Answers are grounded in this paper and cite their sources.
Crop type information is needed well before harvest, yet most classification studies report accuracy only for the completed season. This study asks how early each of eight field crops becomes reliably separable from Sentinel-2 time series, and whether one early date serves them all. The Estonian 2021 subset of EuroCropsML, 52,779 parcels, was represented as ten-day composites of thirteen bands and partitioned by administrative region: three for training, one for validation, one for testing. Random Forest, XGBoost, a one-dimensional convolutional network and a Transformer encoder were trained at eight cut-off dates, the deep models with five random seeds at each. By late July the deep models retained about 98% of full-year macro-F1. Stable recognizability was defined as a validation F1 of 0.70 reached by both deep models in at least four of five seeds and sustained at the next cut-off. Winter wheat and winter rapeseed qualified by 30 April, winter barley, peas and potato by 29 June, while the three spring cereals did not meet the criterion at any cut-off.
No takes yet. Share an insight, caveat, or question.
Adnan Gökten (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: