FieldSeg: A scalable agricultural field extraction framework based on the Segment Anything Model and 10-m Sentinel-2 imagery

• A novel crop field boundary extraction framework was proposed using SAM. • Eight agricultural regions were selected across the world for model assessment. • SAM-derived fields achieved mean dice coefficients ranging from 0.735 to 0.847. • A large-scale demonstration covered an area of 1 million square kilometers. Accurate delineation of agricultural fields from satellite imagery is crucial for digital agriculture and conservation. The Segment Anything Model (SAM), a state-of-the-art image segmentation model, brings new possibilities for this task. However, its feasibility under different agricultural contexts remains unclear, and there are open questions regarding model parametrization, image preprocessing, and integration into an operational framework. This study proposes a new SAM-assisted crop field extraction framework (FieldSeg) using 2022 Sentinel-2 temporal composites and presents the lessons learned using this foundational model in eight agricultural regions across the world. Through rigorous experiments, this study optimized FieldSeg in three stages: input data preparation, model parametrization and patch management, and final fine parametrization. This study explored different bands and temporal metrics combinations and defined a set of optimal configurations for the framework based on performance and processing time. Non-agricultural objects segmented using SAM were removed using an annual crop mask derived from Google Dynamic World. While performance was low to moderate in regions with small fields (<5ha in China, South Africa, and Spain), FieldSeg achieved a promising performance in the study areas with medium-large fields (≥5 ha in Argentina, Australia, Brazil, USA-California, and USA-Iowa), with the rates of correctly extracted fields ranging from 0.541 to 0.814. The extracted fields showed a good segmentation quality, with mean dice coefficients ranging from 0.735 to 0.847. The large-scale applicability of FieldSeg was also demonstrated in four countries (1 million square kilometers), showing promising results and the ability to generalize across different regions.