Extrapolation when gridding data and better interpolators

[leouieda]

A problem we had with grid interpolation was that the matplotlib.mlab.griddata function that was used didn't extrapolate the data outside of the convex hull of the data points. This is good because extrapolated data is usually meaningless and misleading. But this is bad because it looks bad on plots and sometimes you want extrapolated data.

Also, the nearest neighbors algorithm used by matplotlib.mlab.griddata is not very good. A better option is scipy.interpolate.griddata which has a cubic interpolation.

So I've added support for scipy griddata by default. You can still use the nearest neighbors of matplotlib by setting algorithm='nn'.

There is also an extrapolate option (True by default) that extrapolates the data using the nearest method of scipy griddata (takes the value of the closest data point).

Update: Added 2 new functions, interp_at to interpolate on given points (e.g., for extracting profiles) and extrapolate_nans to replace NaNs or masked values using the nearest data value.

[leouieda]

@hbueno You were having trouble upward continuing the loaded data when there were NaNs or masked values. Now you can remove the NaNs from the data and interpolate the missing data using gridder.interp. Then you can use any transform on the data that requires gridded data.

[leouieda]

@hbueno Even simpler now. You can use gridder.extrapolate_nans to fill the masked or nan values of an array. Don't have to run through the interpolation.