Error when using PrismMesh.carvetopo because of matplotlib.mlab.griddata

[hbueno]

The topography/relief defined to mask (remove) prisms from the mesh need to be interpolated. The mesher.carvetopo use by default the Natural Neighbor interpolation in matplotlib.mlab.griddata. However, when I try to run, for example, the cookbook mesher_prismmesh_topo.py, a RuntimeError suggests to intall natgrid or change the interpolation method to linear.

Traceback (most recent call last):
  File "mesher_prismmesh_topo.py", line 18, in <module>
    mesh.carvetopo(x, y, height)
  File "/Users/hbueno/src/fatiando/fatiando/mesher.py", line 1205, in carvetopo
   topo = matplotlib.mlab.griddata(x, y, height, XC, YC).ravel()
  File "/Users/hbueno/anaconda/lib/python2.7/site-packages/matplotlib/mlab.py", line 3443, in griddata
   raise RuntimeError("To use interp='nn' (Natural Neighbor "
RuntimeError: To use interp='nn' (Natural Neighbor interpolation) in griddata, natgrid must be installed.  Either install it from http://sourceforge.net/projects/matplotlib/files/matplotlib-toolkits, or use interp='linear' instead.

I've installed Fatiando with conda as well as all the dependencies and passed in all tests.

[hbueno]

I was using matplotlib=1.4. A possible solution to fix the RuntimeError it was to downgrade to matplotlib=1.3.1 (and consequently to numpy-1.8.2).

[leouieda]

Copying text by @hbueno from #147 here:

@leouieda I'm opening this issue to "discuss" about some dependencies of Fatiando. To be sincere I don't know if it's the case to open another issue or not, but I'm trying to be more organized (and you always can close it!).

As I pointed in this issue #146, I was using matplotlib=1.4 and its dependencies (all installed with conda) without problems until I needed to use the mesher.carvetopo.

Two "possible" solutions I've found:

1. to downgrade to matplotlib=1.3.1 which requires numpy-1.8.2.
2. to install natgrid which requires a separate installation of the natgrid toolkit from the sourceforge download page.

I chose (1) to be easily done with conda and because I don't know another alternative. Anyway, let's suppose this is right, I think we should change the list of Fatiando dependencies. For example, where is matplotlib >= 1.3 change to matplotlib = 1.3 (removing >), or maybe put some temporary warning for who intend to use the matplotlib >= 1.3. Anyway, I'm writing because I've failed when I tried to install natgrid with conda (I now I know why) and as I know, for Fatiando it's interesting that all the dependencies be quickly installed with conda. But please, let me know if you think I'm talking nonsense or doing something wrong ;)

[leouieda]

The problem seems to be that matplotlib no longer ships natgrid with it. I don't know why this only happens in MacOS (I'm using on Ubuntu and it works fine).

Since the problem is the matplotlib.mlab.griddata functions, we can just replace it by scipy.interpolate.griddata. It has no Nearest Neighbors but that is not a very good interpolation for geophysics anyways. That would actually make the code in fatiando.gridder.interp cleaner. It already uses scipy instead of mlab for all other methods and only calls mlab if `algorithm='nn' (which is not the default anymore).

Fixing would be a simple case of replacing all calls to matplotlib.mlab.griddata with the appropriate call to scipy.interpolate.griddata.

@hbueno, do you want to try to fix this?

[hbueno]

Sure, I can try.

[leouieda]

#148 fixed this.