[Talk-GB] Import proposal: Global ML Building Footprints for Nottinghamshire, UK

Mon Feb 6 15:01:22 UTC 2023

Please don't volunteer work for local mappers. I have no desire to go
through validating a couple of hundred thousand new buildings. Many of the
existing ones were mapped quite a long time ago from variable quality
imagery and many could be improved by : a) realignment using the cadastral
layers; and b) checking for extensions etc.

There are currently about 300,000 buildings mapped on OSM in
Nottinhamshire. OS Local provides outlines of 270,000, and there are about
690,000 uprns. As Rob says, splitting buildings by cadastral parcels is an
essential step in preparing any building dataset for use with OSM.
Cross-checking for buildings still containing multiple UPRNs is then a
worthwhile extra step. In general OS Local buildings don't have quite the
level of detail typical of OSM in this area (although they are fine for
things like modern farm & industrial sheds), whereas the Bing set does pick
up quite a bit of this (e.g. returns on terraced houses).

I'm broadly not in favour of a mechanical import, for a range of reasons:

   - Many buildings are misaligned and at a minimum need rotation to be
   placed more accurately.
   - In areas where buildings have been mapped, false positives represent
   perhaps 50% of suggestions this is surely true in other areas.
   - The building data is out-of-date and includes large buildings which
   have been demolished. Without access to the full dataset I can't judge how
   recent the actual acquisition might have been. I suspect this is based on
   Bing imagery which we lost around April last year, which had less parallax
   and better resolution than the current imagery.
   - People are actively mapping buildings at present (particularly in the
   North part of Nottingham
   - Certain buildings types are handled by the algorithm quite poorly
   (particularly link detached houses)
   - There are substantial number of  properties which are still social
   housing (council estates, former mining villages) where there are no
   cadastral parcels. Many of these houses are of common types which allows
   many buildings to be populated quite quickly in a normal edit session if so
   desired (e.g. much of the NE part of Worksop).
   - Building data is not attributed. We have in general tried to use a
   range of building=* tags (specifically for various residential properties -
   detached, semidetached & terraced houses, and apartments)
   - In Nottingham adding buildings should also be carried out in
   conjunction with working out street addresses from the city councils
   Streetlight open data.

On the positive side, in places the Bing buildings are of high quality and
can be added quite quickly with the RapID editor. They do need tweaking
(realigning, rotation), but unfortunately splitting buildings is not
something that the iD code base supports, so this really needs to be done
in Josm, which does have a MapwithAI plugin (not sure of its status, but I
can't seem to download buildings).

If the building set is available, aside from splitting by cadastral parcel
boundaries (which is slightly awkward because of rotation), I think there
are other useful things which could be done:

   - Identify existing buildings which have changed (mainly house
   extensions, or missing conservatories/returns) by comparing footprint
   areas. For Bing buildings with multiple uprns one could sum the areas of
   OSM buildings for comparison
   - Centroids of equivalent buildings can also be compared (better against
   OS Local) to detect buildings which may need significant re-alignment.

tldr: the data as is has a number of limitations which mean that it is not
suitable for import. Further work with the data does offer a range of ways
in which we can improve OSM.

Jerry

PS. I just tried adding <https://www.openstreetmap.org/changeset/132161546>
a couple of terraces in Radford usin RapID, and in practice I'd have been
as fast adding them by hand in JOSM. Apart from splitting (not just the
terraced houses, but an attachment to the industrial building to the
north), they needed re-aligning and not all returns had been captured so
the outlines needed work too.

On Mon, 6 Feb 2023 at 12:35, Andrew Hain <andrewhainosm at hotmail.co.uk>
wrote:

> A potential advantage of importing building outlines where there’s already
> a community is that they can be checked on the ground. That only works if
> the locals are interested in doing that of course.
>
> --
> Andrew
> ------------------------------
> *From:* Philip Barnes <phil at trigpoint.me.uk>
> *Sent:* 06 February 2023 07:44
> *To:* talk-gb at openstreetmap.org <talk-gb at openstreetmap.org>
> *Subject:* Re: [Talk-GB] Import proposal: Global ML Building Footprints
> for Nottinghamshire, UK
>
> I agree with what others have said.
>
> But also why pick Nottingham?
>
> It has a very active local community and missing buildings aren't really a
> problem. Building mapping is pretty comprehensive and the work of splitting
> semis and terraces and adding addresses has largely been done.
>
> Phil (trigpoint)
>
> On 5 February 2023 16:03:02 CET, John Charlton <coolmule0 at hotmail.co.uk>
> wrote:
>
> Hello,
>
> I am interested in using Microsoft's "Global ML Building Footprints
> <https://github.com/microsoft/GlobalMLBuildingFootprints>" for import. It
> is a huge dataset, but I am interested in importing parts locally.
> Specifically around Nottingham, UK.
>
> Where
> ---
> Nottinghamshire, UK. https://www.openstreetmap.org/relation/181040.
> <https://www.openstreetmap.org/relation/181040>
> Strictly within the bounds of this relation
> The import will target built up zones such as residential areas, villages,
> and larger settlements.
> It will not import any building that may overlap with one currently in the
> database, or that may duplicate anything currently within OSM.
>
> What
> ---
> Building footprints. I.e. building shapes in 2D.
> It will not contain any tag data other than buidling:yes
>
> Why
> ---
> There are various places within Nottinghamshire that are without
> buildings. I would like to use this dataset to quickly add missing
> buildings into OSM.
> I feel that it a more efficient use of my time to perform quality
> assurance on the imported data, rather than manually create the building
> footprints from aerial imagery myself.
>
> Accuracy
> ---
> It is a state-of-the-art machine learning database. This does not entail
> that the data should be entire trusted. I have visually examined cases
> around Nottingham to see how this dataset compares to both aerial imagery
> and current building in OSM. I feel that it is to a good quality, accurate,
> and does not have many false-positives within built up areas.
> I will manually check and go through every area imported to ensure that
> the import is sane, looks like it should match with both aerial imagery,
> and with the OSM database with regards to road layout, and any buildings
> that may already be in OSM.
>
> License
> ---
> It uses the Open Data Commons Open Database License (ODbL).
> <https://opendatacommons.org/licenses/odbl/> This is compatible with the
> OSM license. Microsoft have a section in the Readme
> <https://github.com/microsoft/GlobalMLBuildingFootprints#should-we-import-the-data-into-openstreetmap> about
> importing into OSM.
>
> The import process
> ---
> The dataset in available within JOSM with the RapiD plugin. It provides an
> easy way to copy & import data from the dataset into the OSM database.
> The import, if the proposal is accepted, will use a new user, whose
> purpose will be for this import project only. (username:
>  ms-glftp-nottinghamshire)
>
> My Questions
> ---
>
>    - Has the Global ML Building Footprints been imported in another
>    project? I have been unable to find any.
>    - Are there other mailing lists I should reach out on to talk to the
>    community about this?
>
>
> Closing Remarks
> ---
> I have not made an import before, but I am following the guidlines
> <https://wiki.openstreetmap.org/wiki/Import/Guidelines> closely. I
> appreciate I may make mistakes, and look forward to working with the
> community on this.
>
> Kind regards,
> John
>
> _______________________________________________
> Talk-GB mailing list
> Talk-GB at openstreetmap.org
> https://lists.openstreetmap.org/listinfo/talk-gb
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstreetmap.org/pipermail/talk-gb/attachments/20230206/d5a67f5d/attachment-0001.htm>