Shaded Relief Inconsistencies & Artifacts

[geoglrb]

Apologies if this is irrelevant, but...

[geoglrb]

Probably a different issue, but notice building-esque artifacts:

And that said, maybe it's not just HK...

[hyperknot]

If you are talking about the strong grid-like shading of terrains, it is really visible here: http://maps.stamen.com/terrain/beta/#15/47.6879/18.8744

[geoglrb]

Yes, true. Thanks!

It strikes me that there could be three different things going on in the images above, of which that is at least one. Not being enough of a mapnik-er/CartoCSS-er, I don't know if this is 0-3 issues (!).

[mojodna]

There are a few things going on here, none of which we have a whole lot of control over:

• lower-resolution / chunky DEM data available for certain places
• some building-shaped artifacts are artifacts of how the DEMs are processed: structures are removed, leaving flat areas on slopes (particularly notable in cities)
• hard lines near the coast are often from USGS NED and are usually the result of source data edges being hillshaded as steep slopes (we may be able to filter these out in the hillshade code
• strips / grid-like shading artifacts are also the result of how the SRTM DEMs are created (I think)

https://github.com/mapzen/joerd may be a better place to track these issues if we can confirm that they're inherent to the source data.

[mojodna]

Actually, the banding could be the result of the resampling method that Mapzen is using:

http://openterrain.tumblr.com/post/109437899736/the-effects-of-various-resampling-algorithms-on http://openterrain.tumblr.com/post/109435441421/hillshades-generated-from-resampled-srtm-dems http://openterrain.tumblr.com/post/109434986656/resampled-srtm-dems-reprojected-from-wgs84-to-web

Summary (from the first link):

Artifacts aren’t visible when looking at the reprojected DEMs, but notice how the image shifts around slightly when in slideshow mode (all images were taken of the same bounding box at thne same resolution, excepting the zoomed in versions).

The story is quite different when looking at the resulting hillshades. Horizontal and vertical artifacts become quite apparent (along with pixel shift) with methods other than bilinear and cubic. Of these, bilinear has a slight edge (to my eye), as it produces transitions that are less abrupt. You can see the difference in the 2 zoomed images.

[mojodna]

Joerd appears to use nearest neighbor resampling (the default): https://github.com/mapzen/joerd/blob/b590a175fb2f07781ef36bf485117aad6f3e206e/joerd/composite.py#L42-L43

https://github.com/OSGeo/gdal/blob/08b04169ecacc86d1f4c1532e30241cb96c0036d/gdal/swig/python/osgeo/gdal.py#L2589-L2597

[hyperknot]

I can confirm that Mapzen terrain map has the exact same terrain interpolation artefacts: https://tangrams.github.io/walkabout-style-more-labels/#14.84659/47.6803/18.8767