Question

I have a point dataset that I’m trying to interpolate on a grid. These points are aligned in grid fashion with some points missing see below:

To complicate it, it’s possible that other input pointsets may not align on a grid, so I’m trying to use scipy.interpolate.griddata to interpolate these values onto a regular grid. However, sometimes my underlying grid aligns perfectly with the sampling rate of the input point dataset and griddata hangs. According to this question scipy.interpolate.griddata operates poorly if the sample occurs across 3 points that are perfectly aligned which sometimes happens for me.

Is there a high performance alternative for interpolation of point data onto a regular grid in Python?

Answer 1

My solution has been to stochastically perturb the x and y coordinates of each input point by a random factor between 1e-6 * grid_cell_size. I choose 1e-6 as a rule of thumb of several orders of magnitude greater than the minimum delta difference in a 32 bit float, yet small enough that the error introduced by it is still shadowed by the error introduced by the interpolation scheme.

I’m still open to other ideas, but this way I can use the stock interpolation functions provided by scipy by introducing a very small numerical error.