Field Boundaries (Delineate Anything)

This dataset contains national-scale agricultural field boundary delineations generated using the DelAnyFlow methodology, which is a resolution-agnostic, deep-learning approach for the automated extraction of field boundaries from satellite imagery. The products consist of topologically consistent, non-overlapping vector polygons of individual agricultural fields derived from Sentinel-2 imagery with resolutions of 5 m. Currently, the dataset includes complete coverage of Ukraine for 2024, with 4.35 million delineated fields at 5 m resolution. These boundaries were generated using the DelAny instance segmentation model trained on the FBIS-22M dataset (22.9 million field instances across multiple resolutions from 0.25m to 10m). The delineation products are analysis-ready and designed for operational applications including crop type classification, yield estimation, agricultural statistics, subsidy auditing, and food security monitoring. They demonstrate boundary completeness and quality compared to existing operational products, particularly for smallholder and fragmented agricultural systems (0.25-1 hectare fields).