Train datasets and source code of SpiderNet

GLC_L2H: A Global Benchmark Dataset for Land-Cover Mapping with Coarse Geospatial Supervision GLC_L2H is a global benchmark dataset developed for evaluating high-resolution land-cover mapping using existing coarse-resolution land-cover products as supervision. The dataset is designed to support research on learning under spatial and semantic uncertainty caused by resolution discrepancies between coarse supervision and high-resolution imagery. The benchmark covers 97 cities across 44 countries and 6 continents, representing diverse geographic environments and land-cover conditions. High-resolution reference annotations are derived from the OpenEarthMap/xBD-based dataset used in this study, while three existing land-cover products—ESA_GLC10, Esri_GLC10, and GLC_FCS30—are provided as alternative coarse-supervision sources. For the benchmark experiments reported in the associated SpiderNet study, land-cover labels are harmonized into four evaluation classes: Tree canopy, Low vegetation, Built-up, and Water, with non-evaluated regions treated according to the benchmark definition. This repository provides the data organization and benchmark information required to reproduce the experiments, including: geographic coverage and city-level metadata; class-wise statistics for the harmonized land-cover categories; correspondence between the high-resolution reference labels and the three coarse-supervision products; explicit training, validation, and test split definitions; machine-readable split files listing the samples/regions assigned to each subset; preprocessing and label-harmonization information; documentation of the licenses and redistribution conditions of the component datasets. The dataset split is defined explicitly at the sample level rather than regenerated randomly at runtime. The exact membership of each subset is provided in my_dataset_split.json, enabling identical experimental partitions to be reproduced. Licensing and redistribution. GLC_L2H integrates or derives information from multiple existing geospatial datasets. Users should refer to the original licenses and terms of use of OpenEarthMap, ESA_GLC10, Esri_GLC10, and GLC_FCS30. Only components permitted for redistribution under their respective licenses are included directly in this Zenodo release. For components that cannot be redistributed, this repository provides the corresponding source information, acquisition instructions, and preprocessing procedures required to reconstruct the benchmark.