Raster to Vector

gpio process polygonize and gpio process contour turn rasters into GeoParquet, backed by contourrs — a Rust raster-tracing library whose Arrow output flows straight into gpio's write pipeline:

• polygonize traces a categorical raster (land cover, segmentation masks, classification output) into one polygon feature per contiguous class region;
• contour extracts filled contour bands from a continuous raster (elevation, temperature), one polygon per band with min/max break attributes.

The implementation lives in core/process/raster/ (reader.py, polygonize.py, contour.py, output.py), with the lazy dependency shims in core/optional_deps.py.

Installation

Raster support is an optional extra (gpio's first and only raster surface):

pip install 'geoparquet-io[raster]'
# or, for a uv tool install:
uv tool install geoparquet-io --with contourrs --with rasterio

Python 3.12+

contourrs ships wheels for CPython 3.12–3.14 only. On Python 3.10/3.11 the extra installs rasterio but skips contourrs, and the commands explain the floor at runtime. contourrs is Apache-2.0; rasterio reads the GeoTIFF input and supplies the transform, nodata and CRS, which land in the output's geo metadata as PROJJSON.

Polygonize

CLIPython

# Every class becomes polygons; nodata is honored automatically
gpio process polygonize landcover.tif landcover.parquet

# Keep only classes 1 and 3, name the attribute column
gpio process polygonize mask.tif buildings.parquet \
    --values 1,3 --value-column class_id

import geoparquet_io as gpio
from geoparquet_io.api import ops

# Table constructor, chainable
gpio.Table.polygonize('landcover.tif').write('landcover.parquet')

# Pure function returning a PyArrow table
table = ops.polygonize('mask.tif', values=[1.0], value_column='class_id')

Option	Default	Meaning
--band	1	Raster band to trace
--values	all	Comma-separated class values to keep
--value-column	value	Name of the class attribute column
--nodata	file's tag	Nodata value to exclude. An explicit value replaces the tag: pixels equal to a wrong tag become data again
--no-mask	off	Ignore the raster's mask/nodata entirely

The traced polygons come out with counterclockwise exterior rings, as the GeoParquet spec recommends, and round-trip exactly: rasterizing the output reproduces the input array pixel for pixel.

Contour

CLIPython

# Bands every 100 units, spanning the raster's value range
gpio process contour dem.tif contours.parquet --interval 100

# Explicit break values
gpio process contour dem.tif contours.parquet --levels 0,250,500,1000

import geoparquet_io as gpio
from geoparquet_io.api import ops

gpio.Table.contour('dem.tif', interval=100).write('contours.parquet')

table = ops.contour('dem.tif', levels=[0.0, 250.0, 500.0, 1000.0])

Option	Default	Meaning
--interval N	—	Breaks every N units from --base (spans the band's range)
--levels a,b,c	—	Explicit break values (mutually exclusive with --interval)
--base	0.0	Offset for --interval
--band	1	Raster band to contour
--nodata	file's tag	Nodata value to exclude. An explicit value replaces the tag
--no-mask	off	Ignore the raster's mask/nodata entirely
--min-column / --max-column	min / max	Names for each band's break attributes

Each output feature is one band polygon attributed with its [min, max) break values; bands never overlap and together tile the valid data area. Contours are interpolated between grid nodes (marching squares), so a band footprint reaches the outermost pixel centers, not the outer pixel edges.

Output metadata

Both commands write through gpio's normal pipeline: the raster's CRS becomes the geo block's PROJJSON crs, geometry_types and bbox are computed from the actual output, and row groups, compression and GeoParquet version behave exactly as in every other gpio write (all the usual --compression, --row-group-size, --version options apply).