Sweden: wetland collapses between z4 and z3 — the coarse floor absorbing finely-diced mire pieces (R40/R125) #117

New issue

Open

opened 2026-08-13 19:53:07 +00:00 by art · 1 comment

art commented

2026-08-13 19:53:07 +00:00

Owner

From the #116 review (finding 9). Sweden's wetland area (km², 3857) through the coarse cascade:

step	wetland	change
z5	173 392	—
z4	109 008	−37.1%
z3	48 853	−55.2%

nature gains 45 013 km² over the same z4→z3 step (1 877 555 → 1 922 568) — a non-dominant class reclassified into the dominant one, which is what R125's check flags ("flag any non-dominant class whose share drops by an outsized fraction between adjacent zooms").

The mechanism is the documented coarse floor: at z3, min_area_pixels: 16 is ~116 km² of ground at 63°N, and Sweden's 1.39M finely-diced wetland pieces (mean ~0.04 km² real) mostly fall under it — only 103 pieces exceed 100 km². Finland's CORINE wetland is 100× coarser (13 146 pieces) and retains 48% of its z13 area through to z3 vs Sweden's 16.7%, with identical layer parameters. So this is a data-resolution consequence of the absorption floor, not a new bug — but per R40/R125's check it is measured and recorded here rather than left in a PR body.

Options to investigate (if a fix is wanted at all): a per-class area floor for minority classes at the coarse zooms (the R40 remedy), or leaving the largest fens as explicit z3 survivors (they already are — 10 774 wetland rows at z3).

Recorded in docs/country-requirements.md row 10b and the absorption note in conf/common.yaml.

Re-measured 2026-08-15 against the rebuilt cascade (#116's coverage-validity hardening). The figures this issue was filed with — 187 215 → 128 592 → 34 028 km², −31% then −73.5%, nature +139 341 km², 12% retention, 14 404 z3 rows — came from the pre-hardening build. Making the carve a valid coverage improved retention materially: z3 wetland went from 34 028 to 48 853 km². The collapse is real and still worth the R125 record, but it is −55%, not −73.5%.

From the #116 review (finding 9). Sweden's wetland area (km², 3857) through the coarse cascade: | step | wetland | change | |---|---|---| | z5 | 173 392 | — | | z4 | 109 008 | −37.1% | | z3 | 48 853 | −55.2% | `nature` gains 45 013 km² over the same z4→z3 step (1 877 555 → 1 922 568) — a non-dominant class reclassified into the dominant one, which is what R125's check flags ("flag any non-dominant class whose share drops by an outsized fraction between adjacent zooms"). The mechanism is the documented coarse floor: at z3, `min_area_pixels: 16` is ~116 km² of ground at 63°N, and Sweden's 1.39M finely-diced wetland pieces (mean ~0.04 km² real) mostly fall under it — only 103 pieces exceed 100 km². Finland's CORINE wetland is 100× coarser (13 146 pieces) and retains 48% of its z13 area through to z3 vs Sweden's 16.7%, with identical layer parameters. So this is a data-resolution consequence of the absorption floor, not a new bug — but per R40/R125's check it is measured and recorded here rather than left in a PR body. Options to investigate (if a fix is wanted at all): a per-class area floor for minority classes at the coarse zooms (the R40 remedy), or leaving the largest fens as explicit z3 survivors (they already are — 10 774 wetland rows at z3). Recorded in `docs/country-requirements.md` row 10b and the absorption note in `conf/common.yaml`. --- *Re-measured 2026-08-15 against the rebuilt cascade (#116's coverage-validity hardening). The figures this issue was filed with — 187 215 → 128 592 → 34 028 km², −31% then −73.5%, `nature` +139 341 km², 12% retention, 14 404 z3 rows — came from the pre-hardening build. Making the carve a valid coverage improved retention materially: z3 wetland went from 34 028 to 48 853 km². The collapse is real and still worth the R125 record, but it is −55%, not −73.5%.*

art commented

2026-08-15 01:58:05 +00:00

Author

Owner

Updated with the final post-review measurements (the cell-clipped carve absorbs differently): z5 173 392 → z4 109 008 (−37%) → z3 48 853 km² (−55%), 3857; nature absorbs the reclassified area. The phenomenon stands — the coarse floor over finely-diced mires — now with a smaller z4→z3 drop than the first submission (was −73.5%).

Updated with the final post-review measurements (the cell-clipped carve absorbs differently): z5 173 392 → z4 109 008 (−37%) → z3 48 853 km² (−55%), 3857; `nature` absorbs the reclassified area. The phenomenon stands — the coarse floor over finely-diced mires — now with a smaller z4→z3 drop than the first submission (was −73.5%).

art referenced this issue

2026-08-15 01:58:18 +00:00

Closes #14 #116

art referenced this issue from a commit

2026-08-15 01:58:20 +00:00

#14 review: the sankmark carve is a cell-clipped coverage — ST_CoverageInvalidEdges 1 415 → 0 on partition 23105

art referenced this issue

2026-08-15 01:58:44 +00:00

Closes #14 #116