When a raster predicate must compare across CRSs, the raster's footprint appears to be reprojected using its corner vertices only. Reprojection bends straight footprint edges into curves, so the corner-to-corner chords under-cover the true extent — dramatically so for polar rasters, where the chords miss most of the cap.
Repro: a 20x20 EPSG:3413 raster with extent (-1e6, -1e6, 1e6, 1e6) (a 2000 km square on the north pole), queried with EPSG:4326 points that transform to well inside the square:
| point (4326) |
in EPSG:3413 |
truly inside? |
RS_Intersects |
POINT(45 82) |
(867972, 0) |
yes |
false |
POINT(180 85) |
(-383228, 383228) |
yes |
false |
POINT(0 85) |
(383228, -383228) |
yes |
false |
POINT(90 85) |
(383228, 383228) |
yes |
false (also via RS_Contains) |
POINT(0 60) |
(2349829, -2349829) |
no |
false ✓ |
POINT(45 82) sits on the +x axis at 87% of the half-width — nowhere near an edge curve subtlety — yet reads false; the four corners transform to lon/lat points near latitude 77-78, and everything poleward of their chords is lost. Densifying each footprint edge before reprojecting (a handful of interior points per edge) makes the reprojected footprint follow the curve; SedonaDB's integration/spark-parity suite (where this surfaced) shows its densified footprint matching the exact point-in-square truth on every probed case.
Related, observed while probing: RS_Within(polar_raster, pole-spanning 4326 polygon) throws rather than answering — the pole/antimeridian degeneracy in flat lon/lat space.
When a raster predicate must compare across CRSs, the raster's footprint appears to be reprojected using its corner vertices only. Reprojection bends straight footprint edges into curves, so the corner-to-corner chords under-cover the true extent — dramatically so for polar rasters, where the chords miss most of the cap.
Repro: a 20x20 EPSG:3413 raster with extent (-1e6, -1e6, 1e6, 1e6) (a 2000 km square on the north pole), queried with EPSG:4326 points that transform to well inside the square:
POINT(45 82)POINT(180 85)POINT(0 85)POINT(90 85)POINT(0 60)POINT(45 82)sits on the +x axis at 87% of the half-width — nowhere near an edge curve subtlety — yet reads false; the four corners transform to lon/lat points near latitude 77-78, and everything poleward of their chords is lost. Densifying each footprint edge before reprojecting (a handful of interior points per edge) makes the reprojected footprint follow the curve; SedonaDB'sintegration/spark-paritysuite (where this surfaced) shows its densified footprint matching the exact point-in-square truth on every probed case.Related, observed while probing:
RS_Within(polar_raster, pole-spanning 4326 polygon)throws rather than answering — the pole/antimeridian degeneracy in flat lon/lat space.