Client Side image preprocessing

[sabman]

Sometime we images from different satellite imagery providers have different color histogram distribution. In order to have comparable images it can be useful to match histograms. Idea from @jo-chemla

• JS port of https://github.com/mapbox/rio-hist

[jo-chemla]

Thanks for opening that discussion, great starting point!

Would be a great enhancement to a class of tools (like historical-satellite) we've been building for the disaster-management and research communities (esp. archaeology, looting detection etc)

It would be useful for such a utility/function to:

• ingest source and target raster data, as well as an option object which could contain downsamplingFactor, returnType etc
• and then return [both or user-selection of] the output raster data of the source matched to target, as well as the arrays representing the mappings, one per channel
• both histogram equalization, and simpler contrast stretching could be useful algorithms, due to how the results could be passed to visualization application for simpler-but-faster live color-matching

The mappings array could probably be used as-is in the maplibre/mapbox fragment shader - or the simpler contrast stretching params, opened a similar discussion here maplibre/maplibre-gl-js#6273 to allwo per-band definition of raster-brightness-min, raster-brightness-max, raster-saturation, raster-contrast

Here is the scikit-image doc and histogram_matching.py implementation based simply on numpy bincount, cumsum, interp. Could probably be made a standalone module if tree-shaking works on geoai-js.

Also sharing interesting notes regarding color-spaces here mapbox/rio-hist#3

[jo-chemla]

Finally took the time to implement this JS script, based on the scikit/numpy implementation. The API works great, even for large images, see screen!
Just need to correctly pipe File input and events to make the input source and target raster images user-modifiable in the demo html.

See the implementation (API and js) here

[jo-chemla]

Now more solid implementation - the base stays the same simple functions, just edited the UX boilerplate around it. Now exports geotiff if source is geotiff, displays histograms via chartjs, reacts to file inputs, has optional settings like maxMpx for faster bins/CDF computation via downsampling, and also color_space + bands like rio-hist to do the matching in RGB, HSV, LAB etc. Test live demo here!

See the implementation (API and js) here

[sabman]

Nice! This makes me think it might be interesting to design a plugin interface for image processing that users could include if they need it and developers can create plugins as needed.