Index type is lost when filtering to empty geo data frame

Description

When filtering a GeoDataFrame with a condition that returns no True value, so that the resulting dataset is empty, the type of the index is lost.

Test

import pandas as pd
import geopandas as gpd
from datetime import timedelta, datetime
from shapely.geometry import Point


df = pd.DataFrame(dict(geometry=[Point(i, i) for i in range(3)], timestamp=[datetime.now() + timedelta(seconds=i) for i in range(3)])).set_index("timestamp")
gdf = gpd.GeoDataFrame(df)
max_date = df.index.max()

df_filtered = df.loc[df.index > max_date, :]
gdf_filtered = gdf.loc[gdf.index > max_date, :]
print(df_filtered.index)
print(gdf_filtered.index)

0.5.1

Output

DatetimeIndex([], dtype='datetime64[ns]', name='timestamp', freq=None) 
DatetimeIndex([], dtype='datetime64[ns]', name='timestamp', freq=None)

Environment

attrs==19.3.0
Click==7.0
click-plugins==1.1.1
cligj==0.5.0
Fiona==1.8.9.post2
geopandas==0.5.1
munch==2.5.0
numpy==1.17.3
pandas==0.25.3
pyproj==2.4.0
python-dateutil==2.8.1
pytz==2019.3
Shapely==1.6.4.post2
six==1.12.0

0.6.1

Output

DatetimeIndex([], dtype='datetime64[ns]', name='timestamp', freq=None) 
RangeIndex(start=0, stop=0, step=1)

Environment

attrs==19.3.0
Click==7.0
click-plugins==1.1.1
cligj==0.5.0
Fiona==1.8.9.post2
geopandas==0.6.1
munch==2.5.0
numpy==1.17.3
pandas==0.25.3
pyproj==2.4.0
python-dateutil==2.8.1
pytz==2019.3
Shapely==1.6.4.post2
six==1.12.0

[jorisvandenbossche]

@simon-keith thanks a lot for the clear, reproducible bug report!

I can reproduce this on 0.6.1, but not with master. So something might have fixed this already. In which case, we should add a test and could backport it to 0.6.x

[jorisvandenbossche]

OK, I found the change that explains why it fails with released pandas but works with pandas master.

Pandas 0.25:

In [18]: df = pd.DataFrame({'a':[1, 2, 3], 'b': [.2, .3, .4]}, index=pd.date_range("2012-01-01", periods=3))

In [20]: df_empty = df[:0] 

In [21]: df_empty.index 
Out[21]: DatetimeIndex([], dtype='datetime64[ns]', freq='D')

In [22]: df_empty['a'] = np.array([], dtype='int64')  

In [23]: df_empty.index   
Out[23]: RangeIndex(start=0, stop=0, step=1)

In [24]: pd.__version__ 
Out[24]: '0.25.3'

With pandas master:

In [1]: df = pd.DataFrame({'a':[1, 2, 3], 'b': [.2, .3, .4]}, index=pd.date_range("2012-01-01", periods=3))

In [2]: df_empty = df[:0] 

In [3]: df_empty.index  
Out[3]: DatetimeIndex([], dtype='datetime64[ns]', freq='D')

In [4]: df_empty['a'] = np.array([], dtype='int64')   

In [5]: df_empty.index
Out[5]: DatetimeIndex([], dtype='datetime64[ns]', freq='D')

In [6]: pd.__version__ 
Out[6]: '0.26.0.dev0+822.g6b06f4342.dirty'

So assigning to a column of an empty frame resets the index on released pandas, but this has been fixed to preserve the index on pandas master.

And with GeoPandas 0.6, we started to rely on this pattern to write the geometries into the array (and thus also do this for empty geodataframes), hence introducing this regression for released pandas versions.