Geo Polygonize

A native Rust port of the JTS/GEOS polygonization algorithm. This crate allows you to reconstruct valid polygons from a set of lines, including handling of complex topologies like holes, nested shells, and disconnected components.

Features

• Robust Polygonization: Extracts polygons from unstructured linework.
• Iterative Grid Noding (Unchecked): Splits and snaps dirty linework, without claiming certified snap-rounding guarantees.
• Certified Fixed-Precision Noding: Optional hot-pixel snap rounding with an independent full-noding postcondition check.
• Hardware Acceleration: Uses SIMD instructions (via wide crate) for critical geometric predicates like Point-in-Polygon checks.
• Wasm Optimized: Tailored for WebAssembly with talc allocator and binary GeoArrow support.
• Performance: SIMD, spatial indexing, and optional parallel execution, with checked-in benchmark tooling.
• Geo Ecosystem: Fully integrated with geo-types and geo crates.
• GeoArrow Support: Arrow C Data Interface and Arrow IPC integration with GeoArrow metadata.

Engineering Roadmap

For delivered foundations and the next evidence-gated work, see ROADMAP.md.

Support policy

Supported APIs, toolchains, targets, feature combinations, compatibility windows, and failure/threading contracts are defined in SUPPORT.md.

Usage

Library

use geo_polygonize_core::{polygonize_line_strings, PolygonizerOptions};
use geo_types::LineString;

fn main() {
    let ring = LineString::from(vec![
        (0.0, 0.0),
        (10.0, 0.0),
        (10.0, 10.0),
        (0.0, 10.0),
        (0.0, 0.0),
    ]);

    let result = polygonize_line_strings([&ring], &PolygonizerOptions::default())
        .expect("Polygonization failed");
    let polygons = result.into_multi_polygon();

    for polygon in polygons {
        println!("Found polygon: {polygon:?}");
    }
}

polygonize_line_strings accepts owned or borrowed geo_traits::LineStringTrait values. Use polygonize when source IDs or Z values must be supplied explicitly, and keep its full result when dangles, cut edges, invalid rings, or diagnostics are needed.

Rust consumers should import the supported facade from the crate root. Graph, noding, containment, utility, mutable-builder, and tiled APIs are internal or experimental and may change without compatibility guarantees before 1.0. TiledPolygonizer validates its grid/options and propagates per-tile errors, but callers must still choose a sufficient buffer and verify untiled equivalence for their workload. Its result includes per-tile topology counts and merge/dedup counts; these are observational and do not certify buffer sufficiency. Canonical mode applies the same final ordering pass after tile merge; the equivalence gate covers one, two, four, and many boundary crossings plus concave, hole, dirty-intersection, and dangle cases.

Choosing noding and precision

Polygonization quality is heavily influenced by input noding strategy.

• node_input = false (default): Fastest path. Use this when your input linework is already noded (all intersections are explicit vertices).
• node_input = true: Enables unchecked iterative grid noding. Use this for real-world datasets that may contain slight misalignments, overlaps, or self-intersections, and validate outputs when correctness must be certified.
• PrecisionModel::Floating (default) preserves input coordinates and uses floating-point intersections.
• PrecisionModel::FixedGrid { grid_size } rounds topology coordinates to an explicit positive grid, even when input noding is disabled. Choose the grid in the units of your data; oversnapping can collapse narrow features.

Practical workflow:

1. Run with node_input = false first on trusted data.
2. If you observe missing polygons, sliver artifacts, or unresolved intersections, enable node_input.
3. Use a fixed precision model only when your application has an explicit coordinate grid contract.

Canonical options now use precision_model instead of snap_grid_size. Legacy positional Python, Wasm, and C APIs still translate their grid argument when noding is enabled and ignore it on the non-noding fast path.

Set options.noding.guarantee to NodingGuarantee::Validate to run an independent full-noding check before graph construction. Validation reports the first pair with an interior intersection or unnormalized collinear overlap. Use NodingGuarantee::CertifiedFixedPrecision with node_input, FixedGrid, the Snap backend, and SnapStrategy::Grid for hot-pixel snap rounding plus that validation. Certified coordinates must fit exact integer grid indices.

Z semantics

Topology is always computed in XY. ZPolicy::InterpolateAlongEdge is the default and reconstructs split-vertex Z on each source edge; PreferNearestEndpoint uses the nearer endpoint, and Ignore emits zero Z. When several source edges meet at the same XY node, the graph chooses the Z from the lowest line ID (then the lowest Z) and reports conflicts through diagnostics. ErrorOnConflict instead returns a typed error when the difference exceeds z.conflict_tolerance.

Use Line3D or the Python/Wasm typed-buffer APIs with stride=3 for Z-aware processing. The GeoRust, GeoJSON, and current GeoArrow adapters are XY-only.

Output semantics

The polygonizer intentionally returns only valid polygonal areas that can be formed from closed cycles:

• Dangles are removed: dead-end edges do not appear in output polygons.
• Cut edges are excluded: edges that are connected but cannot bound a face are ignored.
• Holes and nested shells are preserved when enough boundary information is present.

This behavior matches classical JTS/GEOS polygonization semantics and is useful for cleaning linework before area analysis.

GeoArrow Integration

Rust Arrow and GeoArrow adapters live in the geo-polygonize-arrow crate; the Wasm crate includes it automatically.

For the Arrow C Data Interface ABI, including ownership and error retrieval, see docs/C_ABI.md.

use geo_polygonize_arrow::{polygonize_arrow, PolygonizerOptions};
// ... create Arrow array ...
// let result = polygonize_arrow(&array, &field, options);

The FlatGeoBuf file adapter lives in the geo-polygonize-flatgeobuf crate.

Python

The Python package is published as geo-polygonize-py and imported as geo_polygonize.

pip install geo-polygonize-py

import numpy as np
from geo_polygonize import polygonize, import_probe

# 1. Using Shapely LineStrings or coordinate lists directly
lines = [
    [(0, 0), (10, 0), (10, 10), (0, 10), (0, 0)],
    [(0, 0), (10, 10)]
]

# return_polygons=True returns a list of shapely.geometry.Polygon objects
polygons = polygonize(lines=lines, return_polygons=True)
for p in polygons:
    print(p.area)

# 2. Using High-Performance Flat Arrays
# Flat buffers avoid Python object-per-coordinate overhead
coords = np.array([
    0.0, 0.0, 10.0, 0.0, 10.0, 10.0, 0.0, 10.0, 0.0, 0.0,
    0.0, 0.0, 10.0, 10.0
], dtype=np.float64)

# Start indices for each line segment.
# The final closing offset is computed implicitly.
offsets = np.array([0, 5], dtype=np.uint32)

# `report` preserves the compatibility result with objects, buffers, fingerprint,
# diagnostics, and per-binding phase timings.
result_dict = polygonize(coords=coords, offsets=offsets)

# Request only the representation you need.
buffers = polygonize(coords=coords, offsets=offsets, output="buffers")
objects = polygonize(coords=coords, offsets=offsets, output="objects")

# Execution budgets are non-semantic and stay separate from topology options.
bounded = polygonize(
    coords=coords,
    offsets=offsets,
    output="buffers",
    execution_limits={"max_output_coordinates": 1_000_000},
)

# Native-extension probes are cheap and safe for optional integrations.
ok, error = import_probe()

CFB/autograder integrations should use the versioned production profile rather than assembling caller-side knobs or using legacy polygonize(..., node=True, snap=0.5) calls:

from geo_polygonize import cfb_robust_options, polygonize_with_options

result = polygonize_with_options(
    coords=coords,
    offsets=offsets,
    options=cfb_robust_options(),
)

The default report return shape remains a stable dictionary with polygons as SimplePolygon values. buffers omits Python geometry objects and the fingerprint; objects omits flattened arrays and the fingerprint. Use return_polygons=True only when you want Shapely Polygon objects.

When provenance is enabled, coincident and partially overlapping input lines are dissolved into one topology edge while every contributing nonzero line_id remains in the polygon's sorted boundary_line_ids.

SnapStrategy::Grid keeps topology and output coordinates on a configured fixed precision grid. The CFB profile uses GeosCompat: the grid establishes robust topology, then output nodes regain deterministic source XY coordinates to better match Shapely snap plus full-precision noding. It is not set_precision emulation, and exact parity is not guaranteed for many-to-one snaps.

For Shapely parity checks, compare report-mode outputs with the built-in mismatch helper:

from geo_polygonize import explain_mismatch, polygonize_with_options

options = cfb_robust_options()
result_a = polygonize_with_options(coords=coords_a, offsets=offsets_a, options=options)
result_b = polygonize_with_options(coords=coords_b, offsets=offsets_b, options=options)
result_a["options"] = options
result_b["options"] = options

mismatch = explain_mismatch(result_a, result_b)

For a minimal Shapely smoke comparison, use area signatures:

from shapely.ops import polygonize as shapely_polygonize

rust_polys = polygonize_with_options(lines=lines, options=cfb_robust_options(), return_polygons=True)
rust_areas = sorted(round(poly.area, 6) for poly in rust_polys)
shapely_areas = sorted(round(poly.area, 6) for poly in shapely_polygonize(lines))

WebAssembly (WASM)

This library supports WebAssembly with an ergonomic dual-build configuration that automatically utilizes SIMD instructions where available.

Installation:

npm install geo-polygonize

Standard Usage (Quick Demos): The default entry point automatically handles feature detection (SIMD) and lazy-loading of the Wasm binary. The Wasm is inlined as a Base64 Data URI, so no extra bundler configuration is needed. For app builds, prefer the slim entry point below so your bundler keeps the Wasm assets out of the JavaScript chunk.

import init, { polygonize, polygonize_geoarrow } from "geo-polygonize";

async function run() {
    await init();

    const geojson = {
        "type": "FeatureCollection",
        "features": [
            // ... your line features
        ]
    };

    // Returns a GeoJSON FeatureCollection string
    // Pass explicitly matching backend configuration if desired
    const result = polygonize(
        JSON.stringify(geojson),
        true, // node_input
        0.5   // snap_grid_size
    );
    console.log(JSON.parse(result));

    // Or use Arrow IPC bytes
    // const ipcBuffer = ...;
    // const arrowResult = polygonize_geoarrow(ipcBuffer, false, 1e-10, false);
}

Slim Usage (Apps / Manual Loading): For Vite and other app bundlers, import from geo-polygonize/slim and pass explicit Wasm asset URLs.

import { cfbRobustOptions, initBest } from "geo-polygonize/slim";
import scalarUrl from "geo-polygonize/geo_polygonize.wasm?url";
import simdUrl from "geo-polygonize/geo_polygonize_simd.wasm?url";

async function run() {
    const wasm = await initBest(
        { module_or_path: scalarUrl },
        { module_or_path: simdUrl },
    );

    const result = wasm.polygonizeWithOptions(
        JSON.stringify(geojson),
        cfbRobustOptions,
    );
}

For cancellable browser work, use the worker-backed asynchronous entry point. Aborting terminates its disposable worker; the direct Wasm exports above remain synchronous and cannot be interrupted with an AbortSignal.

import { polygonizeWithOptionsAsync } from "geo-polygonize";

const controller = new AbortController();
const result = await polygonizeWithOptionsAsync(
    JSON.stringify(geojson),
    cfbRobustOptions,
    { signal: controller.signal },
);
// controller.abort();

Multithreaded Usage (Experimental): This library provides a multithreaded build powered by wasm-bindgen-rayon.

import init, { initThreadPool, polygonize } from "geo-polygonize/threads";

async function run() {
    await init();

    // Initialize thread pool (e.g., with navigator.hardwareConcurrency)
    await initThreadPool(navigator.hardwareConcurrency);

    // ... use polygonize as usual
}

Important: Multithreaded WebAssembly requires SharedArrayBuffer, which is only available in secure contexts. You must serve your page with the following headers:

Cross-Origin-Opener-Policy: same-origin
Cross-Origin-Embedder-Policy: require-corp

CLI Example

The repository includes a CLI tool to polygonize GeoJSON files.

# Build the example
cargo build -p geo-polygonize-core --example polygonize --release

# Run on input lines
cargo run -p geo-polygonize-core --release --example polygonize -- --input lines.geojson --output polygons.geojson --node

Visualization

You can visualize the results using the provided Python script (requires matplotlib and shapely).

python3 scripts/visualize.py --input lines.geojson --output polygons.geojson --save result.png

Examples

Below are some examples of what the polygonizer can do.

Nested Holes and Islands

The algorithm correctly identifies nested structures (Island inside a Hole inside a Shell).

Incomplete Grid / Dangles

The algorithm prunes dangles (dead-end lines) and extracts only closed cycles.

Touching Polygons (Shared Edges)

Using robust noding (--node), it can reconstruct adjacent polygons that share boundaries, even if the input lines are not perfectly noded.

Self-Intersecting Geometry (Bowtie)

Self-intersecting lines are split at intersection points, and valid cycles are extracted.

Complex Geometries

The polygonizer can handle complex, curved inputs (approximated by LineStrings) such as overlapping circles and shapes with multiple holes.

Overlapping Circles: Note how the intersection regions are correctly identified as separate polygons.

Curved Holes: A complex polygon with multiple circular holes.

Benchmarks

This library includes a "severe" comparison suite against shapely (GEOS).

See BENCHMARKS.md for detailed results and instructions on how to run them.

Architecture

This implementation moves away from the pointer-based graph structures of JTS/GEOS to a Rust-idiomatic Index Graph (Arena) approach.

See ARCHITECTURE.md for a deep dive into the optimization strategies.

Key optimizations include:

1. Noding: Unchecked iterative grid noding with spatially dispatched intersection detection.
2. Vectorization: SIMD-accelerated Ray Casting for efficient Hole Assignment.
3. Memory Layout: Structure of Arrays (SoA) for graph nodes and talc allocator for Wasm.

License

MIT/Apache-2.0