Fix split buildings

[maning]

LA City data was split to the parcel boundaries resulting to small polygons that should be part of the larger building (see: #71). We fix this during the import by using the Auto-tools plugin in JOSM but there are instances (prior to using Auto-tools) when this wasn't fixed.

We ran a detection for split buildings by analyzing size and shapes of buildings using OSM-QA-tiles, turf and tilereduce. Basic steps below:

• Filter buildings that has lacounty:bld_id tag.
• Find all adjacent polygons (sharing at least one edge).
• Compute area, perimeter and shape factor:

Properties	Description
_area	Area of the polygon (meter^2)
_perimeter	Perimeter (meter)
_shape	Shape factor. Circles - approach a value of 100.00, squares 78.00, thin thread-like object have the lowest shape factor approaching 0. (4 * PI * Area) / (Perimeter^2)

The general idea is that small (_area) buildings are more compact (circles and squares) while large buildings are complex and wiggly. The reverse should be invalid buildings.

After several trials, we came up with an acceptable thresholds for split buildings.

return (p.properties._area <= 13 && p.properties._shape <= 76) ||
    (p.properties._area > 13 && p.properties._area <= 30 && p.properties._shape <= 52) ||
(p.properties._area > 30 && p.properties._area < 150 && p.properties._shape <= 17);

Trial Results

A for ~800 samples have the following results:

Category	N	%
👍 (valid detection)	661	81.60
👎 (invalid detection)	147	18.15
¯\_(ツ)_/¯ (Don't know)	2	0.25

Valid detection

Thin/narrow rectangles

Triangles

Diagonal cuts

Invalid detection

Roofs that connect other building

Small buildings the slightly touches another building

Small square building that connects two apartment building complexes

These are just initial results and we can tweak the thresholds as we continue. I think the ~81% valid detection is good enough for the community to review and fix.

Dump of split buildings here (~4k buildings).

Next actions

• [x] Prepare maproulette task. @jschleuss
• [x] Announce the task to community for fixing. @jschleuss
• [ ] Review thresholds based on the initial run and create another data dump for another fixing task if necessary. @maning

[jschleuss]

Here's the MapRoulette challenge: http://maproulette.org/map/419

I think it's easier to edit the data in JOSM (if you allow remote control).

[maning]

Tested the maproulette task, some observations:

1. Its hard to see the error at z19 (ticket).
2. After loading in JOSM, its hard to see which of the buildings was identified as an error. Can we configure the task to select the way when loaded in JOSM? In the GeoJSON, the way id is in the @id property. AFAIK, this is possible.
3. Using iD, we cannot do what our JOSM plugin does for merging tags with Auto-tools. Is this OK for the user to decide which tag to maintain?

[maning]

Posted a diary for the community to help in fixing: http://www.openstreetmap.org/user/manings/diary/39230

Current metrics as of today, let's watch if this improves after a week.

[maning]

@jschleuss @cityhubla Let's share this maproulette project to MaptimeLA.

[jschleuss]

Yep! Meetup this Thursday will promote fixing split buildings.

[maning]

Lol! Just saw that this was done.

72% positive detection rate is not so bad.