GeoRepair

OGC geometry repair and validation for Rust. Passes the GEOS XML validation suite.

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.

Detects and fixes invalid GIS geometries (self-intersections, unclosed rings, degenerate shapes, NaN coordinates) using algorithms selected by geometry type. Built-in I/O for WKB, WKT, and a custom binary batch format with no extra dependencies.

The Structure strategy (default) mirrors GEOS ST_MakeValid: planar graph extraction, face walking, winding-number assembly. The Arrange strategy is a CDT-based fallback for complex topologies.

Performance on the 1,579,030-polygon production dataset (i5-12400F, release, parallel batch, GEOS 3.14.1 conda-forge as reference):

Dataset	GeoRepair	GEOS	vs GEOS
Validation (1.58M)	0.8-2.3 s	3.6-3.9 s	0.2-0.6x (2-5x faster)
Invalid subset (1 poly, 2026-08-03)	11.9 ms	2.2 ms	5.4x
Full dataset (1.58M polys)	3.5-4.3 s	3.3-3.5 s	1.0-1.2x

Performance

Real-world dataset (1,579,030 polygons)

Structure batch on a production GIS dataset, read from the original GeoPackage (its geometry blobs are WKB; 813 features flattened to 1,579,030 polygon parts). i5-12400F (6C/12T), release profile (LTO), mimalloc (default feature), Rayon 12-thread batch on both sides. GEOS is conda-forge libgeos 3.14.1, MSVC, serial per-call, no LTO; "par batch" means many GEOS C calls run concurrently via Rayon. GEOS geometries are built via CoordSeq direct construction, no WKT round-trip; the one-time pre-build of 1.58M GEOS geometries (1.4-1.5 s) is excluded from the timings.

Dataset	GeoRepair (par)	GEOS (par batch)	vs GEOS
Validation (1.58M)	0.8-2.3 s (0.5-1.5 µs/poly)	3.6-3.9 s (2.3-2.5 µs/poly)	0.2-0.6x
Invalid subset (1 poly, 2026-08-03)	11.9 ms	2.2 ms	5.4x
Full dataset (1.58M polys)	3.5-4.3 s	3.3-3.5 s	1.0-1.2x

Two settled runs per source; the bands cover the run-to-run spread. The GeoRepair column is measured in the same process as the GEOS column (co-residency inflates it 5-15% vs a standalone run, where the full pass is ~4.1-4.2 s).

Synthetic benchmarks

Structure strategy, same machine/profile, CoordSeq direct GEOS construction. Both columns measured in the same process; the GeoRepair column is the Rayon 12-thread batch (a standalone run measures 10-30% lower). Ratio is GEOS / GeoRepair, i.e. how many times GeoRepair is faster. Stars: *** >= 100x, ** 10-100x, * 1-10x, blank means GEOS is faster.

Polygons:

Benchmark	GeoRepair (µs)	GEOS (µs)	Ratio	★
valid polygon 4v	0.055	0.263	4.78x	*
valid polygon 10v	0.168	0.324	1.93x	*
valid polygon 50v	0.413	0.443	1.07x	*
valid polygon 100v	0.624	0.576	0.92x
valid polygon 500v	2.362	1.581	0.67x
valid polygon 1000v	5.204	2.556	0.49x
valid polygon 5000v	38.76	11.27	0.29x
valid polygon 10000v	57.32	19.04	0.33x
invalid bowtie 4v	0.708	22.70	32.1x	**
invalid star 100v	9.035	8.155	0.90x
self-touch poly	0.869	19.78	22.8x	**
collapsed poly	0.495	15.86	32.0x	**
near-collinear	1.247	18.52	14.8x	**
large coord 1e12	0.081	0.260	3.2x	*

LineStrings & MultiLineStrings:

Benchmark	GeoRepair (µs)	GEOS (µs)	Ratio	★
valid line	0.011	0.053	4.8x	*
zero-length line	0.008	0.412	51.5x	**
valid ls 4v	0.032	0.111	3.5x	*
valid ls 10v	0.078	0.088	1.1x	*
valid ls 50v	0.107	0.082	0.8x
valid ls 100v	0.235	0.125	0.5x
valid ls 500v	0.760	0.588	0.8x
collinear ls 4v	0.027	0.042	1.6x	*
collinear ls 10v	0.057	0.052	0.9x
collinear ls 50v	0.121	0.087	0.7x
collinear ls 100v	0.208	0.133	0.6x
collinear ls 500v	0.607	0.604	1.0x	*
zigzag ls 10v	0.041	0.046	1.1x	*
zigzag ls 50v	0.110	0.092	0.8x
zigzag ls 100v	0.248	0.123	0.5x
zigzag ls 500v	0.576	0.473	0.8x
spiral ls 10v	0.044	0.075	1.7x	*
spiral ls 50v	0.123	0.099	0.8x
spiral ls 100v	0.230	0.142	0.6x
self-int ls 5v	0.031	0.232	7.5x	*
dense self ls 10v	0.043	0.203	4.7x	*
dense self ls 50v	0.117	0.158	1.4x	*
dense self ls 100v	0.239	0.155	0.6x
duped ls 100v	0.492	0.356	0.7x
mls 50x3v	1.671	0.468	0.3x
mls 50x4v	1.610	0.580	0.4x

Special shapes:

Benchmark	GeoRepair (µs)	GEOS (µs)	Ratio	★
star-burst 10sp	0.058	0.126	2.2x	*
star-burst 50sp	0.196	0.169	0.9x
star-burst 100sp	0.258	0.192	0.7x
star-burst 500sp	1.063	0.613	0.6x
collinear ov 10seg	0.073	0.173	2.4x	*
collinear ov 50seg	0.225	0.242	1.1x	*
collinear ov 100seg	0.440	0.302	0.7x
collinear ov 500seg	1.556	1.008	0.6x
x-scale 10v	0.035	0.075	2.1x	*
x-scale 50v	0.123	0.112	0.9x
x-scale 100v	0.375	0.219	0.6x
ringing 100v	0.199	0.095	0.5x
ringing 500v	1.014	0.376	0.4x
hilbert 256v	0.272	0.370	1.4x	*
hilbert 1024v	1.210	1.281	1.1x	*
lissajous 200v	0.295	0.295	1.0x	*
lissajous 500v	0.602	0.484	0.8x
lissajous 1000v	2.313	2.355	1.0x	*
lissajous 7:4 500v	0.470	0.660	1.4x	*
spoke 10sp	0.056	0.086	1.5x	*
spoke 50sp	0.147	0.127	0.9x
spoke 100sp	0.222	0.171	0.8x
spoke 500sp	1.647	0.980	0.6x
star-comb 20sp	0.063	0.095	1.5x	*
star-comb 100sp	0.121	0.174	1.4x	*
star-comb 500sp	0.548	0.683	1.2x	*

Holes, overlaps & grids:

Benchmark	GeoRepair (µs)	GEOS (µs)	Ratio	★
hole hier 5h	1.016	1.143	1.1x	*
hole hier 20h	5.994	8.025	1.3x	*
hole hier 50h	47.57	31.20	0.66x
overlap mp 5sh	2.128	29.90	14.0x	**
overlap mp 20sh	5.411	454.3	84.0x	**
overlap mp 50sh	31.17	7330.2	235x	***
dense grid 5x5=25	6.870	1785.7	260x	***
dense grid 10x10=100	44.08	14367	326x	***
dense grid 20x20=400	757.3	103750.3	137x	***
sliver 100v	1.765	7.294	4.1x	*
sliver 500v	14.93	18.88	1.26x	*

Arrange pipeline (CDT fallback):

Benchmark	GeoRepair (µs)	GEOS (µs)	Ratio	★
arrange valid 4v	0.046	0.263	5.7x	*
arrange valid 10v	0.135	0.324	2.4x	*
arrange valid 50v	0.433	0.443	1.0x	*
arrange bowtie 4v	0.928	22.70	24.5x	**
arrange star 10sp	0.068	0.126	1.9x	*
arrange star 50sp	0.244	0.169	0.7x

Pattern: GeoRepair wins big on invalid repair and MultiPolygon unification (GEOS hits quadratic worst cases: dense grids 137-326x, overlaps 14-235x, bowtie/collapsed/self-touch 15-32x). GEOS wins on large valid polygons (0.3-0.7x on 500-10000 verts) where mature C++ optimization dominates. Sub-µs rows fluctuate with Rayon dispatch overhead and should not be read as precise.

Run benchmarks

# Real-world dataset benchmark with GEOS comparison (system GEOS)
GEOS_LIB_DIR='D:\Miniconda\Library\lib' GEOS_INCLUDE_DIR='D:\Miniconda\Library\include' \
GEOS_VERSION=3.14.1 cargo bench --features bench-geos-system,arrange,structure,parallel,simd,io-shp --bench real_world

# Real-world dataset, no GEOS (load + validation + repair + output gate)
BENCH_FILE=benches/real_world/data_0.bin ./target/release/deps/real_world-*.exe --fast

# Synthetic benchmarks with GEOS comparison
GEOS_LIB_DIR='D:\Miniconda\Library\lib' GEOS_INCLUDE_DIR='D:\Miniconda\Library\include' \
GEOS_VERSION=3.14.1 cargo bench --features bench-geos-system,arrange,structure,parallel,simd --bench bench

# Synthetic benchmarks, serial + parallel columns only (no GEOS)
cargo bench --features arrange,structure,parallel,simd --bench bench

Full GEOS setup lives in benches/AGENTS.md. Measurement rules that matter: always take the settled second run (first-run-after-build is inflated ~18% by Windows Defender + cold LTO code); never trust a bench binary you cannot trace to a source file; examples/sp_diag.rs is the invalid-subset diagnostic (acceptance counts + per-stage timing on the biggest giant).

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 gate (deliberate policy, real-world impact ~1 poly). Our validator runs Shewchuk exact predicates (agreeing with GEOS on the OGC definition, 934/934 GEOS XML suite pass) plus one relative noise gate: edges whose exact orientation is nonzero but within ~32 ulps of the pair's own length scale are treated as coincident (see src/validation/mod.rs). Measured on the 1.58M real-world dataset (2026-08-03): 1,579,029 parts are winding-only (CW, GEOS-valid), and exactly 1 part carries a non-winding defect. Earlier counts (2,298 via arrange::validate_polygon, 1,855 via the full validator) were inflated by a real bug: the product-form proper-crossing test o1 * o2 < 0.0 treated a -0.0 orientation (an exact collinear touch, common on snapped real-world vertices) as a crossing, flagging GEOS-valid geometry. Fixed 2026-08-03 with a zero-safe strict opposite-sign predicate in edges_intersect_general and in the sweep (segments_properly_cross, has_no_intersections), which also closed the --fast bench gate artifact: arrange::validate_polygon previously flagged all 2,298 real-world Structure outputs while the full validator and GEOS accepted them; it now agrees with GEOS (the zero-orient fix plus an inclusive hole-containment check, GEOS-aligned for boundary-touching holes). Repair contract is unchanged: a repair ships only validator-clean geometry and degrades to an empty GeometryCollection otherwise. Earlier 2026-08-03 differential-fuzz fixes still in force: exact-collinear micro-edge overlaps below the length gate, -0.0 pinches, closing-edge backtracking (the pair (0, n-1) was skipped outright), segment-local vertex-on-edge tolerance (a pair-max eps inflated past micro segments in mixed-magnitude rings).
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. Standalone, the parallel check measures 96 → 53 ms; in the batch the per-giant serial chain (~370 ms for the biggest) is Amdahl-bound.
4. Giants (>4,096 ring edges) route to the boolean pipeline. Single-pass noding costs 168 ms on a 200k-edge giant vs 36 ms for the boolean path; SP_MAX_EDGES = 4096 in src/core/mod.rs is the gate.
5. Mass-overlap repairs are the slowest synthetic class (dense grid 20x20: ~0.76 ms/poly parallel) but remain 100x+ faster than GEOS on the same shapes.
6. Python bindings: tests/test_python.py covers the WKT surface (18 tests). GeoJSON bindings were removed by design; the test file now imports only repair_wkt / repair_wkt_batch / is_valid_wkt / validate_wkt / validate_wkt_batch / validate_and_fix_wkt.
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 is intentionally absent; -C target-cpu=native regresses the full pass ~25%. The one AVX2 kernel kept is the bbox scan, runtime-dispatched via is_x86_feature_detected!, bit-exact vs scalar.
8. proj requires native PROJ; io-gpkg + proj are mutually exclusive (sqlite3 link conflict).
9. GEOS XML suite coverage: 934/934 dispatched cases pass; 1,565 overlay/relate cases are skipped (overlay operations are out of scope); 209 masked divergences (documented tolerance gates, e.g. even-odd area 1e-6 for island-in-hole). The suite's WKT/WKB readers are our own: external wkt/wkb crates are disallowed because they measure slower than the built-in readers.
10. Sub-µs synthetic rows are noise (Rayon dispatch overhead). The trustworthy metrics are the real-world batch numbers and the larger synthetic rows.

Integration with the geo ecosystem

geo-repair is built on geo types and plugs into the georust ecosystem two ways:

• geo-traits sources (geo-traits feature). The interop module runs validation and repair over geo_traits::GeometryTrait / geo_traits::GeometryCollectionTrait, the trait layer implemented by geo, geoarrow, geozero, and wkb. Any such source can be validated or repaired in one call (interop::is_valid_geometry, interop::make_valid_geometry, interop::make_valid_geometries, ...) without materializing geo types; results come back as geo::Geometry<f64>.

• geo's Validation trait (always available). GeoRepairValidation(&geometry) wraps any &geo::Geometry<f64> and exposes geo_repair's validator through geo's Validation trait (.is_valid(), .check_validation(), .validation_errors()) with geo's Invalid* error taxonomy. The orphan rule prevents implementing geo's trait for geo's own types, so the adapter is the bridge. Mapping is best-effort: geo_repair's stricter gates (32-ulp collinear, T-junction) surface through it, and classes geo does not model (ring closure, orientation, duplicates, ...) are omitted from the geo view.

use geo::algorithm::validation::Validation;
use geo_repair::GeoRepairValidation;

let adapter = GeoRepairValidation(&geometry);
assert!(!adapter.is_valid());

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)	yes
simd	Runtime-dispatched AVX2 bbox kernel (the one SIMD kernel that beat auto-vectorized scalar, 4.5x); all other kernels auto-vectorized scalar	yes
validate	OGC validation predicates	yes
mimalloc	Use mimalloc global allocator	yes
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 (placeholder)	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-csv	CSV format backend	no
io-gml	GML/XML format backend	no
io-gpkg	GeoPackage format backend (not WASM)	no
io-all	All opt-in backends except gpkg	no
io-all-native	All opt-in backends including gpkg	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

License

Apache-2.0