Lfmc dataset processing question

Hi
I note that the truth annotation is sparse points via https://storage.googleapis.com/ai2-olmoearth-projects-public-data/projects/lfmc/20251029/annotation_features.geojson

But the data config [defines a raster](https://github.com/allenai/olmoearth_projects/blob/main/olmoearth_run_data/lfmc/dataset.json) for labels (example below). This conversion looks straightforard but could you share the processing of advise if there is any nuance?

I view a prediction (below) and it is dense point predictions, I wonder if you could comment on the confidence in dense predictions based on sparse truth? Thanks