GeoRepair

OGC compliant geometry repair and validation for Rust.

This crate is experimental. The API is actively evolving: expect breaking changes between 0.x releases. Core algorithms, I/O backends, and feature flags are all subject to change as we improve correctness and performance.

GeoRepair repairs invalid OGC geometries and answers whether a geometry is valid. It fixes the common damage: self-intersections, unclosed rings, degenerate shapes, and NaN coordinates. Input that is already valid comes back unchanged.

Two strategies, chosen by geometry type:

• Structure (default) mirrors GEOS ST_MakeValid: planar graph extraction, face walking, even-odd assembly.
• Arrange is a CDT-based fallback for complex topologies.

WKB and WKT reading and writing, plus a custom binary batch format, are built in with no extra dependencies.

WKT keeps Z and M when you ask for it: read_wkt_zm returns the geometry, its Dims and the extra ordinates in flat coordinate order (the same layout as read_ewkb), and write_wkt_zm writes the modifier back. read_wkt reads the same files and returns x/y, so a Z-enabled layer from OGR or QGIS loads without error.

Everything is evaluated against a large test set ported from GEOS and CGAL, plus the crate's own fuzz tests. The result should always come out valid per GEOS, but the strategy differences can make the output slightly differ.

Performance

A test dataset (1,579,030 polygons)

Structure batch on a GIS dataset (813 features from a GeoPackage, flattened to 1,579,030 polygon parts), measured 2026-09-18. i5-12400F (6C/12T), release LTO, Rayon 12-thread batch; GEOS is conda-forge 3.14.1, serial per-call, run concurrently via Rayon, geometries built via CoordSeq direct construction (the one-time 1.5 s pre-build is excluded).

Dataset	GeoRepair	GEOS	Ratio
Validation (1.58M)	2.1 s (1.3 µs/poly)	3.0 s (1.9 µs/poly)	1.4x
Full pass (1.58M)	2.1 s (1.3 µs/poly)	3.2 s (2.1 µs/poly)	1.5x

Synthetic benchmarks

Full table, parallel batch (µs); ratio = GEOS / GeoRepair, >1 means we win. Methodology: docs/BENCHMARKS.md. Regenerate: python scripts/readme_bench_table.py --update <bench.json>; CI gate: python scripts/readme_bench_table.py --check <bench.json>. The overlap and dense grid rows were re-measured 2026-11-05; the rest are the 2026-09-18 set.

Benchmark	GeoRepair	GEOS	Ratio
valid polygon 4v	0.051	0.237	4.7x
valid polygon 10v	0.137	0.319	2.3x
valid polygon 50v	0.245	0.415	1.7x
valid polygon 100v	0.408	0.578	1.4x
valid polygon 500v	1.4	1.7	1.2x
valid polygon 1000v	2.4	2.7	1.1x
valid polygon 5000v	11.1	8.4	0.76x
valid polygon 10000v	21.4	17.4	0.81x
invalid bowtie 4v	0.661	19.9	30x
invalid bowtie 50v	7.7	65.8	8.5x
invalid bowtie 100v	25.1	118	4.7x
invalid bowtie 500v	56.9	511	9.0x
invalid bowtie 1000v	109	1001	9.2x
star poly 100v	8.6	6.6	0.76x
star poly 500v	172	91.4	0.53x
star poly 1000v	241	292	1.2x
spaghetti 500v	1589	2633	1.7x
spaghetti 2000v	11391	16251	1.4x
self-touch 100v	0.493	0.759	1.5x
self-touch 500v	1.5	1.6	1.0x
self-touch 1000v	2.8	3.2	1.1x
collapsed 100v	21.3	137	6.4x
collapsed 500v	113	554	4.9x
collapsed 1000v	189	1154	6.1x
near-collinear 100v	73.1	191	2.6x
near-collinear 500v	874	1184	1.4x
near-collinear 1000v	2880	2857	0.99x
large coord 1e12 100v	0.438	0.472	1.1x
large coord 1e12 500v	1.2	1.1	0.91x
large coord 1e12 1000v	2.6	2.6	1.0x
valid line	0.009	0.039	4.1x
zero-length line	0.007	0.247	36x
valid ls 4v	0.022	0.039	1.7x
valid ls 10v	0.037	0.044	1.2x
valid ls 50v	0.078	0.095	1.2x
valid ls 100v	0.146	0.126	0.86x
valid ls 500v	0.406	0.495	1.2x
valid ls 1000v	1.4	2.6	1.8x
collinear ls 4v	0.022	0.563	26x
collinear ls 10v	0.037	0.58	16x
collinear ls 50v	0.067	0.72	11x
collinear ls 100v	0.121	0.943	7.8x
collinear ls 500v	0.477	2.0	4.1x
collinear ls 1000v	1.3	8.6	6.5x
zigzag ls 10v	0.042	1.3	31x
zigzag ls 50v	0.074	4.7	63x
zigzag ls 100v	0.126	10.0	79x
zigzag ls 500v	0.493	56.0	113x
zigzag ls 1000v	1.1	118	108x
spiral ls 10v	0.096	0.835	8.7x
spiral ls 50v	0.704	4.3	6.1x
spiral ls 100v	1.6	16.1	10x
spiral ls 500v	73.7	551	7.5x
spiral ls 1000v	270	1861	6.9x
self-int ls 100v	1.2	2.6	2.2x
self-int ls 500v	5.1	3.8	0.75x
self-int ls 1000v	9.1	6.5	0.72x
dense self ls 10v	0.043	1.2	28x
dense self ls 50v	0.094	5.3	57x
dense self ls 100v	0.177	11.7	66x
dense self ls 500v	0.671	122	181x
dense self ls 1000v	1.1	349	305x
duped ls 100v	0.268	0.759	2.8x
duped ls 500v	0.492	1.4	2.8x
duped ls 1000v	0.738	3.1	4.2x
mls 50x3v	2.7	30.2	11x
mls 250x3v	15.2	132	8.7x
mls 500x3v	36.5	252	6.9x
self-int mls 50x4v	17.6	77.5	4.4x
self-int mls 250x4v	130	466	3.6x
self-int mls 500x4v	402	1034	2.6x
star-burst 10sp	3.0	14.6	4.8x
star-burst 50sp	0.707	249	353x
star-burst 100sp	1.3	1085	814x
star-burst 500sp	7.6	36693	4838x
star-burst 1000sp	14.2	188647	13304x
collinear ov 10seg	1.1	13.1	12x
collinear ov 50seg	7.4	60.5	8.2x
collinear ov 100seg	13.5	117	8.7x
collinear ov 500seg	62.2	628	10x
collinear ov 1000seg	182	1426	7.8x
x-scale 10v	2.4	21.1	8.8x
x-scale 50v	63.4	436	6.9x
x-scale 100v	306	1689	5.5x
x-scale 500v	24135	59625	2.5x
x-scale 1000v	1935	371599	192x
ringing 100v	7.5	66.9	8.9x
ringing 500v	27.9	364	13x
ringing 1000v	65.6	991	15x
hilbert 256v	35.5	171	4.8x
hilbert 1024v	219	1344	6.1x
lissajous 200v	37.0	174	4.7x
lissajous 500v	58.6	419	7.2x
lissajous 1000v	99.9	854	8.5x
lissajous 2000v	230	1726	7.5x
lissajous 5000v	671	6131	9.1x
lissajous 7:4 500v	46.2	49.8	1.1x
spoke 10sp	3.4	14.8	4.3x
spoke 50sp	0.816	250	307x
spoke 100sp	1.5	1098	739x
spoke 500sp	10.7	36970	3442x
spoke 1000sp	20.4	201201	9844x
star-comb 20sp	0.386	2.7	7.0x
star-comb 100sp	5.6	33.0	5.9x
star-comb 500sp	78.4	679	8.7x
star-comb 1000sp	668	2607	3.9x
hole hier 5h	1.5	1.8	1.2x
hole hier 20h	5.3	7.5	1.4x
hole hier 50h	19.4	26.1	1.3x
hole hier 100h	44.5	50.9	1.1x
overlap mp 5sh	7.5	452	61x
overlap mp 20sh	70.4	2727	39x
overlap mp 50sh	451	6972	15x
overlap mp 100sh	1484	15482	10x
dense grid 5x5=25	9.7	1751	180x
dense grid 10x10=100	45.8	13017	284x
dense grid 20x20=400	341	97618	286x
dense grid 30x30=900	996	396084	398x
sliver 100v	1.3	2.5	1.9x
sliver 500v	7.3	17.6	2.4x
sliver 1000v	14.0	27.4	2.0x
arrange valid 4v	0.059	0.275	4.6x
arrange valid 10v	0.102	0.353	3.4x
arrange valid 50v	0.426	0.484	1.1x
arrange valid 100v	0.498	0.719	1.4x
arrange valid 500v	1.9	2.0	1.1x
arrange valid 1000v	3.5	3.6	1.0x
arrange bowtie 4v	0.981	16.3	17x
arrange bowtie 100v	23.1	116	5.0x
arrange star 10sp	2.7	14.7	5.5x
arrange star 50sp	1.0	261	250x
arrange star 100sp	2.4	1189	491x
arrange star 500sp	7.4	42434	5747x

Limitations

1. GEOS comparison is against conda-forge MSVC GEOS (serial per-call, no LTO, no mimalloc) - a static LLVM-built GEOS would improve the GEOS side of every table.
2. Validator strictness is deliberate: eps-class predicates (1e-12 x scale tolerance, fast-FP-first with exact escalation). On the 1.58M real-world dataset it agrees with GEOS 0/0 (winding-agnostic); it diverges on borderline inputs (213 baselined XML cases, mostly the OGC winding contract - docs/BENCHMARKS.md). Repair ships only validator-clean geometry and degrades to an empty GeometryCollection otherwise.
3. W/12 pool-saturation floor: the parallel batch fills all 12 workers with giants; nested intra-poly rayon finds no idle threads in-batch (Amdahl-bound; standalone the parallel check measures 96 ms vs 53 ms in-batch).
4. Giants (>4,096 ring edges) route to the boolean pipeline - on a 200k-edge giant single-pass noding costs 168 ms vs 36 ms for the boolean path (SP_MAX_EDGES in src/core/mod.rs).
5. Mass-overlap repairs are the slowest synthetic class (~1.5 ms for a 100-shell overlap) but still beat GEOS makeValid by 10x to 61x on the overlap rows and 180x to 398x on the dense grid.
6. Python bindings: tests/test_python.py covers the WKB and WKT surface (41 tests); GeoJSON bindings were removed deliberately.
7. simd-portable is nightly-only (3 E0554 on stable, expected); hand-written AVX2 beyond the bbox scan measured slower than auto-vectorized scalar (point_in_ring 8.4x, is_ring_ccw 2.8x), and -C target-cpu=native regresses the full pass ~25%.
8. proj requires native PROJ and is mutually exclusive with io-gpkg (sqlite3 link conflict). io-gpkg is opt-in, so a proj build enables proj and leaves io-gpkg off.
9. WKT reads a dimension modifier, it does not infer one from the ordinate count: POINT Z (1 2 3) parses, POINT (1 2 3) stays an error, where GEOS infers Z from the third ordinate.

Integration with the geo ecosystem

geo-repair plugs into georust two ways: a geo-traits interop module that validates or repairs any GeometryTrait source (geo, geoarrow, geozero, wkb) in one call, and a GeoRepairValidation adapter that exposes our validator through geo's own Validation trait. Details and examples: docs/INTEROP.md.

Embedding surface

Three pieces exist for a host application: a batch entry that builds with or without the thread pool, an opt-in output policy, and error classes.

• batch::{fix_polygon_batch, fix_polygon_batch_owned, fix_polygon_batch_chunked} run on rayon when parallel is enabled, and on a serial loop when it is not.
• output::{OutputPolicy, apply_policy, make_valid_with_policy} reduce a result to one homogeneous geometry for a host that has no collection type. OutputReport counts the parts kept and dropped.
• GeometryValidationError::code() gives a stable key, label() a short name, and class() one of six ValidationClass values. ValidationResult::class_counts() totals them per class.

The default build carries no thread pool and no bundled SQLite. Enable parallel and io-gpkg, or the full set, when the host wants them.

Python bindings

The geo-repair PyPI package exposes the full validation and repair surface over WKB bytes and WKT text, single geometry and batch (including a rayon-backed parallel batch). Install with pip install geo-repair; abi3 wheels (cp38-abi3) serve Python 3.8+ and typing stubs ship in the wheel.

Every function exists for WKB and WKT (repair_*, repair_*_batch, par_repair_*_batch, repair_validate_*, is_valid_*, validate_*, validate_and_fix_*, plus batch forms), with method (auto/arrange/structure) and keep_collapsed parameters. A QGIS Processing script (qgis/qgis_geo_repair.py) streams features through the WKB batch API. Mind that due to overhead it might suffer from performance issues. Examples and full API reference: docs/BINDINGS.md.

C API

The ffi feature exposes a panic-safe C API over WKB and WKT, single geometries and parallel batches. Build with cargo build --release --features ffi,parallel; the output is target/release/geo_repair.{dll,so,dylib} plus libgeo_repair.a and the include/geo_repair.h header.

Every result carries a GeoRepairErrorCode (None/Parse/InvalidInput/ InvalidGeometry/Encode/Panic); batches report per-item outcomes without failing as a whole. All results must be freed with the matching geo_repair_free_* (double-free safe). The ABI (struct layouts, error codes) is fixed from 0.14.2; panic containment requires the release profile's panic = "unwind". Prebuilt libraries for Windows, Linux, and macOS are attached to every GitHub release. Full API reference: docs/BINDINGS.md.

Features

Feature	Description	Default
arrange	CDT-based polygon repair (requires spade)	yes
structure	Structure-based fast path repair	yes
parallel	Rayon parallel processing (non-WASM); the batch entries fall back to a serial loop without it	no
simd	No-op on stable (kernels compiled unconditionally, LLVM auto-vectorized; hand-written AVX2 measured slower, kept only for the nightly simd-portable path)	no
validate	OGC validation predicates	yes
rstar	R-tree index used by noding (enabled transitively by arrange/structure/validate)	yes (transitive)
mimalloc	Use mimalloc global allocator	yes
full	The desktop set: default plus parallel and io-gpkg	no
std	Standard library + file I/O. Disable for no_std builds.	yes
simd-portable	Portable SIMD via core::simd (nightly only)	no
memmap	Memory-mapped binary file loading	no
wasm	WASM browser fetch (synchronous XHR)	no
proj	CRS transformation via geo/proj (requires native PROJ)	no
serde	Geometry serde support (geo/serde)	no
ffi	C-compatible FFI bindings	no
geo-traits	Interop surface over geo_traits::GeometryTrait / GeometryCollectionTrait (geo, geoarrow, geozero, wkb sources)	no
python	Python bindings via PyO3	no
io-shp	Shapefile format backend	no
io-wkt	No-op (WKT is built-in, kept for CI compatibility)	no
io-wkb	No-op (WKB is built-in, kept for CI compatibility)	no
io-geojson	No-op (GeoJSON backend removed, kept for CI compatibility)	no
io-csv	CSV format backend	no
io-gml	GML/XML format backend	no
io-gpkg	GeoPackage format backend (gated out on wasm32)	no
io-all	All opt-in backends except gpkg	no
io-all-native	All opt-in backends including gpkg	no
hotpath	Dev-only: hotpath::measure instrumentation on hot paths	no
hotpath-alloc	Dev-only: hotpath instrumentation plus allocation counters	no
bench-geos	GEOS comparison benchmarks (build from source, MSVC, no LTO)	no
bench-geos-system	GEOS comparison benchmarks (link against system GEOS, conda-forge MSVC)	no
bench-gdal-system	GDAL/OGR I/O comparison benchmark (link against system GDAL)	no

License

Apache-2.0