Part of MONSTER: https://arxiv.org/abs/2502.15122.

S2Agri is a land cover classification dataset and contains a single tile of Sentinel-2 data (T31TFM), which covers a 12,100 km² area in France [1, 2]. Ten spectral bands are used, and these are provided at 10m resolution. The dataset contains time series of length 24, observed between January and October 2017. The area has a wide range of crop types and terrain conditions.

The original S2Agri dataset is designed for parcel-based processing and contains data for 191,703 land parcels, with data for each parcel provided in a separate file. We have reorganised the data for pixel-based processing, leading to a dataset containing 59,268,823 pixels. Two sets of land cover classification labels are provided, one with $19$ classes and the other with 44 classes. However, some of the 44-classes are only represented by one land parcel. We have removed the pixels in these land parcels from the dataset. This means there are only 17 and 34 classes respectively that are represented in the final dataset. The class label of each parcel comes from the French Land Parcel Identification System. The dataset is unbalanced: the largest four of the 19-class labels account for 90% of the parcels.

We thus provide two versions of the S2Agri dataset, S2Agri-17, which uses the 17 class labels and S2Agri-34, which uses the 34 class labels. Additionally, we have created smaller versions of the datasets consisting of data for a randomly selected 10% of the land parcels, each containing 5,850,881 pixels. These are S2Agri-10pc-17 and S2Agri-10pc-34 for the 17-class and 34-class labels, respectively.

To create the folds used for cross-validation, we split the data based on the original land parcels, thus ensuring that all pixels in a land parcel are allocated to the same fold. Splits are stratified by class labels to ensure an even representation of the classes across the folds.

[1] Vivien Sainte Fare Garnot, Loic Landrieu, Sebastien Giordano, and Nesrine Chehata. (2020). Satellite image time series classification with pixel-set encoders and temporal self-attention. In IEEE/CVF Conference on Computer Vision and Pattern Recognition.

[2] Vivien Sainte Fare Garnot and Loic Landrieu. (2022). S2Agri pixel set. https://zenodo.org/records/5815488. CC BY 4.0.