Scalable field boundary refinement from satellite time series using deep learning

Purpose: Accurate spatial management units are fundamental to precision agriculture because they directly influence crop classification, input optimization, and decision-support systems. However, administrative agricultural parcels often aggregate multiple crop units within a single polygon, introducing structural uncertainty into parcel-based crop mapping and management analytics. This study addresses inaccuracies in existing agricultural parcel boundaries that introduce upstream errors that cascade into downstream analyses, including crop mapping, yield estimation, and decision-support workflows. Consequently, the proposed framework is designed for parcel refinement rather than semantic crop classification. Methods: We selected parcel refinement as a polygon-conditioned semantic segmentation problem, leveraging existing parcel geometries to generate more accurate, agronomically coherent management units rather than performing field boundary detection from scratch. Implemented at the national scale in Israel, the framework uses a computationally efficient single-sensor Sentinel-2 time-series approach based on harmonic modeling to generate phenology-aware composites that improve robustness to noise, cloud contamination, and temporal gaps. Geometry-preserving synthetic data augmentation was incorporated to improve boundary learning under limited annotations, and outputs were evaluated using both pixel-level segmentation metrics and polygon-level operational correctness. Results: We compared a zero-shot Segment Anything Model (SAM) pipeline with supervised deep learning architectures (U-Net, DeepLabV3, and SegFormer). U-Net achieved the strongest boundary performance (mean IoU = 0.76) and improved polygon-level correctness from 75.16% to 87.8% (absolute gain: 12.64% points; relative improvement: 16.8%). Conclusion: The proposed framework provides a scalable pathway to reduce structural uncertainty in agricultural parcel databases and strengthen the spatial foundation of precision agriculture and data-driven agricultural management systems.