Fast file-based format for geometries with Geopandas

pip install geofeather==0.3.0


DEPRECATED - geofeather

Build Status Coverage Status


The core functionality in geofeather has been integrated directly into GeoPandas version 0.8.0. See docs for instructions about how to use to_feather / read_feather in GeoPandas.

You are encouraged to use GeoPandas for this functionality. According to early benchmarks, it is even faster!

NOTE: you are not able to read geofeather-created files directly into GeoPandas via read_feather; you will first need to use geofeather to read into a GeoDataFrame, then you can write that to a new feather file.

I may release an updated version to help migrate from geofeather to the new functionality in GeoPandas. GeoPandas uses a metadata schema stored within the feather file to hold the CRS information and other details, which makes the new representation more compact (no more sidecar files for CRS info).

If you need help converting files created using geofeather for use in GeoPandas, please create an issue.


A faster file-based format for geometries with geopandas.

This project capitalizes on the very fast feather file format to store geometry (points, lines, polygons) data for interoperability with geopandas.

Introductory post.

Why does this exist?

This project exists because reading and writing standard spatial formats (e.g., shapefile) in geopandas is slow. I was working with millions of geometries in multiple processing steps, and needed a fast way to read and write intermediate files.

In our benchmarks, we see about 5-6x faster file writes than writing from geopandas to shapefile via .to_file() on a GeoDataFrame.

We see about 2x faster reads compared to geopandas read_file() function.

How does it work?

The feather format works brilliantly for standard pandas data frames. In order to leverage the feather format, we simply convert the geometry data from shapely objects into Well Known Binary (WKB) format, and then store that column as raw bytes.

We store the coordinate reference system using JSON format in a sidecar file .crs.


Available on PyPi at:

pip install geofeather



Given an existing GeoDataFrame my_gdf, pass this into to_geofeather:

to_geofeather(my_gdf, 'test.feather')


my_gdf = from_geofeather('test.feather')


pygeos provides much faster operations of geospatial operations over arrays of geospatial data.

geopandas is in the process of migrating to using pygeos geometries as its internal data storage instead of shapely objects.

Until pygeos is fully integrated, there are shims in geofeather to support interoperability with pandas DataFrames containing pygeos geometries. If you are already using pygeos against data you read from geofeather, using the following shims will generate 3-7x speedups reading and writing data compared to geofeather reading into GeoDataFrames.

Internally, the feather file is identical to the one created above.

pygeos is required in order to use this functionality.

WARNING: this will be deprecated as soon as pygeos is integrated into geopandas.

from geofeather.pygeos import to_geofeather, from_geofeather

# given a DataFrame df containing pygeos geometries in 'geometry' column
# and a crs object

to_geofeather(df, 'test.feather', crs=crs)

df = from_geofeather('test.geofeather')

Note: no CRS information is returned when reading from geofeather into a DataFrame, in order to keep the function signature the same as above from_geofeather


Right now, indexes are not supported in feather files. In order to get around this, simply reset your index before calling to_geofeather.



  • add crs attribute to pandas DataFrame containing pygeos geometries


  • allow serializing to / from pandas DataFrames containing pygeos geometries (see notes above).
  • use new CRS object in geopandas data frames (#4)
  • dropped to_shp; use geopandas to_file() instead.


  • allow reading a subset of columns from a feather file
  • store geometry in 'geometry' column instead of 'wkb' column (simplification to avoid renaming columns)


  • Initial release


Everything that makes this fast is due to the hard work of contributors to pyarrow, geopandas, and shapely.