Mapping small landscape features in agricultural lands at the national level using CNN-based semantic segmentation

Spatially explicit information on small landscape features is required for biodiversity assessment and implementation of the European Union (EU)’s Common Agricultural Policy (CAP). However, mapping these features is often done manually, making it time-consuming and difficult to scale. This study evaluates a deep learning-based segmentation approach for automated mapping of field islands, a locally defined class of small landscape features composed of heterogeneous woody and non-vegetated elements. We trained a U-Net model using high-resolution orthophotos and reference polygons from a government registry. Across test sites, slightly reducing input image resolution, increasing patch size, and using higher patch overlap produced the best results, achieving an average F1-score of 0.72 while reducing computational demand compared to native resolution. Applied at national scale, the model identified 34% additional field islands than the government registry. The proposed methodology and dataset can support CAP implementation and improve habitat and carbon assessment.