Hacktoberfest 2026: the issues maintainers tagged for October, open and beginner-friendly. Browse Hacktoberfest issues

False-positive buildings on farm dams in Western Australia

Open
  • Is something wrong?
#152 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Assessment

Difficulty
5/5
Estimated time
Over a week
Newbie friendliness
30/100
Issue type
Bug
Clarity
Mostly clear
Activity status
Active
Tech stack
python
Domain
data, machine-learning

Research direction

Start by reading the repository README and the related reports #43 and #94, then review the attached wa_building_dam_hits.zip and its frac_in_dam and geometry_source fields. Compare the proposed water mask filtering and negative training examples against the reported dam and lake samples. Done means reducing these false positives while checking that real buildings are retained; the issue does not identify a code entry point or test.

Written by the indexing model from the issue text.

Description

Region: Western Australia, agricultural zone
Observed in: Overture Maps release 2026-09-23.1, where these footprints are attributed to ml_buildings version 2026-08-11

Summary

The building model regularly detects farm dams as buildings in Western Australia. About 15,000 footprints in the state's agricultural zone trace dam water or dam banks. In a random manual review of 100 footprints that overlap a mapped dam, 99 were not buildings.

  • 15,109 Microsoft footprints lie at least 50% inside a mapped farm dam. That is 2.3% of the 646,522 Microsoft footprints in the areas checked.
  • Manual imagery review of 100 random cases: 99 not buildings, 1 real (99%, 95% CI 94.6–99.8%).
  • From this one error type alone, about 2.3% of footprints in the checked areas are false. That is above the 0.1–2.2% corpus false-positive rates published in this repo's README for other regions; the README gives no figure for Australia.

What the errors look like

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

Image Lat, lon Area What it is
Image −33.46269, 117.93659 1,021 m² Shallow dam water
Image −33.73576, 118.71789 601 m² Dam water surface
Image −33.36167, 120.10574 225 m² Half on dam water, half on the bank
Image −35.08762, 117.92193 166 m² Dark dam water

Two patterns account for the errors we reviewed:

  1. Water surfaces, from pale and shallow to turbid green and dark. These are often rectangular, because farm dams here are dug as rectangles.
  2. Dam banks. The bright spoil ring around the excavation, often with part of the footprint on the water's edge.

A second, smaller error class shows up on inland salt lakes: footprints on bare lake beds, found by running the same check against DEA Waterbodies. Of 53 sampled footprints inside waterbodies over 100 ha, 51 were not buildings.

These dams and lakes are common across dryland Australia, so the same errors likely occur outside WA.

How this was found

  1. All 175,143 polygons from DPIRD's Farm dams of Western Australia dataset were tested against every building footprint in the 40,745 zoom-14 tiles they touch, 829,146 buildings in total, as published by Overture Maps.
  2. A footprint counts as a hit when at least half of it lies inside a dam polygon.
  3. 100 hits were drawn at random and labelled one at a time over Esri World Imagery.

Suggestions

  • Post-filter with a water or dam mask. Drop or down-weight detections that mostly overlap a known waterbody. For Australia, two open sources:
    • Farm dams of Western Australia (DPIRD). 175,143 dam polygons, CC BY 4.0, attribution: © State of Western Australia, Department of Primary Industries and Regional Development.
    • DEA Waterbodies (Geoscience Australia). National, but too coarse for most farm dams, and its small polygons cover real buildings in towns. Its waterbodies over about 100 ha are reliable: in a review of footprints inside them, 51 of 53 were not buildings.
  • Add negative training examples of farm dams: water surfaces and banks. The attached geometries give 15,000 labelled-negative candidates with locations.

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

Attachments

wa_building_dam_hits.zip contains all 15,282 footprints that overlap a dam, 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. The 173 OpenStreetMap rows can be ignored; filter on geometry_source.

Related

  • Reported to Overture Maps, where these footprints are conflated: OvertureMaps/data#580
  • Proposal to add the DPIRD farm dams layer to Overture as a mask and water source: OvertureMaps/data#581
  • #43 (buildings in a river in Kamloops) is the same error on mapped water. You asked there for coordinates to remove them; the attachment here gives coordinates for all 15,109.
  • #94 mentions that some regions were reprocessed with a confidence attribute. The WA footprints in Overture carry none. If WA could be reprocessed with one, or a water mask applied, these would be much easier to filter.
Dominant language
Python
Stars
2k
Forks
278
No merged PRs in 30d

Getting set up

This project ships no dev container, Dockerfile or contributing guide, so setting up is up to you: start from its README, and see our first-contribution guide for the general steps.

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from microsoft/GlobalMLBuildingFootprints

All issues in microsoft/GlobalMLBuildingFootprints

Similar issues

More Python issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.