jiayuasu opened a new issue, #3255:
URL: https://github.com/apache/sedona/issues/3255
## Expected behavior
LineString segments, and polygon-boundary segments when `allTouched = true`,
should be rasterized whenever their intersection with the raster window has
positive length, regardless of whether either original endpoint is inside the
window. Reversing a segment should produce the same band.
Segments that are disjoint from the window, or that only touch it at one
corner point, should burn no cells. A segment clipped to a window edge should
retain the exact side of nearby internal pixel grid lines.
## Actual behavior
`Rasterization.clipSegmentToRasterBounds` returns `null` whenever both
endpoints are outside the raster extent. This drops valid segments that cross
the raster from one side to another.
The current one-endpoint clipping path also stores an interpolated endpoint
as ordinary `double` coordinates. If the unconstrained coordinate rounds onto
an internal grid line, rasterization can select the wrong adjacent cell even
though the segment remains strictly on one side of that line inside the raster.
## Steps to reproduce
On a 6 x 6 north-up unit raster, this horizontal line crosses all six
columns but both endpoints are outside:
```sql
SELECT RS_BandAsArray(
RS_AsRaster(
ST_GeomFromWKT('LINESTRING (-1 2.5, 7 2.5)'),
RS_MakeEmptyRaster(1, 'B', 6, 6, 0.0, 6.0, 1.0, -1.0, 0.0, 0.0, 0),
'B', true, 1.0, 0.0, false),
1);
```
Expected: all six cells in zero-based row 3 are `1`.
Actual: every cell is `0`.
A secondary clipping case demonstrates loss of endpoint sidedness:
```sql
SELECT RS_BandAsArray(
RS_AsRaster(
ST_GeomFromWKT('LINESTRING (-1 3.0000000000000004, 1 3.0)'),
RS_MakeEmptyRaster(1, 'B', 6, 6, 0.0, 6.0, 1.0, -1.0, 0.0, 0.0, 0),
'B', true, 1.0, 0.0, false),
1);
```
Within the raster, the nondegenerate part of this line is strictly above `y
= 3`, except at its final endpoint. It should therefore burn `(row 2, column
0)`. The clipped coordinate rounds to `y = 3`, and Sedona instead burns `(row
3, column 0)`. Reversing the input must preserve the correct result.
## Root cause
The clipper has three related numerical/topological limitations:
1. It treats "both endpoints outside" as equivalent to "no intersection."
2. It returns only rounded coordinates, losing whether an interpolated
coordinate is exactly on, just below, or just above an internal grid line.
3. Once both-outside crossings are admitted, direct differences such as `x2
- x1` can overflow for valid finite endpoints such as `-1e308` and `1e308`.
## Acceptance criteria
- A nondegenerate segment with a positive-length window intersection is
rasterized even when both source endpoints are outside.
- Disjoint and parallel-outside segments burn nothing.
- Corner-only tangencies burn nothing; boundary-coincident positive-length
segments retain the established half-open side.
- Clipped endpoints preserve the side of internal grid lines, including
exact, `Math.nextDown`, and `Math.nextUp` cases on both axes.
- Forward and reversed segments produce identical bands.
- An originally degenerate in-bounds LineString is distinguished from a
nondegenerate segment collapsed to a tangent point by clipping.
- Clipping valid finite endpoints does not produce NaN-driven omissions or
traversal streaks. Include horizontal and vertical `+/-1e308` crossings, or
explicitly validate and document a smaller supported coordinate range.
- Tests use an independent exact-rational segment/window oracle. The oracle
must not exclude intersections near grid lines.
- The replacement adds no arbitrary-precision or per-segment helper
allocations on the ordinary finite-coordinate path; any exact fallback is
limited to unresolved or overflow-prone cases.
## Related work
- #3118 replaced fixed-step sampling with exact cell traversal, but did not
change segment clipping.
- #3120 and #3251 address endpoint-cell selection after coordinates are
already in pixel space; #3251 deliberately leaves clipping unchanged.
## Settings
- Sedona version: PR #3251 head `72d9cc26541` / `2.0.0-SNAPSHOT` (the
both-outside crossing omission also reproduces on current `master`)
- API type: SQL and Java
- JRE: 17
- Environment: local development checkout
I searched open and closed issues and did not find a duplicate.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
