MAINT: Reduce warnings in CI

[m-richards]

No description provided.

[martinfleis]

@m-richards what is the state of this? With pandas 2.1, some of the warnings are propagated to a user (categorical dtype checks). Would be good to resolve that in 0.14.

[m-richards]

@m-richards what is the state of this? With pandas 2.1, some of the warnings are propagated to a user (categorical dtype checks). Would be good to resolve that in 0.14.

I'll have another look, I think it's left as draft as there was more that could be done, but there's probably enough changes in this PR as is already.

Edit: I think this is fine to review as is, although I can add to it if there are other warnings that are missing and particularly noisy.

[m-richards]

Just for clarity since this is a bit ugly.

In pandas 2.1, convert dtype is deprecated and the future behaviour is the same as convert_dtype = True (pandas-dev/pandas#52786)

So this just preserves the existing signature and forwards what the user provides. If we have the default of None, we pass nothing through on pandas 2.1 to get the convert_dtype=True behaviour without the warning

[martinfleis]

I pushed the same change of categorical dtype from plot to explore but the rest looks great! Thanks for this!

[jorisvandenbossche]

Should we actually be warning here? (this warning is coming from ourselves, right?) Since the keyword is explicitly set

[m-richards]

That's what you'd expect right? That's not how its implemented though (xref #1832), there is no way to silence the warning currently if the index is unaligned.

[jorisvandenbossche]

This is actually something we should ideally fix in overlay itself, as the user shouldn't see this warning (it will not change any behaviour, because the resulting geometry collection will also get dropped afterwards).
So I would suggest to not hide this warning here in our tests (and just leave it as a warning for now for this PR)

[m-richards]

Agree that makes sense. I'll add the warning back in, and I might move this to another issue since it's actually bubbled up from a dissolve call within overlay, so needs some examining to work out where to fix. (Or if we just roll over the deprecation and change the behaviour that won't be necessary).

[jorisvandenbossche]

In the last commit (09189ce), I reverted a change to silence a warning in a groupby test. Because with this change, the test is now failing with latest geopandas (maybe related to the _constructor changes?). In any case something to further investigate, but let's do that in a follow-up such that this PR can be merged (I also want to do a general follow-up for fixing more warnings).

[jorisvandenbossche]

Related to that last commit, a reproducer from that test:

from shapely.geometry import Point
from geopandas import GeoDataFrame

crs = "EPSG:4326"

df = GeoDataFrame(
    {
        "geometry": [Point(0, 0), Point(1, 1), Point(0, 0)],
        "value1": np.arange(3, dtype="int64"),
        "value2": np.array([1, 2, 1], dtype="int64"),
    },
    crs=crs,
)

# dummy test asserting we can access the crs
def func(group):
    assert isinstance(group, GeoDataFrame)
    assert group.crs == crs

# original code, warning with latest pandas about group key being included in applied group
df.groupby("value2").apply(func)
# proposed fix to avoid the warning, now failing
df.groupby("value2")[df.columns].apply(func)

This is failing since #3080, but I think this is rather a bug on the pandas side (not calling finalize when doing the column selection, and thus the active geometry name is not passed on), and essentially similar as the other follow-up issue from #3080: this was previously only working because of passing the data to GeoDataFrame which then defaults to "geometry" column. If changing the above example to use a different column name, it was already failing before #3080