This dataset contains the final crop type classification map for the agricultural zone of the Limarí River Basin (Coquimbo Region, Chile), covering approximately 30,000 ha of cultivated area. The map was produced using an ensemble of Transformer (TRF) neural network models trained on Sentinel-2 Level-2A (MSIL2A) time-series imagery from tiles T19JBG, T19JBF, T19JCG, and T19JCF, spanning the 2021/2022 agricultural season (69 acquisition dates after cloud filtering). Input features are the 10-m resolution Sentinel-2 bands B02, B03, B04, and B08. The classification distinguishes 16 classes: almond, avocado, industrial grape, lemon, mandarin, olive, orange, table grape, walnut, forage, short cycle crop, barren, barren shadowed, riverside vegetation, urban, and water. Ground-truth samples (2,683 points) were derived from the Fruit Cadastre of Coquimbo Region (CIREN, 2021) and supplemented with field surveys. The final map was generated by combining predictions from three independently trained Transformer models (one per cross-validation fold) using a Product of Experts ensemble strategy, which multiplies per-class probabilities across folds and assigns the class with the highest normalized probability. The best model (using reflectance-only input) achieved an overall accuracy of 0.944. The output is a single-band GeoTIFF at 10 m spatial resolution where each pixel value represents a crop or land cover class encoded as an integer (1–16). The probability cells derived for each TRF model are also included in a zip file, which are available to reproduce the final map. Code used to train the models and produce this map is available at: https://github.com/aldotapia/S2-ClassifierComparison