Wrong elevation rendering

[Pazapas]

Hello,

I used maplibre-gl to display a map of Ireland. I produced my vector tiles from OSM data and elevation tiles from SRTM data (both in mapbox tiles format). All worked fine. But when I tried to do the same for France, elevation data rendering seems to be corrupted. I checked sample points from encoded PNG tiles and data are well encoded. Using the rule height = -10000 + ((R * 256 * 256 + G * 256 + B) * 0.1), I find correct elevations. It seems that for all mountain areas above 3000m (I cannot prove it, but i think it's 3048m or 10000 feet), something is going wrong. The problem is the same with Firefox and Chromium.

Does anyone have an idea to solve that ?
(I discovered this 2 years ago, and still not found neither clue nor solution on the web)

In the following screenshot, you can see that elevation rendering is corrupted in 3D and 2D.

[HarelM]

From my experience, the problem is usually the data.
Having said that you should be able to create a tile with some "static" values including > 10000 feet and see what the terrain looks like.
You can check the numbers demo in the demotiles repo to see an example of generated tiles that are easy to look at.

[Pazapas]

Hi HareIM,
Thank you for your fast answer. Do you have any shortcut to do the test you suggested. Creating false SRTM to generate mbtiles files is a process not so fast.

Here is an example of tile from Firefox cache with elevation data above 3000m.

[acalcutt]

I don't have an answer for your specific problem, but I have generated France with RGE_Alti sources with the process here https://github.com/acalcutt/terrain_merged/tree/main/datasets/RGE_ALTI (using download_france.sh , convert_france_images_EPSG4326.sh , create_terrainrgb_EPSG4326.sh and it seemed to work.

I know with the RGE_Alti source files I used they had multiple projections that I had to convert into a common projection so I could combine them properly, which is why there is a convert script in that terrain.

You can download the merged torrent file I generated based on that here
https://techidiots.net/downloads/open-base-map-torrents/mbtiles/merged-terrainrgb
(France_Merged_2024_z0-Z15_cubic_webp_mbtiles.torrent)

I am also using modified version or rio rgbify which I have added some code to clip elevation below the minimum value before converting, because I found that was one reason for spikes. In this case though I am not sure that is what is causing your spikes

[Pazapas]

Hi Acalcutt,

Many thanks for your links, as torrent seem not to be available, I will follow your process using your shared process, scripts and tools.

[HarelM]

Firefox might have fingerprinting prevention that I've seen to cause spikes, so I would stay with chrome at the moment to reduce noise.
As to creating an image, you should be able to use some basic paint like program to create a 256x256 tile with some colors corresponding to different elevation and use that to test your 10000ft theory.
another aaproach is to simply take the DEMdata class from the code here and output the elevation when you load the relevant tile - reading a png image should be fairly simple, in any language you choose, you can have some AI generate the code for you to make sure the data is valid and that you don't have the spikes inside it...

[Pazapas]

Hi HareIM,
During last year, both Chromium and Firefox added a regular pattern of "negative" spikes. This problem disappeared few updates later.
As you suggest, I will make a little program to convert each pixel from an extracted PNG tile and to find spikes.

[Pazapas]

Hello HarIM and Acalcutt,
I made a small C++ program to check a tile in Alps area and the problem is definitely in the generation processus. I now have to understand why the RGB encoding of elevations produces negative values...
Thank you for your help

[Pazapas]

...Finally, I obtained a correct terrain RGB mbtiles with the use of Acalcutt work. Its scripts work like a charm. I just reduce the max zoom level to 12 for a reasonable rendering time (nearly 3.5hours with a Ryzen 7745, for the the last 10 tiles, 1.5 hour). Now I have to understand why my process introduced weird negative elevation values...
To conclude, Maplibre is a great project.
Thank you

[larsmaxfield]

Do you know which downsampling filter is used in your process when generating the tiles? If it's a filter that interpolates like bilinear or bicubic, that can cause issues with RGB encoded values. I recommend trying nearest neighbor if possible. It ensures the downsampling always uses a real encoded value, not an interpolated one.

[Pazapas]

Hello larsmaxfield
I selected for gdal_translate the same resampling argument as with gdalwarp: lanczos. I understand that such a filter create a smoother result than with rounding to nearest sample. But I don't see why elevation values are suddenly shifted with a big negative offset at around 3000 meters (or 10000 feet).
Acalcutt scripts and his modified rgbfy version saved me :-)