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

TiSeLaC (Time Series Land Cover Classification) was created for the time series land cover classification challenge held in conjunction with the 2017 European Conference on Machine Learning & Principles and Practice of Knowledge Discovery in Databases [1]. It was generated from a time series of 23 Landsat 8 images of Reunion Island, acquired in 2014. At the 30m spatial resolution of Landsat 8 images, Reunion Island is covered by 2866 x 2633 pixels, however only 99,687 of these pixels are included in the dataset.

Class labels were obtained from the 2012 Corine Land Cover (CLC) map and the 2014 farmers' graphical land parcel registration (Régistre Parcellaire Graphique) and the nine most significant classes have been included in the dataset. The number of pixels in each class is capped at 20,000. The data was obtained from the winning entry's GitHub repository [2], which includes the spatial coordinates for each pixel.

We provide training and testing splits designed to give spatial separation between the splits, which should lead to realistic estimations of the generalisation capability of trained models. We first divided the original pixel grid into a coarse grid, with each grid cell sized at 300 x 300 pixels, then computed the number of dataset pixels in each cell (the cell size). These cells are processed in descending order of size, and allocated to the fold with the fewest pixels.

[1] Dino Ienco. (2017). TiSeLaC : Time Series Land Cover Classification Challenge. https://sites.google.com/site/dinoienco/tiselac-time-series-land-cover-classification-challenge (via Internet Archive).

[2] Nicola Di Mauro, Antonio Vergari, Teresa M.A. Basile, Fabrizio G. Ventola, and Floriana Esposito. (2017). End-to-end learning of deep spatio-temporal representations for satellite image time series classification. In Proceedings of the European Conference on Machine Learning & Principles and Practice of Knowledge Discovery in Databases.