Yeah, I'm surprised they released the demo about multi-temporal crop prediction. Their accuracy is, frankly, pretty terrible. It's basically what I managed the first time I tried to run a classifier against the CDL dataset across years.