Water polygons cross national borders and two partitions draw the same lake — 1 438 km² doubled at the SE/FI border #197

New issue

Open

opened 2026-08-20 19:58:24 +00:00 by art-bot · 0 comments

art-bot commented

2026-08-20 19:58:24 +00:00

Collaborator

Measured

Area of each country's map_water_bodies_14 partition falling outside that
country's own map_country_mask_0, over polygons larger than 5 km²
(EPSG:3857 areas, the unit these tables are in):

country	polygons outside	km² outside	straddling (part in, part out)	wholly outside
pl	3	3 934	1	2
fi	74	857	27	47
se	44	613	0	44
gr	3	470	3	0
is	0	0	0	0

Iceland is the control: no land neighbour, nothing outside. Poland's 3 934 km²
is almost entirely Morze Bałtyckie, the Baltic slab mislabelled as a lake —
that is #43 and not this issue.

The two remaining classes are different failures and want different fixes:

Straddling — one real feature shared with a neighbour. Greece: ΛΙΜΝΗ ΜΕΓΑΛΗ
ΠΡΕΣΠΑ (493 km², 423 outside — Albania and North Macedonia), ΛΙΜΝΗ ΔΟΪΡΑΝΗ (69,
39 outside), ΛΙΜΝΗ ΜΙΚΡΗ ΠΡΕΣΠΑ (83, 8). The WFD publishes a lake water body as
one polygon regardless of which state it lies in, so the whole lake arrives.
Verified against gr_perifereies (the administrative extent, inland water
included), not just the mask; interior lakes come back 0 km² outside, which is
the control for the measurement. Finland's 27 straddlers (116 km²) sit at
68.7°N / 25.2°E — the Torne–Muonio–Könkämä border chain.

Wholly outside — a polygon entirely on the other side of its own mask.
Sweden 44 (613 km²) at 67.4°N / 23.1°E, Finland 47 (736 km²) at 67.0°N /
24.9°E: the same corridor, mirrored. This is not shared water, it is each
country's source and each country's mask disagreeing about where the border
runs, or coastal/archipelago water that the mask does not cover.

It is already visible, and it is not additive

In the SE/FI border box (19–27°E, 65–70°N), 745 polygon pairs from the two
partitions overlap, 1 438 km² of doubly-drawn water. Decoded tile
zxy_map(11, 1142, 475) over Kilpisjärvi: water_bodies features=49 — 38 from
se, 10 from fi, of which 24 pairs overlap inside that one tile, 94 km².
The largest overlaps are at 68.5–69.0°N / 20.7–22.0°E, the Treriksröset corner.

Water is one colour, so two stacked polygons look like one — which is exactly
why this has survived. What it costs: every such feature is paid for twice in
tile bytes and in the generalization cascade, and the moment a shared feature is
named it is labelled twice by two partitions that cannot see each other.
Greece's lakes are all named; Albania is simply not imported yet, so the label
collision is scheduled rather than absent.

Why it is a shared-table question, not a per-country one

There is one zxy_map_water_bodies reading every country's parent tables, so
"whose lake is this" has no answer inside a single country config — which is the
same shape as the problem docs/unification.md Workstream B exists for, and the
water form of the question the world carve asks for land (#186: at a coast we
must carve, at a land border we must not, and the data cannot tell them apart).

Options, none of them free

Clip every layer to its own country mask. Correct ownership, and it puts a
straight edge through open water where no edge exists — Prespa would end
mid-lake at the border. Also needs a mask good enough to clip against, which
is the #186 problem again.
Keep whole features, dedupe at the tile query. Nothing to invent
geometrically, but it needs a rule for which partition wins, and the tile
function currently unions partitions without looking at them.
Declare an owner per shared feature at import time (the country whose
intersection is largest, say), and let the others drop it. Cheap to compute,
survives a neighbour being added later, and is a real decision recorded in the
data rather than a rendering trick.
Do nothing deliberately, i.e. write down that shared water is drawn by
every country that has it. Honest, and it makes the label duplication a known
cost rather than a surprise.

Whatever the answer, it wants an acceptance test that fails when a country's
partition carries area outside its own mask beyond a documented allowance —
today that test would go red for four of the five countries, which is the point.

Not in scope

#43 (Poland's Baltic slab as a lake) — a classification bug that this
measurement happens to rank first.
The cover, roads, railways and waterways layers were not measured. The
same question applies to every one of them; a border river's centreline is the
obvious next case.

Measured 2026-08-20 against the live database, after !194 and !195 merged
(main ebdaa34).

Found while reviewing !195 (Greece's water_bodies), but it is not a Greece problem and not a water_bodies problem — it is what happens when a feature that crosses a national border meets a table partitioned by country. Greece is only where it was visible, because Prespa is 86% not Greek. **Sweden and Finland are already drawing 1 438 km² of the same water twice, today, on the live map.** ## Measured Area of each country's `map_water_bodies_14` partition falling outside that country's own `map_country_mask_0`, over polygons larger than 5 km² (EPSG:3857 areas, the unit these tables are in): | country | polygons outside | km² outside | straddling (part in, part out) | wholly outside | |---|---|---|---|---| | pl | 3 | 3 934 | 1 | 2 | | fi | 74 | 857 | **27** | 47 | | se | 44 | 613 | **0** | 44 | | gr | 3 | 470 | **3** | 0 | | is | 0 | 0 | 0 | 0 | Iceland is the control: no land neighbour, nothing outside. Poland's 3 934 km² is almost entirely `Morze Bałtyckie`, the Baltic slab mislabelled as a lake — that is #43 and **not** this issue. The two remaining classes are different failures and want different fixes: **Straddling — one real feature shared with a neighbour.** Greece: ΛΙΜΝΗ ΜΕΓΑΛΗ ΠΡΕΣΠΑ (493 km², 423 outside — Albania and North Macedonia), ΛΙΜΝΗ ΔΟΪΡΑΝΗ (69, 39 outside), ΛΙΜΝΗ ΜΙΚΡΗ ΠΡΕΣΠΑ (83, 8). The WFD publishes a lake water body as one polygon regardless of which state it lies in, so the whole lake arrives. Verified against `gr_perifereies` (the administrative extent, inland water included), not just the mask; interior lakes come back 0 km² outside, which is the control for the measurement. Finland's 27 straddlers (116 km²) sit at 68.7°N / 25.2°E — the Torne–Muonio–Könkämä border chain. **Wholly outside — a polygon entirely on the other side of its own mask.** Sweden 44 (613 km²) at 67.4°N / 23.1°E, Finland 47 (736 km²) at 67.0°N / 24.9°E: the same corridor, mirrored. This is not shared water, it is each country's source and each country's mask disagreeing about where the border runs, or coastal/archipelago water that the mask does not cover. ## It is already visible, and it is not additive In the SE/FI border box (19–27°E, 65–70°N), **745 polygon pairs from the two partitions overlap, 1 438 km² of doubly-drawn water.** Decoded tile `zxy_map(11, 1142, 475)` over Kilpisjärvi: `water_bodies features=49` — 38 from `se`, 10 from `fi`, of which **24 pairs overlap inside that one tile, 94 km²**. The largest overlaps are at 68.5–69.0°N / 20.7–22.0°E, the Treriksröset corner. Water is one colour, so two stacked polygons look like one — which is exactly why this has survived. What it costs: every such feature is paid for twice in tile bytes and in the generalization cascade, and the moment a shared feature is *named* it is labelled twice by two partitions that cannot see each other. Greece's lakes are all named; Albania is simply not imported yet, so the label collision is scheduled rather than absent. ## Why it is a shared-table question, not a per-country one There is one `zxy_map_water_bodies` reading every country's parent tables, so "whose lake is this" has no answer inside a single country config — which is the same shape as the problem `docs/unification.md` Workstream B exists for, and the water form of the question the world carve asks for land (#186: at a coast we must carve, at a land border we must not, and the data cannot tell them apart). ## Options, none of them free - **Clip every layer to its own country mask.** Correct ownership, and it puts a straight edge through open water where no edge exists — Prespa would end mid-lake at the border. Also needs a mask good enough to clip against, which is the #186 problem again. - **Keep whole features, dedupe at the tile query.** Nothing to invent geometrically, but it needs a rule for which partition wins, and the tile function currently unions partitions without looking at them. - **Declare an owner per shared feature at import time** (the country whose intersection is largest, say), and let the others drop it. Cheap to compute, survives a neighbour being added later, and is a real decision recorded in the data rather than a rendering trick. - **Do nothing deliberately**, i.e. write down that shared water is drawn by every country that has it. Honest, and it makes the label duplication a known cost rather than a surprise. Whatever the answer, it wants an acceptance test that fails when a country's partition carries area outside its own mask beyond a documented allowance — today that test would go red for four of the five countries, which is the point. ## Not in scope - #43 (Poland's Baltic slab as a lake) — a classification bug that this measurement happens to rank first. - The `cover`, `roads`, `railways` and `waterways` layers were not measured. The same question applies to every one of them; a border river's centreline is the obvious next case. Measured 2026-08-20 against the live database, after !194 and !195 merged (main `ebdaa34`).