Zero-shot parcel instance delineation in fragmented agricultural landscapes: Cross-source evaluation and structure-aware validation

Agricultural parcel and boundary delineation (APBD) remains challenging in fragmented agricultural landscapes, where small, densely adjacent fields increase merge and split errors even when overlap-based scores appear acceptable. This study addresses the problem as a zero-shot, cross-source, and structure-aware segmentation task for operational use without site-specific retraining or annotated training data. We combine vegetation-enhancement (VE) preprocessing with foundation-model-based instance segmentation and evaluate the pipeline across Sentinel-2 (S2, 10 m), PlanetScope (PS, 3 m), and Sentinel-2 Deep Resolution 3.0 (S2DR3, 1 m). Evaluation uses a structure-aware protocol that complements standard detection metrics with explicit merge and split quantification, enabling assessment of geographic information system (GIS)-ready parcel outputs beyond overlap alone. On a manually annotated benchmark of 424 parcels in Trentino (Italy), the best configuration (Vegetation-Enhancement + Segment Anything Model (VE+SAM) with S2DR3) achieves F1 0.79 and recall 0.78, with the lowest Global Under-Segmentation Error (GUSE 0.21). At source level, moving from S2 to S2DR3 nearly triples recall and halves GUSE, confirming that input resolution is a dominant factor in highly fragmented small-parcel settings. External validation on three Dutch tiles (300 km²), where parcels are larger and fragmentation is lower, shows a weaker resolution effect: S2 is no longer fundamentally inadequate and gains from finer imagery are smaller. In this setting, VE+SAM still maintains a consistent structural advantage over the domain-trained baseline, with systematically lower split and merge error rates, indicating that structural gains from the proposed pipeline persist when resolution is less limiting.