The LUCAS dataset revisited: enhancing spatial representativeness for machine learning land cover mapping

Accurate land cover mapping is essential for environmental monitoring, sustainable land management and georisk assessment. However, the limited spatial representativeness of in situ training data can constrain the performance of machine learning in remote sensing applications. This study improves the positional and spatial accuracy of the Land Use and Coverage Area Frame Survey (LUCAS) dataset, a key European source of land cover mapping data—by replacing individual observation points with automatically generated representative areas that reflect local land cover homogeneity more accurately. When applied to the LUCAS 2018 dataset, the region-growing approach generated spatially representative areas for 86.2% of the points, compared to the 19% coverage provided by existing LUCAS Copernicus polygons. Multitemporal Sentinel-2 classification experiments conducted across five European countries demonstrated consistent accuracy enhancements, with F1 scores increasing from 53.1% to 76.5% when using original LUCAS points to 92.7%–97.3% when using representative areas. The method also outperformed the Sen4Map benchmark dataset. These results emphasize the crucial role of spatially representative training data in enhancing the reliability of machine learning-based land cover classification and change detection. It also provides a robust framework for developing and fine-tuning next-generation Earth observation foundation models.