Adding postprocessing to WorldCereal jobs

[GriffinBabe]

@kvantricht Made some good progress on this issue today

A few things to discuss:

• We cannot go through the GFMAP connection methods as CDSE firewall is blocking our long UDFs. For the moment I commented that part and set to https://openeo.creo.vito.be/openeo/, which does backdoors the firewall. We need to find a solution for that before merging this PR of course
• There is one part missing for the postprocessing: whenever performing reclassification, we get the index from np.argmax(...). However, we do not have access to a LUT as in the catboost inference. One way could be to read this information from the catboost model outside the UDFs in the generate_map function and then provide the LUT to all UDFs from the parameters, I propose we discuss this tomorrow. (At the moment I'm simply leaving the argmax indices as new labels, which is consistent with the default catboost model we provide, but is not future proof)
• I uploaded new unit-tests which are passing well. Could you provide me the preprocessed inputs NetCDF that was set-up here https://github.com/WorldCereal/worldcereal-classification/blob/98-add-postprocessing/tests/worldcerealtests/conftest.py#L16 ? They are not tracked in the repository, I tried with my own files, with which I had a good crop-land but a bad crop-type (only other crops predicted). Before trying on a different location I would like to test it with the files that you used here.

[kvantricht]

• Before trying on a different location I would like to test it with the files that you used here.

• this URL used to be the one we were using, I forgot why. This was indeed the reason. Let's see how it can be solved ...
• TBD tomorrow, not fully sure what the situation is
• Added the resource for cropland. For crop type I didn't have one. Should be simple to generate one though. TBD.

[kvantricht]

@GriffinBabe in fact I don't think these are very useful right now. Bands are still only ['classification', 'probability']. We need updated ones with all probabilities right?

[kvantricht]

is save_result() now always triggered when requesting postprocessing? I can imagine e.g. for the demo that users just want the end product. The only use case for saving intermediate results is probably the global processing where we still want to play with the probabilities afterwards.

[GriffinBabe]

Let's make it parametrizable in generate_map with default=False?

[GriffinBabe]

@kvantricht

• I added an exception in the generate_map function to connect to the backdoor URL if the users wants to connect to CDSE
• For the output labels of the model, and the index to class code LUT I decided to load those informations in a separate util function and pass those values in the UDFs parameters. Anyways before we needed to load the model locally in CropTypeClassifier to determine the number of output bands to output and specify in the computing graph, now I do it outside the UDF, but I also pass it to the PostProcessor so it can reclassify with the correct labels instead of the argmax indices.