scipy ND Interpolating over NaNs

Question

I have been trouble working out how to use the scipy.interpolate functions (either LinearNDInterpolator, griddata or Preferably NearestNDInterpolator)

There are some tutorials online but i am confused what form my data needs to be in. The online documentation for nearestND is terrible.

The function asks for:

    x : (Npoints, Ndims) ndarray of floats
         Data point coordinates.
    y : (Npoints,) ndarray of float or complex
         Data point values.

I have data in the form: lat,long,data,time held within an xarray dataset. There are some gaps in the data I would like to fill in.

I don't understand how to tell the function my x points. i have tried (lat,long) as a tuple and np.meshgrid(lat,long) but can't seem to get it going.

Any help on how i can pass my lat,long coordinates into the function? Bonus points for time coordinates as well to make the estimates more robust through the third dimension.

Thanks!

Answer 1

i have tried (lat,long) as a tuple

If lat and long are 1D arrays or lists, try this:

points = np.array((lat, long)).T # make a 2D array of shape Npoints x 2
nd = NearestNDInterpolator(points, data)

The you can compute interpolated values as nd(lat1, long1) , etc.

Answer 2

Scipy provides multivariate interpolation methods for both unstructured data and data point regularly placed on a grid. Unstructured data means the data could be provided as a list of non-ordered points. It seems that your data is structured: it is an array of size (480, 2040). However, the NearestNDInterpolator works on unstructured data. The flatten method can be used to transform the array to a list (1d) of value (of length 480*2040). The same have to be done for the coordinates. meshgrid is used to have the coordinates for every points of the grid, and again flatten is used to obtain a "list" of 2d coordinates (an array of shape 480*2040 x 2).

Here is an example which go from structured data to unstructured:

import numpy as np

lat = np.linspace(2, 6, 10)
lon = np.linspace(5, 9, 14)

latM, lonM = np.meshgrid(lat, lon)  # M is for Matrix

dataM = np.sin(latM)*np.cos(lonM)  # example of data, Matrix form

from scipy.interpolate import NearestNDInterpolator

points = np.array((latM.flatten(), lonM.flatten())).T
print( points.shape )
# >>> (140, 2)

f_nearest = NearestNDInterpolator(points, dataM.flatten())
f_nearest(5, 5)

Working with NaNs should not be a big problem in this case, because it is just a missing point in the list, except that the coordinates of the missing points have to be removed from the list too.