[Buildings] ~15,000 Microsoft ML footprints on farm dams in Western Australia

Start by reading the existing water exclusion filter referenced in the issue and the prior work in #199; the issue proposes adding a farm dam source to that filter, with a 50% overlap threshold. Review #458 and #478 for the data source work and confirm how the DPIRD polygons are supplied. Done means the false building footprints on mapped farm dams are excluded across the dataset without removing real buildings.

Release

2026-09-23.1

Theme

buildings

Feature type

building

GERS ID

9cb1e315-5385-4b24-812f-84516b589efe, 71a11bb5-466c-438e-afc8-f73a4cfee4d9, 7fcac848-2aa7-43e8-9ae2-9ccb9f99432b, 93aca956-4cb4-4045-9b54-4bd3ee2949fc (examples; all 15,109 are in the attachment)

Location

Western Australia's agricultural zone. Most of the hits fall in bbox 115.1, -34.8, 122.1, -29.5 (lon/lat), with a scattering across the rest of the state. Example: https://explore.overturemaps.org/#18/-33.46269/117.93659

What kind of problem is it?

Feature doesn't exist in the real world

What's wrong

About 15,000 Microsoft ML building footprints in Western Australia are farm dams that the model mistook for buildings. They trace dam water and dam banks.

• 15,109 Microsoft ML footprints lie at least 50% inside a mapped farm dam, 2.3% of the 646,522 Microsoft footprints in the areas checked. Only 173 OpenStreetMap buildings do (0.09%); a sample of those were real structures under a dam polygon that is wrong or out of date.
• In a manual imagery review of 100 random Microsoft hits, 99 are not buildings (99%, 95% CI 94.6–99.8%).
• They sit on 14,129 distinct dams, 8.1% of all mapped farm dams in the state.

#199 established that ML buildings crossing water are already excluded during conflation. These most likely get through because farm dams are almost entirely missing from the water data: only 289 of the 15,282 hits (1.9%) have their centre inside any polygon in base/water.

Footprints (red) inside DPIRD farm dam polygons (cyan), over Esri World Imagery:

Example	GERS ID	Lat, lon	Area	What it is
	9cb1e315-5385-4b24-812f-84516b589efe	−33.46269, 117.93659	1,021 m²	Shallow dam water
	71a11bb5-466c-438e-afc8-f73a4cfee4d9	−33.73576, 118.71789	601 m²	Dam water surface
	7fcac848-2aa7-43e8-9ae2-9ccb9f99432b	−33.36167, 120.10574	225 m²	Half on dam water, half on the bank
	93aca956-4cb4-4045-9b54-4bd3ee2949fc	−35.08762, 117.92193	166 m²	Dark dam water

The footprints' own attributes don't separate them. Their sources entry has confidence: null. Missing height is the strongest signal, but 94% of footprints with no height are not on dams, so dropping them would remove about 15 other buildings for every false one. Area doesn't separate them either: median 192 m² against 198 m².

What you expected

These footprints should be excluded, like ML buildings that cross mapped water. The fix that scales is to add a farm-dam source to the existing water exclusion filter, so it works across the dataset rather than per feature (#458, #478). Two open sources cover this:

• Farm dams of Western Australia (DPIRD). 175,143 dam polygons for WA, CC BY 4.0. This is the layer that found these errors. Its metadata estimates about 1,800 features in the South West are not dams, so a ≥50% overlap threshold is sensible. I'm proposing it as a data source separately.

• DEA Waterbodies (Geoscience Australia). National, but built from 30 m Landsat pixels, so it misses most farm dams: only 215 of the 15,109 dam hits (1.4%) fall inside a DEA polygon. Run on its own across WA it flags a largely different set of 2,245 buildings, and a 100-hit review found 76 not buildings and 24 real. Waterbody size decides it:

  DEA waterbody size	Not a building	Real
  Over 100 ha	51	2
  10–100 ha	9	5
  1–10 ha	13	16
  Under 1 ha	3	1

  The large ones are mostly inland salt lakes, with footprints on bare lake beds. The small ones include polygons over Perth suburbs and industrial sites that cover real houses. If DEA is used, restrict it to waterbodies over about 100 ha.

Evidence

All 175,143 DPIRD farm dam polygons were intersected with every building in the 40,745 zoom-14 tiles of buildings.pmtiles (2026-09-23.1) they touch, 829,146 buildings in total. A building counts as a hit when at least half its footprint lies inside a dam, estimated from a 6 × 6 grid of points inside the footprint. 100 random hits were then labelled one at a time over Esri World Imagery: 99 not a building, 1 real.

wa_building_dam_hits.zip holds all 15,282 hits as GeoJSON (with footprints) and CSV. Both have the same columns: overture_id, geometry_source, height_m, area_m2, frac_in_dam (share of the footprint inside a dam), dpird_dam_objectid, lat, lon. overture_id is the GERS ID.

I'm the custodian of the DPIRD farm dams dataset and happy to help with access or questions about it.

wa_building_dam_hits.zip

Source attribution

[{"property":"","dataset":"Microsoft ML Buildings","license":"ODbL-1.0","record_id":null,"update_time":null,"confidence":null,"between":null,"provider":"microsoft","resource":"ml_buildings","version":"2026-08-11"}]

How you accessed the data

buildings.pmtiles and base.pmtiles for 2026-09-23.1 from the overturemaps-extras-us-west-2 bucket, read tile by tile at zoom 14 (buildings) and 13 (water) with the pmtiles JS library; overlap tested against the DPIRD and DEA polygons in Node. Example GERS IDs confirmed in the 2026-09-23.1 GeoParquet with DuckDB.

Before submitting

• I searched existing issues and didn't find this one.
• I confirmed the problem is in the latest release, or I noted the release above.