BELDE: Building a Large-scale Earth-observation Land-cover Dataset for Europe

BELDE (Building a Large-scale Earth-observation Land-cover Dataset for Europe) has been introduced as a publicly available dataset for RGB-based remote sensing semantic segmentation, comprising 1,088,385 image-segmentation map pairs with 7 land-cover classes at a 10 m spatial resolution. The dataset, constructed from Sentinel-2 images and ESA WorldCover annotations, also includes BELDE-K and BELDE-CA-NV for cross-region studies. Initial evaluations show models trained on BELDE achieve an F1 score of 83.0% in Europe, but performance drops significantly in out-of-distribution tests, emphasizing the dataset's utility for improving the robustness of Earth observation models.