API: how to expose coverage_union?

[martinfleis]

I am wondering how (and if) do we want to expose shapely.coverage_union because the way it behaves on shapely side is super weird. The function assumes two arrays, that are of the same length but does not return a row-wise union as shapely.union does but a single (Multi)Polygon as a union of all of it. Geometries from array a, geometries from array b, all merged to a single geometry.

What I don't understand is the use case for such a function. The assumption here is that polygons form a coverage. Then when would I use coverage_union with its expectation of aligned arrays over coverage_union_all on concatenated arrays? The outcome is exactly the same? Why does shapely.coverage_union exists, especially since it is implemented as return coverage_union_all([a, b], **kwargs)?

My suggestion is to not expose it on geopandas side and expose only coverage_union_all within union_all via a keyword.

[theroggy]

I would guess it was just incorrectly implemented in shapely (pygeos) and nobody ever noticed?

[jorisvandenbossche]

Indeed, I think the way the two arrays in shapely.coverage_union are put together should be transposed to work as expected.

Anyway, for the topic of "unary_union", the relevant function is shapely.coverage_union_all (as equivalent of shapely.union_all, t which shapely.unary_union is aliased), and that one is working correctly.

I agree that for this use case, we can add some "method" keyword to union_all() (in a next shapely release, there might be another method once GEOSDisjointSubsetUnion (GEOS 3.12+) is exposed)

---

I think there theoretically can be a use case for the element-wise coverage_union, although that's clearly not implemented correctly right now. But let's discuss that on the shapely side. And then if that would be fixed (and not removed), we can still discuss whether we also want to add a coverage_union() method here.

[jorisvandenbossche]

Opened shapely/shapely#1963 on the shapely side