hotosm / osma-health

HOT Analytics for Health
12 stars 3 forks source link

Looking for known, well-mapped areas in Botswana #12

Closed rossbernet closed 6 years ago

rossbernet commented 6 years ago

The Azavea team is building a model that will produce population estimates based on building densities from OSM. This will be compared to the worldpop raster to find areas that are incompletely mapped.

To train the model, it will be necessary to provide locations that are known to be accurately mapped. We are starting the efforts in Botswana, so ideally locations in Botswana, but countries with similar land use or population distributions would work. A range of settlement patterns would be useful: rural <> urban.

@geohacker @jpolchlo @moradology @kamicut

awright commented 6 years ago

If understanding correctly, this relates to #5 AOI scale? Apologies if you have to break it down for me.

rossbernet commented 6 years ago

My impression is that #5 is concerned with prioritizing starting locations and determining the scale of analysis.

This issue is concerned with finding training data that can be used to train a model with locations that are predetermined to be accurately mapped. We would like to be able to say, "this is an example of 100m x 100m area or this zoom level 12 tile is known to have buildings that have been accurately and completely mapped" And ideally we would have several different areas representing different settlement types. I.e. an example of a fully mapped rural area, small town, bigger city.

PaulUithol commented 6 years ago

As we'll be ground truthing a lot of data shortly in Botswana - we're currently doing a verification pass over Chobe district, https://tasks.hotosm.org/project/4317

geohacker commented 6 years ago

Thank you for the ticket @rossbernet! I'm tagging @rub21 - he might have some ideas of locations that are worth looking.

smit1678 commented 6 years ago

The malaria map viz might help or osm-analytics:

https://www.mapbox.com/malaria-mapping/#5.21/-18.390/25.539 http://osm-analytics.org/#/

But for #5 we'll prob hone in on some other districts in Zim or Cambodia where there is well mapped areas.

ridixcr commented 6 years ago

With a quick search in Africa I could detect, good places with OSM coverage and DigitalGlobe Standard.

* Link Digital Globe OSM Data
Niamey https://www.openstreetmap.org/#map=14/13.5340/2.1157 Good Image Medium
Dosso https://www.openstreetmap.org/#map=15/13.0442/3.2002 Good Image Medium
Gasau https://www.openstreetmap.org/#map=14/12.1775/6.6819 Good Image Low
Kano https://www.openstreetmap.org/#map=13/11.9848/8.5537 Good Image Low
Namiya https://www.openstreetmap.org/#map=13/10.5000/7.4247 Good Image Low

the data and the image are not aligned because they worked with Bing.

cc. @Rub21 @geohacker

jpolchlo commented 6 years ago

I've made a viewer that can be used to pick out tiles from the set of zoom 12 vectortiles over Africa. It's described in this PR: https://github.com/azavea/hot-osm-population/pull/1. This may be useful to generate lists of tiles for the model fit. It does require the manual input of a Mapbox API key here: https://github.com/azavea/hot-osm-population/blob/master/view/index.html#L101.

Perhaps it will be useful?

kamicut commented 6 years ago

Thanks @jpolchlo! @smit1678 could we perhaps add HOT's Mapbox key and deploy this to s3?

smit1678 commented 6 years ago

👍 sent via Slack

jpolchlo commented 6 years ago

FYI: found a small bug. Fixed it in azavea/hot-osm-population#2. If you've pulled the source, please re-get it.

jpolchlo commented 6 years ago

For frame of reference, the above simple viewer suggests a fairly manual process:

  1. Load the web page.
  2. Scroll around looking for areas of good coverage.
  3. Click on the tile, write down the index
  4. Repeat until there are enough samples.

Mapbox gl is also a bit glitchy when you don't have a complete pyramid of vectortiles. The sample implementation draws from vector tiles that Azavea produced, which start at zoom 12 and go up to zoom 15, so vector data might not appear until you zoom in far enough. Might be worth switching over to the Mapbox QA tiles for display.

The feeling is that there has to be a better way. I just mocked something out that worked, but isn't apt for general consumption. @geohacker: @kamicut suggested that I ask you if you've got suggestions for improving this.

echeipesh commented 6 years ago

Ended up generating the training set in QGIS by overlaying WorldPop/OSM/MapBox layers. More info here: https://github.com/azavea/hot-osm-population/blob/de930c3879690efd28200cae71ec9b92e9faf8ce/README.md