Revisit active cropland marker in the frame work of Phase II

[kvantricht]

• Translate current active cropland marker to the WorldCereal UDF using GFMAP
• Test on globally representative dataset different season detection parameters

[kvantricht]

@GriffinBabe, task will be explained when we get at it, near the end of March.

[jdegerickx]

@jdegerickx --> Prepare demo notebook showing how to apply the active cropland marker on the following toy dataset:

/vitodata/worldcereal/data/openeo/inputs_presto/preprocessed_merged/belgium_good_2020-12-01_2021-11-30.nc

NOTE: we should investigate the impact of moving to MONTHLY composites on the performance of the algorithm!

[jdegerickx]

@kvantricht, could you point me to the code you used to generate the inputs located in:
/vitodata/worldcereal/data/openeo/inputs_presto/preprocessed_merged/belgium_good_2020-12-01_2021-11-30.nc

I'd like to generate a dekadal version and extend the time period a bit to test active cropland marker...

[kvantricht]

@jdegerickx it's here: https://github.com/WorldCereal/wc-classification/blob/openeo/notebooks/OPENEO/extract_inference_patches.ipynb

[kvantricht]

@jdegerickx FYI, just pushed something that may help parsing the compositing window as processing option: https://github.com/WorldCereal/wc-classification/blob/openeo/src/worldcereal/openeo/patches.py#L48

you could now try something like:

# Initialize the extractor
extractor = ExtractCDSE(gdf_cdse, 
     endpoint='https://openeo.dataspace.copernicus.eu/',
     patch_size=patch_size,
     target_dir=raw_extractions_dir,
     parallel_jobs=2,
     composite_window='dekad'
     )

and let me know if that works.

[jdegerickx]

Note to self: active cropland benchmarking results/conclusions are here:

https://vitoresearch.sharepoint.com/sites/21717-ccn-world-cereal-VITO_internal/_layouts/OneNote.aspx?id=%2Fsites%2F21717-ccn-world-cereal-VITO_internal%2FShared%20Documents%2FVITO_internal%2FWorldCereal%20Notes&wd=target%28General.one%7C786E6814-6C2C-4B41-8E1B-A4DB8CBBBC4F%2FActive%20cropland%20benchmarking%7CEC6C2DFC-144A-4F6E-B12D-5E98459C0AF9%2F%29
onenote:https://vitoresearch.sharepoint.com/sites/21717-ccn-world-cereal-VITO_internal/Shared%20Documents/VITO_internal/WorldCereal%20Notes/General.one#Active%20cropland%20benchmarking&section-id={786E6814-6C2C-4B41-8E1B-A4DB8CBBBC4F}&page-id={EC6C2DFC-144A-4F6E-B12D-5E98459C0AF9}&end

[jdegerickx]

Decided to postpone this activity to after System v2 release.
We are inclined to abandon the previous approach.

This requires a new benchmarking experiment:

• set up a benchmarking dataset, globally representative, of patches with crop type information
• run an experiment with Presto, in which you feed a partially masked timeseries (everything out of season of interest is masked out) and just run the default crop/no-crop model --> does this result in meaningful results? @cbutsko to define a separate issue for this
• develop a simple unsupervised approach using expert features (min and max EVI, presence/absence of sharp increases and decreases, ...) (cf Sen2Agri features)
• Compare both previous approaches with our traditional way (which is already implemented in a separate script based on openeo extractions)