Nuth & Kaab coregistration spreads NaNs

[adehecq]

The Nuth & Kaab coregistration (and I assume other types of coregistration as well) spread NaNs because of the resampling.
An easy way to see it is to run the Nuth & Kaab example and to add some random gaps in the DEM. See attached script (rrnamed to txt so that GitHub does not complain) and figure below.

plot_nuth_kaab.txt

[adehecq]

A simple solution is to replace the whole _apply_matrix method by something like this:

dem_to_be_aligned_copy = dem_to_be_aligned.copy()
dem_to_be_aligned_copy.shift(-nuth_kaab._meta['offset_east_px']*nuth_kaab._meta['resolution'], -nuth_kaab._meta['offset_north_px']*nuth_kaab._meta['resolution'])
aligned_dem = dem_to_be_aligned_copy.reproject(reference_dem, resampling='cubic')

Note that one can check that the coregistration works well by running Nuth & Kaab a second time and checking the residual translation. The result is best if resampling with cubic. This is probably because in Nuth & Kaab fit function, the raster is resampled using RectBivariateSpline with degree 1.

[adehecq]

Actually, cubic interpolation spreads the NaNs even more, so bilinear is preferable, see below.

With bilinear interpolation

With cubic interpolation

See attached script for tests.
plot_nuth_kaab.txt

[erikmannerfelt]

This is most likely due to the dilate_mask argument, which we deemed was a preferable behavior over potentially introducing weird artefacts. I'm not saying this isn't an issue, just that the effect is an intentional "better of two evils"!

#87 (comment)

[adehecq]

The NaN spread is even worse with two coregistrations in a row, like coreg = xdem.coreg.ICP() + xdem.coreg.NuthKaab() (see attached script and figure).
Turning off the 'dilate_mask' option reduces the gaps, but it is not perfect yet, because currently the apply method is run twice with two items in the CoregPipeline.

test_coreg.txt