From the source
Ai2 has released PLANTEUR, a crop-mapping benchmark that extends the PASTIS dataset from metropolitan France to six tropical regions, so teams building geospatial models can test them on smallholder, tropical, topographically complex farmland.
- PASTIS is widely trusted for its high-quality labels and geospatial metadata, but its coverage stopped at France, while crop-type mapping underpins food security planning, water allocation and tracking farmland change.
- PLANTEUR draws on France's Land Parcel Identification System and spans Réunion, Guadeloupe, Martinique, Guyane, Mayotte and Saint-Martin.
- Its format closely resembles PASTIS, with 2,750 aligned Sentinel-2, Sentinel-1 and Landsat-8 time series and 4,506 km² of parcel-derived labels across seven crop classes, including sugarcane, banana, pineapple and tropical tubers such as cassava, yam and taro.
- Ai2 reports baseline numbers for linear probes and supervised methods on a standard split, a leave-one-island-out split, a few-shot setup, and one that includes PASTIS samples in training.
- Every method evaluated, including Ai2's own, scores substantially lower on PLANTEUR than on PASTIS, and each model loses roughly half its accuracy on an island it never saw in training.
