[OSM-talk] Map of Population Density vs. OpenStreetMap density

[Christoph Hormann]

On Sunday 07 July 2019, Darafei "Komяpa" Praliaskouski wrote:
>
> We're using GHS population grid in Switzerland.
> https://ghsl.jrc.ec.europa.eu/data.php
> Methodologically, they use radar data to find "houses". It means on
> their dataset people also live along roads with asphalt, and - may
> happen - bare rocks are also populated. You can drop them a line on
> jrc-ghsl-data at ec.europa.eu to say thanks.

I am familiar with that data - they use census based or otherwise 
estimated population numbers per admininstrative unit and distribute 
this population among areas they identified as "built-up" using rather 
questionable processes (what we in German tend to describe 
as "Kaffeesatzleserei").  There is no identification of houses - source 
data used is way too low resolution for that.

I am not aware of any serious overall evaluation of the quality of this 
or any other global population density data sets.  If you read 
literature on the matter the quality/validation part is usually just 
some superficial "throwing around numbers to make the results look 
good" without actually looking at how the data compares to the 
geographic reality it is meant to represent and where and how it fails 
to do so.

I am sorry for the negativity - i just know all too well how these kind 
of publicly financed research projects work in Europe and how detached 
from reality they often become.

> To fix it we can get "unpopulated areas" polygons from OSM.

Not really - you would have to reproduce the population distribution 
process described above based on corrected data of builtup areas.  If 
you just remove populations that are obviously wrong locally you'd 
underestimate the overall population.

> > And i am not a fan of deliberately pixelated visualizations where
> > the data is shown in a pixel grid at a coarser resolution than what
> > the display offers.
>
> Can you point to a better visualization which we can learn from?
> Map is supposed to be used on settlement level, where our grid is "4
> pixels per screen" - to highlight a settlement without trying to
> predict its boundaries.

You are essentially visualizing a classification map (with ten classes).  
The most common way to do this would be on a per pixel basis.  See for 
example the "NLCD Land Cover" layer on https://viewer.nationalmap.gov/. 
If this is too noisy (which is also very much influenced by the choice 
of colors) you can denoise and geometrically generalize the 
classification for the target resolution.  Just subsampling at a coarser 
grid does not really work for this - you just get coarser noise and 
less information.

> We've built such map initially, and it's not significantly different
> from this one in disaster mapping perspective. People don't map
> physical geography far from their home much in OSM [...].

The thing is that statement is correct to very different degrees in 
different parts of the world.  Looking selectively at the mapping of 
physical geography would allow evaluating those differences.

Of course you are right that just counting features would not really 
work for analyzing that.  Counting features works well for things with a 
fairly defined amount of information per feature (like buildings, 
addresses, POIs) but not for geometrically sophisticated geometries.

-- 
Christoph Hormann
http://www.imagico.de/