rusterize 0.6.0

High performance rasterization tool for Python built in Rust

rusterize

High performance rasterization tool for Python built in Rust. This repository stems from the fasterize package built in C++ for R and ports parts of the logics into Python with a Rust backend, in addition to some useful improvements (see API).

rusterize is designed to work on (multi)polygons and (multi)linestrings, even when they are nested inside complex geometry collections. Functionally, it takes an input geopandas dataframe and returns a xarray, a numpy, or a sparse array in COOrdinate format.

Installation

rusterize is distributed in two flavors. A core library that performs the rasterization and returns a bare numpy array, or a xarray-flavored version that returns a georeferenced xarray. This latter requires xarray and rioxarray to be installed. This is the recommended flavor.

Install the current version with pip:

# Core library
pip install rusterize

# With xarray capabilities
pip install 'rusterize[xarray]'

Contributing

Any contribution is welcome! You can install rusterize directly from this repo using maturin as an editable package. For this to work, you’ll need to have Rust and cargo installed.

# Clone repo
git clone https://github.com/<username>/rusterize.git
cd rusterize

# Install the Rust nightly toolchain
rustup toolchain install nightly-2025-07-31

 # Install maturin
pip install maturin

# Install editable version with optmized code
maturin develop --profile dist-release

API

This package has a simple API:

from rusterize import rusterize

# gdf = <geodataframe>

# rusterize
rusterize(
    gdf,
    like=None,
    res=(30, 30),
    out_shape=(10, 10),
    extent=(0, 10, 10, 20),
    field="field",
    by="by",
    burn=None,
    fun="sum",
    background=0,
    encoding="xarray",
    dtype="uint8"
)

• gdf: geopandas dataframe to rasterize
• like: xr.DataArray to use as template for res, out_shape, and extent. Mutually exclusive with these parameters (default: None)
• res: (xres, yres) for desired resolution (default: None)
• out_shape: (nrows, ncols) for desired output shape (default: None)
• extent: (xmin, ymin, xmax, ymax) for desired output extent (default: None)
• field: column to rasterize. Mutually exclusive with burn (default: None -> a value of 1 is rasterized)
• by: column for grouping. Assign each group to a band in the stack. Values are taken from field if specified, else burn is rasterized (default: None -> singleband raster)
• burn: a single value to burn. Mutually exclusive with field (default: None). If no field is found in gdf or if field is None, then burn=1
• fun: pixel function to use when multiple values overlap. Available options are sum, first, last, min, max, count, or any (default: last)
• background: background value in final raster (default: np.nan). A None value corresponds to the default of the specified dtype. An illegal value for a dtype will be replaced with the default of that dtype. For example, a background=np.nan for dtype="uint8" will become background=0, where 0 is the default for uint8.
• encoding: defines the output format of the rasterization. This is either a dense xarray/numpy representing the burned rasterized geometries, or a sparse array in COOrdinate format good for sparse observations and low memory consumption. Available options are xarray, numpy, sparse (default: xarray -> will trigger an error if xarray and rioxarray are not found).
• dtype: dtype of the final raster. Available options are uint8, uint16, uint32, uint64, int8, int16, int32, int64, float32, float64 (default: float64)

Note that control over the desired extent is not as strict as for resolution and shape. That is, when resolution, output shape, and extent are specified, priority is given to resolution and shape. So, extent is not guaranteed, but resolution and shape are. If extent is not given, it is taken from the polygons and is not modified, unless you specify a resolution value. If you only specify an output shape, the extent is maintained. This mimics the logics of gdal_rasterize.

Encoding

Version 0.5.0 introduced a new encoding parameter to control the output format of the rasterization. This means that you can return a xarray/numpy with the burned rasterized geometries, or a new SparseArray structure. This SparseArray structure stores the band/row/column triplets of where the geometries should be burned onto the final raster, as well as their corresponding values before applying any pixel function. This can be used as an intermediate output to avoid allocating memory before materializing the final raster, or as a final product. SparseArray has three convenience functions: to_xarray(), to_numpy(), and to_frame(). The first two return the final xarray/numpy, the last returns a polars dataframe with only the coordinates and values of the rasterized geometries. Note that SparseArray avoids allocating memory for the array during rasterization until when it's actually needed (e.g. calling to_xarray()). See below for an example.

Usage

rusterize consists of a single function rusterize().

from rusterize import rusterize
import geopandas as gpd
from shapely import wkt
import matplotlib.pyplot as plt

# Construct geometries
geoms = [
    "POLYGON ((-180 -20, -140 55, 10 0, -140 -60, -180 -20), (-150 -20, -100 -10, -110 20, -150 -20))",
    "POLYGON ((-10 0, 140 60, 160 0, 140 -55, -10 0))",
    "POLYGON ((-125 0, 0 60, 40 5, 15 -45, -125 0))",
    "MULTILINESTRING ((-180 -70, -140 -50), (-140 -50, -100 -70), (-100 -70, -60 -50), (-60 -50, -20 -70), (-20 -70, 20 -50), (20 -50, 60 -70), (60 -70, 100 -50), (100 -50, 140 -70), (140 -70, 180 -50))",
    "GEOMETRYCOLLECTION (POINT (50 -40), POLYGON ((75 -40, 75 -30, 100 -30, 100 -40, 75 -40)), LINESTRING (60 -40, 80 0), GEOMETRYCOLLECTION (POLYGON ((100 20, 100 30, 110 30, 110 20, 100 20))))"
]

# Convert WKT strings to Shapely geometries
geometries = [wkt.loads(geom) for geom in geoms]

# Create a GeoDataFrame
gdf = gpd.GeoDataFrame({'value': range(1, len(geoms) + 1)}, geometry=geometries, crs='EPSG:32619')

# rusterize to "xarray" -> return a xarray with the burned geometries and spatial reference (default)
# will raise a ModuleNotFoundError if xarray and rioxarray are not found
output = rusterize(
    gdf,
    res=(1, 1),
    field="value",
    fun="sum",
).squeeze()

# plot it
fig, ax = plt.subplots(figsize=(12, 6))
output.plot.imshow(ax=ax)
plt.show()

# rusterize to "sparse" -> custom structure storing the coordinates and values of the rasterized geometries
output = rusterize(
    gdf,
    res=(1, 1),
    field="value",
    fun="sum",
    encoding="sparse"
)
output
# SparseArray:
# - Shape: (131, 361)
# - Extent: (-180.5, -70.5, 180.5, 60.5)
# - Resolution: (1.0, 1.0)
# - EPSG: 32619
# - Estimated size: 369.46 KB

# materialize into xarray or numpy
array = output.to_xarray()
array = output.to_numpy()

# get only coordinates and values
output.to_frame()
# shape: (29_340, 3)
# ┌─────┬─────┬──────┐
# │ row ┆ col ┆ data │
# │ --- ┆ --- ┆ ---  │
# │ u32 ┆ u32 ┆ f64  │
# ╞═════╪═════╪══════╡
# │ 6   ┆ 40  ┆ 1.0  │
# │ 6   ┆ 41  ┆ 1.0  │
# │ 6   ┆ 42  ┆ 1.0  │
# │ 7   ┆ 39  ┆ 1.0  │
# │ 7   ┆ 40  ┆ 1.0  │
# │ …   ┆ …   ┆ …    │
# │ 64  ┆ 258 ┆ 1.0  │
# │ 63  ┆ 259 ┆ 1.0  │
# │ 62  ┆ 259 ┆ 1.0  │
# │ 61  ┆ 260 ┆ 1.0  │
# │ 60  ┆ 260 ┆ 1.0  │
# └─────┴─────┴──────┘

Benchmarks

rusterize is fast! Let’s try it on small and large datasets.

from rusterize import rusterize
import geopandas as gpd
import requests
import zipfile
from io import BytesIO

# large dataset (~380 MB)
url = "https://s3.amazonaws.com/hp3-shapefiles/Mammals_Terrestrial.zip"
response = requests.get(url)

# unzip
with zipfile.ZipFile(BytesIO(response.content), 'r') as zip_ref:
    zip_ref.extractall()

# read
gdf_large = gpd.read_file("Mammals_Terrestrial/Mammals_Terrestrial.shp")

# small dataset (first 1000 rows)
gdf_small = gdf_large.iloc[:1000, :]

# rusterize at 1/6 degree resolution
def test_large(benchmark):
  benchmark(rusterize, gdf_large, res=(1/6, 1/6), fun="sum")

def test_small(benchmark):
  benchmark(rusterize, gdf_small, res=(1/6, 1/6), fun="sum")

Then you can run it with pytest and pytest-benchmark:

pytest <python file> --benchmark-min-rounds=20 --benchmark-time-unit='s'

--------------------------------------------- benchmark: 1 tests --------------------------------------------
Name (time in s)         Min      Max     Mean  StdDev   Median     IQR  Outliers     OPS  Rounds  Iterations
-------------------------------------------------------------------------------------------------------------
rusterize_small       0.0791    0.0899   0.0812  0.0027   0.0803  0.0020       2;2  12.3214     20          1
rusterize_large     1.379545    1.4474   1.4006  0.0178   1.3966  0.0214       5;1   0.7140     20          1
-------------------------------------------------------------------------------------------------------------

And fasterize:

library(sf)
library(raster)
library(fasterize)
library(microbenchmark)

large <- st_read("Mammals_Terrestrial/Mammals_Terrestrial.shp", quiet = TRUE)
small <- large[1:1000, ]
fn <- function(v) {
  r <- raster(v, res = 1/6)
  return(fasterize(v, r, fun = "sum"))
}
microbenchmark(
  fasterize_large = f <- fn(large),
  fasterize_small = f <- fn(small),
  times=20L,
  unit='s'
)

Unit: seconds
            expr       min         lq       mean     median        uq        max neval
 fasterize_small 0.4741043  0.4926114  0.5191707  0.5193289  0.536741  0.5859029    20
 fasterize_large 9.2199426 10.3595465 10.6653139 10.5369429 11.025771 11.7944567    20

And on an even larger datasets? Here we use a layer from the province of Quebec, Canada representing ~2M polygons of forest stands, rasterized at 30 meters (20 rounds) with no field value, pixel function any, and dense encoding. The comparison with gdal_rasterize was run with hyperfine --runs 20 "gdal_rasterize -tr 30 30 -burn 1 <data_in> <data_out>".

# rusterize
--------------------------------------------- benchmark: 1 tests --------------------------------------------
Name (time in s)         Min      Max     Mean  StdDev   Median     IQR  Outliers     OPS  Rounds  Iterations
-------------------------------------------------------------------------------------------------------------
rusterize             5.9331   7.2308   6.1302  0.3183  5.9903   0.1736       2;4  0.1631      20           1
-------------------------------------------------------------------------------------------------------------

# fasterize
Unit: seconds
      expr      min       lq     mean   median       uq      max neval
 fasterize 157.4734 177.2055 194.3222 194.6455 213.9195 230.6504    20

# gdal_rasterize (CLI) - read from fast drive, write to fast drive
Time (mean ± σ):      5.495 s ±  0.038 s    [User: 4.268 s, System: 1.225 s]
Range (min … max):    5.452 s …  5.623 s    20 runs

In terms of (multi)line rasterization speed, here's a benchmark against gdal_rasterize using a layer from the province of Quebec, Canada, representing a subset of the road network for a total of ~535K multilinestrings.

# rusterize
--------------------------------------------- benchmark: 1 tests --------------------------------------------
Name (time in s)         Min      Max     Mean  StdDev   Median     IQR  Outliers     OPS  Rounds  Iterations
-------------------------------------------------------------------------------------------------------------
test                  4.5272   5.9488   4.7171  0.3236   4.6360  0.1680       2;2  0.2120      20           1
-------------------------------------------------------------------------------------------------------------

# gdal_rasterize (CLI) - read from fast drive, write to fast drive
Time (mean ± σ):      8.719 s ±  0.063 s    [User: 3.782 s, System: 4.917 s]
Range (min … max):    8.658 s …  8.874 s    20 runs

Comparison with other tools

While rusterize is fast, there are other fast alternatives out there, including GDAL, rasterio and geocube. However, rusterize allows for a seamless, Rust-native processing with similar or lower memory footprint that doesn't require you to leave Python, and returns the geoinformation you need for downstream processing with ample control over resolution, shape, extent, and data type.

The following is a time comparison on a single run on the same forest stands dataset used earlier.

rusterize:    5.9 sec
rasterio:     68  sec (but no spatial information)
fasterize:    157 sec (including raster creation)
geocube:      260 sec (larger memory footprint)