tags:
- satellite
Part of MONSTER: https://arxiv.org/abs/2502.15122.
TimeSen2Crop consists of pixel-level Sentinel-2 data at a 10m resolution, extracted from the 15 Sentinel-2 tiles that cover Austria [1]. The dataset contains 16 classes representing different land cover types. The original data contains all Sentinel 2 images covering Austria acquired between September 2017 and August 2018 plus images for one tile acquired between September 2018 and August 2019. As the tiles are from different Sentinel-2 tracks and have been processed to remove images with cloud cover greater than 80%, the image acquisition dates for each tile differ and are irregular. This version of the dataset has been processed to interpolate each pixel to a daily time series representing one year of data (thus each pixel has a time series length of 365) and removing the "other crops" class. The processed dataset contains 1,135,511 multivariate time series, each with 9 channels (representing 9 spectral bands) and 15 classes. Classes are unbalanced and unevenly distributed across the Sentinel-2 tiles. The dataset has been split into cross-validation folds based on geographic location by Sentinel-2 tile (i.e., such that, for each fold, time series from a given location appear in either the training set or test set but not both).
[1] Giulio Weikmann, Claudia Paris, and Lorenzo Bruzzone. (2021). TimeSen2Crop: A million labeled samples dataset of Sentinel 2 image time series for crop-type classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:4699–4708.