Brandenburg-S12-CropVal: A Sentinel-1/2 Multi-Modal Cropland Segmentation Dataset for Cross-Regional Transferability Assessment (Brandenburg, Germany, 2024)

Brandenburg-S12-CropVal is a large-scale, multi-modal remote sensing dataset covering the entire federal state of Brandenburg, Germany (~29,500 km²), designed for evaluating the cross-regional and cross-sensor generalization capability of cropland segmentation models. Developed as part of the CrossModMamba study, it complements the training dataset HSR-MFCS (https://doi.org/10.5281/zenodo.17054418) from Sichuan, China. The dataset comprises Sentinel-2 Level-2A surface reflectance imagery (RGB bands B4, B3, B2) and Sentinel-1 GRD SAR imagery (VV and VH polarizations, IW mode, descending orbit) acquired during the 2024 growing season, all processed on Google Earth Engine. Cloud contamination was mitigated using the S2_CLOUD_PROBABILITY dataset and QA60 quality band, retaining only clear-sky observations. Reference labels were derived from official parcel-level crop type vector polygons for 2024 provided by the Brandenburg State Office for Surveying and Geoinformation (LGB), standardized according to the EuroCrops HCAT3 hierarchical taxonomy and consolidated into binary classes: cropland (all cultivated agricultural parcels) and non-cropland (forests, water bodies, built-up areas, and bare land). All layers are co-registered at 10 m spatial resolution and projected to EPSG:32633 (WGS 84 / UTM zone 33N). This dataset enables a stringent assessment of model transferability across five simultaneous domain shifts: (1) geographic region (East Asia → Central Europe), (2) sensor platform (Gaofen-1/3 → Sentinel-1/2), (3) spatial resolution (4 m → 10 m), (4) landscape structure (fragmented hilly terraces → large-scale consolidated plains), and (5) crop phenology and management practices. In the CrossModMamba study, the model trained exclusively on Sichuan HSR-MFCS data was applied to this dataset, demonstrating robust cross-domain generalization.