It extends the PASTIS benchmark beyond metropolitan France to Reunion, Guadeloupe, Martinique, Guyane, Mayotte, and Saint-Martin.
PLANTEUR contains 2,750 aligned Sentinel-2, Sentinel-1, and Landsat-8 time series and 4,506 km2 of parcel-derived labels. The labels cover seven crop classes, including sugarcane, banana, pineapple, and tropical tubers. Ai2 says the data are derived from France's Land Parcel Identification System and that the dataset is available through Hugging Face and GitHub.
The accompanying baselines cover linear probes and supervised methods under standard, leave-one-island-out, few-shot, and PASTIS-augmented training setups. Ai2 reports that every evaluated method scores substantially lower on PLANTEUR than on PASTIS, and that models lose roughly half their accuracy when tested on an island absent from training.
The source does not establish which approach is best for operational crop monitoring, how results transfer beyond these six territories, or whether benchmark performance predicts production outcomes.