Misc Issues PR

[herbiebradley]

This PR contains several miscellaneous features:

• Minor linting and type hinting improvements, including making type hints of class labels consistently Union[int, str].
• Pin Rasterio to 1.1.8
• New _add_nodes function in geograph.
• Fix the merge_classes function by detecting neighbouring polygons of the same class and merging them. Closes Merge polygons in the GeoGraph merge_classes function #44
• Support for barrier_classes in add_habitat, with basic pathfinding functionality. This will only detect pathfinding problems if the polygon of the barrier node completely cuts the buffered polygon of the main node in half, preventing all access to the neighbouring polygon. This works by using Shapely's polygon difference function, see the comments for the actual algorithm.

The pathfinding code slows add_habitat down quite a bit, as might be expected, but it seems to make some difference to the final edge count:

When this one is merged I will create the 0.0.2 release, upload it to PyPi and add it to Zenodo.

[Croydon-Brixton]

Looks already very good to me, thank you for the effort to add a first iteration of the path finding algorithm and getting the ball rolling!
I added a couple of questions regarding edge cases. You documented the code amazingly well, it was a pleasure to read as always @herbiebradley ! 🎉

[Croydon-Brixton]

Very nice description! This is incredibly helpful, thank you!

With regards to edge cases, how would the following case be handled:

Here if we consider classes 7 and 13 as barriers separately, they do not cut off the access between 2 and 4. If they are considered as a joint barrier however, they will cut off the access. Do I understand it correctly that in this algorithm the barrier polygons would be processed sequentially (i.e. first 2, then 4, ...) and therefore the edge would be added?
Or am I overlooking sth?

If the former is the case, we could go around this by performing a difference between the buffered original polygon and all barrier polygons that intersect with the buffered original polygon first and then checking, as you did, whether the piece that contains the original node will still intersect with the neighbour (or in fact any of its neighbours, this might give us a quite a speedup?)

[herbiebradley]

Good idea, I implemented this in the latest commit. I got around a 60% speedup, so it's looking good!

[rdnfn]

Not much more to add beyond Simon's comments... looks great! 🚀

[rdnfn]

I think this is a good enough assumption but note that I don't think this will not always hold. For large travel distances in particular - for example imagine a circular road around a patch and a large travel distance. If we want to be certain we could also check if it overlaps with the original node, but it might not be worth it in terms of computational cost.

[herbiebradley]

Good point - I realised that it is easier to check which polygon in the MultiPolygon contains a point in the original, like this:

diff = buff_poly - barrier_poly
if isinstance(diff, shapely.geometry.MultiPolygon):
    if len(diff) > 1:
        bools = [poly.contains(repr_point) for poly in diff]
        node_poly = diff[bools.index(True)]
        if not node_poly.intersects(nbr_polygon):
            add_edge = False

This should be of similar speed.

[herbiebradley]

@Croydon-Brixton @rdnfn Thanks for the feedback - all issues have hopefully been resolved now! 🚀