Usage of croptype maps for global models training

[cbutsko]

Now that we've done a nice exercise demonstrating the usage of sampled croptype maps, it'd be useful to:

• consolidate the results (Lucile is already working on that for the webinar)
• make necessary adjustments to either datasets (confidence score?), or global model training (downsampling these datasets?.. introducing weights depending on data source?)
• make adjustments to user-facing side of things: (1) if we don't have an explicit attribute, we might want to store a list of such datasets; (2) make adjustable parameters like upper cap for samples from map (the simplest)
• outline possible future experiments

[cbutsko]

This is a duplicate issue at this point, more on that is discussed and will be resolved in this issue WorldCereal/worldcereal-classification#508

Suggestion for a separate follow-up issue:

For downstream model training, implement a cap setting for map-sampled data that is allowed to be used. Our experiments with Jeroen showed that overwhelming the targeted local training with map-sampled data in the presence of not-too-abundant true reference data does not improve the results, while balancing that out in proportion helps.
‼️This might become redundant if we introduce quality-based sample weights in the downstream training, and then even large amounts of map-sampled data will not affect training as much as high-quality data.