Fill in missing values with nearest neighbour in Python numpy masked arrays?

Question

I am working with a 2D Numpy masked_array in Python. I need to change the data values in the masked area such that they equal the nearest unmasked value.

NB. If there are more than one nearest unmasked values then it can take any of those nearest values (which ever one turns out to be easiest to code…)

e.g.

import numpy
import numpy.ma as ma

a = numpy.arange(100).reshape(10,10)
fill_value=-99
a[2:4,3:8] = fill_value
a[8,8] = fill_value
a = ma.masked_array(a,a==fill_value)

>>> a  [[0 1 2 3 4 5 6 7 8 9]
  [10 11 12 13 14 15 16 17 18 19]
  [20 21 22 -- -- -- -- -- 28 29]
  [30 31 32 -- -- -- -- -- 38 39]
  [40 41 42 43 44 45 46 47 48 49]
  [50 51 52 53 54 55 56 57 58 59]
  [60 61 62 63 64 65 66 67 68 69]
  [70 71 72 73 74 75 76 77 78 79]
  [80 81 82 83 84 85 86 87 -- 89]
  [90 91 92 93 94 95 96 97 98 99]],

• I need it to look like this:

>>> a.data
 [[0 1 2 3 4 5 6 7 8 9]
 [10 11 12 13 14 15 16 17 18 19]
 [20 21 22 ? 14 15 16 ? 28 29]
 [30 31 32 ? 44 45 46 ? 38 39]
 [40 41 42 43 44 45 46 47 48 49]
 [50 51 52 53 54 55 56 57 58 59]
 [60 61 62 63 64 65 66 67 68 69]
 [70 71 72 73 74 75 76 77 78 79]
 [80 81 82 83 84 85 86 87 ? 89]
 [90 91 92 93 94 95 96 97 98 99]],

NB. where "?" could take any of the adjacent unmasked values.

What is the most efficient way to do this?

Thanks for your help.

Answer 1

You could use np.roll to make shifted copies of a, then use boolean logic on the masks to identify the spots to be filled in:

import numpy as np
import numpy.ma as ma

a = np.arange(100).reshape(10,10)
fill_value=-99
a[2:4,3:8] = fill_value
a[8,8] = fill_value
a = ma.masked_array(a,a==fill_value)
print(a)

# [[0 1 2 3 4 5 6 7 8 9]
#  [10 11 12 13 14 15 16 17 18 19]
#  [20 21 22 -- -- -- -- -- 28 29]
#  [30 31 32 -- -- -- -- -- 38 39]
#  [40 41 42 43 44 45 46 47 48 49]
#  [50 51 52 53 54 55 56 57 58 59]
#  [60 61 62 63 64 65 66 67 68 69]
#  [70 71 72 73 74 75 76 77 78 79]
#  [80 81 82 83 84 85 86 87 -- 89]
#  [90 91 92 93 94 95 96 97 98 99]]

for shift in (-1,1):
    for axis in (0,1):        
        a_shifted=np.roll(a,shift=shift,axis=axis)
        idx=~a_shifted.mask * a.mask
        a[idx]=a_shifted[idx]

print(a)

# [[0 1 2 3 4 5 6 7 8 9]
#  [10 11 12 13 14 15 16 17 18 19]
#  [20 21 22 13 14 15 16 28 28 29]
#  [30 31 32 43 44 45 46 47 38 39]
#  [40 41 42 43 44 45 46 47 48 49]
#  [50 51 52 53 54 55 56 57 58 59]
#  [60 61 62 63 64 65 66 67 68 69]
#  [70 71 72 73 74 75 76 77 78 79]
#  [80 81 82 83 84 85 86 87 98 89]
#  [90 91 92 93 94 95 96 97 98 99]]

---

If you'd like to use a larger set of nearest neighbors, you could perhaps do something like this:

neighbors=((0,1),(0,-1),(1,0),(-1,0),(1,1),(-1,1),(1,-1),(-1,-1),
           (0,2),(0,-2),(2,0),(-2,0))

Note that the order of the elements in neighbors is important. You probably want to fill in missing values with the nearest neighbor, not just any neighbor. There's probably a smarter way to generate the neighbors sequence, but I'm not seeing it at the moment.

a_copy=a.copy()
for hor_shift,vert_shift in neighbors:
    if not np.any(a.mask): break
    a_shifted=np.roll(a_copy,shift=hor_shift,axis=1)
    a_shifted=np.roll(a_shifted,shift=vert_shift,axis=0)
    idx=~a_shifted.mask*a.mask
    a[idx]=a_shifted[idx]

Note that np.roll happily rolls the lower edge to the top, so a missing value at the top may be filled in by a value from the very bottom. If this is a problem, I'd have to think more about how to fix it. The obvious but not very clever solution would be to use if statements and feed the edges a different sequence of admissible neighbors...