    hx = gt[1]/2.0
    hy = gt[5]/2.0
    # Calculate raster lower bound indices from point
    fx = (px - (gt[0] + hx))/gt[1]
    fy = (py - (gt[3] + hy))/gt[5]
    ix1 = int(floor(fx))
    iy1 = int(floor(fy))
    # Special case where point is on upper bounds
    if fx == float(nx - 1):
        ix1 -= 1
    if fy == float(ny - 1):
        iy1 -= 1
    # Upper bound indices on raster
    ix2 = ix1 + 1
    iy2 = iy1 + 1
    # Test array bounds to ensure point is within raster midpoints
    if (ix1 < 0) or (iy1 < 0) or (ix2 > nx - 1) or (iy2 > ny - 1):
        return no_data
    # Calculate differences from point to bounding raster midpoints
    dx1 = px - (gt[0] + ix1*gt[1] + hx)
    dy1 = py - (gt[3] + iy1*gt[5] + hy)
    dx2 = (gt[0] + ix2*gt[1] + hx) - px
    dy2 = (gt[3] + iy2*gt[5] + hy) - py
    # Use the differences to weigh the four raster values
    div = gt[1]*gt[5]
    return (band_array[iy1,ix1]*dx2*dy2/div +
            band_array[iy1,ix2]*dx1*dy2/div +
            band_array[iy2,ix1]*dx2*dy1/div +
            band_array[iy2,ix2]*dx1*dy1/div)

Note that the result will be returned with an apparent higher precision than the source data, since it is up-classed to NumPy's dtype('float64') data type. You can use the return value with .astype(band_array.dtype) to make the output data type the same as the input array.

Answer 2

I tried it locally to similar results, but I'm on a 64-bit platform so it didn't hit the memory limit. Perhaps instead try interpolating small pieces of the array at a time, like in this example.

You could also just do this with GDAL, from the command line it'd be:

gdalwarp -ts $XSIZE*2 0 -r bilinear input.tif interp.tif

To do the equivalent operation in Python, use ReprojectImage():

mem_drv = gdal.GetDriverByName('MEM')
dest = mem_drv.Create('', nx, ny, 1)

resample_by = 2
dt = (gt[0], gt[1] * resample_by, gt[2], gt[3], gt[4], gt[5] * resample_by)
dest.setGeoTransform(dt)

resampling_method = gdal.GRA_Bilinear    
res = gdal.ReprojectImage(source, dest, None, None, resampling_method)

# then, write the result to a file of your choice...    

Answer 3

I have had the exact issue in the past, and never resolved it using interpolate.interp2d. I have had success using scipy.ndimage.map_coordinates. Try the following:

scipy.ndimage.map_coordinates(band_array, [ax,ay]], order=1)

This seems to give the same output as bilinear.

Answer 4

scipy.interpolate.interp2d() works fine with more modern scipy. I think older versions assume irregular grids and don't take advantage of the regular grids. I get the same error as you do with scipy.version = 0.11.0, but on scipy.version = 0.14.0, it happily works on some 1600x1600 model output.

Thank you for the hints in your question.

#!/usr/bin/env python

from osgeo import gdal
from numpy import array
import argparse

parser = argparse.ArgumentParser()
parser.add_argument("filename",help='raster file from which to interpolate a (1/3,1/3) point from from')
args = parser.parse_args()

# Read raster
source = gdal.Open(args.filename)
nx, ny = source.RasterXSize, source.RasterYSize
gt = source.GetGeoTransform()
band_array = source.GetRasterBand(1).ReadAsArray()
# Close raster
source = None

# Compute mid-point grid spacings
ax = array([gt[0] + ix*gt[1] + gt[1]/2.0 for ix in range(nx)])
ay = array([gt[3] + iy*gt[5] + gt[5]/2.0 for iy in range(ny)])

from scipy import interpolate
bilinterp = interpolate.interp2d(ax, ay, band_array, kind='linear')

x1 = gt[0] + gt[1]*nx/3
y1 = gt[3] + gt[5]*ny/3.

print(nx, ny, x1,y1,bilinterp(x1,y1))

####################################

$ time ./interp2dTesting.py test.tif 
(1600, 1600, -76.322, 30.70889, array([-8609.27777778]))

real    0m4.086s
user    0m0.590s
sys 0m0.252s