Make classification layers (SCL, cloud masks) a full multiscale pyramid #274

Activity

1. emmanuelmathot

   assigned

   shollendEODC

   on Sep 28, 2026
2. emmanuelmathot

   added

   enhancementNew feature or requestNew feature or request

   multiscalesMultiscale pyramids and overviewsMultiscale pyramids and overviews

   sentinel-2Sentinel-2 MSI products (L1C, L2A)Sentinel-2 MSI products (L1C, L2A)

   sample-serviceRoadmap with EODC for the EOPF Sentinel Zarr Samples Service (#277)Roadmap with EODC for the EOPF Sentinel Zarr Samples Service (#277)

   on Sep 28, 2026
3. shollendEODC commented on Oct 8, 2026

   

   shollendEODC

   on Oct 8, 2026

   Collaborator

   More actions

   i adressed this topic in a pre-PR to investigate how easy/difficult it would be to add the associated logic. Its currently available here: https://github.com/shollendEODC/data-model/tree/integrate/cpm-geozarr-main-feat-generic-conversion-s2-cld-snw-scl-MS and is branched off of 039f38c

   Based on your task list i defined the code and implemented it, samples can be accessed here: aws s3 --profile eodc-cm --no-sign-request ls s3://continuous-integration/eopf-geozarr/branches/cpm-geozarr-main-feat-generic-conversion-s2-cld-snw-scl-MS/

   Write r120m, r360m and r720m for l2a_classification (and the other classification masks). Use classification resampling (_coarsen_variable already supports it).

   This should work, given the restructuring of the code into a generic_multiscales(...) function

   Write a multiscales block on l2a_classification. MultiscaleMeta(..., resampling_method="average") is hard-coded (s2_multiscale.py L864). Make the resampling method a parameter, and use nearest (or mode) for classification.

   MultiscaleMeta is not used, as the type validation of MultiscalesAttrs and LayoutObject should be enough. resampling_method is now extracted and used as a parameter.

   Use the same chunking and sharding as reflectance (spatial_chunk and sharding) for every raster group on x/y, not create_original_encoding.

   chunking and sharding is derived from create_uniform_encoding(...) which should satisfy this demand.

   Add CF flag_values and flag_meanings to scl, as quality/mask has flag_masks and flag_meanings.

   these are added, altough i noticed differences between l2a_classification/20m and l2a_classification/60m, which carry different metadata (seems to be passed down from cpm?)

   here it might be relevant to note, that the l1c_classification variable b00 triggers the average handling for coarsening, as the current flagging system detects variables starting with b as reflectance. This should be adapted from my pint of view..

   Add classification:classes to the SCL asset in the STAC item.

   Should this not be done in eopf-stac? Seems more logical to me.

   add a test that opens SCL at every level with titiler's GeoZarrReader.

   This should be done outside of this repo right? Tests should be added, but only focus on eopf-geozarr logic

   This is not a PR yet, just an aggregation of insigths to report progress and trigger discussions. The current development on the branch is also not complete yet, especially the tests need to be modified and adapted.
4. shollendEODC commented on Oct 13, 2026

   

   shollendEODC

   on Oct 13, 2026

   Collaborator

   More actions

   PR opened in #313