Conformance

terrio's results compared with GDAL's (and laspy's and PDAL's for point clouds) on real-world files: the files of scripts/fetch_test_data.sh (Natural Earth, USGS 3DEP, Sentinel-2, NASA ECOSTRESS and OPERA, OpenStreetMap and more, with copies GDAL derives in other formats), and synthetic rasters for the driver round trips. Each case is identical to the reference, within tolerance (the suite says which), or differs. Regenerate this page with:

cargo build --release
python scripts/conformance.py ../terrio-data --output docs/conformance.md

Generated on 2026-10-14 with terrio 0.1.1 (9e7f7f5), GDAL 3.12.2, PROJ 9.7.1, PDAL 2.9.3, laspy 2.7.0; the coordinate transformation suites on 2026-10-21 with terrio 0.3.2, GDAL 3.12.2 and PROJ 9.8.1, with 50 reference points of each kind per EPSG CRS.

Summary

Dataset summaries and pixels

Every raster and vector file of the test data: size, type, CRS, geotransform, nodata, band statistics, layers, fields, feature counts and extents as GDAL reports them (numbers within 1e-6 relative, statistics within 1e-4), and every raster pixel exactly.

Vector conversions

A file converted by terrio convert into each format, then read by GDAL: feature counts, fields, CRS, extent and a sample of geometries and attribute values must match what GDAL reads from the original, but for the differences GDAL's reading of ogr2ogr's own conversion into that format has too (KML naming the name field Name, shapefiles truncating field names, MapInfo rounding coordinates to its grid, CSV files without a CRS).

Raster drivers

Synthetic rasters of every data type a format holds (several bands, NaN and finite nodata, geographic, projected and user-defined CRSs). Read: GDAL writes, terrio reads; write: terrio writes, GDAL reads. Pixels, type, CRS and nodata must match exactly, geotransforms within 1e-9.

Windowed and overview reads

Windows of every GeoTIFF of the test data, at full resolution and from each overview, read in place by terrio convert --srcwin --overview: pixels identical, geotransforms within 1e-9. Remote: the same on public Cloud Optimized GeoTIFFs over HTTP.

Raster operations

The outputs of the same command lines on real files: size, type, CRS and geotransform must match, then pixels are compared. Identical: every pixel equal. Within tolerance: every pixel within the case's tolerance (1e-2 for interpolating resamplings).

Cases that are not identical:

Terrain analysis and raster algebra

The outputs of the same command lines on real DEMs and land cover: size, type, nodata and geotransform must match, then pixels are compared. Within tolerance: every pixel within the case's tolerance (float32 rounding of NumPy's SIMD functions, for example).

Cases that are not identical:

Vector operations

Conversions with ogr2ogr's options (where, spat, select, sql, nlt...) compared feature by feature; SQL expressions, DISTINCT, summaries and attribute joins against OGRSQL compare values, nulls, logical field types, geometry WKB and CRS, including duplicate and null join keys. Integer widths normalize to Terrio's Integer model. Rasterized pixels compare exactly; polygons compare by count, holes and validity, and by rasterizing them back.

Arrow and GeoArrow

Every layer of the test data exported with Layer.to_arrow() (WKB and native GeoArrow) against GDAL's Arrow stream of the same layer, through pyarrow and GeoPandas, and imported back with terrio.from_arrow() (and by GDAL's WriteArrow): rows, geometries, attribute values and CRS. Within tolerance: geometries equal up to vertex order.

Cases that are not identical:

GeoPackage updates and spatial filters

--append, --overwrite and added layers compared with ogr2ogr (features, extents, counts, sequences, R-trees, integrity_check, the GeoPackage validator, rasters kept), and --spat through the R-tree compared with ogr2ogr -spat.

Point clouds

Every dimension of every point equal to laspy's; headers equal to laspy's and PDAL's; written files read back identically by laspy and PDAL; PLY, E57 and EPT points equal to PDAL's; COPC and EPT queries returning the same points as PDAL's; waveform samples equal to the file's packets; filters giving the same classes or points as PDAL's; grids equal to PDAL's writers.gdal. Reprojections are within 2e-7 degrees or 2 mm of PDAL's; TIN grids differ from gdal_grid's where points share an x and y.

Cases that are not identical:

Coordinate transformations

50 reference points per EPSG geographic and projected CRS (the centre of its area of use and random points of it), transformed from WGS 84 (forward), 50 more back (inverse), and 50 between the CRS and another whose area holds them, compared with GDAL's results, neither using datum shift grids, a case per CRS (per pair of CRSs from CRS to CRS) judged by its worst point; then projections EPSG has no code for, against PROJ (crates/terrio/tests/geostationary.rs: points from nadir to near the limb, forward within 1 mm and inverse within 1e-8°, and points the satellite cannot see refused; crates/terrio-crs/tests/ellipsoidal_projections.rs: points over the globe on WGS 84 and on a sphere, forward within 1 mm and inverse within 1e-8°, and for the gnomonic points beyond the horizon refused; crates/terrio-crs/tests/proj_strings.rs: every PROJ projection with typical parameters, on WGS 84 and on a sphere, as a PROJ string and as pyproj's WKT1_GDAL and WKT2_2019, 20 points each, matching PROJ or refused). Identical: within 1 mm. Within tolerance: within 1 cm (forward, inverse) or 1 m (CRS to CRS, where the datum transformations GDAL picks can differ by area of use). Known differences: deprecated CRSs whose own datum PROJ uses, where terrio uses their replacement's (#153).

Cases that are not identical:

Datum shift grids, geoid models and coordinate epochs

The EPSG reference points again, GDAL and terrio both using the grids of scripts/proj_grids.txt (crates/terrio-crs/tests/epsg_accuracy_grids.rs); then random points through every horizontal shift grid (NTv2 and GeoTIFF, forward and inverse), geoid models (GeoTIFF and GTX), NADCON5 and geocentric translation grids, compound CRSs and time-dependent transformations with coordinate epochs (scripts/compare_crs_grids_with_gdal.py), with the largest difference of each case. Identical: within 1 mm. Within tolerance: within 1 cm, or 1 m from CRS to CRS. Known differences: deprecated CRSs whose own datum PROJ uses, where terrio uses their replacement's (#153).

Cases that are not identical: