An automated mixed-sample generation strategy for land cover classification in heterogeneous river basins

Land cover (LC) classification in heterogeneous areas is challenging because pixels are often influenced by multiple LC types and show less distinctive feature responses. This study proposes an automated mixed-sample generation strategy for improving LC classification in heterogeneous river basins. The strategy combines cross-product full-agreement pixels and partial-agreement pixels from multiple LC products to balance label reliability and sample diversity. Full-agreement pixels provide reliable labels, while partial-agreement pixels add information from low-agreement areas that may include heterogeneous and transitional LC conditions. The method was tested in two Chinese river basins using random forest classification. Compared with a baseline using only full-agreement pixels, mean Intersection over Union (mIoU) increased from 0.762 to 0.809 in one basin and from 0.729 to 0.782 in the other. In cross-product non-full-agreement areas, mIoU further increased by 0.083 and 0.129. The results suggest that the strategy provides a simple automated option for regional LC mapping.